# Productlab: everything an AI needs to know > Last updated: 2026-10-06 > Productlab is a product-management community and conference company in Berlin (ProductLab UG), founded by Daniele Ronca. This single file covers current information about Productlab, the Productlab Conference 2026 (15-17 September 2026: talks, learnings and approved transcripts) and the Productlab Conference 2025 (talks and learnings). Website: https://productlab.app ## How to use this file with your AI Give this file (or its URL) to Claude, ChatGPT, Gemini or any other assistant and ask questions about Productlab and what was said at its conferences. Session summaries and key takeaways are editorial; transcripts are edited from machine captions and keep the speakers' own wording. Some 2026 sessions show a summary only, because the speaker's review of the transcript is still pending or the speaker asked for it not to be published. Example prompts: - "Tell me the top 10 learnings from Productlab Conference 2026." - "Which talks are about AI-native teams?" - "Summarise Jana Waldschmidt's talk and give me 3 actions for my team." --- ## Contents 1. About Productlab: organisation, leadership, contact, Productlab Conference 2027, recap of Productlab Conference 2026 (Leaders Lab, Craft & AI Labs, Conference Day, partners), Leaders Studio, community, newsletter 2. Productlab Conference 2026: talks, learnings and transcripts: index, people, then every Conference Day session by stage with summary, key takeaways, video and (where approved) transcript and Q&A (20 sessions) 3. Productlab Conference 2025: talks and learnings: overview recap plus 14 talk write-ups 4. Contact and canonical URLs --- # About Productlab > Productlab is a product-management community and conference company based in Berlin, founded by Daniele Ronca. It runs the annual Productlab Conference, Leaders Studio (executive coaching and peer circles for product leaders) and a year-round community of product events. The 2026 conference took place on 15-17 September 2026 in Berlin; the next edition is on 8-10 September 2027 in Berlin. Productlab exists to strengthen Europe's product leadership ecosystem through depth, craft and peer exchange. It has three pillars: - Conference: the annual Productlab Conference in Berlin, built for in-house product teams and senior practitioners. - Leaders Studio: 1:1 executive coaching, Leaders Circles and team coaching for product leaders. - Community: year-round events in Berlin and across Europe, plus the Productlab newsletter. ## Organisation facts - Legal entity: ProductLab UG (trading as Productlab) - Founded: 2023, Berlin, during a period of tech layoffs, as a place for product people to learn from each other and find their next role - Founder: Daniele Ronca - Community: 5,000+ members from 300+ companies - Newsletter: 3,000+ subscribers - Productlab Conference 2025 (first edition, Berlin, September 2025): 250+ attendees, 4.8/5 average rating - Productlab Conference 2026 (second edition, Berlin, 15-17 September 2026): around 1K+ product people in Berlin and online across the three days and the livestream - Productlab Conference 2027: 8-10 September 2027, Berlin ## Leadership - Daniele Ronca: Founder. Runs the conference and the community, and handles onboarding and matching for Leaders Studio. - Lucie McLean: Head Coach, Leaders Studio. Product leader with 20+ years leading product at the BBC, Omio, Zalando and Google; trained as an executive coach at Henley Business School. Leads the Leaders Circles and coaches senior product leaders one-to-one. ## Contact - General: team@productlab.app - Company and group packages: team@productlab.app (subject: Custom Company Package) - Website: https://productlab.app - Conference 2027 pre-registration: https://productlab.app/waitlist - Newsletter: https://newsletter.productlab.app - Community events: https://luma.com/productlab-community - LinkedIn: https://linkedin.com/company/productlab3 - YouTube: https://www.youtube.com/@ProductlabCommunityConference - Instagram: https://www.instagram.com/productlab_conf/ --- ## Productlab Conference 2027 - Dates: 8-10 September 2027 - Location: Berlin, Germany - Format: three days, following the 2026 structure (a leaders day, a hands-on workshop day and the main Conference Day), in person and online - Pre-registration and waitlist: https://productlab.app/waitlist - Every 2026 workshop sold out, so the waitlist lets people pre-book 2027 places. Workshops are allocated first. A minimum of 5 years in product is required. The leaders day is by approval, for people actively managing product people. - Venue, speakers, agenda and prices for 2027 are not yet published. For anything not listed here, see https://productlab.app or write to team@productlab.app. --- ## Productlab Conference 2026 (past event) The second Productlab Conference took place on 15-17 September 2026 in Berlin. Around 1K+ product people took part in Berlin and online. - Theme: "Love, Aim, Build_". Conference Day focused on how to adopt AI, and what keeps us relevant as humans. - 15 September: Leaders Lab, a working summit for senior product leaders, at The Social Hub Berlin-Mitte - 16 September: Craft & AI Labs (AI-Craft Day), five full-day hands-on workshops, at The Social Hub Berlin-Mitte - 17 September: Conference Day across three stages at CIC Berlin (Contentsquare Stage, Miro Stage, Doctolib Stage), plus roundtables in the Sponsor Lounge. The Contentsquare Stage and the Miro Stage were livestreamed. - Audience: senior PMs, Group PMs, PMMs, Heads of Product, Directors, VPs and CPOs, with 5+ years in product. - Agendas: https://productlab.app/agenda/15sept26, https://productlab.app/agenda/16sept26, https://productlab.app/agenda/17sept26 The full Conference Day programme, with a summary, key takeaways and (where approved by the speaker) the edited transcript of every recorded session, is in the section "Productlab Conference 2026: talks, learnings and transcripts" of this file. ### 15 September: Leaders Lab A full-day working summit for senior product leaders (Heads of Product, Directors, VPs, CPOs), by application, led by Ravi Mehta (ex-CPO Tinder, Product Director Facebook, VP Consumer Product TripAdvisor; co-creator of Reforge's AI Strategy and Product Leadership programs). Overview: https://productlab.app/leaders-lab The day ran as four 70-minute sessions, each with 35 minutes of teaching by Ravi, 25 minutes of roundtable discussion at three tables led by captains, and a 10-minute group recap: 1. How Customer Expectations Have Changed: which AI-driven shifts are already showing up in your data, and where your product is most behind. 2. Evaluating Your Market Position: where you are strong, where you are exposed, and where the opportunity is. 3. The AI-Powered Product Only You Can Build: combining AI with what is uniquely yours (data, customers, workflows), with cases from Granola, Adobe and Descript. 4. Building an AI-Native Team: how the PM role is changing, who to hire next, and where AI creates new bottlenecks. Participants left with four written takeaways: what customers want in the AI era, a candid read on their strategic position, a strategy for an AI-powered product only their company can ship, and a 90-day plan for their AI-native team. The planned roundtable captains were Lucie McLean (Google) and Florian Bonnet (Veriff). The evening closed with Axel Hecklau, World Vice-Champion of Magic, on "Surprise: The Play with Expectation": why breaking an expectation at the right moment makes products memorable. ### 16 September: Craft & AI Labs (AI-Craft Day) Five full-day, hands-on workshops with capped seats (about 20 per room), each with a morning and an afternoon block, a shared lunch and a closing recap in which each workshop shared its takeaways. All five workshops sold out. Overview: https://productlab.app/labs-day 1. Building Agentic AI Applications, with Zsuzsanna Tamas (Product Lead AI/ML; ex-Amazon Music, CoachHub, DKB, Adidas Runtastic). The shift from AI as a feature to AI as an actor: agent anatomy (LLMs, memory, tools, retrieval, APIs), scoping, constraints, human-in-the-loop, designing for failure, evaluation and observability, when to extend an agent versus add a second one, and the "Agentic Mesh" for governing many agents. 2. Discovery & Validation in the AI Era, with Birte Loeckel (Fractional Director/CPO; ex-eBay, Delivery Hero, Kleinanzeigen, mobile.de). Pragmatic discovery under real constraints, common discovery traps, using AI as an accelerator without replacing judgment, validation in messy environments, and influencing stakeholders so insights turn into decisions. 3. From Strategic Choices to Measurable Execution, with Tim Herbig (product coach and author). Product strategy as plain-English choices connected to company strategy, metrics beyond revenue KPIs that represent strategic success, and execution guardrails to track progress week to week. 4. Build your Product AI OS, with Tamer El-Hawari and Zihan Wang (Productbench). Setting up tools, a knowledge base and stack connections for predictable AI output, then turning common PM use cases into repeatable workflows and skills. 5. From Prompt to Product. A Functional MVP., with Vito Bica (Founder and CPTO, Arcate). A live solo build with AI from an empty repo to a deployed app: design systems, connecting real data, skill files for consistent output, debugging patterns, and a deploy to a live URL. ### 17 September: Conference Day Three stages at CIC Berlin, run in parallel: the Contentsquare Stage (main stage: keynotes, talks and a firechat), the Miro Stage and the Doctolib Stage (interactive deep dives for smaller groups). Twelve roundtables of about 10 people ran in the Sponsor Lounge; they were not recorded. Overview: https://productlab.app/conference-day Stage sessions (details, takeaways and transcripts in the 2026 section below): - Welcome: Daniele Ronca, with Wilfrid Kroath (Contentsquare) - Opening keynote: Jana Waldschmidt (GetYourGuide), "When Everything Looks Finished" - Csaba Tamas (Parloa), "Execution Is Cheap. What now? How to think of AI native companies" - Malte Landwehr (Peec AI), "How to make LLMs love your product" - Diana Pulnar (Babbel) with Susanne Dubuisson (Doctolib), "Everyone Personalizes the Answer. Nobody Personalizes the Learning." - Jane Austin and Mariusz Cieśla (Contentsquare), "How Contentsquare Ships Product" - Guannan Li (ablefy), "What the AI-Native Playbook Doesn't Tell You" - Firechat: Jan Oberhauser (n8n) and Luca Rossi (Refactoring), "The Future of AI Workflows" - Prasad Gupte (tonies), "Against Simplicity" - Rohan Garg (AI consultant, Kleinanzeigen), "Building 0 to 1 AI products" - Closing keynote: Michelle Parsons (Lora; ex-CPO Hinge), "Still Human" - Miro Stage: Joe McLean and Vihar Parikh (Miro), "From AI Assisted to AI Native: Rewiring how your teams innovate"; Cassiano Vailati and Matteo Risso (WeRoad), "WeRoad: An AI Journey"; Lucie McLean (Google), "Stop micromanaging, start scaling"; Ravi Mehta, "Building AI Products Competitors Can't Match" - Doctolib Stage: Dajana Aleksic (Mercedes-Benz), "The Interface Is Not Your Product. Now What?"; Florian Bonnet (Veriff), "From Feature Factory to Economic Ownership"; Vito Bica (Arcate), "For product builders: Let('s) build."; Alexandru Bleau (E.ON), "Start the Hard Talk" (a LEGO Serious Play session) Roundtables (Sponsor Lounge, not recorded): Alex Allan (leading product through M&A), Ravi Mehta (strategic thinking for product managers), Silvanus Alt of UXCam (an AI-first rebuild), Christian Geißler of MYNE Homes (why domain depth beats the wrapper race), Jeff Grang of Purchasely (the growth loop from install to renewal), Birte Loeckel (company politics), Rory Woodbridge of The Product Marketer (what AI can't do for your GTM), Jorge Aboytes of UserTesting (the confidence trap), Victoria Sheer of 12BRAVE (side businesses and first customers), Anastasiia Kostrub of Notabene (orchestrating systems instead of building features), Zsuzsanna Tamas (the business case for agents), Susanne Dubuisson of Doctolib (where AI actually made a difference). Stage hosts: Matteo Gamba, Francesca Cortesi, Veronika Morozová (Contentsquare) and Andreea Lungulescu (The Principal Recruiter). Artists and performances included Alex Reyes ("The Last Human Memory", an immersive AI installation), Harshini Karunaratne (opening spoken-word performance on AI and being human), Levan Shanshiashvili (plant music) and a closing performance by Marco D'Avila. Conference board 2026: Francesca Cortesi, Florian Bonnet, Csaba Tamas, Salvador Restelli, Pranav Pathak and Fabio Muroni. ### Partners 2026 - Vision partner: Contentsquare (named the main stage) - Strategy partner: Miro (named the Miro Stage) - Discovery partners: UserTesting, UXCam, Purchasely - Employee experience partner: Parloa - Doctolib (named the Doctolib Stage) and Delivery Hero - Venue partners: CIC Berlin (17 September) and The Social Hub Berlin-Mitte (15-16 September, also stay partner) - Community partners: Talent Crunch, Birra&Build, Bonial, Employed.world, Product Meetup Amsterdam, Weavii, KoRo - Media partner: Handpicked Berlin - Partner page: https://productlab.app/partners ## Productlab Conference 2025 (past event) The first Productlab Conference took place in Berlin in September 2025 with 250+ attendees, three parallel tracks and a 4.8/5 average rating. Partners 2025: Delivery Hero, UXCam, Rows, thermondo, Untitled Data Company. Talk write-ups are on the Productlab blog (https://productlab.app/blog) and in the section "Productlab Conference 2025: talks and learnings" of this file. --- ## Productlab Leaders Studio Coaching for product leaders who are moving from strategy to impact faster while navigating more complex organisations. Three formats: - 1:1 Leadership Coaching: work with an experienced executive coach who has also been a product leader. Free, no-obligation chemistry sessions to find the right coach; a fixed package or an ongoing cadence of sessions; corporate packages for several team members. https://productlab.app/leaders-studio-coaching - Leaders Circles: a hand-matched group of 10 to 12 product leaders at a similar stage, meeting online (Google Meet) every two weeks over 13 weeks. Includes a 90-minute introduction circle, 5 themed circles, a 90-minute closing circle, two 60-minute private coaching sessions, self-reflection exercises, frameworks and resources, a private WhatsApp group and a certificate. Morning, lunchtime and evening cohorts, designed for European time zones with options for APAC and the US East Coast. Every participant signs a confidentiality agreement. Process: apply, a short call within a week, matching with a circle and coach, then the 13-week program. https://productlab.app/leaders-studio-circles - Your Team: coaching support for product teams. Coaches: - Lucie McLean, Head Coach (ex-BBC, Omio, Zalando, Google; executive coach trained at Henley Business School) - Ravi Mehta, product strategy advisor (led product organisations at Tripadvisor, Facebook and Microsoft; most recently CPO at Tinder) - Elena Luneva, product, technology and business leader (most recently CPTO at Madison Reed and Braintrust) - Daniele Ronca handles onboarding, support and matching. ## Productlab Community Year-round events for product people in Berlin, with meetups hosted at tech companies (past hosts include N26, Google, SumUp, Contentful, Babbel, GetYourGuide, Amplitude and HelloFresh), speaker series, masterminds and roundtables, plus community nights in other European cities. Events: https://luma.com/productlab-community ## Productlab Newsletter A newsletter for product people with event news, conference learnings and talk write-ups. Subscribe: https://newsletter.productlab.app ## Organisations represented Speakers and attendees across editions have come from companies including Amazon, Google, Booking.com, Just Eat Takeaway, Zalando, Mercedes-Benz, Cisco, HelloFresh, Delivery Hero, N26, SumUp, Stripe, GetYourGuide, Peec AI, Parloa, Veriff, Personio, n8n, ablefy, tonies, Kleinanzeigen, E.ON, Babbel, WeRoad, Doctolib, Miro, Contentsquare, UserTesting and Purchasely. ## What Productlab is not - Not a course, certification or framework program - Not a beginner event: the conference expects 5+ years in product, and many attendees have 10+ - Not a vendor expo or badge-scanning event - Not a US Big Tech echo chamber: the reference frame is European companies and European context --- # Productlab Conference 2026: talks, learnings and transcripts ## The 2026 recap: AI lowered the floor. Few extended the ceiling. By Daniele Ronca, founder of Productlab. Also on https://productlab.app/blog *What I brought home from three days in Berlin with 1,000+ product people.* [CTA: See all the photos](https://drive.google.com/drive/folders/1jAY4p4C6KrmdzE-hAN59T9dwHaQfKE12?usp=drive_link) At Leaders Lab I told the story of A+T, the architecture magazine, and of 2008. Standardisation arrived, and many architects decided the point was a better facade. The craft had always lived elsewhere: in stakeholders, laws, budgets and people. Product is having its 2008. AI lowered the floor, so anyone can prototype and ship. At Productlab Conference 2026 (Berlin, 15 to 17 September) we found out who moved the ceiling at the same time. *Source: "Lower the floor, raise the ceiling" is a phrase from MIT's early computing and learning work. I found it in [John Maeda's LinkedIn post "Floor vs ceiling"](https://www.linkedin.com/feed/update/urn:li:ugcPost:7467981309320028161/), which links it to William Morris and the Arts & Crafts movement: when machines took over the floor, artisans raised the ceiling. The 2008 story comes from my Leaders Lab session on A+T, the architecture magazine.* ### The highlights - **AI writes the code; humans still set the pace.** Parloa: **97%** of changes AI-built, a "marginal" speed-up. - **Polish stopped being a signal.** At GetYourGuide, **99%** of take-homes now pass. - **Your analytics can't see AI discovery.** Flow Agency: **0%** in analytics, **21%** when asked. --- ### Part 1 · Polish is free, so judgment is what you read now. Ravi Mehta, Instructor & Advisor @ Reforge, named the floor: **slop is "the accumulation of unexamined defaults".** #### A finished document no longer proves the thinking happened. Jana Waldschmidt, Director of Product @ GetYourGuide, sees 99% of take-homes pass. Ten minutes into the follow-up, candidates can read their slides but can't explain how they got there. So she interviews for what you can observe: grounding in the customer, and hunting for evidence against your own case. #### The counter-view: is your judgment even yours? Jana believes judgment is earned. Eight hours later, Michelle Parsons, CEO & Co-founder @ Lora, closed with: "I'm not even sure our taste or our judgment is ours anymore." **So what:** stop grading polish. Ask people to defend one choice. ### Part 2 · Everyone has the model. The harness is the product. Ravi again: Claude Code broke through "not because the model got better but because the harness got better". #### The harness is a product decision. Guannan Li, Director of Product Growth @ ablefy, lived through six months of engineers pushing back on PM-built pull requests. The fight ended when a harness ran **14 automated checks** on every PR before a human looked. #### The counter-view: how much context? Matteo Risso, Senior Product Manager @ WeRoad, called connected company context "the most powerful thing that happened to my profession this year". Ravi: "less context is actually better than more in most cases, because of something called context rot." I'm closer to Matteo. Our year, like many, was about interconnecting, and everything leads to a company brain. But that brain shouldn't stop us discovering new ideas, or going against the data to rethink the whole thing from first principles. **So what:** checks, context and where AI goes belong on the product roadmap. ### Part 3 · Velocity is not revenue. Our own admission: Florian Bonnet deserved a bigger stage. He was our facilitator last year and sits on our conference board, yet his talk on money overflowed its room and people stood. I was biased in the room assignment. I wanted to give Matteo and Cassiano, younger speakers, their time to shine, and they delivered. Honestly, it was hard to pick this year. #### Shipping fast is not a moat. Florian, VP Product Solutions & Growth @ Veriff, charted velocity against economic impact: "they don't correlate". He calls it busyness versus business. #### Nobody showed AI revenue yet. Rohan Garg, AI Consultant @ Kleinanzeigen, said monetisation is "not there yet" and expects the ROI question "6 to 12 months down the line". The cost side can jump overnight: when WeRoad asked to move to an enterprise plan, Cassiano Vailati said "in one day our bill went from 200k to 1 million." **So what:** put a revenue line next to every velocity chart. ### Part 4 · The bottleneck moved from building to deciding. Csaba Tamas, CPTO @ Parloa, has AI building almost all of his code. The company is only marginally faster, because product definition didn't speed up at all. #### Decision latency is the number to watch. That's Csaba's term: **decision latency is "the wall-clock time between a question arising and an answer existing".** Parloa runs a decision SLA and cut it "from many months to a couple of weeks". #### People can't absorb what the machines produce. Diana Pulnar, SVP Product & People @ Babbel: "the speed is so high but our mental capacity hasn't improved." Nobody on stage offered a fix, and that worries me most. Off stage, the talk was about people. The market is flooded with candidates, yet people who deliver and go beyond are hard to find. Talking with Matteo Risso about his favourite concept, a *definition of done* for people, I noticed how rare the above-and-beyond characters are. They are what Marty Cagan describes as the rare person who combines design, product and engineering. *Source: the "three skills in one person" framing is Daniele's shorthand for Marty Cagan's writing on strong product people (SVPG).* **So what:** measure time to an agreed spec, not time to a first draft. ### Part 5 · Customers moved while we were building. #### Discovery went dark. Malte Landwehr, CPO & CMO @ Peec AI, asked who gets 10% of users from LLMs: "Nobody." Who tracks it? "You really should track this," he said, and called the blind spot dark chat. #### Nobody asked for AI. Ravi: "Customers don't want AI." They want the problem solved, which is why Babbel went back to talking to learners. Did Kleinanzeigen's AI shopping assistant beat classic search? Rohan: "No. I think normal search still works better; it generates more leads." **So what:** if your funnel can't see where people come from, ask them. --- ### What to do on Monday **PMs** - Sit through three customer interviews; don't synthesise ten (Jana). - Ask every signup where they found you, ChatGPT included (Malte). **Product leaders** - Track decision latency and time to an agreed spec (Csaba). - Interview for judgment you can observe, not polish (Jana). **Founders & execs** - Chart velocity against revenue before celebrating either (Florian). - Default to deterministic code; Ravi says it wins "95% of the time". > **Where I might be wrong** > - Merged roles may be a phase: WeRoad sees "every role take back its position". > - No AI revenue on stage may be timing, not absence (Rohan: 6 to 12 months). > - Ravi may be right about context rot, and a company brain could feed the defaults I want teams to question. > - If Michelle is right, judgment can't be the ceiling, because it isn't ours. ### Who gets the time back The ceiling is a choice as much as a skill. If the floor gives you back an afternoon, does it go to three customer interviews or ten more PRDs? A year ago I told the room "we're going to work more, not less," and it happened. My answer to Jana and Michelle: judgment is the sum of experiences outside our comfort zone, with high achievers, where taste and ideas flourish. So surround yourself with, or build, a professional sounding board of people who read a lot, exchange, and keep rethinking the status quo. We ended with ice cream and a community selfie on stage. **Productlab Conference 2027 · 8-10 September 2027 · Berlin.** [Join the waitlist](https://productlab.app/waitlist). --- ### How this was made - Source: transcripts of the 20 Conference Day sessions across three stages at The Social Hub and CIC Berlin. - Summary only (under speaker review, no quotes): the opening welcome, Contentsquare (Jane Austin & Mariusz Cieśla), Miro (Joe McLean & Vihar Parikh), Prasad Gupte (tonies). Lucie McLean (Google): approved without the audience Q&A. Dajana Aleksic (Mercedes-Benz): her approved summary only. - Doctolib Stage was not filmed; those sessions are covered by approved summaries and slides. - Roundtables and Leaders Lab conversations were not recorded and are not quoted. - Also on the programme: Harshini Karunaratne (spoken word), Alex Reyes (*The Last Human Memory*), Levan Shanshiashvili (music with plants), Marco D'Avila (closing performance), Axel Hecklau (magic at Leaders Lab). *Disclaimer: we do our best to get every word and number right, but the transcripts, summaries and graphics on this page are produced by Productlab with the help of AI, from the recordings and the speakers' slides. They can contain mistakes, and they are our reading of the talks, not necessarily the views of the speakers or their companies. Use them as examples and inspiration, not as business advice or a substitute for learning: put every idea in the context of your own company, team and customers. Spotted an error, or you're a speaker who would like something corrected or removed? Write to [team@productlab.app](mailto:team@productlab.app) and we'll fix it quickly.* --- **17 September 2026 · CIC Berlin · Theme: Human/Skills — how to adopt AI, and what keeps us relevant as humans.** Cleaned transcripts of every session on the three stages of Conference Day: the Contentsquare Stage (main stage), the Miro Stage and the Doctolib Stage (deep dives). Roundtables in the Sponsor Lounge were not recorded. ## About these transcripts - One section per stage, one sub-section per session. Every session starts with its metadata (speakers, roles, companies, stage, time, format, host, official abstract), then **Summary**, **Key takeaways**, **Transcript** and, where there was one, **Q&A**. - Transcripts come from machine captions (YouTube for the two streamed stages, Granola for the Doctolib Stage). They have been edited for readability: filler words, false starts, repetitions, breaks, music and stage logistics are removed, and names and terms are corrected. The wording stays the speakers' own. The Summary and Key takeaways are editorial. Check a quote against the recording before you publish it. - The Miro Stage recording starts a few minutes into the first session. Agenda: https://productlab.app/agenda/17sept26 · More: https://productlab.app ## Index | Time (CEST) | Stage | Track / format | Session | Speakers | |---|---|---|---|---| | 09:30 | CSQ | Welcome | [Welcome to Productlab Conference](#09-30-welcome-to-productlab-conference-daniele-ronca-wilfrid-kroath) | Daniele Ronca & Wilfrid Kroath | | 09:50 | CSQ | Keynote | [When Everything Looks Finished](#09-50-when-everything-looks-finished-jana-waldschmidt) | Jana Waldschmidt | | 10:30 | CSQ | Build & Craft | [Execution Is Cheap. What now? How to think of AI native companies](#10-30-execution-is-cheap-what-now-how-to-think-of-ai-native-companies-csaba-tamas) | Csaba Tamas | | 11:10 | CSQ | Build & Craft | [How to make LLMs love your product](#11-10-how-to-make-llms-love-your-product-malte-landwehr) | Malte Landwehr | | 12:05 | CSQ | Build & Craft | [Everyone Personalizes the Answer. Nobody Personalizes the Learning.](#12-05-everyone-personalizes-the-answer-nobody-personalizes-the-learning-diana-pulnar-susanne-dubuisson) | Diana Pulnar & Susanne Dubuisson | | 12:45 | CSQ | Build & Craft | [How Contentsquare Ships Product](#12-45-how-contentsquare-ships-product-jane-austin-mariusz-cieśla) | Jane Austin & Mariusz Cieśla | | 15:00 | CSQ | Leadership | [What the AI-Native Playbook Doesn't Tell You](#15-00-what-the-ai-native-playbook-doesn-t-tell-you-guannan-li) | Guannan Li | | 15:40 | CSQ | Leadership | [Firechat: The Future of AI Workflows](#15-40-firechat-the-future-of-ai-workflows-jan-oberhauser-luca-rossi) | Jan Oberhauser & Luca Rossi | | 16:40 | CSQ | Strategy | [Against Simplicity](#16-40-against-simplicity-prasad-gupte) | Prasad Gupte | | 17:20 | CSQ | Build & Craft | [Building 0→1 AI products](#17-20-building-0-1-ai-products-rohan-garg) | Rohan Garg | | 18:00 | CSQ | Closing keynote | [Still Human](#18-00-still-human-michelle-parsons) | Michelle Parsons | | 18:30 | CSQ | Closing | [Final Remarks](#18-30-final-remarks-daniele-ronca) | Daniele Ronca | | 10:30 | Miro | Build & Craft | [From AI Assisted to AI Native: Rewiring how your teams innovate](#10-30-from-ai-assisted-to-ai-native-rewiring-how-your-teams-innovate-joe-mclean-vihar-parikh) | Joe McLean & Vihar Parikh | | 12:05 | Miro | Build & Craft | [WeRoad: An AI Journey](#12-05-weroad-an-ai-journey-cassiano-vailati-matteo-risso) | Cassiano Vailati & Matteo Risso | | 15:00 | Miro | Leadership | [Stop micromanaging, start scaling](#15-00-stop-micromanaging-start-scaling-lucie-mclean) | Lucie McLean | | 16:40 | Miro | Build & Craft | [Building AI Products Competitors Can't Match](#16-40-building-ai-products-competitors-can-t-match-ravi-mehta) | Ravi Mehta | | 10:30 | Doctolib | Build & Craft | [The Interface Is Not Your Product. Now What?](#10-30-the-interface-is-not-your-product-now-what-dajana-aleksic) | Dajana Aleksic | | 12:05 | Doctolib | Strategy | [From Feature Factory to Economic Ownership](#12-05-from-feature-factory-to-economic-ownership-florian-bonnet) | Florian Bonnet | | 15:00 | Doctolib | Build & Craft | [For product builders: Let('s) build.](#15-00-for-product-builders-let-s-build-vito-bica) | Vito Bica | | 16:40 | Doctolib | Leadership | [Start the Hard Talk](#16-40-start-the-hard-talk-alexandru-bleau) | Alexandru Bleau | ## People | Name | Role, company | LinkedIn | Sessions (17 Sept) | |---|---|---|---| | Daniele Ronca | Founder, Productlab | https://productlab.app | 09:30 CSQ; 18:00 CSQ (host); 18:30 CSQ | | Wilfrid Kroath | Enterprise Transformation Leader & Automotive Sector Lead, Contentsquare | https://www.linkedin.com/in/wilfridkroath/ | 09:30 CSQ | | Matteo Gamba | Stage host, ex-Wayfair | https://www.linkedin.com/in/matteogamba/ | 09:30 CSQ (host); 09:50 CSQ (host); 10:30 CSQ (host); 11:10 CSQ (host); 12:05 CSQ (host); 12:45 CSQ (host) | | Jana Waldschmidt | Director of Product, GetYourGuide | https://www.linkedin.com/in/janawaldschmidt/ | 09:50 CSQ | | Csaba Tamas | CPTO, Parloa | https://www.linkedin.com/in/tamascsaba/ | 10:30 CSQ | | Malte Landwehr | CPO & CMO, Peec AI | https://www.linkedin.com/in/landwehr/ | 11:10 CSQ | | Diana Pulnar | SVP Product & People, MD Babbel Labs, Babbel | https://www.linkedin.com/in/diana-pulnar-51781b13/ | 12:05 CSQ | | Susanne Dubuisson | Product Director, Doctolib | https://www.linkedin.com/in/susanne-dubuisson/ | 12:05 CSQ | | Jane Austin | SVP Product Design, Contentsquare | https://www.linkedin.com/in/msjaneaustin/ | 12:45 CSQ | | Mariusz Cieśla | Staff Design Engineer, Contentsquare | https://www.linkedin.com/in/mariuszciesla/ | 12:45 CSQ | | Guannan Li | Director of Product Growth, ablefy | https://www.linkedin.com/in/liguannan/ | 15:00 CSQ | | Francesca Cortesi | Stage host, Vend | https://www.linkedin.com/in/francesca-cortesi/ | 15:00 CSQ (host); 15:40 CSQ (host); 16:40 CSQ (host); 17:20 CSQ (host) | | Jan Oberhauser | Founder & CEO, n8n | https://www.linkedin.com/in/janoberhauser/ | 15:40 CSQ | | Luca Rossi | Founder, Refactoring | https://www.linkedin.com/in/lucaronin/ | 15:40 CSQ | | Prasad Gupte | VP, Digital Products, tonies | https://www.linkedin.com/in/prasadgupte/ | 16:40 CSQ | | Rohan Garg | AI Consultant, Kleinanzeigen | https://www.linkedin.com/in/rohangarg/ | 17:20 CSQ | | Michelle Parsons | CEO & Co-founder, Lora | https://www.linkedin.com/in/michcparsons/ | 18:00 CSQ | | Joe McLean | Product Director, AI Canvas, Miro | https://www.linkedin.com/in/josephmmclean/ | 10:30 Miro | | Vihar Parikh | Group Product Manager, Custom Solutions, Miro | https://www.linkedin.com/in/viharparikh | 10:30 Miro | | Veronika Morozová | Stage host, Contentsquare | https://www.linkedin.com/in/veronika-morozov%C3%A1/ | 10:30 Miro (host); 12:05 Miro (host); 15:00 Doctolib (host); 16:40 Doctolib (host) | | Cassiano Vailati | Head of Technology, WeRoad | https://www.linkedin.com/in/cassvail/ | 12:05 Miro | | Matteo Risso | Senior Product Manager, WeRoad | https://www.linkedin.com/in/matteo-risso/ | 12:05 Miro | | Lucie McLean | Product Leader, Google | https://www.linkedin.com/in/luciemclean/ | 15:00 Miro | | Andreea Lungulescu | Stage host, The Principal Recruiter | https://www.linkedin.com/in/alungulescu/ | 15:00 Miro (host); 16:40 Miro (host); 10:30 Doctolib (host); 12:05 Doctolib (host) | | Ravi Mehta | Entrepreneur, Instructor & Advisor, Reforge | https://www.linkedin.com/in/ravimehta/ | 16:40 Miro | | Dajana Aleksic | Head of Product & CX, Mercedes-Benz | https://www.linkedin.com/in/dajana-aleksic/ | 10:30 Doctolib | | Florian Bonnet | VP Product Solutions & Growth, Veriff | https://www.linkedin.com/in/bonnetflorian | 12:05 Doctolib | | Vito Bica | Founder & CPTO, Arcate | https://www.linkedin.com/in/bica/ | 15:00 Doctolib | | Alexandru Bleau | Senior Innovation Manager, E.ON Group Innovation | https://www.linkedin.com/in/alexandrubleau/ | 16:40 Doctolib | --- ## Contentsquare Stage Main stage · keynotes and talks · livestreamed. Hosts: Matteo Gamba (morning), Francesca Cortesi (afternoon). Recording: https://youtube.com/live/_1lTPxog0so ### 09:30 · Welcome to Productlab Conference — Daniele Ronca & Wilfrid Kroath - **Speakers:** - [Daniele Ronca](https://productlab.app) — Founder, Productlab. Founded Productlab in Berlin in 2023: the Productlab Conference, Leaders Studio coaching and a year-round product community. - [Wilfrid Kroath](https://www.linkedin.com/in/wilfridkroath/) — Enterprise Transformation Leader & Automotive Sector Lead, Contentsquare. Works on digital experience and AI-enabled change at Contentsquare; opened the main stage on behalf of the conference main sponsor. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 09:30 CEST - **Format:** Welcome - **Host:** [Matteo Gamba](https://www.linkedin.com/in/matteogamba/) (ex-Wayfair) **Summary.** Daniele Ronca opens Productlab Conference 2026 by framing the year in AI for product teams, explaining how the day is designed for exchange rather than passive listening, and telling the story of how Productlab started in 2023. He thanks the conference board, sponsors and partners and introduces Leaders Studio. Wilfrid Kroath then gives a short welcome on behalf of Contentsquare, the main sponsor. **Key takeaways** - After a year of AI upheaval, the open questions are where product management is heading, and why building faster also means mediocre output arrives faster. - Productlab itself runs on a database (Notion) with Claude Code on top, with the website, email and LinkedIn as different "representations" of the same data. - The day is deliberately built for peer exchange: open Q&A on the main stage, smaller Miro and Doctolib stages (up to 35 people) and round tables of 10. - The conference aims to be "analog" (printed agenda, Polaroids, artists) so attendees reflect instead of answering Slack and email. - Productlab began in 2023 during a period of layoffs as a way for product people to learn and find their next job; its belief is that better products built in Europe create economic and societal outcomes locally. - Productlab has launched Leaders Studio, coaching for leaders and teams, because leaders working at this velocity have less exchange with peers and get lonelier. *Full transcript to follow after the speaker review.* ### 09:50 · When Everything Looks Finished — Jana Waldschmidt Video: https://youtu.be/JaWed5bE7kQ - **Speakers:** - [Jana Waldschmidt](https://www.linkedin.com/in/janawaldschmidt/) — Director of Product, GetYourGuide. Leads the supply-side product teams that help experience creators grow on the GetYourGuide marketplace; 8+ years in product management. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 09:50 CEST - **Format:** Keynote - **Host:** [Matteo Gamba](https://www.linkedin.com/in/matteogamba/) (ex-Wayfair) - **Official abstract:** AI transformation isn't just a skill shift. It's an identity shift. Jana explores the gap between opportunity and anxiety, and what it actually takes to lead your teams through it. **Summary.** Jana Waldschmidt, a product leader at GetYourGuide, describes a growing gap she sees as a hiring manager and leader: AI makes documents and take-home tasks look polished while the thinking behind them hasn't happened. She argues that what she is not willing to give up is judgment, describes the observable behaviours of good judgment, and explains why it can only be built through friction, customer contact and reflected loops. **Key takeaways** - AI is extraordinary at turning half-baked thinking into something presentable, so how finished a document looks no longer signals how finished the thinking is. - Before AI, a good document was hard to fake; the feedback used to be about communication and length, now it is about missing thinking. - Judgment is making good decisions when there is no obvious right one. Observable signs: grounding in the customer, seeking evidence that breaks your own case, and holding a position without gripping it (explaining what would need to be true for it to collapse). - Judgment isn't built by reading books, attending conferences or prompting; it is built by leaning into friction in discovery, validation and debate. - Spend AI productivity gains on getting closer to reality: observe customers directly (sit through three interviews rather than synthesising ten), have a point of view before prompting, and complete and reflect on more build–validate loops. - Her hot take: today's PMs can build judgment faster than any previous generation, because the tools free up time to get back to talking to customers. #### Transcript **Jana Waldschmidt:** Okay, here we go. By the way, anyone from Estonia? Let's try to beat that. Okay, so Finland at least had one extra participant. I'm from Estonia, so I'm extremely unique in that case. If I could have a clicker — wonderful. We have technical difficulties. Okay, meanwhile, no problem. Does anyone want to share a joke? A dad joke? Anyone good at that? I know my husband is in the room — you have to tell a joke now. Well, hey, you're my emotional support; that's why you're here. But I'm personally incredibly excited to be here today. I got invited some months ago after giving a small talk elsewhere, and I thought: we're all really living through AI, everyone is talking about it, so what is the unique thing I can bring? As I was preparing for this, I realized that actually all of us have a very unique twist, taste and story to share. My guidance to you for the rest of the day would be: attend these sessions, listen to different leaders, think about leadership and product management, what changes for us and what doesn't, but collect those puzzle pieces and build them together. Have your own point of view, because I think we're living through an incredible time where we have an opportunity to do so. And with that, I would love to ask if my slides could be on, because as a showman I wouldn't do a really good job. A little pause. Are we on? Right. Thank you. Well, that's one way to wake up: getting energized in a very short period of time by your slides not being accessible. But it is what it is. Let's go straight to the introduction. My name is Jana and I'm a product leader at GetYourGuide. I've been in product for about ten years, and I'm one of those people who stumbled into product management, which I believe many of you did too. I tried multiple roles and then somehow realized that product management combines all of them, so why wouldn't I try? I fell in love from day one, and I've stayed the course. I was an IC for a certain period of time and then eventually grew into a leadership position. If I were to describe what I do now, I would say I have the privilege of watching people fall in love with product management, and that's a great place to be for me. In the past ten years, I don't think there were that many changes, apart from some tools coming up. But now it's a pivotal moment, and it makes us ask big questions. What happens to my job? What happens to my identity, to my skill set? Should I be looking for alternatives? Should I pivot to yoga teaching? — which is a conversation I keep having with my husband. This topic keeps popping up. But I think we're in a better position now to ask better questions, to put those puzzle pieces together, to attend events like this, hear other perspectives, form our own, and see what that means for our careers. Today I will share my perspective on what I believe will stay true, where we can stay the course in our craft. This is just one of the puzzle pieces for you; attend the rest of the day and find yours. For me it's always best to describe the situation starting with a personal story, so I'll take you through my leadership journey of the past couple of years. About two years ago, I get a strategy document. I read it. It reads beautifully: great narrative, very visual, solid recommendation. I hadn't seen such a good-looking document from this particular PM for a while. We go into a conversation, and I ask a typical question: "Why would that be the segment? What choices did you make along the way?" She would scroll back and forth through the document and point me to a bullet point. Okay — very bullet-point, and very atypical behaviour on her part. We spent another 10–15 minutes going back and forth, only to realize that she wasn't really ready. The document looked great, but the thinking behind it hadn't happened. Then, as someone who obviously hires non-stop, I started seeing this happen in take-home tasks. Before, that used to be a good enough prompt to understand, within a couple of hours, what a person is able to achieve, where they focus their time and what kind of questions they like to address. The conversations were much juicier, seeing what assumptions they were making as they unpacked it. And by the way, the submissions were also more human. There were typos, there were debatable choices of visuals, there were fonts you would look at and think, "Why?" There was all of that goodness in there, but you knew there was a human on the other end, and you're not hiring them for design, so that's okay. Now they all come back polished. They're on brand, with the best frameworks, so easy to read. Non-native English speakers — but the language, the sophistication of it, is amazing. That made my life incredibly difficult, because practically everyone, 99%, passes. Everything is green and beautiful. But then, as always, I invite them for a conversation, and ten minutes into it: they can read the slides, but they cannot explain how they got there, because the thinking didn't happen to the extent the polished document suggests. So the gap I started noticing, and that I'd like to discuss with you today, is between how finished something looks and how finished the thinking behind it is. I've observed AI creating more and more of that gap, and for me that is the challenge. Reflecting on my own experience: before AI, a good document was actually hard to fake. Let's go back to that product strategy paper. You get a paper, you look at it. Yes, it's probably 12 pages long. Yes, it has a lot of data packed into it — all those graphs, assumptions baked in, the synthesis of interviews, and so on. It may be difficult to digest, but you get into the conversation and the thinking is solid, there's a juicy conversation to be had, there are baked-in assumptions, and the person really lived through it. The gap I used to have to close was around communication: "How do we fit this into two pages? The work is solid, but 14 pages — I'm so sorry." Then AI arrived, and again and again I get documents that look polished, and far fewer great conversations. I believe AI is extraordinary at taking half-baked thinking and translating it into something that looks very presentable. That posed a question I needed to reflect on, as this was coming so quickly into day-to-day life: what am I willing to accept in this new age, and what am I not going to give up? What will remain true for me, and what am I not willing to lose? In terms of accepting: it's fine, everything is going to look beautiful. I'm not going to spend another day, or another second of my time, wondering why it's filled with em dashes or why it's all so sophisticated. It is what it is; it's most likely 99% generated. But what am I not willing to lose? That required me to reflect on the conversations I was describing. What was I actually asking? What was I trying to poke at? It hardly ever has to do with black and white or the right answer, because most of the time I don't know it, and they are also in the discovery stage. It has a lot to do with the journey. How are you getting to your recommendation? What are you considering along the way? In product management we have a very specific word for that: judgment. I can see a couple of smiles in the room, because "judgment" and "taste" are probably skyrocketing in how many people use them, which is very true. We keep saying we hire for great judgment, senior people have more judgment. In a lot of rooms like this, we mention it a lot but don't really unpack it, while always saying it's so important in the age of AI. So I'd love to offer my perspective on what judgment is, and what the observable behaviours are when I see someone exercising great judgment. To me, great judgment is about making good decisions when there's no obvious right one in front of you. The first thing I observe when working with PMs is being grounded in the customer. A good PM will be able to frame a problem statement — and a lot of the time, unfortunately, there will be a half-baked solution in it. To bring an example from my world at GetYourGuide: "Suppliers are asking for a bulk upload tool." The alternative looks something like a statement where you know the PM actually sat in those conversations. They observed the behaviour, they understood the problem, they know what the day-to-day looks like and where the friction is. This is the hill I'm willing to die on: it's all about customer empathy. So that's the observable behaviour: they bring the customer into the room. That depth matters because product management, in my humble opinion, has nothing to do with collecting support tickets and putting them in the backlog, or listening to interviews and taking statements at face value. It's about depth, it's about discovery, and it's about understanding the surrounding pieces of what makes the problem a problem. The second part is that they go looking for evidence to break apart their own case. It's very easy to create a fan base: get in a room with three people you work really well with and ask, "Does this make sense?" and they say yes. People tend to say yes a lot; they agree. That's not really helpful, is it? It doesn't help you build a stronger argument and evidence. What you will see instead is someone setting the stage and actually inviting that friction — inviting the person from their company, or externally, whom they might not get along with, or who has very opposing views on how to solve the problem. That question will cost you an afternoon, and it might annoy people in the room. The third piece is holding the position without gripping it. When you have a conversation with someone who is honestly in a win-or-lose mindset, what you hear is collapsing or defending: "Sure, we can look at that," or "Whatever, it's in the document. Jana, scroll lower." That is not how I'd love a great product manager exercising judgment to show up. What I'd love to hear is the logic behind it: "This is the assumption. This is what I'm pushing on. This needs to be true for this to collapse." That's a much better conversation to have. Now I imagine the question in the room is: "Okay, great, Jana, thank you — but how do I actually build that, now that we all claim judgment is so important?" One caveat: it's definitely not from reading books, unfortunately. It's not from attending more conferences. It's not about having more conversations. All of that gives you extra information, which is great and important, but it's not how judgment gets built. I believe judgment gets built by leaning into friction, and that friction tends to happen in discovery, in validation, in the debates you get to have. I don't know if anyone wants to volunteer what stands in the way of this happening continuously. At least currently, it's time. There's never enough time. The conversation I have with PMs, or at least used to have up to a certain point, is: "I don't have time to talk to customers. My calendar is filled with meetings. I need to push information from A to B. There's everything to prepare." And I do agree the PM craft has grown to include an enormous amount of extra tasks. But aren't we in an amazing period of time where AI is very helpful with that? Aren't we all saying that we're getting productivity gains from shortcuts, automated workflows and synthesizing information? I would claim we are. So what are we spending that time on? Here's my twist on what I believe you should be spending the time on as a product manager navigating this change. First, get closer to reality. Again, the hill I'm willing to die on is customer empathy. Get out of the office — nothing great happens in the office. Go observe your customers. See what they do, see where they struggle. Sit through three interviews; don't synthesize ten. That depth matters. That is the raw material for your judgment, and it's the only way to sense what the customer is actually struggling with. Second, have a point of view before you ask for one. You don't want to be in a situation where the document looks perfect but the thinking behind it isn't there. When you go into a conversation, go to be challenged. When you go prompting, go to be challenged, not to be given an answer. More and more, what I've been seeing is people putting a very half-baked idea out there, and an LLM, as I mentioned, is great at closing that gap and saying, "Yes, that looks great, let me fill in the gaps for you." Please exercise your own judgment. Third, complete more loops and reflect on them. What I mean is that every product decision you make adds to your playbook: all the wins and losses, all the reflections, the loops of hypothesizing, validating, putting it out there, seeing it picked up or not picked up. That is your playbook, so go through those loops. I appreciate that in some companies the velocity is not there, and that space might not exist for you. But again, aren't we living through a great time where this is possible? The tools are accessible, the data is there, it's cheap, and you can learn something so, so fast. So please invest: build something for yourself, build something for your family, but do create those loops. To summarize on loops: reflect on them. Reflect on all the experiences you have — the conversations with customers, the debates, the products you put out there. Sit and reflect, and make sure it's intentional, because I claim it's not about how many times you go through a loop, but about how much you learn from it every single time. So please invest in it. And I would say: judgment isn't prompted, so don't try. Judgment isn't generated, so don't try. This is not where AI is going to be of help. I truly believe it's earned. But there's good news, so as not to sound too dramatic or leave you hanging. It is all hard work, but — my hot take — we get to build judgment faster than any previous PM generation. You have the tools and shortcuts available to free up that precious, precious time to get back to the roots, which is talking to our customers. Thank you. ### 10:30 · Execution Is Cheap. What now? How to think of AI native companies — Csaba Tamas Video: https://youtu.be/UGGfbEoSZho - **Speakers:** - [Csaba Tamas](https://www.linkedin.com/in/tamascsaba/) — CPTO, Parloa. Leads product management, UX and enablement at Parloa. Previously advised fintechs on AI/ML strategy at AWS and was CPO/CTO at FinLeap Connect. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 10:30 CEST - **Format:** Talk · Track: Build & Craft - **Host:** [Matteo Gamba](https://www.linkedin.com/in/matteogamba/) (ex-Wayfair) - **Official abstract:** AI sped up execution by 15-25%. It sped up judgment by zero. A story about a circus tent, a leaking aircraft, and the product team you'll run next year. **Summary.** Csaba Tamas of Parloa explains why generating 97% of code with AI produced only a marginal increase in overall delivery speed, and argues that AI lowered execution cost but not coordination cost. Using the story of Kelly Johnson and Lockheed's Skunk Works, he lays out how AI-native companies should reduce decision latency, interface surface area and requirements volatility, and align incentives, and he proposes a spec- and prototype-driven workflow plus metrics for measuring return on AI spend. **Key takeaways** - AI generates code many times faster than humans, but company-level throughput gains are around 30%, because coding is only one stage; definition, design, review, testing and alignment are not accelerated, and more review and rework can even make acceleration negative. - Generating code with AI doesn't make you AI-native; that requires reorganizing teams, processes and culture. AI brought down execution cost, not coordination cost. - Lessons from Skunk Works: cut decision latency (co-location, few participants, direct access to decision makers), keep upfront requirements rigid while letting implementation be scrappy, keep systems simple, and pay for impact rather than team size. - Organize teams around software context boundaries (the inverse Conway maneuver) and measure decision response time, since a team blocked on a decision means "five people and 15 AI agents" standing still. - When code is cheap and disposable, the specification, decisions and decision rationale become the asset. His workflow: business value and success metrics → use cases → decision rationale → PM vibe-coded prototype to surface unknowns → detailed specs → tests → code → acceptance testing. - Measure releases reaching production, idea-to-production lead time, time spent waiting, time to an agreed specification, and change failure rate. The organization only moves as fast as its slowest stage, which remains a human (and hybrid human–agent) system problem. #### Transcript **Csaba Tamas:** Good morning, everyone. Are you already energized from the first sessions? Good. My name is Csaba and I work at Parloa. Parloa is a customer experience automation company using AI, primarily in call centers in this first stage. Call center automation, or customer support automation, is the first really validated use case for AI. That's great news for our sales teams: customers are queuing up and pulling the product out of our hands. The bad news is the same thing, because it's bad news for engineering. Everyone wants AI, and everyone wants it now. In a new software category you used to have a decade to get to product-market fit. But today every customer wants customer support automation AI from us — even the laggards. If you think about pioneers, early adopters, early majority, late majority and then the laggards, who used to buy a product only when, like the touch phone, the rotary phone was no longer available: they all want AI today. So we need to deliver fast. That's one of the problems. The second problem is that we're not the only ones who got the memo. A lot of companies, a lot of our competitors, the best technology companies out there, understand the business opportunity. That means there is no real competitive advantage. We don't have more capital than our competition, and we don't have a better secret sauce than our competition. Everyone is cooking with the same water, which means competition is fierce. It feels like a race; it's brutal out there for us. That's why I personally spend a lot of time figuring out how we can be faster and deliver faster, because I believe our intangible long-term competitive advantage is engineering and product operational excellence. We need to be faster than our competitors. We need to meet customer needs faster than some of the large players out there — FAANG or the Magnificent Seven, who are also in this game, who have better market access and a lot more technological foundation in place. So we very quickly got on the bandwagon of AI-driven development. In April this year, 97% of our software changes were already built by AI. So we quickly claimed that we are an AI-native company. If you look at how quickly AI generates code, you would expect overall delivery speed to scale roughly linearly. What you see is AI generating code something like 8 to 100 times faster than a human. But if that's true, where is the throughput? Why don't I see that we are eight to ten times faster as a company, as an overall product organization, than before? The reality is that our increase is there, but it's marginal. When I looked into why it's marginal, I understood that software development is a team sport. Software development is a multi-stage play, and coding is just one stage in it. We accelerated that stage, so we can talk about generating code and about increased throughput in code generation. But what's happening with requirements specification? With integration, acceptance testing, deployment, product enablement — so that the go-to-market teams also understand how to bring this increased feature set to the market? And then how do we do improvements, how do we do maintenance? That's still a question. Engineering organizations, through the DORA metrics, already look at delivery throughput: what is the velocity to deliver valuable changes to the market? But as a product leader I also have to ask myself: what is the business throughput? The role of product is to anticipate customer needs, and the question is: if we have a hypothesis about an anticipated customer need, how quickly do we turn it into business value — retention and cost reduction for our customers, and strategic progress for the company? This is becoming a systems problem. Based on new research, what we see is that AI acceleration is marginal: roughly 30–32% on a company's overall throughput. If you look into the details, there's no surprise there. AI accelerated the implementation stage. If you compare the stages before and after AI, implementation went from five days to half a day, but product definition remains the same, and architecture design remains more or less the same. You've heard it before: AI is not substantially accelerating the human interactions, the product discovery, the consensus-building part. AI can also generate bigger code diffs, which you need to review, so more review is happening. Then more rework is needed, and the quality of the software and of the features can decrease, which can eat into that 30% margin. So you can actually have negative acceleration as well. There is research showing that developers felt faster on some occasions, but when measured in absolute terms they were slower, because generation was faster but everything that came after — testing, cleaning up the results of vibe coding — slowed them down. So what do we do about all this? What's the lesson? How can we still use the benefits of AI to accelerate the overall organization, and how should we really think about AI-native companies? I would argue that if you just generate code with AI — even high-quality code — you're not yet an AI-native company. To become one, we have to think about AI as a tool and think about how we reorganize ourselves as a team: our processes and our social and cultural setup, to make good use of AI. To go deeper, I want to ask you to stop for a moment, take a step back, and go through a story. It's June 1943, and there's a test flight — not of a jet fighter, but of a piston-engine fighter — at one of the airfields in Miami. An army general approaches the chief engineer, Kelly Johnson, and says: "Intelligence says the Nazis have released their jet engine, and that can change the fate of the war." Jet fighters were two to three times faster than piston-engine fighters, and this would completely change how air fights played out. So they needed to do something. There's another thing: the US Air Force at that point had a jet engine, but funnily enough it was half as fast as a piston engine. So it was actually a failure. It existed, but it didn't work; they used it only for testing. So you have this immense pressure and immense competition, and the general says: "Can you build me a jet fighter in 180 days?" 180 days is roughly five months. Imagine: some of us struggle to deliver a complete software feature in five months, let alone, in 1943, a completely new airplane. And not just a new airplane, but one of a kind that hadn't been built yet. You couldn't go and ask other engineers how to build a working jet engine, because the truly working jet engine was over there with the competition, in the Nazi camp, and I think information sharing wasn't so easy between those two camps. So you can see the challenge. Kelly is an ambitious guy, so he says yes on the spot. He was working at Lockheed Martin, based in California, so he flies back from the test flight in Florida to California. When he lands, his bosses are already on the tarmac asking: "What did you do? How could you agree to this?" They repress the urge to fire him and tell him: "You made this for yourself. You're on your own; figure it out." You could argue that at this point Kelly could have asked for more engineers, more money, more structure, more priority. But the only thing he asked for was direct access to the SVPs of the company and to the army generals. He walked through the factory and poached 23 engineers from different teams. These 23 engineers just needed to commit to his motto, which was essentially "get things done" — the words were different, but that was the main idea. And those 23 engineers were a third of the usual team size at the time. What they did was put up a tent next to the wind tunnel, and they worked in that tent, engineers together with the shop workers. So this was a very early form of cross-functional, co-located teams. It was still in the courtyard of Lockheed, but next to a plastics factory, so there was a funny smell. Some of his colleagues came to work in the mornings wearing a gas mask, signalling that the working conditions were not as amazing as we're used to in tech. That was also the main reason they called it the Skunk Works. You've probably heard about Skunk Works in other contexts since, but that's the origin of the story. He also put up a big scoreboard with a countdown and the text "Our days are numbered," because they had 180 days; that's what they had committed to. Every morning one of his colleagues decreased the number, so everyone had a sense of urgency. And they made it in less than 180 days. This whole Skunk Works setup ultimately resulted in many other special airplanes, among others the Blackbird, which you've probably heard of — the spy plane. So the question is: what made them successful, and what can we learn from that success? I wanted to understand the transferable learning from this story for our new AI-driven world. What I saw is that AI brought down the execution cost, but AI didn't bring down the coordination cost. Humans still need to coordinate, and that's less of a technical problem — or maybe it is to some extent — and that's where this story goes. Kelly Johnson essentially focused on reducing decision latency, reducing interface surface area, reducing requirements volatility, and on incentive alignment. Let's go into each of those one by one. What is decision latency? If I make a decision in six months versus two days, that already gives you the difference. What he really looked at was the wall-clock time between a question arising and an answer existing. Co-location helps there. If the engineer and the workshop technician are on the same floor, working shoulder to shoulder, the technician can ask a question and the engineer can answer immediately. The same is true for the leaders. He was very keen to be there with his team every day, to do a check-in, and to cut down misunderstandings and the correspondence needed, so they didn't have to speak over multiple heads and relay messages where a lot gets lost in translation. He also wanted to almost viciously reduce the number of participants. He was really keen not to overload the team — not to just hire more engineers to deliver faster, but to make sure there were as few people on the project as possible. The next point is requirements. They were very rigid on the aircraft's upfront specification. It was very clear what this aircraft needed to do, what the acceptance criteria were, what good looks like, and how it needed to operate once done. And then they said: the engineers designing the thousands of little solutions can be scrappy. They can make a little drawing, try it out, and if it doesn't work, throw it away and redo it. The whole idea was that you can experiment continuously in the implementation while adhering to very rigid, very well-defined, high-level acceptance criteria. He also wanted simple systems. He really asked people to draw simple back-of-the-napkin drawings, so you move very quickly. The last point was incentive alignment: making sure everyone is really connected to the same goal. "Our days are numbered" was one message that aligned a lot of people around the urgency. The second thing was: "If you want to grow, I will pay you based on your impact, not on how big a team you build." So no one had the incentive to grow teams, or to grow any metric other than the end goal. If you think about it, AI itself isn't moving the needle on any of these metrics. But if we get these metrics right, we can accelerate what AI is capable of, and AI might help us here and there. So let's look a bit deeper at what this means for organizational structures for product and product engineering teams. First, what we try to avoid is communication overhead and interdependency overhead. Not having to negotiate decisions with five other teams would be the end goal, the ideal, even if we can't reach it. One way to reduce communication is to identify the contexts and context boundaries in your software, and build teams so that they own the complete context within a given boundary. That means they can make all, or the majority, of the decisions within that context. This could be a product feature, a major product capability, or any structure where you can clearly draw boundaries and give the team total autonomy to call the shots while the principles are clear. This first one is already well known, at least among engineers: it's called the inverse Conway maneuver. It comes from Conway's law, which says essentially that you will ship your org structure — the software architecture will look like your organization. But you can also invert this law and say: wait a second, let's first figure out how the software should look and what its modules are, and then design the organization so that it represents that software architecture. That helps minimize interdependencies and communication in the structure, and it's already implemented in many companies. The next thing that's important is a dedicated decision SLA. How many of you measure the decision or response time of the product manager? How many of you measure the response time of decision makers in the company? I can tell you that once you start measuring it, you'll accelerate a lot more. Suddenly I don't have two or three weeks to organize the next meeting for this decision; I have half a day, or half an hour, and that creates urgency. Besides urgency, it helps teams move forward. What's interesting is that if a team is blocked by a decision, then five people and 15 AI agents are not doing the next step; they're not delivering the next value slice. Or, because they have some other pressure, they make assumptions, and if they make assumptions they might work on things that won't be needed in the future. The next item we need to fix, if we want to learn from Kelly, is the so-called requirements volatility problem. We famously say in product that product management has two major goals: figuring out the right thing to build, and once that's clear, building it right. AI doesn't really help us with either of those. AI just helps us build the wrong product in the wrong way faster. That can feel busy and it can feel like progress, but it's just that: we're accelerating the rework. That's why we need to spend a great deal of energy figuring out what the right thing is and how to build it right. There's some research, the Orchid benchmark, that provides an interesting insight on this. The question was: what happens to code generation with LLMs if the requirements are ambiguous in ways that are specific to humans? Why is this interesting? We do a lot of benchmarking of LLMs' ability to write code and check its quality, but those benchmarks are based on very high-quality, very precise requirements, which isn't what happens in day-to-day operations. That's the ideal scenario: you can measure a model writing extraordinarily good code at high speed based on a high-quality specification. The real question is what happens when the specification looks like what we see every day — ambiguous, unclear, with some contradictions. And they found that AI models are twice as bad at generating code when the requirement specification is ambiguous. By the way, this is not specific to AI agents. We've known since the 70s that 55 to 70% of software bugs come from specification errors and specification ambiguity. So humans had this too. But in the human world it was less of a front-and-centre problem. With AI, because you can now have multiple AIs and burn a lot of tokens on all of this, you'll ask yourself: is the end result worth what we paid in tokens? We should have asked the same question before: was the end result worth the engineering salaries we paid? But back then it was less of a systems problem and more of a social problem. Today it's more of a systems problem. So how does that change the way we think about requirements? What were the old economics versus the new economics? In the old economics, code was expensive. Code was the precious thing the engineering team built; we argued for months about scope, and the codebase was the asset. But what if the codebase is a throwaway thing? What if we can just say, "If it's not okay, let's throw it away again in half an hour"? In that moment, the high-quality specification — the decisions and the consensus you needed to build in the organization — is the asset, because if you have that, you can regenerate the code over and over again. Code is cheap; we discard it freely, and we can throw away all kinds of iterations on how to solve the problem. That prompts the following workflow, which I'm recommending for AI-native companies. First, as we did before, we agree on the business value of the new feature. We want to make sure we're delivering value slices to our customers, and we define the success metrics: once we build this feature, how will we measure that our hypothesis about the value slice is still standing on its feet? Second, we define the use cases, because at the end of the process we want to validate that whatever we built is fit for purpose, and the purpose needs to be defined. The purpose of the software needs to be very clear to everyone — not as one sentence, but as a couple of use cases, so I can validate that it's still real. This can take the form of use cases and acceptance criteria; none of that is new. What is maybe new today, and what was a revelation at least to me, is that instead of typing your conclusions into the requirements specification, it's more valuable to provide the decision rationale: why did you reach that conclusion, and what is important to you? Sometimes the data points and inputs change, and new data points would lead to new conclusions. If you want to give autonomy to your teams, and you want them to be able to continuously iterate the software, its configuration and its specification in the right direction, they need to understand what's important on a higher level. They need the decision rationale, so they can make new decisions in light of new information and still arrive at what is important for your business. The next thing, and yet another change I'm asking our product managers to embark on, is to start vibe coding a prototype. A prototype is a throwaway thing. We don't expect high-quality, secure, resilient, reliable software. What we expect is that the product manager surfaces the unknown unknowns through this process. The LLM will ask clarifying questions about things that were implicit in our specification but not explicit. It makes sense to us, but it might be ambiguous to the LLM — and if it's ambiguous to the LLM, it's also ambiguous to humans. And if you have an ambiguous specification, the results are unpredictable in quality. That's one reason why coding is important. The second reason is that throughout the prototype implementation I might face all kinds of questions or ideas I hadn't thought about, so I can increase the value of my system, and I can iterate freely because it's relatively cheap — certainly not as expensive as five developers over six months. The results of this vibe coding can, on the one hand, inform my use cases and my decision criteria. On the other hand, I can derive very detailed specs for the AI agents from the vibe-coded prototype. Then I can generate the tests, because I will validate the software against them, and then generate the code and the artifacts. And we finalize with acceptance testing. As a product manager, I need to see whether what we built together with the AI agents and the engineers is still fit for purpose and still usable. There's no way around it, at least not as of today, other than really going through the motions and checking whether the software does what we want it to do. In acceptance testing I'm still checking whether the value I promised at the very beginning — the value hypothesis — is there. Lastly, we also need to think about how we fund these projects. I don't know how many of you are already at the stage where your CFO, or someone in your organization, asks: "How much are we spending on this token thingy, and by the way, what is our return on this expense? I clearly see the expense, but I don't really see the return." How do you measure that? Am I measuring lines of code? AI is just generating lines of code. How do I measure return on investment? Here are a couple of things. On one side, we're putting in developer hours and tokens. On the other side, we're looking for return on investment, which is a lagging indicator: one year from now I can see that our NRR increased, or our conversion rate, or our NPS. But that takes time. So what do I measure as a leading indicator, as an input metric? One approach is to measure releases that actually reach production — not story points, not velocity. Then the idea-to-production lead time: from the moment we have a value hypothesis until it goes into production. Then the share of time spent waiting: how long it took to get from the idea to the PRD, from the PRD to the implementation. And not just pull request and lead time, but time to an agreed, fixed specification. This puts a bit of pressure on product managers — not the time to the first draft, because we can use AI to generate that, but the time it takes to go through the motions, align, and create multiple versions of the specification. And lastly, as a control number, we need the change failure rate, because what we're seeing in AI-driven development is that AI-written software isn't as high quality as human-written software was before. We now see a lot of reports on this. At the beginning of the year, Amazon, Meta and many other companies were bullish about how they could accelerate AI-driven development. Now we see them walking those stories back. Amazon said they had so many incidents this year that they hadn't had in the past 20 years. They didn't say it like that — it's written much more nicely — but that's the gist of it. They say they need to bring back friction, which means bringing back people to actually review what the AI generates. The same goes for Meta, who planned larger multi-stage layoffs. They executed the first one, realized it wasn't working and that they need someone to work with the AI, so they cancelled the rest. Most of these topics come from realizing that AI can accelerate the generation throughput, but also the bugs and the incidents. We still need to figure out how to deal with that. So if I want you to have one single takeaway, it would be this: AI accelerates certain select stages, but the value creation velocity of your product organization and your company will only be as high as the slowest stage allows. If your decision-making is slow at the beginning, and it takes six months to get to alignment and a conclusion with your team and your executives about what to build, then you can generate code as fast as you want later: it will still be six months plus the one or two days of code generation. Overall speed remains a (human) system optimization problem. Why did I put "human" in parentheses? Because going forward there is a mélange of humans and AI agents, and the system we're building together is essentially a hybrid system. We need to keep optimizing for AI agents and humans working together shoulder to shoulder. Thank you so much, and if you have any questions, I'm open for that. #### Q&A **Host (Matteo Gamba):** Thanks, Csaba. Do we have any questions for Csaba? **Audience:** [Question not audible on the recording; it concerns how product managers can prototype without depending on the engineering team when working in a complex existing system.] **Csaba Tamas:** If you remember, in this presentation I said we should reduce communication overhead and the interface surface area. That means it's not desirable to have to work together with everyone — not because we don't want to collaborate, but because speaking to people means slowing down, since humans need to align and agree. So the first intent is to avoid the dependency on the engineering team. Now the question is: I have a complex system, and I might not have a complete understanding of it, so how do I vibe code around it? That can be done once I've clarified these little value slices. I handle complexity by breaking it down: I have a complex system, but I can break it into simpler components, and I can vibe code and prototype on the simpler component I'm touching right now. I need to be very clear about what my feature is and where the interfaces around it are, and I can build a prototype that isn't cross-cutting across the whole big complex system but is specific to the area I'm working on. Whether your system is designed for that today is a different question, because in engineering we talk about the "big ball of mud": a system that can't be sliced or modularized because it's intertwined, and technically that's not feasible. I would still argue that when you're doing prototypes, doing something in isolation on your computer, you can at least create a logical boundary for yourself. That's where you create something new, and you don't depend on anyone. When I'm talking about the prototype, it doesn't mean it needs to work or interact with your existing software. It's just a prototype of the new feature you want to build, and you can freely create that boundary for yourself. So, long story short: I would minimize interacting with engineers at this stage — not avoid it, but maybe have a very short conversation to collect some feedback without depending on them — and then create a very small slice for myself and prototype it. **Audience:** You mentioned that product managers are basically vibe coding now. What I'm wondering is how that changes things for engineering, because there are a lot of things like product engineering showing up, and also decisions. I wonder how you are handling that right now. **Csaba Tamas:** I'll abstract away from right now, because we're evolving, and speak about the vision. There's yet another war story. Ford was asked to help aircraft manufacturing become more effective during the Second World War, because they weren't producing as fast as they should. One thing they realized was that there was also ambiguity in the requirements specifications. Engineers weren't defining everything in detail, and the shop workers filled in the gaps. In one of these big bombers they found that nobody had ever designed the toilet. So the shop workers went across the street to a hardware store, bought standard items, and built the toilet on the bomber the way they figured it should look. I think this will continue to be there. On the one hand, I'm asking product managers to write better, less ambiguous specifications, but the ambiguity will still be there because we are humans. On the other hand, I want the engineers to become more product-minded, to be able to cover more of these gaps. Because you're right: the pressure on product managers is increasing. Before, product managers could work with a backlog, because engineering was always slower than product management. Nowadays that dynamic is changing. So the ratio of product managers to engineers is probably changing as well, because yes, I might work with two engineers, but if they use 15 agents, that's like working with a team of 17, and we need to be prepared for that. In that sense I think we need to meet halfway: engineers need to come a little in the direction of product managers, and product managers need to go a little in the direction of engineers. **Audience:** [Question asked without a microphone, about when or where the proposed workflow will break down.] **Csaba Tamas:** So the question is when or where this workflow will break down. I don't know. I'm a big fan of experimentation and a scientific approach, meaning we make a hypothesis, run an experiment and validate. This proposal is something I want to implement right now, and then I can tell you in three to four months how it went. For now, I don't know. **Audience:** [Question not audible on the recording; it concerns how Parloa organizes itself to keep decision response times low.] **Csaba Tamas:** That's a tricky question, in the sense that we are continually iterating on it. One thing I try to do is keep decision response time as low as possible, but the question is whether that's the case everywhere. Since last year we've also designed for a very flat hierarchy, and we're trying to keep it as flat as possible. We ask engineering managers to manage up to 12 or more engineers, so that we don't need very big communication channels. We ask product managers to work shoulder to shoulder with the engineering manager and the tech lead, so that this trio can provide an immediate response to everything. We also still do quarterly planning ahead of time. The idea there is to lock in the big cornerstones of the quarter, and then let the team autonomously figure out the everyday decisions within the boundaries of that quarterly plan. There's a lot more to improve. We're far from having all the decision velocity we want to see, but we have substantially reduced decision latency from many months to a couple of weeks, and I would like to see that at a couple of days, or hours, in the future. There are still some optimization steps to be made. **Host (Matteo Gamba):** Thanks, Csaba, and thanks for the question. ### 11:10 · How to make LLMs love your product — Malte Landwehr Video: https://youtu.be/jMQ5I3CNhpM - **Speakers:** - [Malte Landwehr](https://www.linkedin.com/in/landwehr/) — CPO & CMO, Peec AI. Peec AI reached 1,000 paying customers, $5M ARR and $29M in funding in under 12 months. Previously VP SEO at idealo and VP Product at Searchmetrics. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 11:10 CEST - **Format:** Talk · Track: Build & Craft - **Host:** [Matteo Gamba](https://www.linkedin.com/in/matteogamba/) (ex-Wayfair) - **Official abstract:** LLMs infer products from signals, docs, and reviews — not demos. Learn the PM playbook for getting found and recommended by the models your users trust. **Summary.** Malte Landwehr, CMO of Peec AI, explains how LLMs decide which brands to mention and which websites to cite, and which levers companies can pull to influence that. He focuses on grounding (the web searches LLMs run behind the scenes), walking through content formats, tactics that currently work (and some that are risky), attribution problems, and how brand perception differs depending on how the question is asked. **Key takeaways** - Click-based attribution badly undercounts LLMs: one agency's analytics showed 0% of new customers from LLMs, but 21% said they found them via ChatGPT when asked directly. Malte calls this "dark chat". - Grounding (the web search LLMs run when their model knowledge is not enough) is the easiest lever. SEO is the foundation, but small details, like a paragraph listing competitors on your homepage, can decide who gets recommended. - Study the fan-out queries LLMs generate and write content that matches them. Keep facts consistent across pages, have some core content in English even for local markets, and write explicit, declarative, entity-dense text with summaries and FAQ formats near the top of the page. - Self-promoting listicles used to work very well and are now being discounted. Comparison pages, advertorials, a machine-oriented "AI instructions" page and a reviews page on your own site currently still get cited. - Mass-produced AI content tends to follow a "Mount AI" curve: visibility rises, then crashes. - Brand perception must be measured from both directions. LLMs link Tesla with "fast" in 96% of answers, but recommend Tesla for "fast cars" in only 6%. #### Transcript So, let's see if this works. Yes, it does. Okay. As we already heard, I'm going to talk a little bit about how to make LLMs recommend and love your product. I'm the chief marketing officer at Peec AI. We do this for about 3,000 clients right now. The company is two years old. Maybe you've heard about us or seen some of our advertising in and around Berlin. I want to start with a story from last week. It actually happened, it's not made up, and it fits perfectly with what I want to talk about. Very recently we introduced support for Naver AI Briefings. This is the little AI summary you get if you're in South Korea and you use Naver. It was relevant for a couple of clients of ours that we also track how they appear in there. So we built the parser, tried to roughly understand a little bit of the Korean language, made sure all of our brand recognition and sentiment recognition supports it, got some local IP addresses in South Korea, and we were done, right? Unfortunately not. When I asked any of the LLMs, "Hey, does Peec AI support Naver AI Briefings?", all of them said no, they don't. So I went to Reddit, to the Peec AI subreddit, of which I happen to be a moderator. I asked myself the question, "Does Peec AI support this?", and I wrote the answer in there. The very next day, the majority of LLMs said yes, Peec AI supports Naver AI Briefings. I want to share this to show how simple it sometimes can be to influence how LLMs talk about your brand and your product, and that's what I'm going to talk about for the next couple of minutes. Quick question: can everybody raise their hand who gets at least 10% of their users from LLMs? Nobody. Okay. Can anybody raise their hand who's actually tracking how many users they get from LLMs? Ah. You really should track this. Okay, topic one: clicks versus leads. In Google search, the term "zero-click search" has been around for a long time, because even two or three years ago, 100 Google searches would only lead to 40 clicks on external websites. Roughly 30 clicks went to Google properties, and on 30 searches nothing happened. So a 60% zero-click rate. Now that Google has these little AI summaries on top, the AI Overviews, we are seeing a 70% zero-click rate. And if we look at Google AI Mode, where you're actually in a chat interface, it's 95%. So instead of 40 clicks, only five clicks are left. For ChatGPT there is very little data on this, but all the studies I saw talk about roughly a 99% zero-click rate for search-like prompts. So instead of 40 clicks, there's one click left for the internet. That is of course a problem for some business models. The good news is people still spend money and buy products. They just don't click on links anymore. And there's a little bit of hope that the click rate will increase, because people of course complain about this. If all the media write about how evil it is that they don't get traffic anymore, at some point the LLMs might send a little more traffic in order to avoid lawsuits and a ton of bad press. So there are not that many clicks. Why do we care? I have some data from a client of ours, Flow Agency. It's a marketing agency, and they looked in HubSpot, their web analytics system, and checked how many new customers they were getting from LLMs. The answer was 0%. But they were very convinced that this couldn't be true. So they started asking everybody who signed up, "Hey, where did you hear about us?" And actually 21% of their customers said, "Oh yeah, I found you on ChatGPT." This data is a little bit older now; Gemini and Claude have taken a bigger share since. But if you do click-based attribution and just look at HTTP referrers and UTM parameters, you're going to have a really hard time attributing LLM traffic. I see LLMs like brand marketing or word of mouth: somebody gives you a recommendation, but you often cannot immediately click and buy the thing. You either type it into your Chrome address bar or you search for it on Google, and attribution-wise it ends up somewhere else. I like to call this dark chat or dark search. When I was young, you would still get traffic out of Facebook. I don't know if anybody remembers this, but it actually happened. Back then the Facebook app didn't add any UTM parameters, so that traffic ended up in direct or brand, and everybody tried to attribute it as "dark social". I think with LLMs we are in a similar situation. I also want to quickly talk about two terms: visibility and being the source. This is a typical chat, and you can also see when I actually finished these slides, which was yesterday. There are brand mentions, right? The Productlab Conference is mentioned, but I can't actually click it. Even if I now want to go to the Productlab Conference website, I can't do that. So whenever I say brand mention or visibility, I mean this: the LLM is mentioning or recommending a brand, but it's not necessarily connected to a website. And then of course we also have the sources the LLMs are using. When I talk about sources or citations, this is what I mean. In theory you could click them, but very few people do, because it doesn't really fit the user journey. In order to influence what LLMs say, there are four levers, or attack vectors, we can use. The first is the foundational model. It's trained on data, and we might be able to influence that. The big problem is that it takes forever, because these trainings are not done all the time. Even if a new foundational model comes out today, it wasn't trained on yesterday's data; the knowledge cutoff might be six months or almost a year in the past. So it might work, but it's very hard to test and doesn't get us results quickly. Second, user context: everything the LLM knows about the user, which is primarily derived from their chat history, but potentially also from things they connected to the LLM. This can influence how it answers, but it's difficult to influence at scale. I'm going to show one or two ideas there later, but it's not our main avenue. Third, the prompt. Of course, if we change how people prompt, we could in theory make that very positive for our brands, with out-of-home advertising, TV advertising, whatever. Not my cup of tea, so I'm not going to talk about it. And the last one is grounding. When LLMs can't answer with their foundational model, they do a grounding process, or RAG, or tool calls, or web search. There are many different ways to talk about this. In the end it means they look at your prompt, come up with a couple of search terms, put them into their own search index or into Google or Bing or Brave, and then use the documents they find to produce the answer. This is the easiest thing to influence. It's very similar to SEO, but there are some differences, and I will mostly be talking about this part. Of course, almost everything you do to influence grounding will also influence the foundational model in a year or so. And as I said, I will say a little bit about user context. So let's start with a simple question: is this just SEO? No. This is a case from a client of ours, a financial services company from the US. They've been around for a couple of decades. They rank very well in Bing, they rank very well in Google. When they made an account and set up a couple of prompts, they reported a bug, because Perplexity said that for one set of prompts they had 0% visibility. They were never mentioned. "This is not possible, it must be a bug." So we looked at it together, and the typical answer looked like this: about 10 recommendations of competitors. Wells Fargo was there, Morgan Stanley, and a couple of lesser-known ones that I had to hide in order not to reveal our client. And our client's website was actually the source for a lot of these. What happened is that the LLM did what it's supposed to do. It looked for the most relevant documents on the internet, looked for text snippets in those documents, and used them to come up with the answer. What our client did very well was convincing the LLM, via Google, Bing and other search engines, that they were the most important document. But on their homepage they had a little paragraph saying, "We are pretty elitist. We are not the right solution for everyone. Here are five competitors you can also trust." That was the passage the LLM used. They just changed the headline a little bit and put themselves at number one on that list, and the very next day, when we simulated the LLM conversations again, they were the most recommended brand. So this shows that SEO is in many cases the foundation, but also that very tiny, very stupid mistakes and tweaks can change a lot. One thing that works is content. A case study with Momentum, a client of ours: when we started measuring their visibility by simulating conversations with LLMs, their website was used as a source in only 3% of the chats. So they created expert content and published it, and by the end of the period, I think it was two or three months, 60% of LLM chats used their website as a source. And since they spoke positively about themselves on their own website, their share of brand mentions also went from 3% of chats to 34% of chats. So roughly a 1,000% uplift, and they saw increases in bookings, revenue, clicks and so on in this time period. Now you might ask yourself what content to create. The good thing is that the LLMs give us a lot of advice, because when they perform these fan-out queries, we can actually see them. They are a little bit hidden, but you can see them in the code. So this is the prompt, "What is the best CRM software?", and these are the fan-out queries ChatGPT was using. It added the word "official". It added the word "2026". Here it added a lot of competitor names and the word "comparison". It said "small business" for some reason. And here it added a site search, so it only looked for results on hubspot.com about pricing, and on Salesforce it looked for pricing information. All of this we can use to create content. A very concrete example: I have the prompt "What are the best software inventory tools?", and the LLM adds "IT asset management" to the fan-out query. So if I describe myself as a software inventory tool on my product pages, in my press releases, on my homepage, and in the bio of my CEO and founder that I use when I send them to give an interview, I should maybe use the words "IT asset management", because LLMs seem to think this is related. Of course, this is from one single chat and one single fan-out query. We should do this in bulk and then look at trends. And again we see "2026", so I might want to add a "last updated" date and make it a very recent date. This one is interesting. If I'm HubSpot, this fan-out query tells me that on my website ChatGPT is looking for "CRM pricing Sales Hub 2026". So I should make very, very sure that I have information about this on my website. And if I have multiple pages about this topic, which I probably do, from the help center, product docs and press releases, I should make sure they all say the same thing, because when LLMs do grounding, they look for consensus: multiple sources telling them the same thing. There's another issue with fan-outs. Here I did a prompt in Swedish from a Swedish IP address, and some of the fan-out queries are in English. This is not random. It actually happens quite often in ChatGPT: it doesn't matter what language you use or where you are located. They just assume, or maybe know, that there is more information on the English-speaking internet, so they also pull in English information. So even if you are a company in Germany selling only to the German Mittelstand, and Manfred and Klaus, your target ICP, only speak German, you should still make sure to have a little bit of content about your brand and your product in English. That doesn't mean you should run home and translate your whole website into English. But maybe your homepage, your about page, your product pages, maybe the frequently asked questions should be in English, and maybe motivate the CEO to also give an interview in English every once in a while. This is probably not the room where this advice is important, but I'm sometimes at events where it's super crucial, because nobody even thought about communicating in English. The fact that LLMs use English fan-out queries can also lead to situations like this. Here I ran a couple hundred prompts a couple dozen times in the travel space, and Expedia appears in the sources three times: the German and Austrian websites came up for the German fan-out queries, and the US website for the English fan-out queries. With Google, we have this little thing called hreflang, where I can explain to Google, "Hey, I'm the German website in German, and here is the Austrian website in German." LLMs don't care about that. They find content, and they use it. For Expedia this doesn't really matter, because hotels and flights are the same everywhere; the prices maybe change a bit. But we have clients who are car manufacturers. In Portugal and Spain they sell cars that meet the needs of European consumers and the laws around car safety, so they all have side airbags. But in South and Central America they sell a different car under the same brand and the same model name, because the car has to be significantly cheaper and the safety requirements are different. So they have a Brazilian or Mexican website that talks differently about the product. LLMs don't care. They regularly use the wrong language version to answer questions from European users. So if you have products that differ across languages and countries, or maybe a higher price for US users, this can actually become a problem. It should be monitored. What kind of content should you create for LLMs? We can do the very basic thing of looking at content that is regularly cited by LLMs and content that ranks in Google for similar terms. Then we see: a proper semantic headline structure, so H1, H2, H3, seems to be rewarded by LLMs; having lists; using structured data and schema markup, especially FAQ markup. Very important are self-contained text chunks. When you take a text chunk out of context, it no longer makes sense if there are sentences in it like "For the above reasons, the third option is the best for this person." Out of context, that has zero information. So entities need to be named very explicitly. LLMs also really prefer text that is authoritative, definitive, declarative: something that says "this is the best solution", not "this could be the best solution" or "this might be a viable option". So definitive text chunks are really helpful, and entity-dense content works really well. If you think about it, most information, if you break it down, is actually the relationship between entities. If your text doesn't mention the entities and you take it out of context, it will not answer the question. That also means that on your product documentation, instead of writing "we support this, we support that", you could consider writing "Peec AI supports this, Peec AI supports that". I know it reads horribly for humans, but unfortunately it actually works. For the content, it's also very important to have semantic focus. If you have a product page about how you are a globally multi-language product, just say that. Don't put two paragraphs on top about how your founders met traveling around the world and decided you must be multilingual from the beginning. That's a nice marketing story; you can put it somewhere else. The LLM wants pages that are about one topic; otherwise it gets confused. What also works really well is having a summary at the top of pages. In a summary, super condensed, simple, declarative sentences don't read as cringe, and people will usually forgive you for it. It also forces you to write in this very condensed way. And, as I mentioned earlier, the question-and-answer format works, because again you force yourself to answer a specific question. Listicles. This used to be the most embarrassing part of my presentations. Now it has at least become a little bit better. We go to ChatGPT and ask, "What's the best CRM?", and it says HubSpot and Zoho, and the sources are HubSpot and Zoho. Right now it's a little bit better, but a year ago it was literally like this, and a year ago the best thing you could do was write listicles on your own website and recommend yourself. SuperOffice wrote the listicle "The 10 best CRMs", and the best CRM is SuperOffice. Nightwatch wrote an article about the best SEO tools: the best SEO tool is Nightwatch, and the second-best SEO tool is the Chrome plugin from Nightwatch. Monday wrote an article about the best CRM: the best CRM is Monday. I could go on and on. Why did everybody do this? Why did everybody congratulate themselves? Because it worked. LLMs were eating this up. Especially in the B2B SaaS space, a lot of players noticed very early that LLMs drive a lot of, not traffic, but leads, and they did this at scale. For example, Blaze did this like crazy, and these are just some of their listicles; they have a lot more. Some companies have 200 or 300 listicles on their website, and they always put themselves at number one. And who is number two or three? They always mix it up, in order not to give too much preference to any one competitor. The craziest example is from an agency here in Berlin, Clio. They invented a matcha powder. It has no website, no reviews, no social presence. They just bought three listicles about matcha powder in large German newspapers, put their own matcha powder at position two, and LLMs actually started recommending it. So this actually works. The good news is that the Google LLMs, or rather the systems Google built around the LLMs, now seem to recognize these self-preferential listicles. It seems that if they take your listicle, they ignore you at number one and just use it as a source to recommend your competitors. And Claude and ChatGPT have started to sometimes say in the answer, "Be aware, these rankings are mainly built by companies listing and recommending themselves." So I would no longer do this on my own website. You can of course still do it on other websites like Reddit or Medium. That works surprisingly well. Something that works almost as well are comparison pages. Again, very simple: you compare yourself with a competitor on your own website. It still works very well right now. I believe that in the next six months it will be like listicles, and this will have to be on external websites. The craziest example, which I had to anonymize completely because it's a client of ours: I'm searching for "Red versus Green". These are two competitors, and one of the sources is the article "Red versus Green" on the Red website. That is obviously an article that will conclude that Red is the better one. And they went a little bit crazy, because all the other sources are also from their own website: they also compared Green versus Teal, Green versus Yellow, and so on. They made a ton of comparison articles. It's a very spammy approach. I told them I would not do this. But they are maybe top 10 in their space, and in terms of LLM visibility they are number two. And they said, "I know this is high risk, but it's worth it for us." Because even if they get penalized, what's going to happen? They'll be number 10 again, which they would be anyway. So for them, it was worth it. And everybody's doing this. Dropbox has a "Dropbox versus OneDrive" article. Moosend is even sneakier: they have a "Klaviyo versus Mailchimp" article, and very importantly, they only used honest opinions, and they specifically tell us this. Why would they compare two competitors? Because when you scroll down, it says, "By the way, there are also alternatives to these two," and coincidentally Moosend is number one. Again, I'm not recommending this. It will probably stop working soon, but for now it works very well. If you are a more serious brand and you still want this comparison format on your own website, you can do what Apple is doing: just compare your own products. That's still likely to be picked up by LLMs when they look for comparisons involving your product, and if humans see it, it's not cringe. Nobody thinks you're manipulating anything. Or of course you can be ClickUp and compare Google and Microsoft and something else from Google with ClickUp, and so on. They are one of the more aggressive players in the field. Advertorials. For the non-Germans in the room: this word in the title of these websites basically means advertisement. Bild has this directory, /partner/brand-story, and another publisher has a product story section, I think. Every single article in these directories is an ad. It also says so very obviously; it's even in the title. It couldn't be more obvious. But this screenshot is a list of sources I'm seeing right now when I ask LLMs about a certain subset of insurance in Germany. So LLMs are still citing advertorials. If you have too much money, it's one of the easiest ways to get something into the LLMs. In Germany, Bild and Handelsblatt work very well. There's also Future internationally, who offer this. It's a very easy thing, and it's even marked as an ad, so there's nothing sketchy about it. By the way, there are also many websites willing to do this in a sketchy way, where they don't even say it's an ad. I've been tracking this for one year now. For each LLM there's always some fluctuation, but overall, for the past year, roughly 3% of the sources for this subset of the insurance space in Germany have been paid advertorials. They are so obvious that my Claude Cowork routine can recognize them as advertorials; there's not a lot of magic going on there. llms.txt. There are so many things: WebMCP, ACP, .md, .well-known, llms-full, markdown negotiation. You've probably heard of all of these things that are supposed to be the magical solutions for all our problems with LLMs. Some of them might be really great for agentic workflows, when an agent is supposed to actually explore, use and buy your product. But if we are talking about simple search, somebody saying, "I'm a dentist in Germany, which content management system should I use?", none of these work in any way whatsoever. That doesn't mean I'm against using them. Their focus is just the agentic web: agents acting on your behalf, maybe agentic browsers in a half-agentic way. But when it comes to people just asking LLMs questions, they do not work at all. One little exception there. We created an AI instructions page, just called peec.ai/ai-instructions, and it's linked in the footer. It's really not written for humans. It's written in markdown, and every single sentence is written in this perfect way: "Peec AI supports this, Peec AI supports that." We don't say "Peec AI also supports this", which might read nicer for a human. Everything is as simple as possible, 100% declarative, and the entities are mentioned in every single sentence. If you read this as a human being, you would think, "What is this?" But we released it, and the very next day LLMs started citing it. Right now it's the second most cited part of our website after the homepage. So LLMs really gobble this up. And if you call it "AI instructions", I think humans who find it will also think, "Yeah, okay, makes sense." It worked very, very well for us. Grok. I know everybody ignored Grok for a long time, but it's at least a top-six LLM in terms of users. And with the Grok bot, and Grok being used more and more for vibe coding special user interfaces, I think there could be a bit of a Grok revival. If you want to influence Grok, X is the place to go. I found an example: there was one source on X, a tweet written by me. Sorry, a post; it's a post on X now, not a tweet. There were 25 web sources, two of them agreed with my post and 23 disagreed, and Grok decided to quote my post almost verbatim. So if you want to influence Grok, that seems to be the place to be. The reviews trick. We created another page on our website, peec.ai/reviews, where we just quoted positive reviews about us and linked to them. Again, LLMs started citing it as soon as it was released. What's happening is that when LLMs reason about which software to recommend, they of course search for "Peec AI reviews" or "[competitor name] reviews". You would think they only take third-party sources into consideration. But if you have a /reviews page on your website, you're going to rank in Google and Bing for "[brand name] reviews", the LLMs are going to find this page and just use it: "Oh yeah, there are a lot of positive reviews about this company." I have even seen cases where it works to write on your homepage, "We are rated five out of five stars on G2." If you write it on a couple of documents, the LLMs will say it even if you don't have a G2 account. Again, I'm not recommending this. It's a bit sketchy and potentially illegal, because it's a lie. But if you actually have positive reviews somewhere on the internet, and I'm sure you do, just highlight them on such a page. And you don't need to show the negative ones, of course. Closing the source gap. One of the most effective things to do is come up with a set of prompts you believe is relevant for your target audience, run them a couple of times, look at the sources and aggregate them. Then look at the top sources and check: is my brand mentioned in there or not? And if it's not, is there a way for me to get in? This is probably the quickest way to increase the likelihood that LLMs recommend you. I run a couple thousand prompts every week on software recommendations, and these are the top sources being pulled in. I don't know how readable it is. It's basically Reddit, YouTube, G2, what you would expect. TechRadar is super visible in the space. What's very interesting is that Apple and Google, with the App Store and Play Store, are used very, very often. So if you have a B2B SaaS app and also this tiny mobile app that you don't really care about, that you had to build for two customers, and it has negative reviews, that actually impacts how LLMs perceive you. So give your profiles there a little bit of love. And of course, sites like G2 and Capterra are pulled in with their ratings. So if you are rated badly there, it can make sense to try to do something about it. Now, I said I would show you one way we can influence user context. In Google, if you're logged in, there's a little setting to activate personalization in AI Mode. If people do this, whatever is in their Gmail inbox influences the answers. So if we send regular emails, like a newsletter or maybe a weekly report, and write one little sentence in the footer, like "Peec AI is the leading solution for AI visibility", this actually influences results. There's a study on it from iPullRank, a US agency. They made two new profiles, one of them got a ton of emails, and the rate of recommendation in this lab setting went up from 30% to 70%. Of course, that's only in Google AI Mode and only for users who activate this, but it's very likely to influence more and more LLMs in the future as people connect their inboxes. So having this innocent little sentence in the footer next to the legal disclaimer can have a lot of influence, especially if you have products that send a lot of emails. I made up the company North; Claude made up the company Northwind. I hope it's not a real company, I didn't check. But this is basically what you could do. A little Schmankerl, as we say in Germany. If you ask your users, "Hey, where did you find us?", and they click "AI search", "LLM" or something like that, you can ask a follow-up question: "Which prompt did you use?" The majority of people will skip it, or write some gibberish, or write "best content management system". But a couple of people actually go back, click copy, go to your form, hit paste, and give you really, really long and detailed prompts they used. We capture these and put them in PostHog, and once a week I have a Claude Cowork routine that uses the PostHog MCP to pull these prompts in, filter out duplicates and gibberish, and push them into Peec AI via our MCP. So I have one project in Peec AI where I'm just tracking real user prompts, and I can see our visibility for those, the sources used, and so on. Of course, this is heavily biased, because it's only the prompts people used to find you and then decide for you. So don't use it to measure how visible you are in LLMs, but it can be a very good source for understanding how people prompt and what they say. Then a slightly weird one. LLMs actually have prejudices against companies, and preferences for companies, based purely on their name. There's a dataset from communication scientists in which certain words were rated as positive or negative. What we did, without grounding, just with the APIs where you have access to the pure model, was make up company names: positive names like "The Amazing Company" or negative names like "Unethical Inc." We had the models write descriptions of these fake companies and measured how positive each description was. We ran this hundreds of times to get all the possible biases out. A company name that is 11% more positive actually leads to a 0.8% more positive description of the company. Of course, in the real world the LLM also does grounding and will find lots of positive information about companies like The Boring Company, Discord or Riot Games. So the effect is probably even smaller in the real world. But if there is very little information about you, this can literally influence how LLMs talk about you. We measured this in four different LLMs, so it's a general problem, or feature, that LLMs have. It can of course influence feature naming too, so product marketers can have an impact. I'm a former product marketer; I'm allowed to say that. Brand perception. Everything I said before was about being mentioned as a brand. But you can also ask LLMs, "Tell me about Tesla," and in 96% of cases, the LLMs said fast cars, or speed, or speedy. But when I asked LLMs, "Name me 10 brands that make fast cars," Tesla was recommended in only 6% of cases. So if you think about brand perception, how LLMs perceive your brand, you need to look at it from both directions. The reason this is so different is that if you ask about the brand, the sources are the brand's website or pages on other websites that are solely dedicated to this brand: a profile page, an interview, an article about the brand. But if you ask about fast cars, the sources are independent editorial websites, and the main content format is listicles. Since the LLMs use such different documents for grounding, the perception is also very different. So you might perform very well for one but not for the other. Now, I was a little bit fast, but I have a backup slide, I think. Yes, I will show it in a second. I'm currently hiring two product managers, one with an AI focus, so come talk to me. AI content. You might think, "Okay, we're just going to publish a lot of AI content now and win everything." There's now a bit of a meme in the community where we talk about "Mount AI", because if you publish a lot of AI content, what usually happens is that your visibility goes up, up, up, and then goes down again. No system in the world, not Google, not Bing, not OpenAI, wants to read, crawl, index and recommend content that it could have written itself a minute before. If they wanted AI content as a source, they could just write it on their own in that moment. There's a lot written on this; there are two big influencers in the SEO space, Glenn Gabe and Lily Ray, who regularly publish on it. I could also find a lot of examples, because I went to the website of what I believe to be the largest tool for AI content generation and looked at their customer stories, their case studies. So these are the customers where they thought, "Wow, this brand is using our product so well, we're going to put them on our website with a logo and a video interview." And I have good news: 40% of these customers did not lose their Google rankings. But every fifth customer on the list had this perfect Mount AI shape: visibility in Google and in LLMs went up and then came crashing down. And a third never even had the uplift. They just started publishing AI content, and down went their Google rankings. So do not do this. And one last thing: having your brand recommended by LLMs versus having your website cited by LLMs. Here I looked at a set of keywords. Visibility means the brand is recommended; source means the brand's website is cited as a source. If I'm over time, you can interrupt me at any time. Of course there's a market leader: highest visibility, most cited website. Looking at the top recommended brands, you might think recommendation goes hand in hand with how often their website is cited. But then there are brands that are recommended in 15% of chats while their website is never, ever cited. When this happens, one thing must always be true: there must be positive content about this brand somewhere on the internet. And there must be a second condition, of which there are several options. Either they don't have any content worth citing, or they have massive technical issues, like all the content being hidden behind JavaScript or the website taking forever to load, or they are blocking the LLMs, which in 2026 almost nobody should be doing. And the opposite can also happen: this website here is the second most cited website in the whole dataset, but its brand is never, ever mentioned. And now I'm really done with my presentation. As I said, I'm currently hiring two product managers, one with an AI focus. I hope you have some questions for me. Thank you. And I think I was supposed to tell you to leave feedback under the QR code. Scan the QR code. #### Q&A **Host (Matteo Gamba):** Thanks, Malte. Any questions? We have time for one or two questions and then we can go for the break. **Audience:** Thank you for the presentation. I have a question, because for my product there is never any web search. The agent always replies from memory. It's a bit fundamental, because it's Kotlin, and the question is something like, "Which multiplatform tech stack do you use for your mobile application?" Right now the agents just reply from their training. How do you debug that? Where do they get the information from, and how can you affect it? **Malte:** In some LLMs you can force a web search, or you could add something like "in 2026" to force it to do a web search, and then you would get to the sources. But if there is no web search, it's very difficult to influence the results anyway. There's no really great advice I can give you apart from trying to force web search, by looking for very recent information or by telling it to do a search, do deep research, or use a lot of sources. But of course that influences the results and introduces a lot of biases. **Audience:** And do you think changing this may change the next training period? **Malte:** It really depends on the topic. I would think that LLMs usually lean towards doing a lot of search when giving recommendations. But if they think it's more of an expert-advice topic or a role-play topic, they might just not do any web search. **Audience:** Thank you. **Host (Matteo Gamba):** There was someone raising their hand. **Audience:** Hi, thank you for your talk. My question: any advice on how to create a prompt library? What should we track, branded or non-branded? Where do we get the source? Should we research keywords, understand the customer? **Malte:** The good thing is that it almost doesn't matter how you word your prompt. We did a study with a partner where we asked 144 people to write down a prompt for how they would look for noise-cancelling headphones. Everybody wrote a different prompt, but for 90% of them the LLMs recommended the same brands. So I wouldn't be overly focused on exactly how people write it, because most prompts are unique anyway. Even if you know one of your customers used a specific prompt yesterday, it doesn't really tell you anything about tomorrow. So if you have keyword data from Google Search Console, common questions from Intercom, recorded Gong.io sales meetings, or customer interviews, just extract the words people use to describe their problems and goals, and create prompts out of them. That works pretty well. **Audience:** Quick question: do LLMs differentiate between text that humans can see and text that is kind of hidden? You mentioned something in a footer. **Malte:** It really depends. A reasoning LLM that has some capacity from the user to think about this will usually go for the main content. But if there's something important in the footer, they also use that. And there are some websites that just return markdown instead of the actual visual website for the LLMs, and it becomes harder and harder for them to understand what is actually what. I have seen a lot of success with writing things into the footer. The agency I mentioned with the made-up matcha powder also invented an office in New York. They just wrote in the footer of their website, "We have an office in New York," and the LLMs started using it. So footer content can have influence, but the main content is probably what you will see used most often. The more important thing is that content near the top gets cited more, because they sometimes don't read the whole document, just something like 1,500 tokens, and then decide whether to continue reading. So content at the top is more important. **Host (Matteo Gamba):** Thanks, Malte. Now we have a quick break. We'll see you back in 15 minutes. Thanks, everyone. ### 12:05 · Everyone Personalizes the Answer. Nobody Personalizes the Learning. — Diana Pulnar & Susanne Dubuisson Video: https://youtu.be/cHE_iprdXJE - **Speakers:** - [Diana Pulnar](https://www.linkedin.com/in/diana-pulnar-51781b13/) — SVP Product & People, MD Babbel Labs, Babbel. Building Babbel's new AI-native language-learning product; 25 years building digital products, with formative years at eBay. - [Susanne Dubuisson](https://www.linkedin.com/in/susanne-dubuisson/) — Product Director, Doctolib (moderator). Supports the core appointment experience at Doctolib (25M monthly bookings, 90M patients across Europe); previously search at eBay. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 12:05 CEST - **Format:** Fireside chat · Track: Build & Craft - **Host:** [Matteo Gamba](https://www.linkedin.com/in/matteogamba/) (ex-Wayfair) - **Official abstract:** AI can translate any language instantly and explain any rule in seconds. But knowing about a language still isn't the same as learning to speak it. Diana Pulnar (Babbel) shares what 20 years and millions of learners have taught Babbel about that gap in the AI era. **Summary.** In this fireside chat, Susanne Dubuisson (Doctolib) talks with Diana Pulnar (Babbel) about how AI disrupted language learning and how Babbel reacted. A small team started from scratch without legacy constraints and rebuilt the product as an adaptive, AI-powered, Gen Z-oriented app in six months. They cover the organizational setup, the AI-assisted workflow from PRD to prototype, the team profile it takes, and how the new team coexists with the existing business. **Key takeaways** - AI tutor apps and general-purpose models like Claude turned language learning into a commodity. Babbel chose to compete on the space it can own: real progress through an adaptive curriculum and intrinsic motivation, rather than gamification. - Recent models made it possible, for the first time, to generate an adaptive curriculum and content at scale. Babbel's PhD linguists and didactics experts still tweak and adapt it. - A small, flexible team left legacy behind, connected all its tools (Linear as the single source of truth, Claude, shared skills), and now goes from idea to PRD within an hour and to a working prototype in the codebase within hours. - Discovery still depends on human judgment, taste and strategy, but real, working prototypes make user testing faster and feedback better. - The team of 40–50 is mostly senior ICs with a single leader and a mix of experienced people and Gen Z. It needs comfort with constant change, intentionality and showing up every day. Keeping up with the volume of output is a real challenge. - The new team works as a "speedboat" alongside the existing product, which still earns the money. The two run in parallel, and what the new team learns gets scaled across the company. #### Transcript **Susanne:** Okay. Oh, so you're not too disappointed. That's great. I promise what Diana has to tell is super amazing, because this is what we read about, and now we can actually hear that it is reality. So, Diana, do you want to introduce yourself quickly? **Diana:** Yeah. Hi everyone, I'm Diana. I work at Babbel, the language learning app. I've spent 25 years building digital products, and I spent my most formative years at eBay. That's also where I met Susanne, by the way. A year ago I joined Babbel and deep-dived into language learning. **Susanne:** Yes. So, has anybody here been using a language learning app? Two people, three. Okay, good, be honest. So let's see what the magic is behind it, because some stick and some don't. One sentence about myself: I'm Susanne, and I work as a product director at Doctolib here in Berlin. I've also been working in product forever. But now let's focus on Diana, on Babbel, and on what AI actually shifted for them. Let's start at the beginning. What has shifted the most in your industry? **Diana:** It's a heavily disrupted industry. We actually invented language learning in an app, on the internet, 20 years ago. But ever since then we've been disrupted, first through gamification and now most recently through AI. On the one hand, there are these AI tutor apps that are very, very fluid and offer a great experience. They don't have a curriculum or any plan behind them per se, but they resonate a lot with younger audiences. And then of course you can also go to an AI model provider, to Claude, and ask Claude to set something up for you. So we're disrupted from several directions, and it's somewhat a blessing in disguise, because it actually forced us to really give it a proper go, to reinvent ourselves so that we can compete with all these huge competitors. **Susanne:** But let's stick with the users before we go into that. How have their expectations changed, more concretely? Some examples? **Diana:** Most users expect us to magically know what they need, without us even having interacted that much with them. They expect us to know their context, to somewhat know where they are and how they need to learn the language. They expect us to adapt to their needs very, very specifically, which, by the way, is also a great opportunity. But also, because of these AI tutor apps and what's happening elsewhere, users expect a very fluid and fun experience. And our challenge is that language learning is a little bit hard. It's like sports: you need to do it all the time to make progress. There's a certain cognitive load related to learning, and it needs to happen for you to make any progress. So balancing both, the fun aspect and making real progress, is our main challenge right now. **Susanne:** Okay. So you looked at your product, at the challenge in the market, at the expectations of your users. Then what did you do? You checked your vulnerabilities, your strengths. What came out of that? **Diana:** First of all, we needed to find a space in which we can compete. We cannot be the most gamified app, because that already exists. We are the ones with the curriculum and the cognitive load and so on, which is not really fun, but it's still what we want to do. We actually want to teach a language. We want our learners to make progress. So we looked for a space that isn't taken, and what we found is that there is actually a space in real language learning, but in a fun or engaging way. We asked ourselves, what if we had the most adaptive experience, for example? Most apps still offer one learning path for everyone, but everyone is really different. We know that from school. There are different disciplines in language learning: writing, speaking, listening. If we are the ones who provide the most personalized, adaptive experience, that's where language learning becomes fun, because it fits what you're expecting. The second part is that we're looking into how to motivate learners, because motivation is a big part. Again, as I said, it's like sports: you need to come back every day and do something that is a little bit hard. How can we make this a fun, motivating experience? We don't want to go down the gamification route. We actually want to reflect back the progress you make, and make sure you're proud of what you accomplish with us. **Susanne:** Now that all sounds awesome and super cool, but it's not easy to build, is it? **Diana:** No, it's almost impossible, I have to admit. It was actually impossible a year ago, especially creating an adaptive curriculum. Creating a curriculum is quite a difficult and very nuanced thing. We have PhD linguists and PhD didactics experts who actually know how to teach a language, and they have experimented with this for the last 20 years. That's our big asset: knowing how to actually teach a language and what the steps towards it are. The limitation in the past was that we had people literally writing the content manually. Since then we've started to introduce AI, but it was still a limiting factor. Only with the recent developments and the most recent models is it now possible to create a curriculum and content at scale. Because if you want to adapt to the learner, then for every knowledge component, for every piece that you learn, you need so much more content in order to serve it in the right way. So AI has enabled us while also disrupting us, and we're now trying to make the best of it. **Susanne:** So how did you set it up organizationally? **Diana:** It actually wasn't as easy as you would think. First of all, we took the decision, as I mentioned, that we were going for it, and nothing was out of bounds. We decided that we wanted to walk away from anything legacy. Most of the product people in the room probably know what it means to be tied down by legacy and tech debt, and by what constrains you technically. Our current product is successful and globally available; our strongest market is the US. But to build something new you need to be agile, and you cannot have complex setups. So we chose a few people who were ready, agile and flexible, and we started out. We gave them access to all kinds of tools, left behind what we thought we knew, and went for it. It was a little bit chaotic, positively chaotic, and we improvised our way through the first month or two until we had an alpha version of the app. Then we started to create workflows, and the unlock for us was to really connect all the tools. We use Linear to coordinate, and otherwise Claude and so on. We connected all of them and realized a lot of productivity from it. We still do discovery at our own pace; it's also a creative process. But once we more or less know what we want to do, we create a PRD within an hour or so as a whole team. We use lots of different skills for that. Claude challenges us and so on, and then we create the PRD. Once the PRD is created, the next step is almost just pushing another button, and a prototype is created within our codebase. So within hours we have a fully functioning prototype that we can test with learners. **Susanne:** Let's stick with that for a second. How do you actually co-build with your learners? You have a massive learner base. How do you do it? **Diana:** We use several things. We test everything, and now that you can create prototypes like that, we do a lot of ad hoc testing with just a few learners to see how they react, what they feel and so on. Then, once we had an okay version, the alpha version, we actually A/B tested it for a while, so we put it out there with learners. But we also still do qualitative testing, diary studies and so on. Because in the end, you now have a lot of prototypes and a lot of information, but you still need to interpret it. Just looking at quantitative data didn't give us the answers. We needed to really talk to learners and listen to them in interviews to understand what they feel. And of course there were surprises there. **Susanne:** Wow, that's cool. So you have the prototypes, and you agree on what the next thing to build is. What does that process look like? **Diana:** We actually just changed it yesterday. One of the things I need to mention is that in the AI age, because the models keep changing at such a rapid pace, you also need to adapt how you work. In the past, we decided as a team what we were going to do, broke it down, and then did it. But we realized we're so quick at the doing that we needed to be quicker at that too. So now Linear is our single source of truth, where the whole backlog lives, and whenever someone, or a few people, are free, they just pull the next slice of work. It requires us to slice it super well. But this way we ship whole new features within a week or two, all the time. **Susanne:** What kind of people did you put into this task force, this group? **Diana:** Engineers, of course, the most important part, but also designers and product managers. One thing we found is that the work really converges a lot. Now that a tool can write a PRD, or challenge you, or ask you the right questions, anyone can do this. But it still needs the judgment of product managers. We also have designers, and of course the designers click a button and there's a prototype, but it's not quite right. So we need that taste and the understanding of how user interaction works. And we still need the engineers to put it into production. But I have to admit the engineers already felt that we, as product managers and designers, are now doing a lot of their work up front, because they look at the pull requests and then put things into production. The way everyone works has changed a lot. Long story short, we needed really, really flexible people, because we kept changing the way we worked. We tested a lot. There was no stability, lots of chaos, and you need to be okay with ambiguity. And you need to be open to doing a bit of the other side: leaning into design a little, leaning into engineering a little, and everyone else doing that as well. **Susanne:** So now, six months fast forward. Where are you? **Diana:** Last week we launched our first real version to paying customers. I'm looking at the data minute by minute. Right now it's looking good; we're at par. So within six months we rebuilt the product we had, just in a Gen Z version, with much, much more content behind it and the ability to adapt, and we're now at the stage where we go beyond what we had before. It's been quite a feat, and it was a small team. We're now running both products in parallel to see how it goes. **Susanne:** So what comes next? **Diana:** Now comes the revolution. I'm looking forward to this; this is what's really fun: going beyond and reinventing what language learning is. But this is also where judgment and taste matter a lot, because there are so many options for what you could be doing, and you need to find out what's actually right. On the one hand, we want to provide the best learning experience, the most fun and so on, but we also need to make sure the product is commercially successful. While we were very focused on recreating a product, now the options are exploding. We have an iOS app but not an Android app, so we're building the Android app. We're going into content production at scale. We want to create cool Gen Z experiences that don't exist yet. So now comes a bit of the trouble of prioritization, but also of fulfilling the promise that we can go way beyond what we did before. **Susanne:** We also have a bit of time for questions, so we're wrapping up. If you want to address the audience: if you are in a highly competitive area where AI is really levelling the playing field for anybody coming in, what is your recommendation? What's your tip? **Diana:** Again, we were very, very disrupted. What we built became a commodity overnight, because language learning can now be offered by so many different competitors. What I can recommend is to own up to it, to just accept that everyone can now do language learning, and then to find the space that you can own and go beyond. And the other thing, which I really like and which worked out for us, is to go for it 100%. Not testing a little bit here and seeing whether you can try something with some team, but really saying, "No, let's leave it all behind, go ahead, and try to really make it work." No excuses. **Susanne:** Yeah, it's a good way to do it. It's awesome. Thank you for sharing. #### Q&A **Susanne:** Do we have any questions? Go for it. Wait, hold on, it's not working. This one works. **Audience:** Thanks for sharing your story. As I understood it, you just built one more version of Babbel with the team in six months, right? **Diana:** Yeah, exactly. **Audience:** I'm really looking forward to checking it out and using the lifetime subscription I bought a few years ago, because one life is not enough to learn German. But jokes aside: AI allows you to deliver very fast. When it's clear what to build, it's easier to build with AI. Could you share how you decided what to build and for whom, and how you managed to speed up the discovery process? **Diana:** Thank you for the question, because you're touching on something that, for me, is still something humans do: being creative and judging. There were so many options for where we could go, and I still feel the product strategy needs to be thought through with all the experience and all the taste that you have. So we came up with a product strategy, and for us it is that we want to build the most comprehensive learning experience. Something where you don't just learn to speak, which by the way is very, very difficult, because you need to do all the other things first, but where you also learn all the vocabulary, the grammar and so on, in a way that's really adapted to you. So one pillar of the strategy is comprehensive learning. Another pillar is going for intrinsic motivation: not trying to gamify you into language learning, but helping you be genuinely motivated to do it, to come back, to look forward to coming back. And the third pillar for us was to dramatically reduce time to market. This was set by us, not by Claude, because it needed a lot of thinking, and also the experience we have from 20 years in language learning. Then we broke it down into deliverable chunks. The AI picks up the work, even writes the PRDs and so on, but because there are so many options, you still need to judge what happens afterwards. What we found, by the way, is that these PRDs are the most comprehensive artifacts you can possibly have. On the one hand, that means you really get what you wanted. But we also realized it's really difficult, from a human perspective, to keep up with the reading. Does that answer your question? **Audience:** Kind of. You described the classical product development process as it has existed for decades: building the strategy, validating assumptions, testing those assumptions or testing prototypes. Now you relaunched a product in six months. How did you manage to speed up those processes? How to speed up delivery seems clear; how to speed up discovery is the challenge. **Diana:** While delivery is more or less fully automated, discovery is AI-assisted, but it still needs a bit of human judgment. Creating prototypes is like this: in the past we had 10% prototypes and 90% Figma. Now it's the other way around. We still do a little bit of Figma, and we still talk to each other a bit, but then we have prototypes and can go into user testing immediately. This has sped up the process significantly. It's still a bit unstructured, because it still needs thinking and creativity; you cannot outsource that to a model, in my mind. But because you have prototypes for everything, and very complex prototypes that immediately work within the app, that's where you get the best user feedback, much better than before. Sometimes you would really just show people scribbles and get feedback, but you didn't quite know whether they even understood what you wanted to say. So to me, this has sped up discovery a lot. If you get immediate user feedback, you can react immediately; within hours you have the next prototype, so you can test again. So this has sped up everything, not only delivery but also all the process that happens beforehand. And again, the thinking still needs to be done by someone, because otherwise you have too many options. You still need to settle on something, go for it, and see whether it works. **Susanne:** There's a question over there. **Audience:** I'm going to do a double-decker question. How many people do you have in your team? **Diana:** Around 40 to 50. **Audience:** Okay. And what's the key to success? What kind of profile are you looking for? What kind of people succeed in this new way of working in the age of AI? **Diana:** There's one success factor that applies to all, and other than that it's still a good mix of people, as always. The one success factor that applies to everyone is this: because you have so many agents working for you, you don't need as many leadership layers. We actually only have me as a leader. Everyone else is an individual contributor, very senior ICs with lots of experience and the ability to judge. That means you need to be ready to get into it and do it. We don't need a lot of meetings and coordination. It's either done by a tool, or the team is small enough that we more or less know what everyone is doing. And then we do need a good mix of disciplines. We need very experienced people, but because we're building a Gen Z app, we also need Gen Z. We need the people we're building for. So a mix of experience and young people with new ideas, who think very differently from the older, experienced ones, has been the best mix. The other thing, which I mentioned already: you need to be very, very flexible and agile, because the models have improved so much within the last six months. At the beginning, when we tried to create curricula, it just didn't work, and now it works like a charm. We still have our linguists and didactics specialists tweaking and adapting it, but at least we're now able to create them. So we are constantly changing, much more than before, and you need to be comfortable with that. Not everyone is, by the way. **Susanne:** I would add two more things. I think it also requires you to be extremely intentional. There are so many rabbit holes. While you're waiting for one thread, you're opening the next one, and you've forgotten what the third one was doing. So you need to be extremely intentional, especially in such a team. That sounds counterintuitive to being agile, but it's not; it's required. And the second thing is that you need to show up every single day. You can't say, "Oh, today I don't feel so well because my toenail hurts." You need to show up, because the team depends on every single individual to progress together towards this shared goal, and to adjust the goal as we work. **Diana:** Yeah, and because you deliver so much... I actually wasn't on vacation the whole time, but it was summer and people were on vacation, and what they told us, and this is how I feel about it too, is that it's like a fast-running train. If you get off, the train continues, and then it's really difficult to catch up again. People who were out for two weeks or so told me they never caught up on the PRDs and everything else created in the meantime. So you need to be awake and there the whole time, because the speed is so high but our mental capacity hasn't improved. That's one of our challenges right now. **Audience:** I have a question. A great tutor is someone who not only knows the subject but is also able to motivate you, to understand your intrinsic needs and motivations and everything. How far do you think we are from an AI tutor that can fully replace a human and do things better, not only in knowing the subject but also in motivating you, finding the best way to get you into the learning and keep you there? And is Babbel ready for that, or are you going to be disrupted by it? **Diana:** That's a really good question. I think we're not quite there yet, but we're quite close. The AI tutor apps in the market are largely conversational apps: they just have conversations, and you pick things up that way. The difference with our AI tutor, which we still need to launch, though we're close, is that it is connected to our curriculum and to what the learner has done. So when it has a conversation with the learner, the tutor already knows all the vocabulary the learner knows and all the mistakes the learner has made, but it also knows what the curriculum is and what the learner still needs to learn. To me, that's a good approximation of what you mentioned, but it's not a full AI tutor yet, because our strategy is still to have the curriculum as well: for you to do some lessons, learn a bit of grammar, and practice a lot. Practice, practice, practice is really important. And then to have the AI tutor as additional support. But once this all works really well, we're very close to the state where the AI tutor is doing what a human would do. **Audience:** So also creating the curriculum itself, potentially with external sources that are not yet part of your existing curriculum? **Diana:** Yes, that will be possible down the line. Right now, again, we're challenged enough with just putting this all together and making it work fluently. But it's definitely possible in the future. **Audience:** Hello, Diana, thank you very much for all the details, and for sharing how you did your AI transformation. One thing you emphasized a few times, which I found interesting, is the all-in approach: making a decision that "we want to go for it", leaving no excuses. It sounds very exciting, but how did you build up the commitment to go all in? I'm also working with a legacy system, and we've been talking about whether to rebuild everything almost every year, even before AI. Eventually it comes down to a calculation: your team can spend six months building this, or it can spend them building something new on top of the current system. You have current customers with their demands and a long wish list. So how do you get the commitment, and how do you calculate the investment, to decide that this is the right moment and it will pay off? How did you drive this decision? **Diana:** That's a very good question. For us it was quite obvious that we needed to go this way, because the market was disrupted and our legacy system didn't allow us to pursue the vision we had. It was easy to come up with a vision, but then we didn't know how to implement it, and trying to figure out how to do it in the existing system set us back a year or even more. So we actually didn't need big calculations, because it became clear to us that if we didn't create the next generation of this product, we probably wouldn't be able to compete in all these directions. Again, our competitors are, to some extent, Claude and ChatGPT and so on. It's all the AI tutors. There are several really successful apps in the market; Duolingo also has a super great experience and owns a certain share of the market. So it was clear that if we want to stay relevant, we really need to outperform the market in some way, and our attempt to do this is by being the most agile and hopefully also a very productive company within this market. It was easier than it sounds. But maybe that's also because we got a new CEO who was ready to roll. **Susanne:** Another question. **Audience:** Hi. I want to ask a bit about the team behind it. You said that for six months you had a new team rebuilding the product, but I assume you have a separate team building what's already there. How did you recruit the people for the new thing, and what is the dynamic now between the people building the cool new thing and those working on the existing product? **Diana:** That's a very good question, and one of the challenges we definitely faced. We actually asked whoever wanted to join, because it was already clear that they were going to work very differently. You have people who are really ready to take on everything, and you have people with very deep knowledge in a certain area, and of course you want to preserve that knowledge, because our current app is the app that earns the money, not the new one, which still needs to scale. So the existing team is the team enabling us to go and do things, and our existing product continues to be successful in the market. We pioneer certain things, find out how best to do them, and then scale them across the company. We're a bit like the speedboat in a fleet of boats, trying new things and, again, being comfortable with a lot of ambiguity and change. But we still have the big boat backing us up, continuing to run and iterate on the existing product and continuing to be commercially successful in the market. So we coexist. **Audience:** Hi Diana, I wanted to ask about the team and how they improve their ways of working. You're obviously constantly discovering new things to do, but you also said you have far fewer meetings. So apart from communicating in Linear, how do you decide together what to change about how you work? **Diana:** Thanks for the question. On the one hand we have fewer meetings, but we also have one really important meeting. It sounds inefficient, but we meet as a whole team, all 40 to 50 people, three times a week. So instead of lots of individual stand-ups, we have one stand-up with everyone and discuss everything in there. It's 15 to 30 minutes, mainly about what we want to get done this week. Everyone chips in where they feel there are blockers or something isn't quite working, we figure it out, and then we go. So we have this meeting, but we got rid of a lot of the other ceremonies. And we have a very open culture. You can bring up any improvement at any time, and then we discuss it, debate it, and try it. So innovation comes from several ends. One engineer felt that we needed to connect all the tools, which I mentioned before. Before that, we were just doing all kinds of things, and he said, "Let's maybe connect them all, and let's create skills so that everyone can write a PRD, not only product managers." And he just did it. We looked at it, it kind of worked, but then we didn't know how to actually get the backlog into Linear, and then we figured that out. So it's very fluid. Everyone contributes, everyone makes suggestions, and then we see whether they stick or not. And again, what works today might not work next month, and then we're ready to change it once again. As I mentioned, yesterday we decided to cut the projects differently: instead of different work streams, we now have SWAT teams from different disciplines that take something on, do it within one or two weeks, and are then free again to do something else. We reshuffle the team more and more, because we realized that's how we actually deliver at speed. Long story short, the innovation comes from the team itself. We're small enough to talk to each other all the time, and we're ready to test anything that looks useful. **Audience:** What do you do about physical presence? Is everyone coming together in the office? **Diana:** No, remote works for us. We actually have a nice office, so people like to come in. But we have people everywhere. For the two big releases we had, we invited people to come and have a party, but usually it all works remotely. With these tools, it's so easy to access them from anywhere in the world. **Susanne:** Super. Anybody else? Otherwise I have a question. If you could restart a company from scratch today, without the constraints you have on hiring and firing, resources and anything else, how would you build the product and tech teams for maximum efficiency, to build the same things you want to build but without the constraints and legacy of the company you have today? **Diana:** This is exactly what we did, by the way. This new team we built actually started from scratch, without legacy. **Susanne:** But also people-wise? Or did you keep some of the people? Would you hire completely different people as well? **Diana:** I can't tell, because I was quite happy with the people who helped start it. We started without knowing what good looks like or what we were even looking for. To me it's always a good mix of different people. Because Babbel is quite a big company, we had lots of people, lots of different skills and specializations, and to me that worked out super well. I wouldn't even have known how to look for different people out there. So it worked well, and it's literally the approach we took: trying to start from scratch without being bound by legacy or any constraints. We were allowed to use any tools we wanted, as long as we pulled it off. **Susanne:** And are you also okay with the idea of disrupting Babbel as it is, the original company and product, if the new team and the new direction are successful? **Diana:** Babbel is a company that has been in close touch with its learners all along. We never had the problem of not knowing what learners wanted from us. We knew exactly where the current product falls short; we just didn't know how to get out of it. So the whole company is aligned on what we want to build for our learners and how to disrupt the market. I think we're all going in the same direction, implementing the vision, as you said before. **Susanne:** Nice. Any other questions? Cool. Then thanks a lot. ### 12:45 · How Contentsquare Ships Product — Jane Austin & Mariusz Cieśla - **Speakers:** - [Jane Austin](https://www.linkedin.com/in/msjaneaustin/) — SVP Product Design, Contentsquare. Leads the design vision behind the Contentsquare platform and its AI product suite; previously at startups and scale-ups including Babylon Health. - [Mariusz Cieśla](https://www.linkedin.com/in/mariuszciesla/) — Staff Design Engineer, Contentsquare. Works on core experience and the design system at Contentsquare, and builds internal AI tools for product teams. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 12:45 CEST - **Format:** Talk · Track: Build & Craft - **Host:** [Matteo Gamba](https://www.linkedin.com/in/matteogamba/) (ex-Wayfair) - **Official abstract:** AI can draft anything. Jane Austin shows how Contentsquare builds the context and judgment that decide what actually ships. **Summary.** Jane Austin (SVP Design) and Mariusz Cieśla (Staff Design Engineer) explain how Contentsquare speeds up product delivery with AI prototyping. They describe the "playground", a searchable gallery of prototypes carrying structured metadata, the internal prompt–build–iterate–publish loop, and the cultural and context-layer groundwork that they argue AI needs in order to work. **Key takeaways** - AI prototyping has broken the old link between fidelity and truth: a polished prototype can now be a half-baked idea, so every prototype needs provenance baked in (front matter showing stage, origin, authors, date). - The "playground" is an auditable, searchable single source of truth for prototypes. Every prototype is a GitHub PR, and the gallery stitches them together into a website. - Treat internal tooling as a product: a one-click setup instead of elaborate scripts, and a "game loop" of prompt, build, iterate, publish, with design-system contracts readable by both people and agents. - Culture comes before tools: ask "what did you learn?" rather than "what did you ship?", make it safe to fail, start with small "speedboat" exemplar teams, and push autonomy down to product/design/engineering trios so handovers disappear. - A human-curated context layer (customer calls, research, platform data, field signals, experiment results) is the engine of experimentation. It needs governance, which at Contentsquare sits inside the trios, to resolve contradictions and keep it fresh. - AI should do the heavy lifting so humans can focus on judgment, taste, creativity and polishing the last 20%. *Full transcript to follow after the speaker review.* ### 15:00 · What the AI-Native Playbook Doesn't Tell You — Guannan Li Video: https://youtu.be/J0lqT8W46YQ - **Speakers:** - [Guannan Li](https://www.linkedin.com/in/liguannan/) — Director of Product Growth, ablefy. ablefy serves 70,000+ professional creators and processes over €1M in transactions a day; she has built products across four countries at every stage of scale. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 15:00 CEST - **Format:** Talk · Track: Leadership - **Host:** [Francesca Cortesi](https://www.linkedin.com/in/francesca-cortesi/) (Vend) - **Official abstract:** The AI-native playbook exists. We tested it on a real product, with a real team, under real pressure. Some things flew. Some things broke. All of it was worth it. **Summary.** Guannan Li, Director of Product Growth at ablefy, shares 12 months of running an AI-native "speedboat" team alongside the company's legacy "battleship". She covers what made the three-person team fast, the conflicts it ran into along the way, and how the learnings are being carried back into a ten-year-old codebase with sensitive payment and compliance domains. **Key takeaways** - ablefy's speedboat (one product builder and two engineers) built a branded-community product connected to the core platform via APIs. It went to market within one month and grew to 7,000 active customers in 12 months, with roughly 5–6x the deployment speed of legacy teams of 8–10 people. - Three things made it work: merging roles to kill handovers and "ping pong", using AI to skip or reorder artifacts (spec to build, prototype to code, engineer builds function first and the product builder polishes UX in the same PR), and giving product people access to the codebase as a knowledge base. - Product people shipping production code caused six months of engineering pushback. What resolved it was an agentic harness: guidelines and guardrails that define what good looks like, plus 14 automated checks on every PR before any engineer review. - AI must enable humans, not the reverse. A 50-page generated spec that nobody reads, handed to an agent, produces the wrong feature. Only use AI to the extent you can explain and own what it does. - The legacy system is harder: more specialized roles, no automated-testing safety net, sensitive domains, and messy code that AI understands worse. Still, the destination is the same, and ablefy sees the core product running about a year behind the speedboat. - A speedboat is optional. Its value is real learnings and not missing the market window, but it brings cultural dynamics (job-security fears, "second-class citizen" talk) that leadership should prepare for. #### Transcript Thank you, everyone, for joining my talk. Today I'm going to talk about what the AI-native playbook doesn't tell you. That means I'm mostly going to take you behind the scenes. I won't talk about what an AI-native team or company looks like, but about what actually worked for us in reality through our experiment. We're also going to talk about the actual struggles we went through. Some of those struggles are still there, but luckily there are also a few that used to be relevant and that we managed to overcome. First things first, a little bit about myself. I work as Director of Product Growth at ablefy. I actually have a design background, and that shapes the context of my talk a little. Maybe I feel extra confident or comfortable taking prototyping for granted, so that might be slightly different from the context you're facing. I've been building products for 15 years and leading teams in a few different countries and in very different contexts. Now a bit of context about the experiment. At ablefy we work with the creator economy, which means we work with knowledge entrepreneurs, people who sell their knowledge. They can be people like Daniele, who organizes this conference, but also people who create courses or offer consulting sessions. We help them build all the knowledge products they sell, and we also handle the payment, compliance and operational side for them. Now, about the whole AI transformation. Twelve months ago we asked ourselves: how can we speed up our AI transformation in a company like ours? When I say a company like ours, I mean we are lucky to have around 100,000 customers who rely on us for their business. That's how they make their living and support their families. We also have around one million transactions going through our platform every day. So it's a working platform that people rely on, but it's also ten years of legacy code, which makes it much slower for us to really transform anything. Looking at this challenge, we evaluated different solutions and came up with a proposal: we wanted to build a speedboat team next to the battleship. That way we could still transform our core team, but without missing the timing, for example if a startup in the market simply replicated what we do while we were busy working with our legacy code and all its constraints. So today I'm going to talk about our 12 months of effort and what we learned along the way, in a few parts. First, a bit more detail about how we decided how this team should be built and what product it should build, and what we achieved over the past 12 months. That gives you context for what I mean by "experiments": am I talking about prototyping, or about a real product we wanted to launch? Then we'll deep-dive into what actually helped us achieve this outcome, the things that worked for us in reality, and on the other side, what the biggest challenges were in moving forward as a team like that. Last but not least, we'll look at how these learnings apply to our legacy products. Is this really unique to the experimental team, or can we apply the same learnings? How do things change when we bring them back to a normal system with all its legacy constraints? Let's start with the first part: how we decided what to do. We have an established system, and we decided that the speedboat team would build a new knowledge product: a branded community alongside our system. They're connected in that, for example, if you want to set up a paywall for your community, that goes into our core system, and the paywall is handled in the core. So the two systems work as two independent products but talk to each other through APIs and integrations. We think of it as enhancing the core ecosystem. We serve the same user group, and the team is relatively independent, but it also takes advantage of all the strengths we've built over the years. Now the outcome, not from the commercial perspective, but just to give you a sense of the scale of this experiment. We managed to go to market within one month. By go-to-market we mean we had the first real customer using the community, and they built up a community of 200 members. The system didn't break. It was very basic, but it worked: the classic MVP concept. Over the past 12 months, through organic growth, we now have 7,000 active customers using the system. So it's not a proof of concept. It's a system that can be scaled and used by real people in a real context. That's my answer to the hype question of whether vibe coding can really get us to a real product. For deployment speed, we measured roughly five to six times faster, mostly based on PR releases and also on the direct perception of the engineers. The team size was also hugely reduced. Everything I've described was achieved by a team of three people: one product builder and two engineers. Our legacy teams are mostly around eight to ten people. So that's the outcome. But what really helped us to be faster, go to market much earlier and build with a much smaller team? There were three things we think were critical. The first was really merging roles. If anyone was in this room this morning, you heard the talk from Csaba, I think, saying that the speed of human collaboration hasn't caught up with how fast AI is developing. Whenever you depend on another human, you have a handover, and it slows you down. In the new team we mostly have just two roles. One is the product builder, who takes care of everything from requirements to prototype and also some parts of the functional code, mostly where it relates to UX and the final experience. The second role is the engineer, who takes care of front end, back end, QA and release. That already simplified the whole collaboration process a lot. Here's one example of why that's faster. In almost every company I worked in during my 15 years, I had to play this ping pong. You build or release something and put it on staging, the product manager tests it and realizes something is off. It can be just a few pixels, or error states nobody took care of, or something that could be optimized. It doesn't work to just tell the engineer "something is off, check the error states". It normally needs quite detailed documentation pointing out exactly what's off and how you want it changed. Then you send the documentation to the engineer, and you don't know if it was 100% understood. After some time it comes back to you to test again whether what you reported was fixed, and if something isn't fixed, the whole process repeats. But imagine the new world. You see something is off, and instead of writing documentation and passing it to a human, you go to your coding agent and fix it yourself. You point out exactly what's off, and you see directly in your local environment whether it's fixed. You may also realize that something wasn't in your mind at the beginning, but after playing around you find something new, or after fixing it you realize it's still not what you want. You can change your mind at any time, and because you know the quality standard, you can iterate until it meets your bar. So the charm comes not only from saving handover time, but also from being able to change your mind at any time. Your process is much more flexible. You don't have to think everything through and write down all the edge cases before you even have the first MVP. And this is entirely possible, as you can see in this graph. It's a real GitHub PR graph from our product builder. When we started the team in September last year, for the first six months it was still quite difficult to produce anything real in production. But from March onward, that person has been releasing to production at a very steady pace. The second thing: merged roles are one possibility AI enabled, but not the only one. There are many new possibilities from a process perspective. What I just showed was product builders shaping code directly. But it's also possible to skip artifacts now, because engineers can come up with a design, and designers can come up with a spec. You can go directly from spec to build and skip the design part if you're working on a more logic-heavy feature. You can skip the spec and go directly from prototype to code, or ask AI to translate your prototype to your stack. You no longer have the classic, rigid process of product requirements, then prototype, then engineering. You can really play around with it. That makes something like this possible, and it's one of my favorite ways of working that AI enables. Imagine your support agent tells you that customers really need to be able to configure their roles and permissions. Previously it was just a default, and now we want it configurable. It's not a lot of design; it's a very clear requirement. So you can pass it directly to your engineer. The engineer understands why it's important and starts by building the function. They don't have to be very careful about the user experience, exactly where the panel sits, or whether it's a button or a dropdown. They build the function first and make it work, and then you test this first version together. The engineer can keep building other functions, and when you test the first version and realize something is missing, the product builder can go directly into the same PR and optimize the user experience, deciding exactly where it should be and how it should look. Just in the staging environment, you've already done a few iterations until you feel it meets the standard to be released to customers. The process becomes much more dynamic. No one has to be the bottleneck, and you can proceed in many different ways that weren't possible before. The third and last thing: if you don't have access to the codebase yet, I'd recommend making that the first thing you do. The one thing that changed our lives the most, and it's very simple, was giving product people access to the codebase, because that's basically your knowledge base. It holds all the logic. You can ask how a certain feature works, or ask it to list all the error states. These things used to be very difficult to discover before you started building anything on top, and now it's easy. You can also easily simulate certain states and show how you want something to be, or take screenshots for educational materials. Imagine the codebase becoming context for every step of the classic development process. You can write much better PRDs, you can code much more easily, and engineers can also build much more easily. Now let's move to the challenges. Not everything worked immediately, even among the things I said worked really well for us. Many things didn't work at the beginning, and we had to look for ways to understand why and how to make them work. I picked the top three challenges that I found most interesting. The first one we really struggled with: should product managers or designers really work on production code, or should they focus on prototyping and showing the idea? This was very controversial. As you saw in the graph, we initially tried it for the first six months and got a lot of pushback, so we actually stopped, until we tried again a few months later. Here's what typically happened, and it's a real story from our life. At the beginning, when product people got the power of vibe coding, it was very exciting. So it commonly happened that a product manager came along and said, "Hey, I built a new feature last night. It's perfect. Please review my code, let's release it." That mostly drove the engineers crazy. They would say: you think it works, but it's not working. Your code isn't maintainable, it doesn't truly work. The two roles didn't speak the same language, and it felt like they were just blocking each other. A feature like that might have 4,000 lines of code changes, which made engineers say it doesn't speed them up, because the time spent reviewing the code would be more than writing it themselves. So why does it help? Why can't you just do your job? That's typically what we ran into at the beginning. What really helped us overcome this was, first, what's called an agentic harness. Simply put, it means setting up standards and guardrails for your AI and defining what good looks like, so your quality becomes more controllable. For example, you can set a rule that whenever you code something, there must be no hardcoded strings, because otherwise it doesn't get localized or translated into another language. Whenever that happens, it's a mistake that has to be fixed. You can set up all these standards in your guidelines. In the second screenshot underneath, you can see what it looks like in real life in our GitHub. You can build a process where, whenever you push a PR, in our case 14 checks run automatically. You can't do anything, or even have an engineer review your code, until you pass all of them yourself. The charm of that isn't only that it saves time, for example checking whether it follows the design system, which I no longer have to do manually. It also covers knowledge I don't have. For example, I wouldn't be able to judge whether my code quality is good or not. Even for things I don't know, AI can review and make sure what I deliver meets certain quality criteria and fits the standards the engineering team has. That really helped us own code quality, even as product people, and it resolved the whole conflict once engineers realized they didn't have to fix all the issues the product manager had just created. It's also a loop: engineers have to write more educational feedback and help product people understand what they should do, and then go through this process. The second challenge. When people talk about AI slop, everyone immediately thinks of the amount of documentation we all receive now, much of which the owner hasn't read and doesn't know what's really written in it. In the product world it gets a bit more extreme. This is also a real example from our team. Imagine a product manager writes a spec, and unfortunately it's 50 pages, because it's free to generate more pages. Maybe you just quickly scanned it and didn't read every line, so there are lines you're not even aware of. You pass it to the engineer and say, "Read this spec and build the feature." The engineer also decides they're not going to read a 50-page document, so they hand it to their agent and say, "Go code it accordingly." Then the engineer looks at it, it passes all the checks, and they tell the product manager to check what they built. The product manager says, "No, that's not what I asked for. That's not what I want to build." And the engineer doesn't know what you want to build either, because they never fully understood the product context. This happened in the early phase of our development. We realized we had put all the responsibility on AI and weren't really on top of what we owned. The fix is also very simple: AI has to enable humans, not the other way around. It doesn't matter what tools you use. You use AI to the extent that you can control it, when you're comfortable explaining everything your AI does. You don't have to use it if you can't do that. You have to let AI enable you while still taking full ownership. That means some changes, for example you still have to review manually unless you know 100% that your AI produces accurate results. So it's more a mindset change, and the process then needs to follow it. The last thing I want to share: with all the learnings from the speedboat, we started to think about how much of this still applies to our battleship and to a more generic context. First, we looked at the differences between the two systems. There are four things we think make the job more difficult. The first is that the speedboat started with merged roles already, but on the battleship, on the engineering side, we still have front end, back end, QA and DevOps. So it's not just about how the product builder and the engineer work together. There are four roles; who are you even talking to? Engineers have to go through their own transformation first, before we can talk about cross-functional alignment. The second is the safety net, mostly automated testing. On the speedboat, we can allow product managers to shape code because everything is tested. We have automated tests, and AI can basically evaluate whether the result is reliable. In the legacy product we don't have that. The way people do QA is still by reviewing the actual code to know whether the quality is good, which obviously isn't feasible for product people, or even for engineers when they receive a large amount of code. The third is that we picked this specific product for the speedboat partly because it's not a sensitive domain. In our core products we have payments and compliance, which are more sensitive domains for AI to touch. Last but not least, after ten years of messy code, AI of course understands our codebase much worse than a new codebase that was prepared fresh for AI. So AI makes many more mistakes and hallucinates much more. But none of this means we can't apply these principles. After some exploration, the guiding principle we arrived at was that the two teams will meet in the same place sooner or later. The destination is the same. All the trends we see, smaller teams, merged roles, the agentic harness, access to the codebase, will happen to both teams. None of it is irrelevant to the other context. It's just a matter of your route, the obstacles and the timing. Where we stand now, a few things are already possible for us. First, all our PMs already have access to our codebase, even the legacy code. So you can still use your codebase as your knowledge base to write PRDs and prepare your prototypes. The difference is in actually shaping a feature, but it's still possible for us to make localized changes. A designer can still go and change a button, or decide to optimize a modal. The principle is that you have to roughly understand that you're only changing one part of the codebase, and know that the AI isn't randomly changing hundreds of lines of code. It follows the same principle: you only do what you can be responsible for, rather than hoping someone else will fix your issue. Third, instead of prototyping directly in the real codebase, for bigger features we built a sandbox environment where we prototype everything. I'm showing a screenshot of the sandbox environment. It isolates things a bit from the main codebase, which gives everyone a safer environment to experiment in. The last thing is the agentic harness. It just takes longer, and it's a bit harder to build up all the guardrails, but everything is still relevant. We've also started building a harness around our core codebase, defining how our design system works and defining our code standards. All of this helps us move closer to high-quality code and to the different artifacts. Here's a quick screenshot of our sandbox environment. It really takes just a few prompts and a few skills to simulate what your main production environment looks like. So all in all, this whole story is about trying to find an ideal work environment to test how an AI-native team could work when there are no constraints. We tried to give the team all the freedom, with no constraints, to really play around with all the toolkits, learn from best practice and see what's possible. Then we tried to see what still survives when everything is a constraint again. Luckily, we realized there are many more things we can apply than we initially anticipated. It's just about the route and your speed. So think about how the speedboat experience applies to your specific challenge, and remember: the destination is inspiration, but the route is completely in your hands. Thank you. #### Q&A **Host (Francesca Cortesi):** Thank you for this story of what a speedboat looks like in real life; we always hear about it. I want to remind all of us of the invitation Daniele made at the beginning of the conference: we want this to be a conversation, and all the speakers have agreed to take some questions. We'll take one from YouTube first and then the hands in the room. Thomas Mo asks: if we are encouraging product managers to take an increasing share of creating the product software, won't that mean product managers spend less time with customers? **Guannan:** My personal view is that it's not about asking product managers to build more. It's about taking over the intersection: the customer-facing part, where you translate what you hear from customers into something that's very obvious to you, but where you'd otherwise have to write additional documentation to explain it to an engineer and then check their work. For me, that's the part that makes the most sense. It removes the handover work and the explanation. The quality bar is in your head, so you don't need to explain it; you can just explore until the result reaches the quality bar in your head. I think that's the part that makes the most sense. **Audience:** You mentioned that you now prototype in your sandbox environment. How do you make sure it's still usable when a ton of people are prototyping a ton of things in the same sandbox? **Guannan:** That's a very good question. AI is definitely not a free lunch. One thing we learned is that AI really doesn't help with collaboration. The tools of the last generation, like Miro and Figma, were much friendlier when many people worked on the same thing. So far, even for designers and product managers in this sandbox environment, we more or less follow how engineers work: you create a branch and then merge it back. It's not the best experience, and we're constantly looking at how to make it better. Recently, with Claude artifacts, you can at least share a link directly, and then there's more conversation on Slack to keep each other up to date. That's where we are. But I think this gap will be closed sooner or later, and we'll have real-time collaboration that's much friendlier for everyone. **Audience:** Guannan, thank you for the presentation, it was really interesting. Experiments are always about learning. So two questions: what did you find most surprising when running these experiments? And in the change process, what did the leadership team find most surprising about the experiment and the outcomes? **Guannan:** A few of the surprises are probably what I put in the challenges section. I'd say it's on both sides, more about team dynamics. For example, product people had a lot of passion: now I can actually touch code, I can build something. So we jumped on it quickly, and initially we expected engineers to embrace it as much as we did. We hoped they'd feel it was a much more collaborative way of working. But we experienced a lot of pushback, and as you saw, there was a six-month gap before we even tried again, because engineers were very upset about how we were doing all of this. From the management perspective, maybe not surprising, but I'd put it down to the culture aspect: what it does to how the team perceives things and to the team's sense of security. Starting a speedboat raises questions I didn't cover, because it's a much more dynamic topic about culture. How do other people feel when you start a speedboat like that? Does it mean everyone's job will change? Does it mean we're only going to hire people like that? If it succeeds, does that mean they'll be replaced? All of these dynamics, even "second-class citizen" talk, are things management probably didn't completely anticipate when we first decided to start the team, until those conversations became really prominent. **Audience:** Thank you, I really enjoy the conversation here. What you told us about your speedboat and transferring the learnings onto the legacy product is very cool. I wonder whether you could have achieved as much on the legacy side without the speedboat. And what would you say to a company where a speedboat may not be possible, for example because there's a lot of sensitive data? How can those companies still take much more advantage of this revolution and drive innovation into the legacy side? **Guannan:** As I said, I think it's just a matter of speed. From what I see in my company, the majority of things are applicable, with more or less one year of delay. Right now, for our legacy products, we have the first version of an agentic harness, and we're experimenting with some first features that use AI to code and to test. So it's about a one-year delay, but it's completely possible. You should still build up all these guardrails to make your coding agent more accurate, and the future is still that eventually you're probably not writing your code; your AI is. You can take all the security measures needed to make sure it's secure. So the destination is the same, and we're heading there. We're also discussing that in a few months the teams will get smaller; we want to build smaller teams that move faster. We see the future going in the same direction no matter which domain you work in. Of course, if you work in a more sensitive domain, you need to take security more seriously from the very beginning; it can't be an afterthought. Especially when you already have customers relying on your system, and if you have legacy code, it takes a bit longer. You have to finish your end-to-end testing setup before you let AI control your code in any way. It takes longer to prepare everything, but eventually you want to move in the same direction. So even if you don't start a speedboat in your company, you can still get there. It's just nice to do it in parallel. For us there were two reasons. First, real learnings. Second, not missing the timing, because no one knows how big the window really is: whether you can wait another year to transform your products, or whether a competitor is already in the market doing what you want to do. That's why we started a team to build our own biggest nightmare on the side, at the same time. **Audience:** Thank you so much for the presentation. There's a lot I can take home from this. You already answered the question I first had in mind, which was what's coming next in this journey. But I'm also really interested in your advice for someone like me, whose organization is still very much before any of these steps. What would the first steps be to get this kind of change going, who should own it, and is there any other advice you could give me? **Guannan:** Like I said for the last question, you don't have to build a speedboat. You can also decide to transform your organization in one go. If you feel ready, maybe with strong support from the engineering side, like the Babbel story we heard this morning, where the CEO and leadership decided to go all in, then you can just start building. The AI playbook is out there. Everyone is talking about the same things: smaller teams, merged roles, agentic harness. There are best practices in the market. You can plan your roadmap and start building. For us, a year ago, we picked this route because it was a bit less risky. We wanted to keep the current system that makes money running, figure out how to prepare the people and our route, and not miss the window. So we invested in a small team to test everything, also because we weren't fully ready to go all in and rebuild our system. If you're in an environment like that, you can consider this route too. But be prepared: there will be cultural dynamics, and there will be people who are upset. You need to prepare yourself to face that, and prepare the organization: this is what we want to achieve, and this is what you might see along the way. ### 15:40 · Firechat: The Future of AI Workflows — Jan Oberhauser & Luca Rossi Video: https://youtu.be/uXfLMDNAu4M - **Speakers:** - [Jan Oberhauser](https://www.linkedin.com/in/janoberhauser/) — Founder & CEO, n8n. Founded n8n, an extensible AI workflow automation platform; previously a VFX pipeline engineer at Academy Award–winning studios. - [Luca Rossi](https://www.linkedin.com/in/lucaronin/) — Founder, Refactoring (interviewer). Runs Refactoring.fm, a weekly newsletter and podcast for 175,000+ engineers and managers; previously led engineering at Translated.com and co-founded Wanderio. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 15:40 CEST - **Format:** Fireside chat · Track: Leadership - **Host:** [Francesca Cortesi](https://www.linkedin.com/in/francesca-cortesi/) (Vend) - **Official abstract:** Workflows are turning into voice-driven agents, and n8n sits right in the middle of that shift. Luca Rossi presses Jan Oberhauser on where product is headed next. **Summary.** Luca Rossi, a long-time n8n user, interviews n8n founder and CEO Jan Oberhauser about how the company moved into AI, why it focuses on business-critical workflows that mix AI, deterministic code and humans, and what being profitable and open source means for its product decisions. The conversation then turns to how AI is changing the way n8n itself works, its goal of reaching a billion users with fewer than a thousand employees, and how to take people along through the change. **Key takeaways** - n8n released its first AI features around October 2023, but traction only really came more than a year later. Jan's lesson is that teams often feel late when they are actually early, and that it takes time to iterate on rough first releases. - n8n focuses on workflows that people, teams and companies rely on. It combines AI, deterministic logic (faster, cheaper, 100% reliable) and humans in the loop for edge cases, rather than letting an agent do everything as a black box. - Being profitable, with seed money still in the bank, lets n8n turn down large contracts whose requirements would make the product worse for everyone and think long-term rather than toward the next funding round. - Open source and source availability drove adoption: until recently all of n8n's sales were inbound, from people who brought the free product into their organizations. Jan admits he is "still a bit torn" on open source in the AI era. - Internally n8n runs many agents (sales learnings, HR questions), has an "agent-centric development cycle" (AC/DC) in engineering, and has a labs team that built a product in under a week. The PM-to-engineer ratio has shifted from one PM per six engineers to two. - n8n's goal is "a billion users with less than a thousand employees." Jan sees talent density, transparency and honesty about how roles are changing as the keys to adopting AI without fear. #### Transcript **Luca:** Hi everyone. Thank you so much for being here with us today. **Jan:** Thanks for having me. **Luca:** I'm really excited to have this chat with you today. Not only because you're obviously an incredibly accomplished founder who runs a fantastic product, but also because I've been a big user of n8n myself for many years. As a content creator who runs a newsletter, I have to move stuff around between different tools and create routines and procedures, so n8n has been invaluable over the years. I've also been able to see the different stages the product went through. I remember n8n before AI, so I've seen the whole evolution, even in recent years. I'll ask you a few questions that are a bit selfish, if you like, because they're the questions I'd love to ask you as a user. The first one: it's fascinating to me, or at least that's how it felt to me as a user, how incredibly early you were at transitioning into AI, at a time when it was not as obvious as it is today where AI was going. So how was it for you? Was there a moment when you said, "Okay, we need to go all in on this, our vision is going to go towards that"? And was it as transformational as it felt for me on the other side of the barricades? **Jan:** First, thank you. It's always great to hear when somebody is using us and getting real value out of us, so that's a good start. Apart from that, it's quite interesting that you think we were early. When we started, we thought we were very late. It's like when I first saw Bitcoin, it was at 2k and I thought I was already too late, and look where it is now. So it's an interesting learning for me: very often you think you're late, but you're actually quite early. And between actually working on it, releasing the AI features and then getting traction, another year went by. We released them at the end of 2023, maybe October, and it really took off at the end of 2024. So it took over a year. There is normally still more time than people think. To answer your question: I don't think it felt like we had much of an option, to be honest. We just realized there was this big shift happening, and either we change or we have a problem very soon, because we thought that if we didn't change, there probably wouldn't be a need for our product at all anymore. So we knew we had to do something. The question was just what to do, and that was the harder part. Originally we thought, let's see what we think would be right and what the competition is doing. But we realized that if we just add AI, like most other products, it would definitely make a difference for us, but it wouldn't have the outsized effect we hoped for, especially not in ensuring that we'd be around long-term and really use the opportunity properly. That's why we took a step back and said: not just adding AI, but really thinking about how we can be sure we're part of the revolution, part of ensuring that people actually build AI-powered applications with us. Honestly, when we released the first features, they were still very rough. There was so much still wrong. It didn't work amazingly, but it worked enough to get enough people in. If we had said in the beginning, "We expect X users for this," we probably would have stopped it a million times over, because it took quite a while to iterate and make it better, and for other things to happen, before it really took off more than a year later. **Luca:** It's really interesting to me, because I think people here who run products understand the feeling that you need to find the right trade-offs between your vision for the product, the feedback you get from customers and where the market is going, and try to put everything together. With AI you're saying you thought, "We have no other choice, we have to embrace this." So was it more of a top-down vision, rather than a bunch of people all asking you for the same thing? **Jan:** I'm not sure I would say it was top-down in the end. Maybe a lot of leaders think they're the only ones seeing it, but there's always rumbling across the organization very early on. At least if you've hired smart people, they're going to realize something is wrong too. Again, the problem is the gap between realizing something is wrong and knowing what to do, which always takes a while. It's not that I really had to say, "I think there's something wrong, we have to do something." Literally everybody knew. So it wasn't me saying, "Let's do something," and driving it. Everybody knew we had to do something. The question was just what we were going to do. Once that became clear, well, not everybody was excited about it. There were definitely some people wondering what we had to do with AI, because they never imagined we'd do anything with AI at all. So it was not always the choice for everybody, but most people were very excited about it, and over time everybody came along and got very excited. **Luca:** With respect to the shape of the product you had back then, did it feel like an incremental, easy way to evolve what you already had? Or did it feel like you needed to change things to accommodate this and turn n8n into the best tool to build AI agents and workflows? **Jan:** We were actually quite lucky in that what we already had fit in very well with this whole space. It's always bad to over-engineer something, but honestly, I over-engineered a little bit in the beginning, and it totally paid off. I thought I had made it way too complicated. I had imagined functionality that we never implemented and that I didn't think was needed anymore. Then this thing came along and it was perfect; it just made everything so much easier. There was definitely a big shift in some ways, but we were very lucky that we didn't have to make too many deep changes to the product, and what we already had fit in very well. There's always some luck involved. It's not just that we were amazing; we were also very, very lucky in many regards. **Luca:** For what it's worth, that was my impression from the other side as well. I felt like, of course n8n is perfect for that. It felt like you had to change quite little in terms of philosophy, UX and product. Fast forward to today, and the world of workflows has changed quite a bit. We went from visual workflow builders, the most accessible way to create these kinds of workflows, to agents, where I can send a voice note and tell my agent, "Do this every hour," and it just does it. So how has this picture changed for you? Has your vision for what n8n needs to be, who its users are and its use cases changed now that agents are accessible to everyone, and everyone can create simple periodic procedures that were much more complicated to run before? **Jan:** Yes, definitely, there are a few things there. First, you're totally right: it's very easy to build something these days, especially for personal automations. So we focus more these days on things you really rely on. If you build something like a morning email summary, and you'll survive if you don't get it, then a lot of the other tools are probably great. If you build something that you really rely on, that your team relies on, your department, or especially the whole company, that's where we really shine. What I think n8n does very well is combining the three entities that can do work. We have AI that can do work. Before that we had computers, and in the beginning we always had humans. The thing we do very well is to say that AI is not the solution for literally everything; it's one part of the solution. AI does one piece, and then deterministic logic makes sure it doesn't go totally off the rails. If you want to check whether a company has more than 50 employees and do one thing, and another thing if it has fewer, code is much faster, it's 100% reliable and it's much cheaper. You don't have to use AI for everything. And for the edge cases, you still want a human in the loop. That's where we shine: business-critical use cases. That's where we're focusing the most, and that's where we see a lot of traction, with people who think more deeply about what they really want to build, and with organizations that use us to transform how they actually work internally. **Luca:** I think you touched on a great point about deterministic versus non-deterministic steps in a workflow, and what deserves to be non-deterministic. Because it's so convenient, it feels to me that there's now a tendency to put everything into the agent and let it be in charge of everything, even what's probably best run with code, which is faster and cheaper. So what you're saying is: if you have a workflow, figure out which parts actually deserve that kind of intelligence, latency and cost, versus the parts you can extract and run as a script or through an API, and you'll be better off. **Jan:** Yes, including some other things. If you have an AI agent and you build something, it's a black box. It works, but if there's an edge case or anything like that, you have no idea. Especially for business use cases, you want to be sure that you understand what it's actually doing and why. You want to be able to understand the logic, adjust the logic very easily and be able to rely on it. That's what I think we do very well with n8n, especially with the AI assistant we built and released recently. You describe what kind of workflow you want, and you get the workflow out of it for free. You can actually run it, and if something goes wrong in the middle, you can see where it went wrong, adjust it and continue. There's also the model-independent piece: you can run it on your own infrastructure. You don't have to rely on Anthropic or OpenAI; you can also use open-weight models and fine-tune them for your use case. There are so many things in there that are especially interesting for large organizations, which is honestly where most of our revenue comes from these days. But I think that if you want to keep winning, you have to keep an amazing experience for everybody. The times when enterprise software could be horrible are over; people expect more. If we didn't have a great experience for everybody, I think we would still have a problem. So we have to make sure we provide value for everybody and create a great product that's actually joyful to use. **Luca:** How hard is it to make that experience good for everybody? I think all of us working in product can think of examples of products that have not exactly improved, to put it lightly, when they went to focus on the enterprise, and the experience for the rest of us got a bit worse. How hard is it to find the right solutions, choices and compromises to support the revenue streams that matter while keeping the experience good for everyone? **Jan:** I think some of it is about protecting things. We work with large organizations, and there are much larger contracts we could go after. They very often have very specific requirements, which are totally valid for their organization and their use case, but which would make the product worse for everybody. You have to be very careful to protect the product against those things. If you talk about a $10 million contract, it's a nice contract, but if for that 10 million you make the product worse for everybody, the math doesn't work. If you keep that in mind and make sure you're working with the right customers and that they use the product in the right way, it's great. The good thing is that AI also makes some things simpler, like the building and the research. But you have to protect the product and be okay with the things you're giving up. To be fair, this is probably easier for some orgs than others. We're in the very lucky position of being profitable. Honestly, the money from the seed round is still in the bank, so we can think very long-term. If you have to fundraise again in 6, 12 or 18 months, you have to think about how to maximize your revenue up to that point to raise another great round. We can think about what to do now to be best set up for success in the long term, not for the next funding round, because a lot of the decisions you'd take short-term are not the right thing long-term. **Luca:** You said a very interesting thing about being profitable and not being in a rush to make decisions driven by the need for high growth and chasing the next funding round. How much has being profitable, and being able to go on your own terms, changed decisions you've made on the product, and even your vision, which might be more long-term than it would have been if you were after your Series B and Series C, not profitable, and needing to demonstrate big contracts here and there? **Jan:** Quite a lot, actually. There was definitely a time in between when we were not growing that amazingly, and we didn't have any revenue, or almost none that really mattered. We knew we were going to need more money. So we literally had to turn the company around, from everybody working on making the product better for everybody, to almost everybody in the organization working on revenue-generating features, because we knew that if we didn't show some numbers soon, we wouldn't exist anymore. Did we enjoy that? No. Was it necessary? Yes. And it worked out well. Again, we were very lucky there. Now being in the position to think long-term and be independent of funding rounds is quite freeing. I think it's honestly also the best thing for the community, for the employees and for the investors, because a lot of the short-term decisions are only set up for short-term success. So I'm very happy and lucky that it can work that way. **Luca:** I wouldn't say lucky. You worked hard to get there, so it's well deserved. **Jan:** But I think there's always a big component of luck in there. **Luca:** Of course, always. You spoke about the community and the element of trust that being profitable and here for the long term creates. It makes me think that when I approached n8n for the first time, one big differentiator was that the product was open source, or source available, compared to others out there. There's a big conversation these days about how AI is changing a lot of the optics around open-source and source-available projects. They might become even more important, but there's more burden and a lot of stress on maintainers. Do you feel open source is an important part of your vision for the product, and has anything changed with AI in that respect? **Jan:** It was always important for us to make sure our source code is available, so people can really understand what's happening and be able to self-host it and run it themselves. We also have a labs team, for example, that thinks about other products, how we can expand, and what the next generation of these tools looks like. There too, by default, we assume they'll also be made available. Will all of them be available under the same license? I don't know, but that's the general default assumption. That's how we were able to build this amazing community. We got a lot of adoption because people could start using us for free. That's how we got not just into private users, but into literally all of the large organizations. Until not long ago, all of our sales were inbound; we didn't do any outbound. We got in because people brought us from private use into their organizations. They built more workflows and got a lot of value out of it, and at some point they said, "Now we have 10, 50, 100 or a thousand workflows and so many users, we probably need an enterprise version." That's how we generate our revenue. Regarding open source, I'm still a bit torn. You can obviously say that with AI it's now easier than ever to copy a product. With open source it was always a little bit harder: you could see everything, but reimplementing it was hard. Now you can just vibe-copy something more or less overnight for probably $500. And probably in another half year you won't even need the code anymore; you can just point the AI at the product and say, "Create the same version." But since everything is now so much easier to build, I think you're still as well protected as before. I think people still care about it, and they like this open approach, which I think made us successful. I also think Europe is well set up for this. A lot of US companies are more closed and more protective, and in Europe we think more about being open and inclusive. I think that's probably the more winning recipe. **Luca:** It's also a route that very few products try to go for. I love how you said this creates community and basically creates a top-of-the-funnel channel, because people start self-hosting things and eventually turn into enterprise customers. But I don't feel many SaaS products think strategically that it may be worth exposing an open-source version of the product and using it as part of their go-to-market. So if you were the founder of a product today in a similar space, maybe B2B or dev tools, how would you think about that? Would you consider open source plus a managed version an important option for developing new products, and for building that kind of community and goodwill? **Jan:** I'm not sure you can say it's always the right thing to do. I think it depends on what product you're building, what space you're in, and also who you are personally. For me, I worked at a small open-source startup before I started n8n. We had a community that was not huge, but you already saw the impact there. And with n8n, a day or two after I made it available, the first person contributed the first integration. He built six integrations within a few months, and he was the first person we hired. So for us it definitely made a lot of sense, especially because there were already players in the market. I think we were better, but people probably wouldn't have seen the need to switch to us if we had been another proprietary product, even though I think we had an edge. Being available for free and having this open-source angle, large organizations especially really enjoyed that as well. It made a lot of sense for us. I don't think there's a general right thing to do, but honestly, if I start a company again at some point in the future, I would probably do the same thing again. **Luca:** Love this. I wanted to use the last part of this chat to talk about AI specifically. We've had many great talks today, including the very last one, about how companies are navigating how they work, how they do product development and software engineering, and how many things have changed. You obviously work at the very forefront of this; you build a tool that is knee-deep in it. How has AI changed how the n8n team works in terms of product development and shipping things? **Jan:** There are probably two pieces: how we actually work, and how we think about things. Starting with how we actually work: I recently talked with an employee who had just joined, and they were quite surprised by how many agents we have running for very different things, because we always think about how we can automate something and how we can use AI there. It's anything from sales, where we have a bot going through all the conversations and figuring out learnings, and you can ask it questions and it posts things, to the same thing for product. Then we have agents for HR where people can ask questions. There are a million things running. We also have an initiative inside n8n that our VP of Engineering coined AC/DC: agent-centric development cycle. **Luca:** Not just AC/DC. **Jan:** Yes, exactly. Honestly, it took longer than I think it should have before they really moved fast enough in that direction. But now I really feel we're making big strides in the right direction and moving much faster. As I mentioned before, we also have this internal labs team, similar to the "speedboat" described earlier, which works totally differently because they can start with a completely greenfield product. They don't have to care about any users or an existing product. They can just start and move incredibly fast. Within less than a week, they built an awesome product, including a homepage and everything. Let's see if we can launch that one in the next few weeks, depending on how internal testing goes. It's just amazing how fast they can move. Then, generally, how we think: we set an internal goal to reach a billion users with less than a thousand employees. I'm not sure we're going to get to a billion users, but that matters less. The most important thing is how we can have as much impact as possible with the smallest and leanest team possible. You cannot build an organization the way we built it three years ago; we have to build it the way organizations will be built in a year or two. People literally come to us and expect us to help them transform. If we're not able to do it ourselves, how should we help them? We'd be hypocrites, saying "Do what I say, not what I do." That doesn't work. We want to be at the forefront. That's why we also have an internal AI and automation team. In the first step, I think it did the same thing as in other orgs: picking the low-hanging fruit and automating a lot of stuff for other people. Now that that's done, it's thinking more about how we can empower the rest of the team internally. We talked about engineering with the AC/DC work, but the same thing should happen for literally every part of the organization, finance, HR, wherever. They should all work differently than they did in the past. That's a great effort, and it obviously takes quite a lot of time. But you really have to think about how to empower those people and how to take them on the journey. AI is already really good and doing an amazing job, but it's about getting people to come with you on the journey, making them excited, so that they don't think, "AI is going to take away my role," but rather, "I feel empowered, I can do more of the things I actually enjoy," rather than being threatened by it. I think that's going to be quite impactful. **Luca:** So interesting. I love how you framed it: to serve our customers as well as we can, we need to understand these things ourselves. It's basically dogfooding what we build. Can you already see trends in how the work in your team is changing? We hear things like teams getting smaller, not in terms of overall headcount, but the units of work getting smaller, with fewer people working together. In other cases, people get broader scope because we're becoming more like generalists. Do you see some of these shifts happening, or something specific that makes you think, "This is the direction we're headed"? **Jan:** If you think about product engineering, the ratios have changed. In the past we had one PM, one designer and six engineers. Now we still have one designer and six engineers, but we have two PMs. That's more the direction things are moving. We definitely see designers and PMs, instead of creating nice click demos, implementing a feature that you can play around with. You get a real feeling for it, you can have the conversation with a customer or a user, and you make sure you're not just designing something that sounds great, but that you come with something usable. Then, very often, it has to be reimplemented. Across the org, we look at the size of the organization. You can think about ARR per employee, which I think is a horrible metric, because ARR doesn't say anything on its own; you can have a lot of ARR and still lose a lot of money. But there's still some value in looking at how much people are willing to pay, how much you have per employee, and getting to a good ratio. The same goes for general team size: where are teams normally, how big is a finance team normally at that revenue scale, and then trying to make it definitely smaller in our case. It always gives us a guideline for where we can push more when people ask for headcount. Honestly, we're still not doing an amazing job there. We still grew much more than I would have liked. But there's definitely a lot happening, and a lot of internal discussion about how we can grow headcount as slowly as possible. It's not about saving cost; as I mentioned before, that doesn't matter for us at all. But in the future, work will be different. Working in a much smaller org means a much better culture, people can have much more impact and have much more fun. I think it's literally better for everybody to work in an org with 500 or 1,000 people versus one with 10,000 people. **Luca:** This really fascinates me. Of course, companies hiring less, or hiring freezes, get a lot of bad rep. But I talk with many founders who tell me it's not that they don't want to hire people; they really want to figure out how to make the best of the team they have. We don't know how far we can get, or how much better our teams can work, because we're still in a stage of exploration. Sometimes solving a problem by saying "I'll just hire a couple more people" feels like a lazy way of addressing it. And it's always hard to understand where the boundary is. So I guess it's not easy. **Jan:** Especially because people think that because of AI, people become less important. I actually think it's the opposite. All orgs have access to the same tools and the same LLMs. What makes the difference is the people. People, especially high-quality ones, are the scarcer thing out there, and good people want to work with other good people. The smaller the team, the more important talent density becomes. **Luca:** And the less overhead, too. **Jan:** Yes, exactly. In the end, almost every org could probably work with half the people if they're the right ones. With each person you hire, theoretically you have someone who can do more work, but at the same time it adds overhead. The more people you have, the bigger the finance team has to get, the more reporting overhead you have, the more management layers you need, and the more chance of misunderstanding. You lose so much with every new person you hire. We went from 60 or 70 people at the beginning of last year, to around 240 last year, and now we're at around 380, I think. **Luca:** Wow. **Jan:** And it's not just once they're in. First you have to find them, so instead of people doing the actual work, they have to do a lot of interviewing. They have to onboard people, do performance management; there's so much going on. So really trying to understand how you can do it with the least amount of people is, I think, a big advantage for everybody. **Luca:** Absolutely. Thank you so much. I want to know if there are questions we may want to pick up from the audience. #### Q&A **Audience:** There's a lot of narrative around efficiency, and I think this is really change management: how much of my identity is connected to my work, how much I'm going to fear it and how much I'm going to embrace it. For you, wearing the CEO hat, it's hard to navigate: keeping everyone involved, going through change, maybe pivoting to something else and being three steps ahead. What is your best learning or best tip on how to do that? **Jan:** I think it's really about being very transparent in general. One of our core values is doing the right thing even when it's hard. It's not always easy, but people know that we do things with the right intentions. By default, almost everything is accessible to almost everybody in the organization, except salaries and things like that. And then it's about taking them on the journey: why we're doing things, what they get out of it, and also the risks. If we do it right, it also makes the work more interesting for them. A lot of this is really a people problem: taking away their fear. We had the same in engineering. Honestly, there are probably going to be problems for some people. If the thing you enjoyed most as an engineer is the act of coding itself, you're not going to be that kind of engineer anymore. Sorry, this is not my fault, but we need to be honest about that. But if the thing you enjoyed is building, going from an idea to getting something out, then it's your time. So it's less about us and more about how the role is changing. The role is going to look similar in almost every org, and everybody has to think about what direction their role is going in. Am I still excited about it, or do I maybe have to change? What I really like is that a lot of people seem to be thinking about this much more than in the past. In the past, people always thought the next step was to become a leader, a manager, and go up the ladder. Now people are doing the opposite: they realize they actually enjoy doing IC work, and they go back to it. I think that's great. **Luca:** Careers are not linear, nor should they be. That's my belief. Thank you so much for this amazing chat. It was fantastic listening to you. ### 16:40 · Against Simplicity — Prasad Gupte - **Speakers:** - [Prasad Gupte](https://www.linkedin.com/in/prasadgupte/) — VP, Digital Products, tonies. Owns tonies' digital product portfolio (app, connected hardware, content, e-commerce, CRM) serving 10M+ families; previously growth and product at Babbel and Landmark Group. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 16:40 CEST - **Format:** Talk · Track: Strategy - **Host:** [Francesca Cortesi](https://www.linkedin.com/in/francesca-cortesi/) (Vend) - **Official abstract:** Most product leaders are wired to simplify. At tonies, that instinct would have been wrong. Prasad Gupte makes the case that complexity, when owned with intention, is the most defensible moat a product org can build. **Summary.** Prasad Gupte, who leads digital products at tonies, argues that complexity is not bad in itself, but must be consciously turned into defensibility by managing risk. Using tonies' child users and parent buyers, hardware lead times, Wi-Fi activation, supply-chain shocks, content localization, IP licensing and innovation, he shows where complexity becomes a moat, and offers a simple framework for deciding which complexity to kill, solve, defend or innovate on. **Key takeaways** - Complexity is not a moat by itself; otherwise, as he joked, the Bürgeramt would be the strongest moat of all. It becomes defensible only when you consciously manage the risks it creates. - tonies runs discovery twice: its users are on average five years old and can only be observed, while its buyers are parents, whose idea of value differs by market. - Hardware forces feasibility decisions about a year and a half before customers get the product, and activation (getting a screenless box online on Christmas morning) is a make-or-break moment. - A two-axis framework: does the customer care, and does the competitor care? Kill what nobody cares about, solve table stakes (customer only), defend moats competitors feel but customers don't, and innovate where both feel it. - Brand is the manifestation of the product. PMs should aim for loyalty rather than lock-in, and ask: "if customers had a choice, would they still stay with you?" - In Q&A: tonies runs one innovation pipeline, routed by "blast radius" to a fast digital lane or a slower hardware lane. AI helps internally, including with embedded code, but has not yet been taken directly to consumers. *Full transcript to follow after the speaker review.* ### 17:20 · Building 0→1 AI products — Rohan Garg Video: https://youtu.be/hIALRBRz708 - **Speakers:** - [Rohan Garg](https://www.linkedin.com/in/rohangarg/) — AI Consultant, Kleinanzeigen. In tech since 2006; co-founder and executive at German internet companies including idealo and kununu, working deeply on agentic AI for the last two years. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 17:20 CEST - **Format:** Talk · Track: Build & Craft - **Host:** [Francesca Cortesi](https://www.linkedin.com/in/francesca-cortesi/) (Vend) - **Official abstract:** How do you go from a hypothesis to an agentic product your customers actually use, especially inside a large organization? A real-world case study of the process from start to finish. **Summary.** Rohan Garg, an AI consultant working as a product engineer with Kleinanzeigen, walks through how the marketplace went from investor-facing prototypes to a live conversational shopping assistant, Kleinanzeigen AI. He shares learnings on prototyping the vision on real production data, then on running an orchestrator-plus-agents architecture with small models in production, including a hybrid "reasoning without loops" approach to search, and closes with the open challenges of B2C AI: UX habits, input, infrastructure, cost and monetization. **Key takeaways** - Prototype the vision, not the MVP: B2C AI agents are big investments, and investors buy into a two-to-four-year vision; build prototypes on real production data and proprietary endpoints so feedback is about real results, not just UX. - Kleinanzeigen built a dedicated prototyping infrastructure (a React clone of the app, its own MCP server for authenticated back-end access, own deployment) so any prototype can go live on production infrastructure within a couple of weeks. - An open conversational interface removes UI guardrails, so you can't start narrow: they built a horizontal product with an orchestrator routing to specialised agents (search, general, product, customer support) and then added depth in selected areas. - Small models beat big ones for B2C: larger models cost 20–30x more and are too slow; answers need to arrive in 5–10 seconds. Splitting work into narrow LLM calls gives observability and objective evaluations, making it easy to benchmark new models (they currently use Gemini over Haiku 4.5). - Their search uses a hybrid of workflow and reasoning — a query planner and a backup planner that generate queries but never loop results back — because small models break down in loops. - Open challenges: 20 years of UX habits (only eight results fit in a chat), input as a bottleneck (voice isn't reliable yet at scale), provider downtime and latency, cost, and unresolved monetization and cannibalization; so far normal search still generates more leads than the assistant. #### Transcript The downside of being an individual contributor is that you don't get on stage very often. So it's been a long time since I spoke, but I'm very excited to be here. Thanks for having me, and thanks for the nice words. For the last two years I've been working as an AI consultant. What does this mean? It's a fancy word for a product engineer. I do two things primarily. I do product management, so I propose how AI products should be built, and then I also build them. So I'm also very actively involved in programming, and I think it's a great area for me, where I'm able to influence the product from two different directions. I've been doing this now for two years. One of my clients is Kleinanzeigen, and at Kleinanzeigen I do primarily two things. I build prototypes, because as the largest marketplace in Germany there's a lot of ambition to think about how AI can transform the business. And Kleinanzeigen has a private equity board. So it's always about first building these compelling vision narratives, building these prototypes, bringing them in front of the investors and getting their buy-in. Some of these do get buy-in, and then the next step is to actually go and build them in production. I want to take you through this journey. We only have 20, 25 minutes, so we can't talk a lot, but I'll share the key learnings on how this looks and take a very concrete example to see how we built it. So with that said, let's start. Five stages of a 0-to-1 product, at a very high level. Whenever we want to build something, we build a prototype, and once that prototype finds some sort of resonance inside the company and with the customers, we say, okay, let's bring it live. We bring it live with 1% of customers, then 10% of customers, and slowly we scale it. You know the deal. What we are going to talk about is primarily the first three parts: prototyping and bringing the first stage of the product into the market. I think a couple of years down the line we're going to talk a lot more about unit economics and monetization of these AI agents, especially at B2C scale. We still see very little of it. Most organizations are still in the first phase of finding product-market fit and building the first version of a product that works. So that's the focus today. What is Kleinanzeigen AI? This is the product in a nutshell. It's a shopping assistant, and we're not the first ones in the market with that, so you might have seen it somewhere else as well. What can it do? It can talk to you. It can search products for you. It can answer questions about a specific product. It can compare prices. It can answer any support-related questions you have, and any general shopping advice you need. That's basically the product that we have brought live so far. I think the first question when you look at it is: really, another chatbot? Because there are a lot of chatbots available. Pretty much all the major consumer product companies have built some sort of agentic platform where you can speak and get some work done. That was also the question we had when we started. Why should we build this? What's our unique value proposition? What does a customer gain? There were four areas that we thought about. The first one is a changing buyer role. Compared to traditional retail, on Kleinanzeigen a user is a buyer and a seller at the same time. You might be buying something, you might be selling something, so your agent has to take care of both aspects. Second, at Kleinanzeigen we have a very large range of categories. You can buy a house on Kleinanzeigen, you can buy cars, you can buy cardboard boxes for your next move. You can buy basically everything. What we do not have, however, is a very deep content base. Compared to an Amazon or a Zalando, who are retailers and sometimes also manufacturers, there's much more structured data that the AI can use there, whereas on Kleinanzeigen none of that data is available. The third aspect is human-to-human connection. On Kleinanzeigen, typically 10 to 15 messages are exchanged between a buyer and a seller before they decide to actually get into a transaction. On a retail platform this does not exist, so this is also very different and unique. And the fourth is user location preferences. On a retail platform it doesn't matter where you are; you just buy something and, depending on the shipping cost, it's delivered to your home. On Kleinanzeigen it's different. If you're looking to buy a used couch, and I'm sure some of you have had that experience, you say, "Okay, maybe I can look around in Friedrichshain-Kreuzberg." But if you're looking to buy a car, you say, "Okay, maybe I can take a train and go to Leipzig and find a car there." And if you're looking for an iPhone, you say, "I don't care, somebody can ship it to me," as long as it's a trusted seller. So location is a very fluid term. Those were the main challenges. The way we thought about it, we didn't want it to be a chatbot but a conversational platform that can provide all the capabilities users have today, in the form of a conversation over a period of time. That was the vision, and I would divide the journey into two parts. First is the prototyping. This is a pretty broad vision, so when we started prototyping we were thinking: what exactly should we show our investors? What gets people excited? Over the last few years I've built a lot of prototypes. Some of them land, some of them do not. What I have learned is that the prototype should get some of your users, some of your colleagues and some of your investors, ideally all three, very excited before you start building it in production. I'd like to share three learnings on what gets people excited when they look at a prototype or a new concept. The first, and very important from my perspective, is that you prototype the vision and don't prototype the MVP. The last 20 years of product development have been hardwired into our DNA: we have to build a product, let's start with an MVP, let's think small and scale it over time. When it comes to AI agents, this approach doesn't work, for two reasons. First, AI prototyping has become quite fast, so people expect to see the full thing you would like to do. Programming has become a bit cheaper, especially on the prototyping side. Second, building an AI agent on the B2C side is an expensive proposition. It's a big investment, and big investments typically don't happen on an MVP. They happen on a vision, a two-, three-, four-year vision. That's why it's so important to look at what the vision is and show as much of it as possible. So the approach we took was to think about many different concepts. We thought a user is buying multiple things at the same time, so the user must have multiple conversations running. In one conversation the user is selling something, in another the user is buying something. Some of these conversations would be long-running, because when you're trying to buy a car or rent an apartment, it's a two-month, three-month exercise. It's not one conversation; it's a long-running conversation. We thought about things like web search, because we don't have a lot of structured product data, so sometimes we'll have to go to the web for the details of a specific product. We also thought about agentic follow-ups, where an agent looks on your behalf over a period of time. You say, "These are my requirements," and the agent says, "Okay, I will come back to you when I have something ready." And finally there's something called agent seller contact, where the buyer can tell the agent, "I like this product, but I would like to negotiate the price, and I don't want to talk to the seller for whatever reason. Maybe I can't speak German very well. Just do everything for me." And the agent takes it over and keeps having this conversation with the seller. That's the prototype we actually showed. What we brought live is a much simpler version: one conversation, only one session. It can search, it can give some nice answers, but that's it. It's not doing any agentic stuff. But I think it's very important, if you want to build this over a period of time, to show them the vision. The second key learning is that prototypes have to be built on real production data. I think we are going through what I would call a kind of prototyping slop these days. Almost everybody is able to build prototypes, and what really makes a difference is that your prototype is based on real data: your proprietary data, your proprietary endpoints, your assets. Then it makes a much bigger difference for your users and your colleagues to see how this will work in real life. One example: when we were building this first prototype, we wanted to show how well a conversational search can work. Typically search is lexical, so it uses keywords, and the whole search at Kleinanzeigen was optimized for lexical search. What we did was build our own endpoint for semantic search, which searches based on intent rather than keywords. To do that, we imported a couple of product categories, created our new semantic endpoint on AWS and built a prototype on top of it. The benefit was that we could talk to our users and give them a use case. We could say, "Search for a laptop for yourself," so they weren't giving feedback just on the quality of the UX; they were giving feedback on the results they saw. That was quite interesting for us. And the third one: because we build a lot of prototypes, because our vision is much more agentic and because we continually test new UX patterns, we have built a complete infrastructure for prototyping. How does that infrastructure look? We have a React front end. Typically we build iOS and Android applications, and building on iOS and Android is tough, as a lot of you already know. You need specific developers, especially if you're using native SDKs. That's why we built a React version of the app, which looks about 90% similar to the actual Kleinanzeigen app. Second, we built our own Kleinanzeigen MCP server. In the end it doesn't have to be an MCP server; what we needed was a way to authenticate our system against the back end and access all the capabilities we have there. And we also have our own deployment infrastructure. Once we had that, we were in a position to roll out any sort of prototype within a couple of weeks on working production infrastructure, which makes a much bigger impact, especially if you're going into board meetings. Just as an example, we built this agent which, as I said before, takes instructions from the buyer, talks to the seller, negotiates the price, negotiates pickups, and basically runs as long as it needs to make the deal happen. We did this with our investors, and they were actually able to negotiate with the agent in their own Kleinanzeigen app. It makes a very big difference, and the excitement we can create with it is very different compared to a front-end prototype. Those are some of my learnings on the prototyping side. The next part is more about production. Currently we are at around 5% of users, and I must say it has been an incredible experience looking at how people use the system. We have had some really interesting learnings, and I'd like to share some of them. The first one is about UI guardrails. We talk a lot about LLM guardrails, but in all the products we have built over the last few years, we basically built the UI in a way that allows only a certain set of use cases. On the left side is the traditional search box that Kleinanzeigen has: a small search bar where you put a couple of keywords and apply some filters. On the right side is an open conversational interface. And users are not looking for specific keywords anymore. They're asking questions like: Are MacBooks good? Should I buy an M4 or an M5? What is the price of a new MacBook? Show me some MacBooks. It's everything in one go. They're asking about product discovery, they want guidance on which product to buy, and they also want to see some products. One of the intuitions we've had over the years building MVPs is: start small and grow. But when you build an open conversational agent, it becomes very difficult to start from an MVP, because you cannot limit it to certain use cases. Your agent cannot say "I don't know" 50% of the time. That would be a serious breach of trust between you and the agent. That's why we created a product that is very horizontal, and then we build skills in certain areas only. That's the architecture we use. At the very top we have an orchestrator, and this orchestrator only looks at the user message and decides what kind of answer should be generated and who should generate it. Below it there are multiple agents: a search agent, a general agent, a product agent that can answer questions about a product, and a customer support agent that can answer any support questions. This does two things. First, it gives us a sense of what questions people are asking and what answers we're currently giving. Second, it stops the LLMs from hallucinating, because they always have some sort of baseline to answer from. It's still an MVP. We build a lot more depth on search and on product, but we make sure all the bases are covered. By the way, this is an architecture we'll follow in the future as well, because once you have these agents and every agent has its own context, we can keep adding more agents. The benefit is that more teams can participate. Currently there's one team building all the agents, but in the future another team can contribute an agent, and that agent can sit under the same orchestrator. That's the general principle we follow. Second, because we have so many agents with multiple LLM calls, one of our key learnings is that smaller models are way better than bigger models, especially for B2C use cases. The intuition always says: let's just take GPT-5, let's just take Opus 4.6, one of the best models, and give it all the tasks. The problem is that the larger models are a factor of 20 to 30 times costlier than the smaller models, and they also take much longer. When you're talking to Claude desktop or ChatGPT or whatever your favorite tool is, sometimes you wait a minute and a half for an answer. In a B2C application you cannot make people wait a minute and a half. The answer should be there within 5 to 10 seconds, and that's why we use smaller models. Another big advantage is task focus. We don't want a small model to do many things; it's a chain where every model does a very specific part of the job. A very big benefit of that is that we create a lot of observability in the system. We know what every model is supposed to do, we can check the input, check the output and see how well it's performing. As an example, this is the inside of the search subgraph. The user says something like, "I would like to buy a MacBook under €800 in Berlin." We have to do three things. First, we have to understand that the user is talking about MacBooks as a category, which maps to laptops, and there's a category ID on Kleinanzeigen that we know about. Second, we have to understand where the user is. The user might say "in Berlin," or "in Kreuzberg," or "Friedrichshain-Kreuzberg and Pankow," whatever. In the end we have to take that English and map it to actual locations. And third, for the category MacBooks, we take the filters we have and try to map the user query onto them. These are three different LLM calls for us. Because it's three different LLM calls, we can debug them at a very specific level once we bring these things into production. And because we have these very specific LLM nodes, we can run our evaluations very objectively. We have very well-defined inputs and outputs for every LLM call, which means we have many objective criteria to benchmark the performance of the system. We don't have to go into details, but I want to show you how valuable this is. These are two different models we're comparing here. We were using Haiku 4.5, a small model from Anthropic, and a Gemini model. The Gemini model has almost half the latency of Haiku. P99 for Haiku is almost 25 seconds, which is completely unrealistic for a B2C application. Not only that, Gemini also costs four times less. So, some free advertising for Gemini. I don't want to do that, but it's the model we're currently using. What makes this very powerful is that whenever there's a new model on the market, we can very quickly plug it into our system and run all these benchmarks to see whether it makes a tangible difference to the performance of our system. We've found this very helpful. We have now tested four or five different models, and currently we sit with Gemini. One of the main challenges we face today, and that we think about on a daily basis, is how to orchestrate these LLMs. At a very high level, and these are two screenshots from Anthropic, who wrote an article called "Building effective agents" back in 2024, they proposed two different approaches. One is a more workflow-based approach, on the left side, where the input comes in, you have all these different LLM calls, you define exactly what every call should be doing, and then you have an output. On the right side is what's called reasoning, which is basically what ChatGPT and Claude desktop and everybody uses: there's a big model going in a loop. It has certain tools, it makes those tool calls, it interacts with its environment, it decides whether the work is done or not, and it keeps going in a loop. Of course, from a PM perspective and a user perspective, the second one sounds very good. The problem is that it's very expensive, it takes a minute and a half, and it only works with really big models. So it's completely unrealistic to have something like this running for 5 million users a day. That's why, when we started building, we stuck to the first, deterministic workflow, and we were happy. We actually thought we were very smart and had done a great job. Then we went to production and saw this user query: "I'm looking for a 50cc scooter from a known brand, good condition." What our system did: the message came to the orchestrator, and the orchestrator correctly said, "This is search, search, you solve this." And search just took that query and tried to map it onto the filters it had. The problem is, it searched for "50cc scooter known brand," and on Kleinanzeigen nobody posts an ad called "50cc from a known brand." So we found nothing. Complete junk. If you think of a reasoning flow, how would it solve it? The orchestrator routes to the search subgraph, and search says, "The user is looking for a known-brand scooter. Let me figure out what a known brand for scooters in Germany is." So it does a web search for "Marke 50cc Roller," finds a lot of results and says, "Okay, I understand: Vespa, Piaggio. These are the brands I should be looking for." It formulates multiple queries, runs them in parallel, brings back all the results and checks: am I satisfied? Yes, I'm satisfied. Here's your answer. It took only a minute and a half and maybe $1. The thing is, we want to build the system on the right. That's what the user expects. The system on the left is dumb, and I think that's one of the biggest challenges in a B2C application: how do you actually solve this? We have a solution, and I'll show you how we've done it. It's a hybrid solution, something in between. The orchestrator gets the query, and then we say: we need some sort of reasoning in between. Somebody needs to reason about what a good scooter brand is. So we have a new node called a query planner. The query planner looks at the query, makes a web search if needed and then generates multiple queries. It's a search planner, and it creates five queries. We run these five searches, but we do not feed the results back into this LLM node. If we find enough results, we just show them to the user directly. However, if we don't find enough results, because you could be looking for, I don't know, a Mercedes S-Class for $1,000, and of course you're not going to find anything, then we go to the backup planner. The backup planner gets the information about what the query planner did and how many results we got, and says, "Okay, I understand. Now I can make some modifications. Maybe I can increase the location range, because for this category location isn't a big deal. Maybe instead of $10,000 I can look for $15,000." Or it starts removing some attributes. If you want the best AMG version of the Mercedes, maybe we remove the AMG and just show you the results without it. It creates its own set of queries, and we run those queries. But again, we don't feed the results back into the backup planner. The job is done. So what we're essentially doing is introducing reasoning without creating a loop. The moment we create a loop, a smaller model dies. A model gets very confused once it goes into a loop; it doesn't know what happened in which part of the loop. You need a bigger model for that. That's why we never feed back the results. We assume the results are good enough and show them to the user. If we don't find results, we give it to the backup planner, and the backup planner does multiple queries. And finally, all of this is synthesized by another node, which looks at every result and at the user query and acknowledges whether the results are good or bad. So we've created a system that sits somewhere in the middle. We can generate good results while still using small models and still generating a result in 5 to 10 seconds. That's our current solution, and I think it works quite well. Six months from now I might be describing a different solution, but that's the stage we're at in any B2C application: continuously trying new LLM orchestration mechanisms to see what works best. But it's not all roses. There are a lot of challenges on the B2C side, and some of them you've probably faced yourselves. The big one is 20 years of habit. All the UX patterns across the world have been optimized over the last 20 years. Applications like Kleinanzeigen, Zalando and Amazon have been very well optimized for a certain type of behavior. Now we all want to put AI on top, and sometimes it's very difficult. There's no space. There's no intuitive way to actually integrate it. One of the challenges we see: we do all this interesting work of finding results for the user, but in a conversation window we can only show eight results. And the user asks, "Why are you showing only eight? You only have eight? I don't believe it. You must have 200 of these." Yes, we do have 200. The problem is: how do we show 200 in a conversation window? For that we need a screen with infinite scroll, but then how do we go back to the conversation? It's not just a Kleinanzeigen challenge. I think every B2C application in the world has the same challenge: how do you integrate conversation into the standard shopping experience? Second, input is a bottleneck. People don't like talking to AI; that's very clear. What we've seen is that the more we can personalize, the better it works. People like clicking on suggestions that the AI shows them, and they expect the AI to be super smart. The AI should be telling them what to do. From my perspective, voice is the most natural input users have, but unfortunately voice doesn't work. Anybody who says voice works really well, I have my doubts. Especially at B2C level for 10 million users, at the cost you'd expect, and where people speak Turkish and English and Polish and German, as people on Kleinanzeigen do. I don't think we have that level of reliability yet. Third is infrastructure. We have used all different types of platforms, and everybody has downtimes. Downtimes happen during board demos. Downtime happens whenever it wants to, especially during US working hours. Sometimes the latency is 40 seconds for a model call that's supposed to take one second. That's why I think we're not yet at a stage where we can build really resilient, business-critical flows. The future we see is having our own models and starting to host them very soon, because when working with the bigger API providers, reliability is always a challenge. Fourth is cost and latency. Every week we read in the news about a new model and how good it is, but none of those models are useful for a B2C application, because they're just too expensive and too big for the work we're doing. So we always have to rely on small models, and the moment you start scaling a B2C application, cost becomes a major challenge, and so does latency. That's the trade-off we have to make literally every time. And finally, monetization. That's a very interesting one. I think ChatGPT has maybe a 1% conversion rate from free to premium, maybe 2%, and that makes it very difficult. Right now all these B2C consumer platforms are investing a lot on the AI side. We do have monetization somewhere on the roadmap, but it's not there yet, and I think there will be a watershed moment maybe 6 to 12 months down the line where people will ask, okay, what's the ROI of everything we're doing here? Another challenge related to monetization is cannibalization. Every search happening on this conversational agent is not happening on the search infrastructure that exists today. Does it have a better conversion rate? A worse one? Does it cost less than a conversational agent? All of these questions are still unanswered. That said, it's still fun work. The three things I would always suggest: first, create excitement around your prototypes, and you can only do that by showing the vision and connecting with your data assets. Second, build depth on your use cases and cover all bases. Especially if you're building anything around conversational agents, people can ask anything, and you have to be prepared for every eventuality. And third, LLM orchestration, context management and evaluations. That's basically what I do as a PM the whole day, and I think that's the role of the future AI PM, which is also very exciting. It's just fantastic work. And yeah, that's it from my side. Any questions? #### Q&A **Host (Francesca Cortesi):** Thank you so much, Rohan. We have time for… oh, look how many hands. Now I have no idea who's first. Thank you for helping with the mic. Maybe there in the middle. We have time for a couple, and then I guess you'll be around. **Audience:** Maybe a wide question to open it up. Can you share the success metric you want to achieve with this feature? **Rohan:** Right now we're at a stage where we focus very much on first-degree product metrics, which is how well the product itself is working. We look at how many people are interacting with it. We look at the average length of the conversations. And then, for everybody who is conversing, how many leads are these conversations generating? Those are the three metrics. **Audience:** What's a lead for you? **Rohan:** A lead for Kleinanzeigen is a topic in itself, because of course most of the transactions happen outside the platform. But basically it's a message being sent to a seller, or an ad being deleted after a certain duration. It's an inaccurate metric, but still a standard one. **Audience:** Maybe a quick follow-up: did it work? Yes or no? More leads? **Rohan:** No. I think normal search still works better; it generates more leads. Primarily because in a normal search you're looking at 200 products, and here you're looking at eight. That's the user behavior that limits us right now. **Audience:** Thank you, I have the mic already. A very quick one: super interesting what you're explaining and showing. Did you also build some conversational intelligence into your solution? Meaning, when you see what customers are doing, do you also see where they're struggling, where they're dropping off? Is that built in? **Rohan:** Yes. There are observability frameworks for it. The one we use is called Langfuse. All the data flows into this observability framework, where we can do multiple levels of analysis. I'm personally not a very big fan of LLM-as-a-judge, which the AI industry sells to everybody. You can change the prompt and completely change the metric that LLM-as-a-judge calculates. So I'm still a big fan of having that data anonymized and in a readable form, and a PM or a product engineer actually going through all the different feedback. I think we can still do that at this stage. In the future, I think it will be based mostly on user feedback: users clicking yes or no on a specific conversation, and then us looking into it. **Audience:** Can I go? Okay. Quick question: how easily can you open this infrastructure not just to the internal interface but also to third-party agents, like OpenClaw or something like this? **Rohan:** Correct. We have actually built an MCP server, and that's why we have it, because we prototyped the whole thing. The company is thinking very actively about this. In fact, once you connect it to a massive model like Opus 4.5, it can do a crazy amount of stuff. I was surprised what these models can do once they have full access to the data. They can build you charts, they can do ROI analysis, whatever you want. The problem is monetization. The MCP server is a good solution, but there's no monetization layer built on top of it. The company does want to start experimenting, but currently there's no monetization. It's just cost for now. **Host (Francesca Cortesi):** Last question, we have time for one more, there. **Audience:** Hi, thanks for the talk, very interesting. My question is: why did you decide to build this product? What was the reasoning behind deciding to come up with this AI assistant? **Rohan:** That's a great question. First of all, we are in a hype cycle, so everybody's asking: what's your strategy on AI? Second, we looked at competitors like Amazon, which has been building Rufus for the last couple of years. There's a certain set of use cases that actually work quite well, and for us it was first about finding out what those use cases are. We could discuss theoretically whether people would use it or not, but the easiest thing is to build it, bring it to production and see what people actually use. And the second thing is that Kleinanzeigen has a very strong aspect of people talking to each other. One of the big seller pain points is that once you put up an ad, 20 people message you, you're exchanging all these messages with all these people, and ultimately your product isn't sold. It's a lot of effort for limited reward, and that's where agents could become really powerful. So I think conversational flows integrate very well into the Kleinanzeigen way of commerce. **Host (Francesca Cortesi):** Thank you very much, Rohan. A big applause to you. ### 18:00 · Still Human — Michelle Parsons Video: https://youtu.be/uLFkKX-4pBc - **Speakers:** - [Michelle Parsons](https://www.linkedin.com/in/michcparsons/) — CEO & Co-founder, Lora. Building Lora, an AI-native self-discovery platform. Previously Chief Product Officer at Hinge, with product roles at Spotify and Netflix. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 18:00 CEST - **Format:** Closing keynote - **Host:** [Daniele Ronca](https://productlab.app) - **Official abstract:** She has built for human connection at Spotify, Netflix, Hinge, and now Lora. Her closing keynote asks the question product leaders will keep returning to: what makes us irreplaceable in the age of AI? **Summary.** Introduced by Daniele Ronca, Michelle Parsons, founder and CEO of Lora (launched in the App Store that week) and formerly at Kayak, Spotify, Netflix and Hinge, closes the day with a talk about what it means to be human when AI optimizes everything. She draws on four platform shifts from her career, explains how Lora was built community-first with AI speeding up feedback loops rather than replacing human contact, and argues that empathy, discernment and lived experience are what set product people apart. **Key takeaways** - Change is not new: Michelle lived through desktop-to-mobile at Kayak, curation-to-algorithms at Spotify, the start of the streaming wars at Netflix, and Hinge overtaking Bumble and Tinder by focusing on real outcomes for real users. - Her worry: if social media made us lonelier by optimizing for speed and consumption, AI without guardrails could push us further; she doesn't want agents talking to her friends' agents to set up dinner. - Great products start from a human emotion; whether consumer or B2B SaaS, there is still a human on the other side of the screen. - When building is cheap, what matters is why you build and for whom. Lora was built community-first, with human organic content (one astrology carousel reached 1.8 million views), an ecosystem of touchpoints, and co-building with a 10,000-person Discord; she reports 30M+ organic views, 3,000+ event attendees, a 160,000-person waitlist and 80,000 Substack subscribers before launch. - AI shortened feedback loops so she had time to personally email every beta user, rather than replacing human contact. - Three human edges: empathy (understanding the why behind the data), discernment and true taste in an algorithm-shaped world, and lived experience and intuition coupled with data. #### Transcript **Daniele:** So, are you ready for the keynote? Yes or no? Okay. A quick word, which I want to say not because I must. Last year we did this amazing conference, and we wanted to grow with board members. For one reason or another, they come mostly from Berlin, except Francesca, from the Nordics. And one of the problems she raised was: "Daniele, there are no women leaders on this board except me." And I was like, yes, it's true. Because when I look deep into LinkedIn, there are not many women leaders who support us in the community. They join over time, but I found there's a clear problem we see in the community and in the Berlin tech scene. So when I started to craft this conference, I was thinking: okay, we have Ravi Mehta, everybody knows him. But how can we close with someone who, if I opened the dictionary, would be the picture of a woman tech leader, inspiring a new generation of people in Berlin, from junior to senior PMs to founders like myself? Not because it's mind-blowing or whatever, but because I think diversity, bringing people from many nationalities like we have here in Berlin, is what makes us different. It's quite multicultural, with a lot of talent from all around the world, and that's also part of being different. So I would like to welcome her to the stage. She's coming from New York, and she didn't sleep, like all of us. She has worked at Netflix and Spotify, she built another company and had an exit, and now she has a company that launched last week, this week actually, with amazing traction. With Ravi I made a lot of fun of it, and my pickup line was: "Hey Michelle, we have the ex-CPO of Tinder. If someone from Hinge comes too, maybe we can solve the dating situation in Berlin." And another thing that sparked something for me: I met my partner at a meetup, and we had a newborn 40 days ago. What also sparked me about her new venture is that I discovered things went better in this phase of my life when I started to switch off all these apps and really think for myself: what do I want, what do I want for my future. Take time for myself, pause and move on. The talk will be "Still Human." I would like to open the stage for Michelle Parsons. **Michelle:** All right. Well, we've made it to the end. It's been a very long day. As Daniele said, I have not really slept, but that's what being human is all about, right? Powering through and doing the impossible. My name is Michelle Parsons. I am the founder and CEO of a brand-new company that literally just launched in the App Store on Monday. Lora really is a culmination of the 15 years I've spent building consumer products at some of the biggest companies in the world. I've spent a lot of my time thinking about the problems of human connection, and entertainment really. We want to experience and feel and hear and see things that make us feel something, that make us connect back to ourselves and to others. What I started to observe was that I kept hearing about the loneliness crisis. Why are we all so lonely? I started to really try to unpack this question. And what I got down to was: in a world where we have access to literally everybody, to the most intimate moments of the people closest to us and of people we don't know, there's no reason for us to be lonely. But we've lost the plot a little bit. We've lost the muscle. It's atrophied, because we haven't exercised going deep. So Lora is all about helping you go deep with yourself as an on-ramp to go deep with others. I'll talk a bit more about that later. We've talked a lot today about AI optimizing efficiency and productivity, agents talking to agents, humans nowhere in the loop. So I wanted to talk about what happens when everything is so optimized and productivity is at an all-time high. What happens to us? This is really a talk about being human. My career journey, as I mentioned, is very abbreviated here, because a lot happened before and a lot happened after. I joined Kayak as employee number 100. It was really my first consumer job and the first time I was ever doing product. I remember that when I got to the company I went straight to an offsite, and on the board the CTO had written "the Wild Wild West." And I was like, oh no, what did I get myself into? I spent two amazing years there, and then I joined other companies: Spotify before they were public, Netflix before the streaming wars were a thing, and Hinge as employee number 80. These companies all have one specific thing in common, which I'll come back to in a bit. But if you can, spend 30 seconds thinking about these companies and brands: what might make them different or special in their categories and in the market broadly? So, AI. Let's talk about AI. When I first talked to Daniele, he said, "Let's talk about AI. We'll have a conversation about AI all day. People are scared about AI, AI is changing things. What do we talk about?" And it's interesting where we started. "AI is taking our jobs. No one is safe. All of your jobs will change. They'll be different. They'll be gone." In fact, many companies used this as a reason to do mass layoffs. Then two or three months later we saw the headlines: "Actually, what does AI do again? How much more productive are people? What outcomes are we actually driving?" And then just yesterday it was: "AI is going to kill us." Okay. Another day, another headline. The only consistent thing is that change and acceleration are happening faster than we can comprehend. But I know one thing, and that's that we've actually been here before, many, many times. Raise your hand if you've experienced any kind of shift in your job: the way you work has changed, you've had reorgs, you've gone from IC to manager, the product you were working on became redundant and you had to work on something else, you've been laid off. I know I have. Yeah. Every single one of us has been impacted. We've seen change, and sometimes it feels like it comes on so rapidly, but in fact there are often lots of signals. So I want to talk about four distinct moments in my own career where we saw a before and an after, just like we're seeing a before AI and an after AI. When I went to Kayak, it was all about desktop. Literally 80% of all bookings, especially for hotels and flights, the big purchases, were done on desktop. The mobile team was a pod of six or eight people. Within under two years of my time there, this completely shifted to mobile. Things like HotelTonight launched. Airbnb became way more popular. Online shopping more broadly was becoming something everyone did. And travel became way more accessible. Flights weren't thousands of dollars, hotels weren't super expensive, and you could go sleep on someone's couch. At Spotify, I joined before they were public. A large part of the music industry, and even of Spotify's internal strategy, was all about curation. They had taste curators who literally constructed playlists like RapCaviar, defining who was top of the charts and what culture said. And then all of a sudden the algorithms came in. Now we had things like Daily Mix and Discover Weekly that would predict your taste. Today TikTok defines trends and artists faster than Spotify and the music industry can keep up. From one to everyone. When I joined Netflix, it was before the streaming wars. There was no Peacock and Disney+ and insert-every-other-name-plus here. It was just Netflix. People were racing to license their catalogs to Netflix. And then all of a sudden everybody had a streaming service, and competition was the only consistent thing. Sometimes you're losing, and then all of a sudden you're winning, you're the market leader. When I started at Hinge, Hinge wasn't even on the radar. Bumble reigned supreme. They had the most market share, they were the fastest-growing company. They were really the only dating apps that mattered. And within one year, we put a discrete focus on how you actually drive outcomes for real human users. That simple change propelled the company to become one of the fastest-growing dating apps in the entire world, surpassing Tinder, surpassing Bumble. And literally just two days ago, Blackstone, Bumble's largest investor, was pulling out. AI has completely changed the way we work. It's made us move faster, which is awesome. It's made us smarter and work smarter. It's made a lot of things much easier. We can do things with far fewer people. I literally learned brand marketing in one night. I built an app in a couple of weeks. There are such amazing things we can do with this technology. But what does it mean to be human when everything is optimized? In a world where change is the only constant, where do we as humans, as builders, fit in? I really think our value isn't in becoming more productive. All I hear about today is agents, the ways I can connect my MCP to my Claude, to my ChatGPT or my Codex. It's all about taking away all the extra work that, in so many ways, makes living worth it. I want to text my friend to see when they can get dinner with me. I don't want my agent talking to my friend's agent, setting up a meeting at some random restaurant that happens to have availability. So I sit down and look at the summary from my agent that says, "All right, here are the bullet points." And I'm like, "How are these bullet points for you?" And then, what do we even talk about? If social media made us more lonely because it optimized for speed and entertainment and consumption, my fear is that AI without guardrails is only going to send us further down that hole. Great products are built from human emotion. And one of the amazing things AI does is take away a lot of the grunt and busy work and level the playing field. So when I said that Hinge and Netflix and Spotify and Kayak had something in common, this is what I mean: they build for humans. Because the person is the one booking their dream vacation to hike Machu Picchu. A heartbroken teenager might be sitting in her bedroom, distraught over losing her first love, trying to find a song to make it all better. A dad sits with his young son and says, "This was my favorite show when I was your age," in a bid to bond. And a hopeful romantic is on an app trying to find her soulmate. Every single product decision you make starts with a human emotion, with a feeling, a pull that you need to connect with something and someone. So who are we in all of this? We are humans building for humans in real life. I think that's so important and critical in a world where I can automate my PRD reviews and data analysis. All of those things can be automated, and we can move as fast as possible. But what can't move fast is our need to truly understand ourselves and each other, and the user, the person, the human on the other side of the product. Whether it's a sexy consumer product like the ones I like working on or some B2B SaaS, there's still a human sitting on the other side of that screen. As I mentioned, three days ago I shipped one of the most personal things I've ever built. It was a product that came out of years of going through a really rough breakup, having my career change from under me overnight, and in so many ways feeling so lost, and thinking: how do I build something that I need, that I want, and that I believe would actually help people get off their screens and slow down just a little bit? The night before we launched, I wasn't checking my dashboards or asking Claude Code about optimizations. I was literally reading messages from real users, people in our beta who had been using the product for a couple of weeks. We had created highly automated loops to shrink the feedback cycles with AI, which is fantastic, because it gave me the time to individually email every single one of those people. We got the feedback super rapidly. We use AI throughout the entire process and our workflows to optimize, but we didn't replace human-to-human contact with an agent or some type of automation. We did it ourselves. My approach to building in the age of AI, when building is so cheap, and I was talking to somebody about this at lunch, is that the only thing that matters is why you're building a product and who you're building it for. I think we've been able to take a really differentiated approach, using a bunch of different techniques, tools, frameworks and playbooks that many people have popularized, with my own spin on them. We started with community first. If anybody can build a product in a couple of weeks, and I build an awesome product, someone can go copy it. What is going to ensure that people choose to show up with me, with Lora, with my company? That's really anchored in community building, brand resonance, and building connection between you and your customers and among your customers. We had a super keen focus on organic distribution. How do we use all the top-of-funnel places where people already spend their time to answer the questions our product was solving? And we did it in a way that was really human, without selling the product or showing the product. Again, I'm building a product about human connection, spirituality and self-understanding. One of the ways we do that is through the framework of astrology. Don't laugh. The content piece I'm showing here had 1.8 million views, with demographics split pretty evenly between male and female across multiple age ranges. It was something you could scroll through, 10 slides, and feel seen: "Ah, that's kind of like me," or "Ooh, that's like you, friend." And that created a highway for communication and conversation. We didn't just build a product; we built an ecosystem. We call it the Lora world, where every single touchpoint feeds into the next. You might discover us on social or hear about us from a friend, but when you get into our world, we keep taking you deeper and deeper by providing real value and broadening the ways you can connect with us. We have a Substack, we have a Discord, we have about 10 different mini experiences and microsites, we have our app, and we have in-person and virtual events. All these different ways you can connect and communicate with us. And most importantly, we didn't build in a vacuum. Again, building is really fast. Many investors say: just build a product, put it out and see if people want it. Well, why do I need to do that when I can co-build with people? So we built a huge Discord community of about 10,000 people, where we actually prototyped our very first product in real time, getting feedback instantly as we rolled out changes to algorithms and feature ideas. Building is so different today. It doesn't need to follow the same playbook we followed even a year ago. We moved from testing across all these channels, to getting real signal, to knowing exactly what we needed to build. In a few months we had, and this number on the slide is even wrong, I think over 30 million organic social views. We hosted over 3,000 attendees from our community at virtual and in-person events. We had a waitlist of over 160,000 people in our ecosystem before we even had a product available. Our Substack newsletter has 80,000 subscribers. And our Discord is super active: we have 10,000 people who show up every day to meet each other and engage with me, the Lora team and our practitioners. What we learned is that life doesn't fit neatly into a single box. The people, places, things, milestone moments, the decisions, the questions, the challenges, the successes: all of these make up who you are. But today they live in fragmented places, across different products and surfaces. We're not people who live only on LinkedIn or TikTok or Instagram or at work. We're multifaceted individuals, and each experience and each person defines who we are and how we are. So we built Lora to reflect your life. This was supposed to be a video, but oh well. I'm not going to repeat what all the wonderful speakers said about creating and using AI in their workflows, because again, I use that day in and day out to give me the time and space to focus on the things I think are critically important. The first is empathy, and I think the opening speaker talked a lot about these topics as well. Our number one job as product leaders and product managers is not moving tickets from point A to point B. It's not being engineering managers figuring out how much scope we can fit into a sprint. It's really about understanding who the people behind our products are. Today AI can summarize every single interview or data point you have, but it doesn't tell you why those things are happening. That's where I really encourage everybody to dive in and lean into their own understanding of themselves and of other people, to really put themselves, and I always do this in my head, into someone else's shoes. I heard a lot about discernment when AI was coming in: building is cheap, so taste and judgment are the only things that matter. And I scoff a little bit at taste and judgment, because I'm not even sure our taste or our judgment is ours anymore. We've been so influenced by TikTok and the algorithms that trends rise and fall instantaneously. I always point back to the Labubu craze, if anyone remembers that one. Overnight Labubus were everywhere, Gucci was making clothes for them, people were getting them tattooed. It was a wild time to live, and then almost overnight it was gone. But people want to connect around something. So I do think that figuring out how to use your discernment, your judgment, true taste and creativity will always be a defining factor that sets you apart from everybody else who has access to the exact same tools. And lastly, lived experience. When I was growing up in product, I always heard: don't use your gut. Make decisions from data, from the insights. And that was very true, because you don't want to put bias into things. But I think in this age of AI we're coming full circle in so many ways. Use your intuition, your lived human experience, what you think, what you just know, and couple that with the data and insights. That will set you apart, because it's the one thing no model has. You can upload all the papers you've ever written, all your meeting notes, etc., etc., and it's still going to be extracting based on some non-human framework. The best things I've ever built in my career started here: trying to deeply empathize with people, leveraging my intuition, pairing that with the data, and then figuring out what exactly we should build and why. We are still human, and that is not a limitation. No matter what the articles say, no matter whether you're ever laid off because of technical changes or shifting company priorities, being human is literally the entire job. So that's my talk. If you download Lora, I'd love your feedback. There are a bunch of product people here, so I would very much value it. And let's stay connected. Thank you so much. ### 18:30 · Final Remarks — Daniele Ronca - **Speakers:** - [Daniele Ronca](https://productlab.app) — Founder, Productlab. Founded Productlab in Berlin in 2023: the Productlab Conference, Leaders Studio coaching and a year-round product community. - **Stage:** Contentsquare Stage · CIC Berlin · 17 Sept 2026, 18:30 CEST - **Format:** Closing **Summary.** Productlab founder Daniele Ronca closes Conference Day 2026 by thanking speakers, facilitators, volunteers and sponsors, and sharing a few reflections on where AI is taking the industry: more work rather than less, the rising cost of "intelligence," the data and know-how companies give away, the scarcity of business-driving product talent, and agentic buying. He ends with a community selfie before the party. **Key takeaways** - Last year's prediction held: with AI we work more, not less, because we are asked to deliver far more output. - Referencing Satya Nadella, he argues we now pay more and more for intelligence, while also handing our data, IP and ways of operating to the model providers. - Talent and compute have both become more expensive; product people who truly drive business outcomes are rarer than before. - Agentic buying is still under 1% of transactions (per a conversation with someone at Stripe), but winners will deeply understand the behavior of both users and agents. - Community-led building, brand and experience are what will differentiate; he asks attendees to take home at least three learnings to share with their teams. #### Transcript **Daniele:** Thanks. Hello. Michelle, I think she deserves a much stronger applause, so please help me. I'll try to be super concise in my presentation, because it's been a long day. *[Asks the tech team for the slides.]* As we did last year, from my side there are a couple of reflections on the day, on where we're heading, and on what we found interesting. There's also a little gift for the challenge of the best LinkedIn post of today, which we announced. We'll also do another nice thing: invite all the speakers, the facilitators, the volunteers, all the people involved today, to come on stage, and we'll take an amazing selfie with the community. Then there are around 200 or 300 drinks downstairs, an amazing DJ, and I guess the best talks happen just then. Sorry, and ice cream. Ice cream from Italy, with love. Let's start super fast. Thank you, everyone, for joining today. For us, these have been the most challenging three days we have ever attempted. Of course, we're two years older, but it has been fun, and I want to say thank you to the people who bore with us until this moment. I think the applause goes to all the speakers, the facilitators, the volunteers, whom we doubled, and all the amazing sponsors. So please, an applause for them. The day, and I hope you got some insights, has been all about the implications of AI. From how behavior is changing, as Peec AI showed, and I want to say thank you, because we started implementing things and more people have been finding us through Claude or ChatGPT or whatever in the last two months, and it worked. So AI creates a lot of things. We save time. We make things faster. We can now get to Lenny through the MCP and everything. But it has also changed how we work, whether we like it or not, and it's also a lot about how we lead today. Teams are shrinking. There's the new product builder popping up. We don't know if we should have the designer, or maybe I should be a product manager, maybe I'll combine them, but we want the engineer. So, good luck. A reflection. Last year I was quite right. I said: guys, we're going to work more, not less. And it happened. I don't know about you, but we work more, because we're required to deliver way more outcome. No, sorry, not outcome: output. The next reflection I found super interesting comes from Satya Nadella: we used to work less, and now we pay twice for intelligence. Let me explain a little. Because there's a lot of compute, data centers, infrastructure and politics involved, of course we have to pay, and if you notice, it's going to get even more costly. The €20 plan one year ago was amazing. Now it barely replies to where I should go to a restaurant tonight, and you have to go to the €90 one. And I'm pretty sure next year it's going to double. And intelligence-wise: we buy intelligence, but to become really intelligent, they buy our intelligence. And good luck if they're not learning from us. What he envisions for the very near future is that people will buy LLMs like they buy from AWS, or you name it. I think this is a paradox that should bring some reflection: what kind of data do we give away, and also our way of operating companies. That was our IP: our way of doing things, our way of caring for customers. And we're just giving everything away, because we bought intelligence, but it's not really intelligent, so let's give everything out. I don't want to criticize; it's more a reflection on where we're heading and how the industry is moving. Some things to bring home from our side. From what we see, talent and compute have become more expensive. I do believe product management, and thanks a lot, Michelle, is not dead at all. We still need people. And the more I speak with leaders, the more I hear there's no real talent: people who drive the business. For a business, if I invest X amount in a salary, this person should bring back X plus something to the company, so I can grow and do a thousand things. Those kinds of people are rarer than before. And last but not least: I had a chat with someone from Stripe, and agentic buying is still less than 1%. But I think the companies that will win in the future, whether or not agents are used for small things, like buying a pen on Amazon where you don't need to spend your own time, will also be the ones that deeply understand the buying behavior of both users and agents. These agents are going to have your credit cards, our credit cards, and they're going to be happy to spend more than ever. That makes me a bit cautious: people will be out there saying "I'm the best" to win them over, and it won't be true. This is something to really take care of. Getting back to us as a community: where we are and how we operate. I didn't see Michelle's presentation beforehand, but luckily it's quite aligned. I think community-led building, for us and for others going in this direction, and how we create our brand, our agency and our experiences, is what's going to make you different from others. So again, as I said, I want to say thank you to everyone for sharing and joining. I think you got enough to reflect on, and if you bring home at least three learnings to share with yourself and with others in your team and your company, I think this has been a successful conference. Thank you for showing up and believing in us for the second year. We grew, I guess, double, and this applause goes to you. Meanwhile, all the volunteers and speakers, come on stage for the selfie. Thanks a lot. One, two, three. Thanks a lot. And the drinks are from the sponsors, so please finish them all. --- ## Miro Stage Deep dives · interactive working sessions · livestreamed. Hosts: Veronika Morozová (morning), Andreea Lungulescu (afternoon). Recording: https://youtube.com/live/XyJR7WE0TbM ### 10:30 · From AI Assisted to AI Native: Rewiring how your teams innovate — Joe McLean & Vihar Parikh - **Speakers:** - [Joe McLean](https://www.linkedin.com/in/josephmmclean/) — Product Director, AI Canvas, Miro. A decade at the intersection of software and creativity across design, product and engineering; brought AI into music-creation tools at Splice from 2018. - [Vihar Parikh](https://www.linkedin.com/in/viharparikh) — Group Product Manager, Custom Solutions, Miro. Leads Custom Solutions and Miro's core navigation experience (Home, Spaces, Custom Widgets); a decade on document systems, vision/OCR, big data and now AI. - **Stage:** Miro Stage · CIC Berlin · 17 Sept 2026, 10:30 CEST - **Format:** Talk · Track: Build & Craft - **Host:** [Veronika Morozová](https://www.linkedin.com/in/veronika-morozov%C3%A1/) (Contentsquare) - **Official abstract:** AI made every person on our team faster. It didn't make the team faster. Vihar and I share what actually changed at Miro: the new bottlenecks that show up once everyone is moving fast, and what aligned people, shared context, and fast decisions look like in practice. **Summary.** Joe McLean and Vihar Parikh of Miro argue that AI speeds up individuals much faster than organisations can absorb, creating a "collaboration gap" between fast people and slow company processes. They describe the fractures behind that gap, how Miro rewired its design and prototyping process around vibe coding, and what tooling and culture an AI-native team needs to stay aligned at speed. **Key takeaways** - Joe describes AI capability doubling "over and over again" in his own coding since 2023. Gains come from better models and also from better harnesses, interfaces, context and multi-agent orchestration. - AI adoption is a curve, not a switch: holdout, explorer, scaler, AI first. Individuals move up that curve much faster than company processes, rituals and tool stacks do. - The collaboration gap shows up as four fractures. Silos now include human-to-agent and agent-to-agent work as well as human-to-human. AI amplifies misalignment, creation scales while decision-making doesn't, and the reasoning behind agent decisions becomes invisible. - At Miro, vibe-coded design broke old handoffs such as Figma files for marketing. The fix was one shared prototype repository ("Vibe Lab") built on the shared design system, prototypes brought into Miro boards for commenting, and a new "concept review" step right after kickoff. - Keep a synced source of truth (for Miro, often a table) separate from disposable presentations. Correct outdated decisions quickly, because agents summarising Slack spread misinformation. - "Small increments" means shipping in small amounts of time, not reducing ambition. Every day you don't ship adds to the coordination tax when you merge back. *[The recording starts partway into the session; the opening minutes were not captured.]* *Full transcript to follow after the speaker review.* ### 12:05 · WeRoad: An AI Journey — Cassiano Vailati & Matteo Risso Video: https://youtu.be/uFfM1Ea5xEI - **Speakers:** - [Cassiano Vailati](https://www.linkedin.com/in/cassvail/) — Head of Technology, WeRoad. Leads software and data engineering at WeRoad; joined in 2018 and scaled the team from 5 to 50+ engineers. Focused on AI platforms, AI products and company-wide adoption. - [Matteo Risso](https://www.linkedin.com/in/matteo-risso/) — Senior Product Manager, WeRoad. Leads squads on lower-funnel conversion, ancillaries and post-booking across Europe and the US; led the product side of WeRoad's US launch. - **Stage:** Miro Stage · CIC Berlin · 17 Sept 2026, 12:05 CEST - **Format:** Talk · Track: Build & Craft - **Host:** [Veronika Morozová](https://www.linkedin.com/in/veronika-morozov%C3%A1/) (Contentsquare) - **Official abstract:** How do we evolve a company to AI Native? Cassiano and Matteo will share the two years WeRoad's journey through AI products, engineering, company-wide rollout, the learnings and what it unlocks for product teams. **Summary.** Cassiano Vailati, head of technology at WeRoad, and lead product manager Matteo Risso walk through two years of AI adoption at the Italian group-travel company. The journey has five tracks: AI products, engineering, company-wide adoption and training, a shared "company brain", and automation. For each one they give costs, returns and challenges, and then Matteo describes how the product manager's job changed along the way. The Q&A covers build vs. buy for knowledge tools, the Airbnb investment, stakeholder management and how a PM shipping MVPs alone still works with design and engineering. **Key takeaways** - WeRoad built its own AI stack (RAG search, a tone-of-voice translation system, evals) before launching its chatbot. The chatbot shipped in five languages within weeks. It raised automated resolution from 41% (the old Zendesk bot) to about 80%, and satisfaction rose from 68% to 86%. It costs about 30k a year in LLM spend against roughly 73k in saved sales-team time. - In engineering, the big shift came with Claude Code and Opus 4.6 at the start of 2026. Merged PRs rose 59% year over year, production releases rose 34% and lead time fell from 71 to 19 hours, with a smaller team. The open problems are the learning curve, cognitive overload, and a growing share of code that no human has reviewed. - Giving everyone licences caused "AI sprawl": conflicting dashboards, apps hosted anywhere, security gaps and "AI talking to AI". WeRoad answered with guardrails instead of bans: an internal dashboard-sharing tool, Coolify for self-hosted vibe-coded apps, network controls, and an AI specialist programme of 12 internal champions. - A shared company brain (centralised knowledge, RAG on Onyx, a documented data warehouse and a semantic layer, with an ontology planned) cut answer time by about 81% and tokens by 86% in their tests. Build vs. buy and vendor lock-in stay open questions. - AI is "at a discount": moving to an enterprise plan would have taken one bill from 200k to 1 million. Teams should check whether they could switch vendors and whether they could afford the real price. - For PMs, shared context turned week-long data requests and grooming cycles into same-day analyses and working MVPs. Judgment is now the scarce skill, and other roles, such as design, are catching up to the same speed. #### Transcript **Cassiano:** Hello everybody. I didn't know we were in a cinema, so I think the slides are a perfect fit for this presentation. Today we'll talk about our journey with AI at WeRoad over the past two years. First of all, for those of you who don't know us: we are a travel tech company born in Italy in 2017, and now we're open across the EU and also in the US. We organise group trips for small groups of people who want to travel together, have real-life experiences and make new friends. I'm Cassiano Vailati, head of technology, and I've been at WeRoad for the past seven years. With me is Matteo. **Matteo:** Hi everyone. Lead product manager. **Cassiano:** So let's start. I think this picture represents what landed on our table a couple of years ago: LLMs that were gifted to all of us. Gifted, but they come at a cost. Our journey is divided into five tracks. First, AI products. Second, the engineering side. Third, how adoption and training happened inside our company. Fourth, the AI brain we built and how we handled our data with AI. And fifth, automation. So, AI products, the shiny things. Like most companies, the first question we asked ourselves when we started was: why don't we build a chatbot? That's where everybody started, and we started there too. At the beginning we built an AI search, a RAG system, so retrieval of information to search our catalogue. From that we moved to the chatbot, a more conversational interaction with our customers covering all our information: our travel information, our FAQs and so on. The journey has been quite long. It started in 2024, when we began using vibe coding to build our own stack. While building the search, we decided to do it internally so we could understand the technology and use the knowledge in engineering to build products. The next step was building tools like a translation system that uses our own tone of voice, which is different from just asking for a translation. Then we developed our chatbot. A very important part of that was understanding that the technology was moving fast, so we needed to build evals, testing frameworks for the quality of our outputs. That led us to 2026, the beginning of this year, when we released the chatbot on our website. Since we had built the whole stack underneath it, including the translation system, we were able to release it in all five languages about three weeks after the Italian release. And because we had built this tech, we were also able to build a ton of projects alongside the chatbot. By the time we released the chatbot, we already had 10 or 12 AI products. We'll go through some of them here. Of course there's the search and the chatbot. From the chatbot we spawned other chatbots for our partners and for our coordinators, who are the people who go on the trip with the group. We built internal chatbots for IT and DevOps automation, so they can answer engineers, product managers and everyone in the company when they need to understand where things are not working. We also added AI features inside our products. For example, we generate a lot of content on our web pages: FAQs, SEO tags. We tag all the images we get from our customers or from our own portfolio. And we do CMS translation, again using the same tools we used to build the chatbot. What have been the costs and the return on investment of all this? Here's the example of the chatbot. We already had a chatbot at the beginning, from Zendesk, which we use as our customer messaging tool, and it could handle 41% of chats. As soon as we released our own chatbot, that moved up to 70%, and in the last couple of weeks we're at about 80% in Italy. Customer satisfaction moved from 68% to 80%, and in the last months we reached 86%. So, a big improvement. The conversion rate is an interesting part. It's around 25%, which is high for just speaking with a chatbot. When a human steps in (we have an escalation process, of course), it grows higher, probably because we still cannot replace the empathy of a human with just a bot. That's the real challenge ahead. On costs, we're paying about 30k a year in LLM costs. The return is the time of the sales team, who no longer have to reply on every channel every time, and the chatbot can also answer outside working hours, on Saturdays and Sundays. That's worth about 73k, so roughly double the investment. Of course we had some challenges with this implementation. The first one was tooling. We implemented it internally to build knowledge, but we also looked for external enterprise solutions. What we found is that all the solutions on the market were born at the same moment we were looking to implement one. So it was very hard to trust anything: there was no track record of reliability, or they had bugs or missing features. In the end, running both tracks helped us, and we decided to go with our own. Another important thing is that LLMs change every day. In the past two years we've had a stream of LLM releases, and with every release your chatbot can break, because the LLM changes how it behaves. So you really need a strong evaluation system. You need quality and assurance tests so that the output you expect stays the same no matter how the technology changes. You also cannot get stuck on one model, because models get deprecated and change all the time. Another important concept, one that has come up a lot at this conference, is context, because context is key. We had to give the chatbot our glossary, our knowledge base and our own rules so that it doesn't make mistakes while speaking to our customers. And the last one, which I think is the hardest challenge: it's quite simple to get from zero to 40% quality, maybe even 70 or 80%, with the intelligence of the model itself. But improving beyond 90 or 95%, the effort increases exponentially. So you need to understand where the balance is, when the investment is still worth the trouble, and when handing over to a human agent is the right solution. The next chapter is engineering, the software development life cycle, and how it changed for us engineers. Again, we started in 2024 with just some AI autocompletion. Then we moved to vibe coding in 2025, when Copilot and Cursor came out. But the real big change for all of us happened when Claude Code with Opus 4.6 was released at the beginning of this year. That's the moment we started moving out of our IDEs and delegating work, coding, code reviews and debugging to AI. Soon enough the whole business was adopting it, and in Q2 of 2026 PMs also started contributing to our repositories, which is a really big improvement in productivity. We also started generating documentation to help non-technical people, who didn't have access to our code before, so they could build properly, the same way an engineer builds. The challenge we're working on right now is what we call a software factory. We're trying to find or integrate a tool that lets us delegate an implementation directly from a Linear task to an agent, without a developer in the loop. The code review part is still to be figured out, but this is something we need to invest in. What has the return on investment been on this side? We actually had fewer people in the team during the year, not because of AI but because of normal review cycles and churn. But the outcome of the AI investment has been huge. We had a 59% increase in merged PRs year over year. That's output, of course, not outcome. In terms of outcome, production releases that we communicated to the business, the ones that had an impact on our products, were up about 34%. And the lead time to production, the time from the initial commit in our repository to the release, went down 74%, from 71 hours to 19 hours. That's a huge speed-up in how we do our work. The investment here is, I think, one of the biggest. We invested about 90–100k in Claude seats and Codex seats, and the outcome, measured with that 34%, is worth about 840k. It's like having ten more developers in the team just by using AI. But of course there are challenges ahead. The most important one, I think, is human: the learning curve. We've entered a new era where the job of an engineer has changed completely. As always, there's a steep learning curve at the beginning because everything is changing. Then you gain some confidence, and then you realise there's a really steep growth path ahead before you reach mastery on the AI side. We had to face some doubts in the team. People didn't trust AI. They said: if I cannot check the code myself, how can I trust it? They were afraid of doing the work twice, delegating to AI and then redoing the job. This also creates cognitive overload, because every month there's a new release, a new tool, a new concept to learn. We're still in the middle of that, so it's something we need to care about. How we handle it is by sharing a lot: conversations about AI, how we use it, what someone learned that can be transferred to other people. That way team practices converge. One topic that caused a bit of backlash on our side is metrics. We've always used metrics, like the release time or releases to production we saw before, and also AI adoption: how much a person is actually using AI. But these are just indicators. It's important not to treat AI adoption as a measure of outcome, because that's like saying that if you produce more lines of code you're doing more work. If you use fewer lines of code for the same outcome, you probably did a better job. So it's important to have this discussion with your team. Another thing we're facing is that speed has grown a lot, and so has unreviewed code, meaning code not reviewed by a human. We're trying to put AI automations in place to review it, but about 43% of our code, or 65% on internal tools, is not reviewed by a human. Of course we review the important parts of our product, but this is becoming a trend, and we need to work out how to address it properly so we don't lose control at the end of the day. Chapter three is adoption and training across the whole company. This is quite straightforward. As I said, in 2024 everybody was mostly using ChatGPT. When Claude released Opus 4.6 in Q1, the whole company understood the potential of the tool. Everybody asked for a licence, and we gave licences to all our teams. That's when things really started to scale up and change. We had to build a lot of tools, which we'll see in a moment, and we also ended up creating a training programme for AI specialists inside the company to spread the knowledge. On cost and ROI I'll be quick, because every one of us sees the impact of AI on our daily work, but it's hard to measure. We know that something that used to take days now takes a few hours. We know how much we're investing in licence seats, but it's very hard to put an exact multiplier on the impact AI has had on every person in the company. In development it's a bit easier because you have more metrics, but then you have to look at the outcome, as always. I think this picture represents what happened from the engineering perspective: in one day we had an army of junior engineers. People who didn't know how to code the day before became engineers, and this created a ton of challenges. The first is AI sprawl, which I think Amazon has also talked about happening inside their company. Everybody was building dashboards and reports, everybody was building a new application or a new process to speed up their work. That's 200 people building stuff inside the company. Quality wasn't guaranteed either: we had different dashboards answering the same question with different results, so people didn't know what to trust anymore. Is this data correct or not? On governance and security, as head of technology, it's a big concern, because you don't know where processes are running anymore. We found ourselves with applications hosted on Vercel and all kinds of external platforms, things that were impossible the day before. On security, we have to protect the data we host as a company, with GDPR constraints and so on. But Claude Code doesn't think about that. It just does what you ask it to do. Another very important human problem was AI talking to AI. We get deeply technical questions from people that were written by AI and just copied and pasted to us. Then you either read them or ask your AI to read them and make sense of them. So you end up in a loop where your AI is speaking to their AI, and you hope that someone is actually using their brain to judge whether it's the right thing to do. Cost has not been a problem, because licences are a fixed cost, but you need to be careful if you leave an uncapped token allowance. AI doesn't stop: it will burn all the tokens you give it in five minutes if it can. The solution we came up with was not to block adoption but to use guardrails instead. Keep people working with AI, but build the tools they were trying to find somewhere else. First, we built an internal tool for sharing dashboards, where people can build their own metrics, reports and analyses and share them with others. That gives us a bit of governance, a place where data can be refreshed, and a way to check how people built things, so we can help them improve. We also introduced Coolify, an open-source platform for deploying applications like you can on Vercel. You can vibe code your own app and host it on our servers. Again, it's a bit of governance for us: you can still do it, but we have visibility on what people are building. For security, we introduced Warp, a networking layer around our company systems, so data cannot get out unless there's approval from our IT manager. These are all governance tools. But the most important investment is the AI specialist programme, a training programme we introduced because we couldn't go and speak to 200 people by ourselves. We found the people who were building in BI and building in Coolify, and we asked them: do you want to work more closely with us, so we can teach you how to use AI better and you can go back to your team and spread that knowledge? That way we all use AI in a better way. It's working very well. We found 12 very techy people who are helping us create the governance and the knowledge for the whole company, which is the greatest investment we can make at the moment. Chapter four is our company brain and data warehouse. The goal of this project came directly from the AI sprawl: how can we guarantee that quality and outcomes are the same for everyone using AI in our company? Until Claude, we were using Tableau and Metabase, simple dashboards. We introduced the BigQuery MCP to query our data, but that's not enough, because the AI hallucinates on what you ask it. So we started adding MCP tools for everything, for our data and our internal tools, so the context could grow. We also built the docs to help the AI reason better, and then we introduced a RAG system, which we'll look at in a moment. To build the proper context for AI there are four pillars, and we're at about three out of four. The first is to centralise your knowledge. Your knowledge shouldn't be spread across people's laptops. It should be in a single place that everybody can access, and that every agent can access, whether it's in Claude or the personal AI on your laptop. That gives the AI the proper context to give quality answers. The second is to optimise. You have a huge database of documentation (imagine it like Google), and you need to search it. So you put in a RAG system, a retrieval system that uses a vector store to find the information when you ask a question in natural language. This optimises tokens and also speeds up the AI: faster answers, fewer tokens, less consumption, lower costs. The next step, which we're working on right now, is restructuring our data warehouse documentation for the same purpose: giving AI the right context. Every time you ask for a metric or an aggregation of our data, the AI should immediately know where the data is, how it should be aggregated and which rules not to break. This is done with a combination of documentation and a semantic layer. The fourth step is planned but not adopted yet: the ontology. That means building a graph that connects all the metrics and data points in the documentation, so the AI can navigate faster between all the information you have. So we started from the basics and we're moving to the advanced stuff. This is just an example, so let's not look too hard at the numbers. We ran some tests: AI with nothing connected, then adding MCPs like BigQuery, then adding just our documentation, then adding RAG search on the documentation, and finally everything together. With everything together, answer quality for a data question was four out of four, in 81% less time and with 86% fewer tokens, so 68% lower cost. That also means you save on the amount of AI you use, because one of the main questions we get is "can I get more tokens?" every 15 minutes. So this is super important. Of course, building a brain comes with challenges, and the main one is again buy versus build or self-host. What to choose? Enterprise tools have been developed over the past year. One is Notion; another we looked into was Glean. The bill is quite high. It's probably worth the investment, but you need to understand whether it's worth it for you. On the other side, you can build or use an open-source tool. Then you're moving the cost to an engineer, or half an engineer, who has to maintain and evolve the platform, as with every self-hosted tool. In terms of maturity, we saw that open-source tools lack some features, have limitations or ship breaking changes in their releases. Enterprise tools are a bit more mature because they're enterprise-grade, and maybe they offer more governance. So on that front, buying is the better option. But one very important point is the freedom to move. Notion AI didn't exist, I think, four months ago. Tomorrow Claude will build something, OpenAI will build something, another company will build something. If you make this kind of investment, you're also creating lock-in with these providers, and switching from one to another because a better product comes along can become hard. That's another point to keep in mind. Decisions you take today are hard to sustain as long-term solutions, because all these tools were born in the past two years. It's a huge reminder that all the technology in this field is very, very new. Chapter five is AI automation: how we're scaling our processes with AI. This isn't new for how we work at WeRoad. We've always had what we call a TechOps team, not engineers but tech-savvy people, who were already building processes with n8n, low-code and no-code applications like Appsmith, and all the tools we had before AI. We're still on n8n today, because n8n introduced AI nodes, so you can build a lot of AI workflows inside it. So we kept that tool. What changed is that we're removing the no-code and low-code tools, because now it's faster to build platforms with vibe coding or directly with code. The step we're working on right now is AI data pipelines. We're not only building processes and automations, we're also automating the production of data. We take our data and use AI to build insights that we share with the business automatically. That way the business gets understanding from the data. They don't have to look at the data anymore, because the AI gives them the information when they need it. Here are some examples of AI automation. We've built a lot: about 301 n8n workflows, 27 of which call an LLM. We do invoice approvals, work that used to be done manually by people in finance. We classify our leads as hot, warm and cold. We generate marketing copy for our emails. We read our Facebook posts so we can reply to our customers directly on the social platforms. On the AI pipeline side, the data production I mentioned, we take all the surveys we get from our customers and our coordinators and summarise them, extracting insights. We give those both to the coordinator, to improve their work on the next trip, and to our product team, so they can improve the itinerary if customers are complaining about something, or simply improve the product itself. These are just examples of what you can achieve when you put AI to work at scale. In this case it's not about engineering. It's about how you make a company AI-native. On cost and return, again this has been a huge improvement. The cost is about 1k a year for all these automations; we're very focused on optimising the models we use. The return is the time we save on all the manual or intellectual activities we don't have to do anymore, about 28 times the cost. Of course we had challenges in this area too. One was differentiating product from processes. When we told our stakeholders that we could automate their work, most of the questions we got back were about how to improve the product. It wasn't about what their team was doing that we could really speed up with AI. So making stakeholders understand the difference between speeding up engineering work and speeding up processes has been a challenge. We still face it sometimes, but we're working on it. Replacing no-code and low-code also meant the TechOps team had to close the gap with the engineering team. They had to become a bit more like engineers in using AI, even if they don't fully understand what they're building underneath. On the other hand, this helped us with governance and cost control, as I said, because we can see the tools being built. So what's ahead? This is the roadmap we have for each challenge we're facing. For people, as we said, the AI specialist programme and training. For quality, we're building the semantic layer and the brain we just released. For automation, we're moving to managed agents for AI processes, and to the software factory for engineering. On lock-in and cost, we're introducing an AI gateway: all our AI traffic will pass through a gateway so we can route to cheaper models or open-source models. We're also going to introduce an MCP gateway, mostly to monitor how people in the company use AI and which tools they use most, so we know where to start making improvements and helping them work better. On investment, I'll be quick. We invested 190k over the past year, and the return (some of it estimated, as we said) is about 940k. So it's about a 5x return on productivity, probably even more. Now that we see the return, we're tripling what we invest. We're currently at 5x, but even a 2x ROI on this investment would be a huge gain for a company that wants to scale up. I'm closing with this "precious" AI, because one LLM can rule them all. The biggest challenge I think all of us need to look out for is that AI is at a discount. We went to Claude and asked to move to the enterprise plan, and in one day our bill went from 200k to 1 million. So the question we should be asking ourselves is: are we ready to pay the real price of the AI we get at a discount today, if tomorrow they decide the discount is off? That's why it matters whether you can switch easily. If you're locked in with a single LLM vendor, are you ready to switch if the price becomes too high for your company? Because if you depend on it, your company can't work without it anymore. That's a constraint you could face. We do see the return on investment, but we're on pay-per-use, mostly for automation and for becoming AI-native. So you need to be aware that the more AI-native you want to be, the more money you'll have to put into AI. The last challenge, and probably the most important one, is that the learning curve has changed, because every three months there's something new. We went from prompt engineering to context engineering to harness engineering. That's on the engineering side, but probably on yours too: every three months you have to relearn something from scratch, at a speed we haven't seen in years. I think this is something we should care about. There's no solution, or at least I don't know of one, but I think it's a really important topic to discuss. So hopefully what's next is a brighter place for all of us. I'll leave it to Matteo for the product side. **Matteo:** Thank you, Cassiano. ** We also wanted to bring the product management point of view to this talk, because the whole topic is how we moved from people using AI individually to using AI as a company, with AI used across the whole organisation. If you're a senior product leader asking yourself whether this is the right moment to scale to a company brain, an operating system for your company, I want to give you the perspective of a product manager who has seen his job change overnight at a specific moment this year. This isn't something that happened long ago. It really happened a couple of months ago. Something very important that Cassiano mentioned is the chatbot we released early this year. Last year, one of the main KPIs I was measured on was the efficiency of our commercial team in replying to tickets, and how many tickets they received. That KPI was immediately removed from my scope, because now we have a chatbot. We've solved the scalability of the commercial team, so I can dedicate my time, my judgment and my skills to topics that are more impactful for the business. Those aspects are very relevant for us as product managers now that AI is connected into our environment. Once again, today is a journey, so there are a lot of timelines on our slides. I'm going to go through four stages: from prompts, to the moment we gave everyone AI in our company, to the moment our job really changed in June when we developed the shared context and the company brain, and up to today, where we're delegating everything to our AI agents. Back in 2024 and 2025, all of you will remember what working as a PM looked like. We were in ChatGPT, asking it to review our SQL because a dashboard on Metabase wasn't working anymore, or to write user stories because we were starting to get lazy, or to write an email in response to a stakeholder. Back then we had three main challenges. First, zero context. Every time we started a task in ChatGPT we had to explain everything. Even when they introduced projects inside ChatGPT, you had to create the context for each one. Every time you had a new task, a new product or a new feature, you had to explain everything again, and that took a lot of time. Second, zero memory. Everything you did today is forgotten tomorrow, and you start from scratch. Maybe you can go back to yesterday's conversation, but there's no connection from one conversation to the next. And third, of course, it's not collaborative, so there's no distribution. I remember our leads meetings, where one of the main topics we wanted to solve back then was how to enable collaboration using AI. Our first attempt at a fix was developing personal brains in Obsidian: markdown files on our file system that give context to the agent so it starts understanding the whole picture without us explaining everything every time. But it wasn't scalable. Once again, Obsidian wasn't collaborative, and everyone had a different take on how to approach it. I have a technical background, a bachelor's degree in software engineering, but I felt lost. I needed a lot of help from my peers in the engineering team. It wasn't working for me at my best, everyone took a different approach, and we needed to move on from that. That changed in February, when Opus 4.6 was released. All of you will remember that moment: a lot of talks and articles started spreading around, like "prototypes are the new PRDs", start prototyping your solution, and so on. It changed the way we worked as individual contributors, because we had a powerful new tool. But it still needed time to be perfected in our operations. As Cassiano said, it went company-wide, so everyone got their Claude seat. Our stakeholders also got this new superpower mode. I remember that, from one day to the next, one of my stakeholders came to me with a working prototype. I was like: what is happening, why are you developing a prototype? And she said: because I know I can do it, and I want to. I don't want to send you an email to request a feature; I want to build the feature and show it to you. That was mind-blowing for me as a product manager. But on the other side, some things weren't working anymore. For one metric we had three different dashboards. For one solution we had five different skills. We lost count of how many skills had been developed to create slide decks inside WeRoad. ** At a certain point, thank God, we put them on GitHub and added versioning. Everyone was contributing and everyone was doing something with AI, but we had to deal with the sprawl itself. Something pivotal for us as product managers was that our CPTO made AI fluency a pillar of our competency matrix for individual contributors. So we were pushed to adopt AI in everything we do. He also said: I don't care if you take two weeks to understand how to use AI and stop everything else. The most important thing you can do is learn how to use AI and how to exploit it to do your job better. I think having your C-level, your manager, push the whole department to adopt AI was the best thing that could have happened to WeRoad and to the whole company. We also had an offsite in May where the whole product and tech team gathered in our Milan office, and our CTO gave a speech about this. He said: AI is here, we must adopt it, because there is no other solution. Our market is going in that direction, and we need to be ready and adapt to the change. At that point we started developing some solutions. Cassiano explained BI, which was the place to start putting dashboards so that collaboration could begin. Now we can also host artifacts and prototypes there so people can interact with them. And, as he said, we introduced the shared context. This is the most powerful thing that happened to my profession this year. I still remember the release email for Onyx, our RAG system, which was released to our Claude agents as an MCP. Starting in June this year, everything was connected for us as product managers: the data warehouse, Google Drive, your whole calendar, all the meeting transcripts. Everything you produce and everything you have is now understood, and Claude can analyse it. This completely changed the way we work as product managers at WeRoad. I won't come back to the brain, since Cassiano already explained it, but let me give you two examples. I was in a one-to-one with my manager, and he pointed out that we should look more closely at the data of our checkout funnel. We had seen some payment failures in the funnel, but we were sure we weren't losing customers, because we had always seen that people whose payment failed tried another payment method and then completed the transaction. In the old world, I would have had to write documentation and a request for the data team and the go-to-market tech team, asking for the events and data I needed. It would have taken me at least a week to get an answer for my manager on that specific topic. With this new setup, I just went into Claude and gave it the transcript of the meeting with my manager. No prompt or anything like that. I just said: in this meeting, Simon asked me to do this, let's do it. It understood the context perfectly. It started calling the data warehouse to fetch the data, calling Amplitude to collect all the events of our checkout funnel, and calling the Stripe database to retrieve all the payment-failure information. It understood what had happened and told me that in 2026, 229 customers had failed their payment and never completed their booking. We brought a customer all the way to the end of our funnel, they were about to purchase our product, and they left. That's a huge waste of money for us, including from a marketing-spend point of view. We found we had a big pool of leads that could be recontacted immediately to generate immediate value for the company. So on the analysis and monitoring side, we've sped up our operations a lot. But we can also start building things. As product managers, we're now developing features ourselves: we open PRs to our development team to build MVPs that actually work. Here's an example. We wanted to test an upsell for customers who spend extra days at the destination. I had the final solution crystal clear in my mind. But to get there I would have had to create the user story, run a grooming session with the team, explain why we were doing it, and create context switches for them. A lot of time wasted for everyone. Instead, once again, I jumped into Claude and spoke with my agent (I'm using SuperWhisper, so I'm not typing anymore), and started chatting: I want to develop this specific feature, let's do it. My Claude now understands the whole codebase of our products. It went straight into our repository, found where the feature should go, developed it, opened the PR and created a QA environment for me to test it. I tested it immediately, including the look and feel I wanted, because it also knows our design system; all that information is already in our repository. So in one day I had a full MVP live in a QA environment, ready to be reviewed by my team. This is the most powerful thing I can show you. And now we've reached today, where we're delegating everything we do and amplifying the way we work as product managers. First of all, anything we do once, we immediately convert into a skill. All our agents have tons of skills that let us scale our operations and free up time to work on the things that matter most. My daily operations are scaled down to zero. As I said, the agent has the whole context of all the meetings I run during the day. I have an agent that runs every evening, looks at all the meetings I had that day, analyses the transcripts, creates my to-do list and updates any user story that was discussed in a grooming or another meeting. And the most important thing is that my agent doesn't forget. Back in May I was talking with a stakeholder about an analysis I'd done on a feature we had released. The stakeholder's challenge was: we can see it's working now, but will this feature have the impact we expect in peak season? For us, peak season is August, because a lot of customers travel with us then. So I said: I'll tell my agent to schedule a re-analysis in August, and I'll come back to you. I did that. Then June, July and August passed. I went on holiday and completely forgot about the task. When I came back and turned on my laptop, I saw the agent running the analysis I had asked it to run. That's very powerful too: scheduling a job means your agent will execute it even if you forget. Here I also wanted to share the cost and return on investment of our work as product managers. What's getting harder and harder is using your judgment. Now that we can build everything and run tons of analyses in the blink of an eye, the most important thing is to use your judgment to decide what really needs to be done, and to avoid misjudging at scale. If you can run ten analyses in a day, you need the judgment to evaluate them, and to make sure there's no hallucination behind them before you share them. The return, of course, is time. In this morning's keynote, Jana talked about the time that has finally been returned to us so we can do what's most important: talking with our customers, thinking about the best customer experience we can deliver, and owning the call of using our judgment for the good of the company. One last thing. At the start of this year, on one side we had "prototypes are the new PRDs", and on the other "product managers are stopping being product managers and becoming product builders; there won't be product managers anymore", and so on. I don't think we're becoming mini developers. At the start of the year we simply had a very powerful tool in our hands. And as product managers, we also had the context, the business acumen, the business trajectory and everything else needed to develop a single feature. So we were able to do the job of all the different stakeholders in a company, just with our powerful new tool. But now that the same tool has been given to the whole company, we're seeing every role take back its position in the product life cycle. For example, our designers have integrated the design system into their agent, so they can produce an almost pixel-perfect prototype with our design system in one afternoon. They can speed up their operations and keep pace with us as product managers. The same goes for all the other departments: data and business are all enabled to move at the same speed and produce the outcomes they want. I want to conclude with the very last line of *The Lord of the Rings*: "Well, I'm back." It's what Sam says when he returns to the Shire after a whole journey that changed him completely. It resonates very well with this situation and with my profession, because after this whole journey I've come back to being a product manager with tons of time and tons of possibilities to focus on what really matters. Thank you. ** #### Q&A *[Host opens 15 minutes for questions.]* **Audience:** Thanks for the presentation, amazing. A very specific question. You mentioned that you had some hesitations about Notion because of lock-in or stickiness you don't really want. What's the alternative? How did you build an in-house knowledge management system? It's not that Notion isn't up to par. It's mostly about the investment. Once you go into Notion, you're locked in, and we've already migrated our knowledge base about three times in our company, from ClickUp to Notion and then somewhere else. That was already messy. There's an alternative we looked at, which I think is currently the best company brain, and that's Glean. The cost is very high, enterprise pricing: for 200 people they asked for about 120–130k. They don't let you try the tool; you have to buy a minimum of 70 licences. That's the kind of enterprise deal we said no to. That's why we're careful: this kind of investment, this early, can easily become lock-in, when tomorrow maybe Claude does the same thing, or another company does. Notion has only just come to the market with this. They built a good tool. They have a company brain that indexes all your Notion documents and can connect to Drive, mail, calendar and so on. So it's not a bad tool per se; that wasn't the point. What we did was use Onyx, an open-source tool, because we wanted to validate that it actually had a return on investment. It's open source, so you're not paying for licences; if you want their enterprise brain features, that's an enterprise tool and costs about the same as Notion. So we did it to test it out. Now that we know the return on investment is high, we need to decide when is the right moment to move to an actual enterprise platform, or whether to stay on Onyx and keep using it. I hope that clarifies it. **Audience:** Super interesting talk. I feel like I understand what's happening inside WeRoad, so thank you so much for that. My question is about something that happened in your world this year: your investment from Airbnb. Could you speak to how that is changing things in the company, your strategy or culture, whatever you can share? I don't know if you know more than me, but I don't think it changed much, because we're already in a scale-up phase. WeRoad started in Italy, and we've already opened in Spain, Germany, the UK and so on, so we were already taking the steps to go worldwide. Of course, to do that as a scale-up you need investment. The relationship with Airbnb started when the investment came in, and I know there are conversations at C-level with Airbnb. Also, our CEO now works for Airbnb, so we have a good network there. And now we've opened in the US. The main challenge on the table was exactly going into the US market, which is very different, huge, and very expensive in terms of marketing investment. That was the challenge: making WeRoad a global company. So I think what changed is the speed. Now they're asking us to accelerate a lot, but we've been asked to accelerate every year I've been here, so I haven't seen much difference in that sense. I don't know if you've noticed a difference. **Audience:** Very interesting presentation, thank you so much. I was wondering how you're handling stakeholder management, because you mentioned that stakeholders expect more velocity in the product and that you're working on it. Can you tell us more? **Matteo:** Yes, I think I can take this one. We are speeding up the development phase, and stakeholders are noticing. To tell a fun story: about two weeks ago, someone from the marketing team reached out to me and said: "Hey, can you send fewer release emails, please? My inbox is full of your release emails. Stop that." ** That was the first clear signal that our stakeholders are perceiving the increase in velocity. On stakeholder management itself, I think we've built a very productive relationship with them, because we've always brought them into the loop. We set up a bi-weekly meeting between the product team and our main stakeholders, all the areas involved directly in the customer experience, to show them where we are with our roadmap, what the challenges are and what the main topics are that we're discussing as a team. So we've built a very deep relationship over this last year, also thanks to the speed-up from AI itself. Two years ago, preparing slides for the executive team would have taken me ages to get perfect. Now I have a Claude skill that lets me just dictate to the agent the topic I want to bring to my exec team, and it creates the slides for me while I'm on another task. So stakeholder management has been enabled too, because AI reduces the time spent on operational tasks and on building the relationship itself. **Cassiano:** If I can add something: I think the red thread of my presentation was that humans don't change that fast. The discussion you mentioned, speed and "I want more stuff done", of course we build more, but then capacity fills up again, even with AI as a multiplier, and you still have the same discussion about what's most important. One thing that is changing is the merging of the PM role and the engineering role. We're asking engineers to become more product-savvy, because code has become a commodity, so they have more free time to look at product. And product is coming closer to engineering. That creates more ownership and more responsibility, which actually speeds up the work and puts engineers in contact with stakeholders. That's the kind of relationship Matteo is talking about, and it's improving. But again, when your head is full and your capacity is full, there's nothing you can do about it. Just prioritise. *[The other speaker adds:]* Yes, and to add one thing that comes to mind: it's not Wonderland. Of course we have stakeholders who come to us saying: you can do this with AI in one day, so do it. And you have to manage that and say: yes, we can do it in one day, but is it the right thing to do now? Is there anything more urgent and more important to do right now? So we also have new tools to communicate with our stakeholders, even if they're becoming more demanding. *[Host calls for the last question.]* **Audience:** Matteo, this is for you. You mentioned that you were pushing features to QA yourself. Something I've always found great as a product manager is the collaboration with design and engineering, who bring the feature quality up during discussions, whether design reviews or kickoffs with development. If you're doing this on your own, how are you still incorporating feedback from those different parties to raise the quality of what you're building? **Matteo:** The story I shared was about an MVP. It was something we wanted to validate very quickly, and we didn't know whether customers would adopt that specific feature. Only 60% of our customers spend an extra night at the destination, and we didn't know whether they'd want to stay in the same place where the tour starts or ends. So we had a lot of open questions, and bringing the engineering manager, the design lead of the team and so on into the loop would have required a lot of time. I decided to do it by myself, and since it was a very simple feature added at a very specific point of the experience, I had no doubts that I could do it on my own. Now that we've validated it, we're integrating it properly into both the design and the codebase, and we're evolving the feature itself. So I brought back the design point of view and the engineering point of view. For more complex features, where we already know there's market fit and we want to go live with a version one instead of an MVP, the design team is involved from the very first draft of the requirements, and so is the engineering manager. Most of the time we work as a product trio. And as I said, now that design can also create a design and a working prototype for me in about a day, we're going back to working as a functional product trio. ### 15:00 · Stop micromanaging, start scaling — Lucie McLean - **Speakers:** - [Lucie McLean](https://www.linkedin.com/in/luciemclean/) — Product Leader, Google. Scottish, Berlin-based product leader with 30+ years in media and tech; has led product and engineering teams at the BBC, Zalando and Google. - **Stage:** Miro Stage · CIC Berlin · 17 Sept 2026, 15:00 CEST - **Format:** Talk · Track: Leadership - **Host:** [Andreea Lungulescu](https://www.linkedin.com/in/alungulescu/) (The Principal Recruiter) - **Official abstract:** One of the hardest parts of leadership isn't delegating tasks - it's scaling your intuition. Drawing inspiration from architecture, design and successful product leaders, we'll explore mindset shifts and practical tools to help you 'do' less and 'lead' more. **Summary.** In an interactive working session, Lucie McLean, a group product manager at Google and executive coach, explores why delegating to humans is so hard for product leaders and what to do about it. She walks through four common mindsets that block delegation, then a set of tools organized around themes such as what to delegate, context, intent, taste and quality, autonomy, cadence and recognition. She uses parallels from fashion and architecture (Dior, Alexander McQueen, Le Corbusier, Frank Gehry) and Pixar, and draws many contributions from the audience. **Key takeaways** - Four mindsets block delegation: "it's faster if I do it myself", "if I want it done right I have to do it myself", "I enjoy doing this" (the comfort and dopamine of familiar work) and "my team is already overwhelmed". Each can be reframed: time spent training is an investment, the leader owns the system that produces quality, the goal is to create dopamine for others, and a leader should be a multiplier rather than a protector. - "Keep the steering, delegate the engine": vision, culture, crises and confidential people matters stay with the leader; repetitive work, work below your ceiling and stretch assignments can go to others. Choose who to delegate to by looking at both skill and will. - Context is the leader's main tool: a "file" of inspiration like Raf Simons at Dior, team-owned top-10 problem lists, decision principles, strategy stress-test sessions, laddering individual work up to the company vision, and regular small context updates. - Describe the end state (commander's intent) rather than the method. Sizing frameworks, templates such as PR/FAQs, and asking "would this be on the Google I/O stage?" help people check their own ambition and quality. - Make taste and quality visible through how you work, through occasional IC work and through specific feedback that calls out what good looks like. Be clear about levels of ownership (Michael Hyatt's five levels of delegation), review cadence and preferred artifacts. - Recognition matters: in one study, having work ignored hurt people almost as much as having it shredded. Le Corbusier's Philips Pavilion is a cautionary tale about failing to credit the architect who actually did the work. #### Transcript **Host (Andreea Lungulescu):** Hi. Hello. How many of you are in a food coma? Some of you. Oh, that's bad. You too? Oh, that's really bad. Hi, my name is Andreea. For those of you who will be in this room for the afternoon, I will be moderating. First it will be Lucie, and then we have Ravi. Those of you who've seen me in the other rooms, I'm sorry, because you will probably hear some stuff I've said before. On my t-shirt it says I'm diagnosed. I'm fully diagnosed. So when things go awry, that's why they go awry. I want to tell you something you would have no idea existed. When I heard I would introduce Lucie on stage, I was giddy, because in 2020 I was at Zalando. I'm not a product person. I'm in talent acquisition, and I've been in talent acquisition for 16 years. In 2020 I was at Zalando, in my fifth year, and I was taking care of the product management job family. We were redesigning the career frameworks and the entire interview process, and I was leading that from TA. Then Lucie joined, and she joined the project team. Everybody in the industry already knew how great she was. I had no clue. Then we worked a little bit together, so I learned why everybody was so giddy, and I became very giddy myself. So I'm introducing her today for the first time ever, and I'm very, very happy that I can do that. I learned some other things about her when I stalked her. Hands down, it was really helpful, to be honest. I also found things that you would probably not find online, and I will start with that because it is a completely random fact. Did you know that Lucie led the world's largest female snowboard community for how many years? Eight. Eight years. It's a big thing. Something you would never have found out, right? Lucie McLean is based in Berlin, lucky for us, and is today a group product manager at Google. Her work covers dictionary, translation and sports in Google Search. She started, however, as a journalist, spending nine years at BBC News in TV and web, and then she moved into product management in 2006. At the BBC she led the teams behind BBC News on mobile. Anyone from the UK? You need to tell me if this is true: if you grew up in the UK or raised children there, there is a fair chance you have tapped into something she built. So you need to go and investigate that. At Zalando, she went on to lead the Outfits program. Mind the Product, I think, is something you're very familiar with. Well, this is something I found there: Lucie worked on mobile products before mobile products were a thing, before the iPhone was ever launched, if you can imagine what that means. Lucie also trained as an executive coach with Henley Business School and coaches leaders inside Google and beyond. This afternoon she takes on one question many of us see in our calendars: how do we stop being the final safety net for everyday decisions? I'm really happy to welcome you on stage, Lucie. Can we please get, especially those with food comas, a really big round of applause? **Lucie:** Thank you so much. Amazing. Thank you for joining. I have never done a talk sitting down, so we're going to try it and see how it goes. Enjoy the sofas, but hopefully not too much. We've got about 70 minutes together, so I'll explain a little bit what we're going to talk about and make sure it's meeting your expectations. I know there's lots of choice. My hope is that this is quite interactive as well. The title of the session is "Stop micromanaging, start scaling." Bit of a spoiler: it's a bit of a clickbait title, because I don't think many of us are micromanaging. My experience both as a product leader and as a coach is telling me that. But I think delegation is really hard. From what I've seen this morning, we're getting really good at delegating to AI, to agents, but delegating to humans is really hard, and it's hard for lots of reasons. So the goal of this session is to explore what makes that hard and share a few tips and tricks that can hopefully make it easier. My goal is that you learn a little bit about yourself. As a coach, I'll be asking some questions and asking you to think about some stuff, and we'll see where it goes. Is that what you're expecting? Great. Okay, let's get started then. I'm old, and I love memes and I love metaphors. […] Product leadership, anyone? Yeah. I think today that wobbly board is wobblier than ever. There are more balls being thrown at us, they're changing size, they're multiplying, there are two people off to the side throwing more balls at us, the person on the other side maybe just got made redundant, and our boss is standing there watching. Yeah, right? Is that not anybody's calendar here? A quiet day for some people. This is the reality. This is really hard. So where we spend our time is super important, and I think it's an act of choice. What I want to show you today is that once you're aware of why you're not delegating, you can do something about it, and you can make more active choices about where your time is going. So we're going to talk today about why delegation is hard. I'm going to tell you some stories; some are from product, some are not. It's after lunch, we're tired, stories are good. I'm going to share some practical examples. I'm also going to give you a handout for later, because I know there's a lot of information today. I'll share a QR code a couple of times. If you lose it, just ping me on LinkedIn and I'll send it to you. So the key takeaways and further reading are all available later. If you nod off, it's fine. Please get involved. I will ask you questions, and I'd really like to encourage you to get involved and support and help each other, because we are a community and we're all in this together. Based on that, I'd like us to make a little verbal contract. I really hope people feel they can be honest. You're amongst people who get it, amongst people who are suffering the same problem. If anybody shares something in this room, can we all agree to treat it with confidentiality, with respect, with care? Anything you'd not want shared elsewhere if you'd said it, we keep a level of confidentiality. Call it Chatham House rules, Vegas rules, Berghain rules. Is that okay? Amazing, thank you. I appreciate that. It's kind of obvious, but I'm just going to reiterate why delegation matters. You cannot scale your impact if you're trying to do it all. There is significant research showing that leaders who adopt a delegation mindset are more successful, and their companies are more successful. So your company needs you to do this. Think about your teams as well. They can only grow if they're given opportunities to build skills, and as their leader, it's your job to grow your team. Can I do a quick show of hands? Who currently manages people? Hands up. Who manages managers? Got a few. Those of you who haven't put your hand up: do you have influencing relationships, where you're trying to get people you don't manage to do things? Yeah. So quite a mix here. My hope is that some of this is useful for those of you who aren't managing as well. And let's not forget about our users. We all know that diverse teams and diverse perspectives build better products for diverse audiences. So having more hands, different hands, different eyes, different life experiences on products is good for our products too. Yes? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** I mean, it depends on your style as well. Even when you manage people, there's a way to manage that goes beyond "do this thing because I told you to." We're going to talk about the different kinds of delegation, so my hope is that some of the tools we show today will be useful if you're in an influencing relationship as well. I'll go through this quite quickly. As Andreea mentioned, I've been leading product for 20 years, so I'm wearing both my 20-years-of-product hat today and my coaching hat. I'm thinking about challenges I see people come to me with as a coach and things I've been through myself, and some of the stories I'm going to tell are from my experience as a product leader. I also work with Daniele, so when you see coaching on the slide here, that's me plus a couple of other people. If you're interested in coaching, feel free to come and have a chat. So let's get involved. I'd like to do a quick gut check. How good are you at delegating? Let's imagine one is deeply unhappy, and ten is completely happy, doing great, no notes. Again, no judgment; this is completely your take, and there's no right answer. Who would score themselves at the moment at about eight or above, between eight and ten? A few semi-hands. Who would be maybe six, seven? Yeah. Fours, fives? One to three? Quite a mix. Super interesting. Thank you for sharing that, and we'll come back to it. Let's make it real. Think about a strategically important task that's on your to-do list. It's been there at least a couple of weeks, and you just don't have time to do it. Can you think of something? Maybe it's competitor analysis. Maybe it's doing that Claude Design training. Maybe it's having that hard conversation with a team member. Everybody got something? Something you think, "I could have done that today instead of coming to this talk." Hold on to that idea, and think: if you did that thing, if you found that hour, that afternoon, that week, what difference would it make? Would anybody be interested in sharing what the task was and what difference it would make? Yes. Can we grab a microphone? Is that okay? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Awesome. Thank you for sharing. Anybody else got something they haven't got to that they really think is going to be worth it? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Transform. That's a big outcome, isn't it? Super interesting. Thanks for sharing those examples. It's really clear what we're not getting to. And for everyone here, I'm pretty sure there's something in your day that somebody else could be doing, which would not drive the same outcome as spending your time on that thing. As I mentioned, this talk is called micromanagement, but I'm not really seeing people micromanaging. I feel the current generation of leaders have all been micromanaged. We know how horrible it is. But I'm also seeing some confusion between micromanagement and the opposite end of the spectrum, which is nothing. What is the right amount of management? What is the right amount of directiveness? What is our input? Where are we needed? Where should we give people space? I'm going to share a couple of frameworks later that can hopefully help you explore this and find your version of it. This is one angle I'm seeing on it, and I mentioned I was going to share some stories. There are a couple of industries that offer quite interesting parallels, and that's the creative industries. Think about a fashion house or an architecture practice. They have these really creative figureheads, visionary figures, but they don't design the real detail. They are not sewing the dresses. They're not building the sets at the fashion shows. Their role is vision, goal setting and culture setting, and they rely on a team of people. A brand like Dior has this throughline. It's been going for 80 years, but there's a very clear throughline of what Dior is. Look at architecture by an architect like Frank Gehry. Those two buildings are on completely different continents, but someone who understands architecture could look at them and say, "That's a Gehry." There's a range of people who touch these products, yet Frank Gehry's signature is on them. I'm going to share a few more stories like this, because I think there are really interesting parallels for thinking about what doing the work means, and what our job is as leaders when we're more in this visionary role. How do we get people inspired and bring out the best in them without doing the work for them? We're going to talk about this in two halves. One is about delegation mindsets and the common patterns that can make delegation hard. Then we're going to look at some tools. Based on what we just talked about, I'd like you to have something in mind that you think, "Someone else really could be doing that," or "I wish I could get this off my plate and give it to somebody else." Maybe you're not sure and you want to explore whether you could delegate it, and we'll see if these mindsets are relevant. So what's something on your plate that you think you should delegate? Or maybe something you should have delegated in the past that would be interesting to explore. The first internal narrative that is fairly common is: "It would take longer to explain it to somebody than to do it. It's just faster if I do it myself." And yes, if it's an absolute one-off ten-minute job that will never be repeated, you're probably right. But there aren't that many ten-minute one-offs. There are patterns, there are variants, things that come back. If we're doing that ten-minute job 20 times a week for 20 years of our career, that's a time fallacy. But it's really easy to think that way. If we turn it around, spending time helping people do these tasks feels like a burden, but ultimately it enables them to grow their skills. An example from the design world: there was a British designer called Lee McQueen, who had this incredibly iconic fashion house called Alexander McQueen. Sadly he died in 2010, and the entire brand was built on him and his vision. It was him. There was a bit of a reckoning: can this house go on without him? What do we do? But luckily he'd invested in a person. He'd invested in his personal assistant, Sarah. She'd been with him for 13 years, and she knew what the essence of his work was, what the essence of the brand was, how to continue its legacy, what the culture was. Sarah was so successful that the house continued to exist and grow, and she got the job of designing the Princess of Wales's wedding dress. This was Sarah Burton. She's now creative director at Givenchy, and she started as Lee McQueen's assistant. So there's a real example of what happens when you invest in somebody, when you take the time, when you make the time. It's a sad example as well, but it shows the difference it makes when you invest in people. I think this is also why we're seeing a real shift in the chief of staff role, from something that was maybe more administrative to a real partnership. I just found these examples on LinkedIn, and if you look at the minimum qualifications for this one, it's 11 years of experience in management consulting and product management. That's a really senior role. Leaders at the top are really relying on the people around them to be thought partners. So how are you creating thought partners in your team? Where are you creating space for people to bring their ideas? You have to invest time to do it. So what if the time you're spending isn't a cost, but an investment with a dividend that will pay off? It's a mind shift, a reframe from a short-term mindset to an investment mindset. Think about all those ten minutes and how they add up. Training isn't slowing you down; if you don't train, you're capping your output forever. And think about how you can break it down. Yes, it might take a few hours to train somebody in a whole process, but maybe you train people on one step this week and one step next week, with different ways to approach it. We're going to talk about tools later, because some of the solutions I want to share are relevant for several of the mindsets. But how you train people, how you approach it, making it easy on yourself and not letting it pile up is going to be key here. Next mindset: "If I want it done right, I'm going to have to do it myself. Nobody else can do this. I am the expert." Anybody else feel like that's something they've thought before? Yeah. And you're probably right. You probably are the expert. This doesn't always come from a very positive place. It can come from perfectionism and control. It can come from external drivers, like company cultures that are quite punitive and demand perfection, which can be really hard. It can come from within: looking for certainty amid ambiguity, impostor syndrome, fear of failure. There are lots of reasons why this happens, and it's a completely natural reaction. But in reality there is more than one way to load a dishwasher. It might not be the way you would do it, and in some cases it might be better, and that's okay. I want to show an example from architecture. This is a building by the architect Le Corbusier, the French-Swiss architect. There are lots of his buildings around France, and there's one near the Olympic Stadium in Berlin. Back in the 50s he got a really interesting brief: he was asked to design the pavilion for Philips, the electronics company, for the World's Fair in Brussels. He was kind of busy; he was designing a city in India. So he outsourced the brief to an architect in his firm, a young French-Greek architect who was also an avant-garde composer. He was able to bring these two worlds together, his architecture experience and his music experience, and he brought in more of his composer friends to design the pavilion. They built this incredible cathedral of sound: an exhibition space everyone could go into, wired with speakers to create incredible sound that showed off Philips' capabilities. A really amazing space. They broke new ground technically; they used maths to decide where the speakers went. A super successful piece of work. Could Le Corbusier have done that? Maybe, maybe not. But because he didn't, the firm got to run two large projects, and the other architect got a chance to grow his work. Ultimately, I think it was the fact that somebody else brought a new perspective that made this so successful. So the moral is to think about the unique skills and perspectives your team brings that will add a new dimension to the work if they do it rather than you. So even if you are the best person to do it, what if doing it right is doing it wrong for the team? Think about becoming a delegator. If I do it myself, my team will never learn. Yes, they will make mistakes, but you have to make mistakes to grow, to learn. I really like the last one here. "I'm responsible for high-quality output" becomes "As a leader, I'm responsible for building the system that produces high-quality output." I think systems are a really key concept here, and we'll talk a bit about that. Systems can be really formal or really informal, and I'm hoping later you'll share some of your systems. We're going to come back to this next one: "I really like doing this. This is fun. I feel good doing this." Especially things we used to do. You feel confident. You know what it's like. When you're in a new or more senior role, or things feel really uncertain, it's like, "Oh, it's okay, I know I'm good at this." That's how it can feel. When things feel really uncertain, we regress a little into what we're comfortable with. This is why we watch Friends. We just want something we know, where we feel comfortable. But that dopamine hit it gives us: as leaders, what we really want to do is design the systems that enable other people to get that dopamine. Can we get our dopamine from scaling, from building opportunities for other people? So we're moving from thinking about execution to a strategic mindset. The value for you is seeing people grow and develop their skills and building long-term organizational impact. If you're staying in your comfort zone, you might be neglecting the strategic responsibilities you have. We'll come back to some tools for this in a minute. This is the last one: "My team is overwhelmed. I can't possibly give them any more." We know what it feels like to be overwhelmed. When we've got a team we care about, who are working really hard and already showing signs of stress, we don't want to give them more. It's a really natural response. But we can't be superheroes. We're not superheroes. We have the same number of hours in the day. We're not here to save the world; it's not our job. There are other ways we can make this easier. When we withhold high-value work from our team, we keep them in the lower-value work as well. We're not creating the space for people to grow. So here it's about being less of a protector and more of a multiplier. How do you empower your team? Clear away the noise rather than fully shielding them from work. It's absolutely about prioritization, and about creating opportunities for visibility and growth. Also, if you just absorb it yourself, you're not letting the broader company see that there might not be enough people, or that there are too many things to do. So it doesn't serve anybody to keep absorbing all the work. We're going to come back to this again, and I'm going to share a couple of really practical examples I've seen people use to have that hard conversation. So those are the four common mindsets I've seen. Are any of these resonating? Yeah. Anything we haven't covered? Maybe a pattern you find yourself thinking about as a reason why you don't delegate, or don't delegate more? There's a hand at the back. We're going to grab a microphone. Is that okay? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Yeah, absolutely. A lot of the tools we're going to talk about later are about exactly that. You've absolutely got an initial onboarding piece to do, but once someone's in the company, how do you reduce that workload so the time it takes to get people up to speed is less? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Another really interesting point. I'm going to show a framework later that may be a good way to explore that, because there's something in the when and the how: when to catch that the steering isn't happening. One more reason we haven't talked about today: what if you've got a team that hasn't got the skill level, or you're concerned they haven't? I think this is another reason why delegation skills are so important. If you've got somebody in your team you're not sure about, or someone you feel is having a really difficult time, it's really hard to know what the actual problem is until you can eliminate certain conditions. Is it your delegation that's the problem? Is there a capability issue? Is there a communication issue? Until you're super confident that you've explained things clearly, that your guidance has been clear and the outcomes are clear, it's really hard to figure out what's actually going on with somebody. So that's another reason why this is so important. Sometimes it's not necessarily about skill. Everyone's different; everyone's an individual. Some people have strengths, and some people have things they really don't enjoy doing. If you do enjoy doing them, it veers a little into that productivity dopamine, but it's counterproductive to force people to do things they don't enjoy, maybe aren't good at, and maybe don't want to grow in. We're going to come back to that as the concept of skill and will: both need to be true to delegate. I think you had a question or a comment as well. Andreea, there's someone in the middle. *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** The slide before, about shielding your team, had the word "shield" in it. I think that's where you're shielding. You're not shielding them from the work; you're shielding them from the ambiguity, where it's appropriate. Again, there might be kinds of work, and kinds of people, for whom sorting that out is the actual job. I want to talk a little about this in the context of what to delegate and what not to, but I think this is where your shielding happens. Strategic shielding is important. That's where you need to bring clarity. Now we're going to talk about tools and systems. I've organized these a bit differently from the four mindsets because, as I said, they can be different tools for different parts of that, and they can be used in multiple places. So I've organized them around eight themes. As I said, it's after lunch, these seats are comfy, we've had a big, very delicious lunch. I've got a couple of handouts on this, and one sheet is these eight areas with a checklist of questions. Not with the assumption that you'll check every question every time you delegate, but they might spark some interest. You might feed them into your AI second brain ahead of some meetings. Feel free to play with them. No need to remember all of this. I can also give you the slides and the sources. So absorb what you can, given where we are in the day. There's a QR code if you want to grab it now, and I'll show it again at the end. You can also add me on LinkedIn. Some of these you might do naturally; for some you might think, "Oh, actually, I could probably spend a bit more time on that." I've tried to organize them in a way that hopefully comprehensively covers what I think are the most important themes when we're delegating. So, what to delegate? Lovely AI-generated image here. I think this is a useful way to remember it: your job is to keep the steering but delegate the engine. Steering is: where are we going? Where's the destination? That's me. What's the vision? What culture am I setting? If there's a crisis, if we're heading towards an iceberg, you take the steering. Confidential matters, people things: these are things you don't delegate. Things that are core to your role and where your leads expect to see you. These are your chance of visibility, and the expectation is that you're setting the culture. Then the engine. It's not necessarily the most complimentary concept, but hopefully it gives you a sense that there are areas where others can shine. Repetitive tasks, hopefully mainly, and using AI for a lot of these. But also work that is clearly below your ceiling, especially where others are more skilled, even if they're only 70% there. Creating spaces for others to learn is really important. Stretch assignments are key; creating stretch assignments is super important. That's how I think about it. What do you need to steer? It's possibly less than you think. Again, it's this aspect of vision and culture. You might create a partnership where you're still steering and you've got a great first mate, but ultimately you're responsible, and it's clear who is responsible and accountable. Who to delegate to? There's a really nice framework in this book. I hate the title of this book, *High Output Management*; I've crossed that out and put "High Impact Management." But there's this concept to think about when you're deciding who to delegate to. Skill: what evidence have I got that they're probably going to be good at this? Have they done it before? Have they done something similar? Have they got the training and the knowledge, or can I get that for them? And will: how well would they cope if it goes wrong? What proactiveness have I seen them demonstrate? How do they cope with pressure? Are they eager? Do they want it? Have they got confidence? These are things to think about when deciding who to give it to. Here's my first example. I stole this from a colleague, and this is what she used to talk to her boss about where she was spending her time. Remember that last mindset, the feeling of "I can't give my team more"? It's important to have the conversation about what that person is not going to do in order to do this. I found this framework really simple. It's quite literal, but writing down the things I'm going to spend time on and the things I'm not was really interesting, because the things I'm not working on make people really uncomfortable. It's a real forcing function, as is the expectation that I've only got 10% on Project C. I think it's a really nice, simple way to do it, and my colleague found it super effective. It created lots of debate, but ultimately she got clarity on what she was spending time on. She told me once that the 10% on the P4 work, the things that just need doing, was some weeks 70%. She was able to surface that she wasn't getting time for the important things, and through that they were able to figure out what that 70% was. Maybe she could delegate it, maybe they should stop doing it. It really surfaced where the tensions were, which helped reduce her workload and the operational overhead. Context and intent. Let's talk about context. Those of you who attended Ravi's Leaders Lab session the other day will have heard him talk about how, as product managers, we're setting context all the time. Context is really what we do, whether we're giving it to other people, to AI or to stakeholders. So whether you're concerned it's going to take too long to delegate something, or concerned someone won't do it right, you set them up for success by providing the right kind of context at the right time. A lot of the time we're creating artifacts: strategies, visions, feedback. But how we do it in a way that has lasting value is really important. I imagine we'll have a January "ta-da" of the 2027 strategy. How often are we going back to it? How are we helping people really internalize it, so they understand where we're going and everything they do is seen through the lens of our direction? So again, let's talk about fashion. Let's talk about Dior, and a couple of big fashion examples. At the House of Dior, a Belgian designer called Raf Simons took over the house, and there's a really cool full documentary you can find on YouTube showing what happened. He had eight weeks to design this incredible collection. What's really interesting about Raf is that he doesn't draw. He doesn't do those beautiful line drawings of fabulous dresses. He doesn't draw anything; he has designers who do that. What he does is create a file. He calls it a file. He sends it out, and it's his inspirations, pulled from the archives, from now, from things that have nothing to do with fashion. He creates this context of vision and hands it over to designers, who add their own layers on top. I think he said they come up with over a hundred designs and whittle it down to about 20, and that's what they build the collection from. So what's your version of the file? What are you handing over to people in the team that gives a clear sense of direction, bringing the best of your experience and your vision, but creating space for them to draw and design and present, and then co-create together what you actually choose? A few examples now of where I see product leaders doing this really well. I don't know if anybody knows Abia. She used to work at Netflix and now works at Google. One thing she does that I absolutely love is that she has the team write down the top ten problems the team is solving. They write it down; she does not. They own it. They update it, they debate it, they discuss it. It's a living briefing document that the team creates. She showed me the YouTube one once when we were working together. I think it's a great example of holding the team accountable for setting the vision and really making them feel like they own it. Another thing I find really useful is giving people artifacts that help them make decisions. It's very common for a strategy to cover all the things. What are we not doing? Am I doing A or B? How do I make the decision? The more people you're delegating to can feel confident moving things on themselves, without coming to ask you whether you'd do A or B, the better. I found this example of Medium's early design principles. "Direction over choice": you want people to focus on writing, not choosing. You can imagine that when there were debates about features and design choices, that could have been quite helpful. So when they bring things back for review and sign-off, they've maybe narrowed down the options. "Appropriate over consistent": we've all been in those rooms where some people say, "No, it needs to be consistent," and others say, "No, I wouldn't do it that way." Principles like these can really help people make decisions. So if you think about the artifacts you're sharing, how are you helping people narrow down the choices? That way, the things you're discussing with your team have already been through iterations of decisions before they come to you. I also like the work Roman Pichler does on how to run a quarterly strategy review. It's not rocket science, but he's got a really interesting article, which you can find via the link in the sources, about just having a session with your team, even if you didn't write the strategy, and really pulling it apart and exploring it. Stress-testing it and living inside it is really important. The more often you look at it, touch it and play with it, the easier it is to remember it and attach it to your work. One more example, something I've actually done everywhere I've worked, is a very interactive session taking the themes of the strategy and connecting them to team members. A very long time ago, the BBC director general was called Tony, and I had a tester in my team called Tony. At our offsite we ran an exercise for Tony the tester to ladder his work, testing children's games, right back up to the BBC's vision. I've run versions of this everywhere I've worked since. It's a really good exercise if your employee engagement scores show that your team doesn't feel very connected to the strategy. Helping them connect to it helps them remember it and put it at the heart of their work, and that influences the work you get back when you're delegating, especially more complex, ambiguous pieces of work, which ultimately are the interesting things people want to do. The last thing I want to mention is that sharing context is quite time-consuming, especially if you've got a task that requires a lot of explaining. You can reduce that if you're doing it constantly. Oversharing is never a bad thing in your team. There'll be some people who don't want it, but chances are lots of people in your team will feel more engaged, will have new ideas and will be more ready to take tasks on if you share with them regularly. So maybe at your weekly team meeting, if you have one, you have three bullet points: "Here are three things I know now that I didn't know last week that might be useful to you." You may be able to generate your first draft from the notes of the meetings you've been in. Let people in. Let people understand what's happening around you, and they will surprise you. So I'd like to open the floor and see if anybody has thoughts they'd like to share. How are you keeping context alive? How do you minimize the briefing work when you want to delegate or create an opportunity for someone? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Yep. Can you tell us more? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** What kind of things are you communicating in your one-pager? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** And how do you keep on top of that in your workload? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Great, thank you. Thanks for sharing. Anybody else got anything they'd like to share? Yes, at the back. We'll come back to you in the room. *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Yeah, thank you. *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Thank you. That was my hypothesis as well, so it's great to hear it's working for you. Any more examples? One more. Getting your steps in today, Andreea. *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Super interesting, thank you. I've also seen that: encouraging the full range of team members to come to testing, to see people hitting the button, getting confused, getting lost. Helping people really live in it and see that problem. *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Definitely, thank you. That reminds me of another good example: just be sharing all the time in your Slack channels. Keep showing stuff, keep telling stuff. Not everyone's going to see it all, but get in the habit of sharing. Thank you. I think one of the keys to really successful delegation, especially for the more complex, ambiguous, hard things, is giving people space to use their skills and bring their creativity. In the briefing, you focus very much on the why and the end: what good looks like, what will be different if we've achieved this. I want to share a couple of examples from my world, but before that, one from the military, which is very much not my world. There's a framework called commander's intent, and it gets to this point of not telling somebody what to do, but telling them what the goal is. I really like this statement about the end state, what success looks like: as a commander or a leader, your job is to describe what the world looks like after the mission is successful. I think that's really important. It gives people the space to bring their own creativity and ideas, while being very clear on where we're going. I've got a couple of examples of where I've seen this used. At Google, a few times we've been asked to come up with fresh ideas, and one problem we were having was figuring out what is big enough. We've got these targets to hit, but how much of the target does each initiative have to deliver? Are we going to do 50 things, or five? So our leadership came up with a framework to help us figure out if something was big enough, and if it wasn't that size, don't bring it for discussion and review. We had two axes, the number of visits happening and the number of users. Either you could impact a large number of users and have a small impact, or impact a smaller number of users and have a big impact. The version we had had numbers on it, based on metrics we knew, and it meant people were able to size-check themselves. We didn't need to go back and ask, "Is this big enough?" You could do that work yourself and really see how you could push and stretch yourself. Another way I've done it: let's assume we're keeping users at the heart, but I've found it helps, in terms of communicating with different people, to help people in my team picture where they're going to be and what their role is. We give people work partly because they get something from it: it's going to help them get a promotion, grow their visibility, build new skills. Picturing the stage for your product is quite interesting, especially at a large company. At Google I've definitely said, "Would you see this at Google I/O? Think about the things that are on that stage. Go back and watch last year's. Can you imagine the thing you're proposing or designing being on that stage?" When are we hitting that quality bar? At Zalando we might have talked about Partner Day: would this be at Partner Day? At the BBC we'd talk about: am I going to stand on a stage and do a press conference about it? So thinking about what motivates your team members can be another way to bring the challenge and the opportunity to life and really grow the ambition. I think templates can be really helpful as guardrails as well. At Zalando we used the PR/FAQ process. I left three years ago, so I'm not quite sure what version is being used now, but these templates can really help people help themselves in shaking out the work they're doing and making sure it meets a quality bar. I definitely found this when I was trying to open up the opportunity to propose ideas to all members of our team: product, engineering, design, analytics. People would say, "I've got an idea," and I'd say, "Great, fill in the discovery template." The challenge of explaining what problem we're going to solve, what the value of solving it is, and how it aligns with the strategy tripped people up quite a bit. They realized that having an idea is the easy part; figuring out if it's the right one is the next piece. I found that templatizing the work through all these stages really helped raise the bar and helped people figure out what was going to be good enough. Let's talk about taste and quality as well, because I think that's what can be at the heart of "is it going to be good enough?" and "only I can achieve this," which we talked about earlier. As leaders we've got very, very high quality bars, and being confident that the output of our teams is going to meet them is definitely a concern for lots of people. Let's go back to the fashion industry. I talked about the thread that many design houses have. Interestingly, it can be quite vague or quite specific. For Dolce & Gabbana, for example, it's about Sicily, it's about widows, it's about lace. You can tell what Dolce & Gabbana products look like because they have this brand DNA. Marimekko, the Finnish house, actually has very specific designs: there are three silhouettes for dresses, A-line, straight and fit-and-flare. So there are some very strict guardrails, but a ton of creativity available in the fabrics and the colors. That combination of constraints and space is really interesting. Another example I really liked, about how you live in this taste and create a shared sense of it, came from a video I saw: a day in the life of the designer Thom Browne. I saw the screen and thought, okay, there's a gray, multi-textural palette. Then I saw a picture of them working in their office, and they are living in this world. It's very clear what the taste, the style and the tone of this brand are, and how they work. As we go back into our offices and spend more time together, how are we creating a sense of what our shared throughline is? What is our version of Sicily? What is our version of this gray palette? How are we building this common sense? I don't know if taste is the right word, because taste is a little subjective, but think of taste as an intentional approach to what our brand is, what it stands for, who our users are, all that good stuff. How are we doing that in our organizations? You can be very, very specific about it with design principles. I think Apple's are a really good example. If anybody's ever designed an app for Apple and tried to get it through the App Store, they will know how hard this is. But it's very specific, very intentional; you know what good looks like. That's a very prescriptive version. You might also use AI as your own second brain, for course correction and for inspiration. I think Matteo from WeRoad showed a great example earlier of how WeRoad is building collective shared knowledge that's helping shape the quality of the PMs' work and their output. As leaders, a really important thing we can do is be very intentional and visible in how we lead, so people understand what we like. Yes, it can be in documents and templates, but how we move through the world is really important. This is also from the Raf Simons video. You'll see him sat in the background, and they're working live on an archival dress, a beautiful dress from the 50s. They're pinning it and adding things and changing it live, and you can see everyone is paying real attention to what he's saying. Obviously they've got their backs to him at the moment, but you can see them thinking, "Okay, I'm learning about this guy. I'm learning how he thinks, what he likes, what he appreciates." I had a similar experience at Zalando, where the SVP of product wrote a PR/FAQ in about 24 hours, and it was phenomenal. That's how I learned the benchmark. I saw his work, I saw his output, and I thought, "Blimey, this is a high bar." But I learned so much from that. It really helped me understand what the quality bar was. So think about how you move through your world and around your team. How are people picking up on what matters to you and what your quality bar is? One way to do it is by doing some IC work. I really enjoyed the Lenny's Podcast episode with Tom Verrilli, where he talks about the value of doing some IC work as a leader, because you're keeping your skills up to date. And, like the example I just shared, it also helps others see what your standards are, what good enough is. That might be a bit of motivation when you do have to do a hands-on piece of work: how can you use it to help others see your quality bar? A really simple example is to be very specific and intentional in how you give feedback. This is a fake example. You might leave a comment like, "Yeah, it's great, move on." Or you leave a comment that calls out what you like about it, what the person did that you really appreciated. The person who wrote this work will see it, and ten more people will read it and think, "Okay, now I know what the boss likes." So think about where and how you can call out what you like, what good looks like, how you praise and what you praise, and how that creates a shared sense. Again, it should hopefully lighten the load when you're delegating, because it helps people absorb your standards day to day. I'm going to keep going; we can come back to these at the end. Let's talk quickly about autonomy. An example I really like here is from Pixar. There's a really fantastic book about how Pixar works, and the reality is that all Pixar films suck at first. The first draft is always terrible, and there's a process to move them from suck to non-suck. Part of it is the Braintrust process, where the director gets a group of peers together, they watch the film and give feedback. But what I love is that the director can completely ignore all of it. It's super clear who is responsible for the quality bar. My assumption is they take quite a lot of the feedback in. The feedback is on the work, not about the people; it's not personal. But it's very clear what your job is. When you're delegating, this is really important to think about: what is the person responsible for? What decisions have been made? How do they need to work with you? So let's look at a couple of examples of how to do this. The executive coach Michael Hyatt came up with a framework, the five levels of delegation. At level one we've got "do as I say": just do the thing, don't have opinions. There's not much room for creativity or putting your own stamp on it. At the other end you've got "act independently": you have full ownership of the outcome, all the decisions and the execution. My assumption is that most of the work we do day to day is at levels two, three and four. The further down you go, the more complex and ambiguous it gets and the more risk there is, but there's also more reward for the people doing the work. So I think it's really important to help people understand what they are responsible for and what they're not. That creates real clarity about who the final owner of the decision is and where people have to focus their time. This is one of the things I've got as a document for you, and I'm going to add a few more slides to it, so hopefully it lands as a starting point. Another thing to think about is how you give feedback. I think we've all been in those meetings where you're reviewing a document or a proposal and people just say stuff because they want to be heard. People throw in landmines. I worked with a VP at Google who had a very strongly worded way of approaching this, probably the right way to think about it: if you're going to contribute feedback, your feedback should endorse or improve. Create a culture where feedback is useful and valuable, where it helps people grow, where it's welcome, and where you encourage people to show work early and iterate on it. Cadence and escalation. We'll come back to this framework. If you're going to delegate work, be really, really clear about how to engage with it. Again, I've got a copy of this for you to take away. This is an example of where and how you have to partner closely with the people you're working with. Make it clear when you want to be bothered and when you don't. How often do you want to discuss this stuff? Figure out a rhythm that helps them get what they need, that helps you feel as in control as you need, but also gives people time and space. What kind of artifacts do you want? Do you want to see first drafts, or final versions? How far in advance do you want things? These are all really important conversations to have when you're delegating work, so that everyone knows what to expect and when. This is super important. Who you are as a person also matters. Do you want drafts a day in advance, or a week? Do you want documents? Really high-fidelity documents? How do you work? We've all had bosses who work in completely different ways. I've had managers who want things 24 hours in advance, and managers who want things three days in advance. I've had managers who want documents and managers who want presentations. Managers who read things in advance, and managers who will only read things in the room with me. How do people know what you want? Give them the best chance of being successful. Minimize the risk and ambiguity where you can, so more of their time and energy is spent on solving the problem, not on figuring out how to communicate with you about solving the problem. The last topic is growth and recognition. As you give people tasks, give them feedback on the themes that actually make sense. If you give someone a simple task, give feedback on that. If you give them a really complex strategic decision to make, focus on their judgment, the quality of the decision and how effectively they communicated it to stakeholders, so the feedback helps figure out where they have grown. Any German speakers in the room? Anybody like to translate this for me? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** I saved it till last, but I think recognition is super important. Celebrating people's work, whether through promotion or just verbalizing your appreciation, is really important. This saying does not work anymore. It absolutely does not work. It's super important to celebrate people, and I think it's a really key thing. There's a really interesting study; I've got a link to it in my slides. People were given a task to do: they were given a piece of paper and were looping pairs of letters. They handed it in at the end, and it was a scientific experiment. In one room, people handed it in and were told, "Yep, great, thanks very much." In the next group, they handed it in and it was put in a shredder. In the next group, they handed it in and were ignored; nothing was said. The negative impact on the people who were ignored was almost as much as having their work shredded. So remember to do this. It's super important. Remember to celebrate people. Remember the story from earlier about Le Corbusier? It went a bit wrong. When the piece of work was revealed, there was a press release, and it talked about this incredible piece of work from the Le Corbusier organization. No mention of the architect who actually did the work. He was not very happy. There was a bit of a fight. He got his name on it eventually, but he left the company. It's an interesting reminder to say thank you and to recognize people. So that was a bit of a smorgasbord of things that might help with the challenges we talked about, whether you're finding it difficult to delegate because of time constraints, concerns about quality or concerns about overload. Is there anything you picked up today that might help you increase your score? I know we had some people at eight to ten. Anybody got something they might want to try? What's an action you could take away? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Yeah. And I would say as well: we're product people, we can run experiments. Let's treat how we lead as an experiment. Try something, see what happens, and evolve it. Thank you. I would love to hear from anybody else. *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Super interesting, thank you. Anybody else? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** Thank you. And being aware that the imprint is being formed with every breath you take, both in what you say and in how you move through the world. Thank you. Anything else resonate? Or not resonate? Have we missed anything? We've got a couple more minutes. **Host (Andreea Lungulescu):** Can I add one thing? **Lucie:** Absolutely. Of course you can. **Host (Andreea Lungulescu):** I did say it very openly: I'm actually really proud of the ADHD thing. Even though people think it's hype, it's not hype, and it's not fun all the time. It's very fun sometimes. But I'm a solopreneur. I no longer identify as corporate; I identify as solo, with products that I built and run. And the dopamine part is the one that got me the most, because it's definitely there. I would not get a freelancer to do stuff, because I'm having so much fun doing it, and I'd never thought about it that way. So for me, from what you shared at the beginning, it would be this dopamine issue. **Lucie:** Yeah. It's a very hard thing for me to let go of as well. Thanks for sharing, I appreciate that. Yes? *[Audience contribution not published: this session ran under Chatham House rules agreed with the room.]* **Lucie:** That's super interesting, thank you. There are two things in there I found really interesting. That feeling when you turn your out-of-office on and you've done a good handover: imagine you could feel a version of that every day, because you've got one less thing on your list. And when you come back from holiday, you think, "Oh, what's happened? It's all been a disaster." It rarely is. Everything was fine. It can be fine if you delegate more as well. Then you've got more time for what you want to spend time on, whether that's being on vacation or writing that strategy you haven't got round to yet. *Note: no audience Q&A in the video or the transcript, at the speaker's request.* ### 16:40 · Building AI Products Competitors Can't Match — Ravi Mehta Video: https://youtu.be/f35orrLSScU - **Speakers:** - [Ravi Mehta](https://www.linkedin.com/in/ravimehta/) — Entrepreneur, Instructor & Advisor, Reforge. Former product leader at Tripadvisor, Facebook and Tinder (CPO); his frameworks on product strategy and leadership are widely used. - **Stage:** Miro Stage · CIC Berlin · 17 Sept 2026, 16:40 CEST - **Format:** Talk · Track: Build & Craft - **Host:** [Andreea Lungulescu](https://www.linkedin.com/in/alungulescu/) (The Principal Recruiter) - **Official abstract:** Everyone has access to the same AI models. The real question is how you build products competitors simply can't copy. Ravi shows how combining AI with what's uniquely yours creates defensible features no one else can replicate. **Summary.** Introduced by host Andreea, Ravi Mehta, product strategy advisor, co-creator of Reforge's AI strategy program and former CPO of Tinder, argues that in the most intense competitive environment in tech history, the advantage does not come from the model everyone can buy but from what is uniquely yours. He presents a Lego-block framework (AI capabilities, proprietary data, product functionality, connected by context, data flywheels, orchestration and tools), illustrates it with Perplexity, Claude Code, OpenClaw, Miro, Descript, Midjourney and Granola, and runs the session interactively, collecting audience examples along the way. **Key takeaways** - The AI wave has not produced the blue oceans earlier platform shifts did: big tech, OpenAI and Anthropic, and VC-funded competitors make everything feel like a red ocean, and product-market fit is not just decaying but collapsing (Chegg, Stack Overflow). - Two failure modes: reinventing the wheel by training your own model, and treating AI as a black box bolted onto the product (e.g. adding chat). - Build from three blocks: AI capabilities everyone has, data only you have (real-time, user-specific, domain-specific, human-judgment and reinforcement data), and functionality only your product has (workflows, algorithms, business rules, integrations). Your data gives the AI context and its output feeds a proprietary data flywheel; your product orchestrates the AI (deterministic software is the right answer 95% of the time) and gives it superpowers as tools. - The breakthrough is often the harness, not the model: Perplexity added web search, Claude Code ran as a command-line tool, OpenClaw's code was all harness. - Start problem-first: good AI features sit at the intersection of unmet customer problems and what AI can do. Descript ships AI everywhere without labelling it AI, because customers want problems solved, not AI. - Granola won a red-ocean market with off-the-shelf models (Deepgram, Anthropic, OpenAI) by choosing amplification over autonomy, combining your notes with the transcript, and orchestrating through a native Mac app. On "slop": it is the accumulation of unexamined defaults, avoided through sharp, well-sized context. #### Transcript **Host (Andreea Lungulescu):** What were you doing in high school? Were you dabbling in anything that would have potentially revolutionized the world? No? Me neither. However, Ravi, our final speaker, founded a successful game company while he was in high school, because why not? He then went on to become one of the first members of the Xbox team at Microsoft. Since then, Ravi Mehta has led product at brands that reach hundreds of millions of people. At TripAdvisor he was VP of Consumer Product, where he grew the product team from 7 to 70. At Facebook he was product director, studying how Gen Z spends their time online. Can we do a talk only on that, please? At Tinder he was Chief Product Officer, leading product strategy for what he calls the world's number one monetizing app. I had no idea, and I have probably put money into it at some point. Many of you use his thinking without even knowing it. You might have heard his talks on Lenny's podcast or read his articles in Lenny's newsletter. He co-created Reforge's AI strategy program with Brian, who is with us in the photo, not here today. His newsletter, Ravi on Product, now has over 32,000 subscribers. On Claire Vo's *How I AI* he showed why he builds AI prototypes by starting with the data model. Today he advises founders and invests as an angel, and this week he already led the Leaders Lab here at Productlab. Who was at his session? Come on, you can be proud of it. Nice, you have fans in the room. This afternoon he answers the question every team using the same AI models should ask: what can we build that nobody else can copy? I know you may be tired, but I want the loudest applause you can fathom for Ravi. **Ravi:** Thank you, everyone. Really excited to be here. It's been a fantastic trip to Berlin so far. I started on Tuesday with the Leaders Lab, then did some of the workshop stuff. I did Berte's improvisation workshop, which was a lot of fun; I'd never done improvisation before. I hope everyone has had an amazing day today. I want to talk about one of the most important topics for all of us: in a world with these incredibly powerful AI models, how can we build products that competitors can't match? This is a section of the AI strategy program that Brian and I launched about 18 months ago. As I was getting ready for our leadership day, which takes a much bigger part of that program, I realized that a lot has changed in those 18 months, so the program needed to be updated. I did a bunch of updates, we went through them on Tuesday, and we've got some updates here as well. If any of you are part of Reforge, I highly recommend the AI strategy program. We're doing our next cohort in October with a great set of guests: the head of product from Descript, the CPTO of Miro and the head of consumer product at ElevenLabs will all be visiting. We'll also be doing another cohort of AI prototyping, a course we launched in April of this year. For those who don't know me, I'm Ravi. Today I'm a product strategy advisor. I work with three or four companies at a time as a fractional product executive, helping them think through their product strategy and through these tough questions about how to make the most of this moment, which is both challenging and full of opportunity. I also write Ravi on Product, so please subscribe to the newsletter. I'm going to start doing more events outside of Reforge for newsletter subscribers. Our first one is next Friday: many of you may have read my article about the 12 competencies of product management, and I'm releasing the first update in five years, with new competencies designed for this era of AI. As we talk about the most important moment in the history of technology, I think it's worth looking back at what people thought the future might look like. Around the year 1900, an artist was commissioned to draw pictures of what he thought the world would look like a hundred years later, in the year 2000. He imagined mail being delivered in a very different way: why would it not be delivered by plane? I'm still waiting for that one. He imagined a much better way to buy clothes: why would you not just walk into a shop where some sort of steampunk robot would make your clothes for you? And he imagined a much better way to get ready in the morning: sit back and let your little robot do your makeup and your hair. What's interesting is that our future looks nothing like that. You know what our future looks like? It looks like this: everyone is staring at their phone. The amount of time and attention we spend on software products eclipses everything else. The reason is that the defining technology of our age has been information technology. It hasn't yet been robots, and it hasn't yet been air travel technology. Information technology has evolved in three waves. The first was personal computing. The idea that you could have a powerful computer on every desk and in every home was a dream when Bill Gates started Microsoft, and today it's a de facto reality. That was the first information technology wave. Then came the internet: we've got all these devices, so what happens if we connect them together? We can do a whole set of new things, and that led to the Web 1.0 boom. Then we had the mobile shift: we have all these devices and they're all connected, so what if we detach them from the desk and from on top of the TV? We unplug them, make them small enough to fit everywhere and efficient enough to run on batteries alone. That was the mobile revolution. Each of these revolutions created companies that are household names today. Microsoft, Intel and Oracle were all able to take the fertile ground established by the personal computer revolution and build such a strong position that years later they are still household names. Microsoft just turned 50 years old. We don't think about 50-year-old companies in tech; we think things are constantly churning, old companies die and new companies emerge. But in actuality, if you use one of these moments to establish a strong position, that position is not just durable for the moment. It's durable across several of these transitions. The internet created Facebook, Google and Salesforce. Mobile gave Apple, which had actually lost the personal computer platform shift, another crack at the nut, and they won the mobile shift. That also gave rise to Uber and TikTok, both of which depended on the idea that you could have a small connected device in your hands. One of the reasons this happened is that past platform shifts created blue ocean opportunities. One of the frameworks I really like is from *Blue Ocean Strategy*. I did a session this morning on strategic thinking, we had a great conversation, and I think *Blue Ocean Strategy* is one of the seminal books on how to think strategically. The authors talk about two different types of markets. There are red ocean markets, where a bunch of competitors fight each other for market share. How do they get more market share? They lower their prices, and as they lower their prices, their margins go down, until eventually all those margins go away. That leads to a red ocean, to red lines on your accounting statement, to losses, and to markets where all the value is delivered to consumers rather than to companies. Blue oceans are completely different. They are untapped markets with limited competition, and ideally those are the markets you want to be in. Each of these prior shifts created blue oceans for new companies to emerge. We're now in the fourth wave of technology, and I believe it will be the most important one. We started with personal computers, with this power to compute deterministically. Then we connected all those things together. Then we shrank them down so they could fit in our pocket. And now we're adding the next fundamental capability, which is natural intelligence. We're in this fourth wave, but I'm not seeing blue ocean opportunities the way we saw them before. Products find themselves competing in a lightning-fast environment. Everybody has learned: fool me once, shame on you; fool me twice, shame on me. So the big tech companies are moving in fast, because they don't want to lose the AI platform shift. We've got fast-moving competitors like OpenAI and Anthropic, who not only provide the services we use to build our products but also compete with all of our products. For the first time in history, a single product competes with online therapy companies and data analytics companies, and that product is ChatGPT. We are all competing with ChatGPT. Then there are the venture capitalists, who think, "This is when we make all our money, during a platform shift, so let's put money into all the companies." All of a sudden, every one of us is in what feels like a red ocean full of competitors. We're also seeing something we've never seen before in the history of tech. Product-market fit has never been a given: once you achieve it, you start to lose it, because customers have escalating expectations of what your product can deliver. But now we're not just seeing product-market fit decay; we're seeing it collapse. Here's Chegg's stock price. This chart is a little old and it looks even worse now. Stack Overflow is another example: the number of questions asked on Stack Overflow, a developer Q&A forum, has fallen by 99% since ChatGPT launched. Some companies are collapsing and others are growing faster than ever. ChatGPT grew to a million users in just five days. Midjourney achieved, I think, a $500 million run rate in something like 12 months. This is by far the most intense environment I've seen in the history of tech, and I think we can all agree that's what we're seeing out there. It started a few years ago and it's continuing: innovation after innovation, new technology after new technology. How do you win in the most intense environment in the history of tech? That is the most important question for all of us as product builders. In talking to companies and trying to understand how people answer this question, Brian and I encountered a couple of failure modes as we put this program together. One, which came from the early innings of AI, was companies thinking they need to reinvent the wheel and train their own model. A lot of venture capitalists felt you could only have a durable, defensible advantage if you trained your own model. The problem is that everyone else is training models too, and a new model can come out. The two months, two years or three years you spent building a model can quickly be eclipsed by a new release, and suddenly the durable advantage you were creating evaporates. So that's one failure mode: reinventing the AI wheel. The other failure mode is treating AI like a black box: not really thinking about the new fundamental capabilities, and just saying, "We're going to layer AI into our product, we'll add chat." A lot of companies did this a year, a year and a half ago. They bolted it on, it may or may not have made sense for customers, and even though they added AI, they didn't really get much of an advantage from it. I want to talk about a different way to approach AI, which is to think about it through a Lego-block analogy: you assemble your product features from a set of fundamental building blocks. Those include the AI capabilities we all have available to us, the data that only your company has, and the functionality that only your product has. We'll talk about how you can use this very simple but very powerful framework to identify opportunities to build features within your product, or new products, that can win in this really intense market. It all starts with the AI capabilities; of course, that's where AI features begin. It's important to remember that we're living in a somewhat magical time. There are problems that were intractable computer science problems for three, four or five decades. Just telling whether or not a photo contained a bird was an intractable problem for 30 years of computer vision, and now it's as trivial as calling an API. The same goes for translation, for generation and for the reasoning we've all gotten used to. I've been building products for 30 years, and one of the things I'm most excited about as a product builder today is that we have this expanding, almost magical set of building blocks to work with. That's one of the reasons I love the Lego analogy: we can snap these blocks together in ways that are really powerful for our customers. This is only a sample of what's available to us, and it keeps growing. We can detect people's emotions with shocking accuracy. We can do search that is miles better than what Google could do just a few years ago. We can do complex reasoning. We can do style transfer, so you can take a photo and apply a Picasso style or a cartoon style to it. Many people in this room have played with retrieval and knowledge graphs, ways to bring large bases of knowledge into the answers you provide. Classification was another of these really tough problems, where models would continuously get things wrong, and its quality has increased fundamentally. The same applies to audio, video and any type of media you can create: we can identify it, catalog it, understand it and generate it from scratch. With that, I want to make this a little more interactive rather than just standing up here. We'll go around with the microphone: what AI capabilities are you leveraging today to solve customer problems? I know it's the first question and a cold room, so it's hard to be the first to raise your hand, but go ahead, someone jump in, or I can cold call if that's easier. Go ahead. **Participant:** I run a voice-of-customer program, and we use it a lot for analyzing customer feedback at an incredible scale that just wasn't possible even a couple of years ago. **Ravi:** Yes, this is one of the biggest changes. It used to be very hard to get signal out of unstructured data. Now any unstructured data you get from customers, you can turn into gold, and everyone working on products should be doing exactly this. I do it with every one of my clients: I ask, "What open-ended data do you have? Let's go in, analyze it and figure out what we haven't learned from it yet." Fantastic example. What else? What other AI capabilities are you using? **Eva (Eyeable):** Hi, I'm Eva from Eyeable. We're an accessibility SaaS platform that checks how accessible websites and apps are. Traditionally it's a yes-or-no answer, for example whether a picture has alt text. We have a manual team who test as well; that was how it worked back then. Now we additionally have AI tests that simulate the behavior of someone who really has a disability and recognize, for example from the context of the website, whether a picture should have a description or not. **Ravi:** This is another great example. We've all probably done accessibility studies and audits on our sites, and it used to be incredibly time-consuming to even get a basic initial evaluation. Now you can literally pull up Codex, ask it to do an accessibility evaluation, and it will go through your site and give you an initial report. You should probably work with a company like yours to do a better job of it, but you can start that process, where before it would have been impossible. Let's do one more. What other AI capabilities are we using? These are both fantastic examples, not the standard chatbot. **Participant:** It's actually inspiration coming from Reforge. We use synthetic users to do UI testing, for debugging, and that works pretty well. It's the only mature use case for synthetic users so far, for me. **Ravi:** Amazing. The idea with synthetic users is that you can give an AI a persona and context and have it use your product like a new user, or like an experienced user who uses particular workflows, and suddenly, even without talking to customers, you generate a whole bunch of information. For us as product leaders, voice-of-the-customer work and customer research have fundamentally changed, and I'm not seeing teams use enough of this: synthetic users, voice-of-the-customer analysis, or taking the transcripts of all your screen recordings and getting information out of them. These are gold mines of data that were previously locked away because they were unstructured, and now they're available to us. So everyone has access to the same AI capabilities. The models that can do the accessibility audits, the voice-of-the-customer work or the synthetic user testing are available to everyone who is willing to pay the token costs. Your competitive advantage is probably not going to come from the model. If it does and you have a durable advantage there, great; that's a fantastic part of building a competitive product. But as we'll see, the most successful AI companies don't necessarily have the best model. They're figuring out other ways to build competitive products. Your competitive advantage is going to come from what is uniquely yours, and there are three things that define that. The first is your data: what data is uniquely yours and not available in the training data sets? The second is your functionality, and I'll explain what that means in a second. The third, which is not to be underestimated, is your understanding of unmet customer needs. As a company with established customers, you have an advantage that every seed-stage AI-native startup doesn't have: customers who are constantly telling you secrets. Can you learn those secrets and use them to move faster than a new company possibly can? Let's go through these three pieces in a little more detail. Your data is important because it provides context to the AI models, and context is the first step in taking an off-the-shelf model that everyone has and turning it into something bespoke for your product. The importance of context cannot be overstated. I know we've all spent a lot of time on prompt engineering and context engineering. I encourage you to think about it from a fundamental strategic perspective: if you can give the AI model context that your competitors don't have, you can create a capability your competitors never could. The last piece, which is really interesting especially as we move into a more agentic world, is that your functionality determines how AI behaves. The way I like to think about it is that your product can give AI superpowers. You've spent years, in some cases decades, building product functionality that is only available within your product, and now AI can learn to use it. An AI agent could say, "I want to use your refund system to automatically give a refund to a customer," or "I want to use the Figma surface to create a UI design directly on the board that a designer can then manipulate." If we start to think about our products not as something dated and obsolete that we have to reinvent, but as a set of superpowers for the AI, we can really lean into that and understand how to get ahead of any competitor in the market. So we have these three building blocks, and just as important is how they connect. As I mentioned, the first and most important connection is that your data feeds the AI capabilities; it's how you inform the AI's understanding. In many cases AI is going to generate something, and the output is only as good as the context. Better context leads to more differentiated, better output. That output then becomes part of your data, and this creates a flywheel that is incredibly important to your business. It's your proprietary data flywheel that only you can own, building up your understanding of the customer's needs so that you can use AI in a way others can't. There's a second flywheel around functionality. Your product determines the orchestration of AI: when do we bring AI capabilities in, and when do we use traditional software? That's an important question, because traditional deterministic software is the right answer 95% of the time. It has zero marginal cost. It works in a predictable way. It can be fully debugged. It's not going to drift over the next couple of months, it's not going to change because of some change in the model, and it's not going to fail when the model isn't available. So wherever possible, use deterministic software to solve the customer's problem. If you can't, use deterministic software to figure out which model to use and when to use it. That's the orchestration side. And AI has now gotten incredibly good at using your product, or the set of features within it, as a tool. Because you can surface what your product can do to the AI, the AI can learn it. These are the superpowers your AI can execute that only your product can make available. I know this is a little conceptual right now; I'll go through specific examples. One thing I think is really important, and I've just added it to the presentation because it's so key, is that the breakthrough is not always the model. Yes, the models are getting better, and that's great; it gives us more capability. But if we look at the innovations that have been really disruptive over the last couple of years, it usually comes down not to the model but to the harness. The first example, which seems trite now, is Perplexity. In the early days, ChatGPT had a training cutoff date, and if you asked it, "What's the weather like in Berlin tomorrow?", like I did yesterday, it couldn't answer. It would say, "Sorry, my training data ended in January 2023." Perplexity said, "Let's use the off-the-shelf models from OpenAI and add a tool: web search. Once we've searched the web, we'll pull that data in, add it to the context and answer based on it." It was the first at-scale example of retrieval-augmented generation, and they created an AI innovation not through the model but through the harness, by giving it web search. One of the best examples is Claude Code. Claude Code led to a revolution in the ability of AI to develop software. But the same model Claude Code uses was available in Cursor; you could have used it the same day and it would have had the same coding capabilities. What was different about Claude Code is that it was implemented as a command-line tool, and as a command-line tool it could execute the hundreds of other command-line tools on your computer. It could search files, read files, write files, load tools, run tests, compile code, start a browser. Because it could run all of those tools, the AI coding agent can say, "I'm going to write some code, try it out, test it. It didn't work, so I'm going to fix the bug," and do a loop that combines a set of tools with a set of capabilities. That created a really robust harness that fundamentally changed the trajectory of AI-based coding, not because the model got better but because the harness got better. The last example is OpenClaw. OpenClaw was the first example of agentic systems working really well. In a short period of time, Peter, the engineer who built OpenClaw, wrote 300,000 lines of code via AI. Not a single line of it was AI model code. All of it was harness code: giving the AI access to your calendar, to your apps, to whatever was on your system. It gave the AI hands, so it could act within your computer and within the set of tools available. If we think about it this way, these three fundamental building blocks plus these four ways of connecting them enable you to create something that, by default, competitors can't match. I want to look at one example of that; I don't know if we'll be able to get the sound. This is a clip from Miro. At the start it goes by really quickly: "The canvas is the prompt." Based on some notes on the canvas from a shared brainstorm, a person selected those notes and said, "I want to create a prototype." It created a prototype and put it onto the Miro board as standard Miro objects, which you can then comment on and work with like any other content on the board. This is an application of AI that is very different from ChatGPT. ChatGPT can't do this, because it doesn't have access to the same surfaces, the same superpowers, that Miro has. If we look at what Miro is doing through this model: first, they take their proprietary data, which in this case is all the information on your Miro board, and use it to provide situational awareness to the AI, so that when the AI answers, it answers in the context of what's on your board. That's really interesting, not just because it works nicely within the product, but because it's one of the first examples of shared context engineering. The problem with a lot of the work we've been doing over the last year and a half is that all our context lives in isolated tools. It's in chat, it's in Claude, and suddenly we have all these PRDs and prototypes that people created in isolation and somehow have to merge together. Miro by default enables context engineering to happen in a shared workspace, which is incredibly powerful. Then, instead of creating some artifact that lives elsewhere, it output data directly onto the board, which became part of the flywheel. AI not only let you work with the data on your board, it added to the data on your board, making Miro's core use case even more vital for teams. From an orchestration standpoint, it works within Miro's tools: the user decides when to activate the AI features based on their use case, and the AI features can then use the board as a tool, like any other user. Another feature they have is the ability to get comments from different personas. You can have a Scrum Master persona or a product leader persona comment on your board, and you can actually see the product leader moving around the board adding comments. It's just using Miro's tools to act like any other user within your Miro experience. So a key question for us is how we can use this approach in our own products, and what's really important is that it starts with the customers. One of the downsides of the last couple of years is that AI has been so compelling as a tool that it has gotten all of us thinking solution-first: "The AI can do this, so let's figure out what we should do for customers with it." That's the wrong way to think about it. We need to think problem-first. A good AI feature sits at the intersection of customer problems, particularly unmet ones, and what the AI can do. That's where we'll find effective AI features. So we need to work backwards from customer problems that haven't been solved. Those can be unresolved pain points in your experience. They can be repetitive tasks customers have to do over and over again, which they're quickly losing patience for. They can be cognitive burden, where the tool is too hard to understand or use. Customers are also quickly losing patience with spending lots of time learning your tool; I think the age of tool certifications is probably over. We don't want complicated interfaces; we want something easy to use. There may also be inefficient workarounds, where your customer works outside your product to get something done. That's a great signal of an unmet customer problem you can solve. So my next question for everyone: given that good AI products and features start from unmet customer problems, what unmet customer problems can you solve with AI? **Participant:** The Berlin housing problem. **Ravi:** You could probably make a little bit of progress; I don't know how much. I'm from Los Angeles, and we also have a housing problem. So how would you use AI to solve the Berlin housing problem? **Host (Andreea Lungulescu):** Okay, here's what I've done. I run a huge community, and unlike you people, my community in talent doesn't really commune. So I vibe coded a mobile app that prompts people to connect with each other. They literally get "you should connect with so-and-so," and they started connecting on LinkedIn, and now they do their own thing, which is wonderful for me, because I don't have to carry them around saying, "Hey, meet so-and-so." It's an app built with Lovable. It's that easy. **Ravi:** Oh, that's great. And I know from my Tinder days that one of the hardest things, once two people connect, is getting them to talk. Providing a little context is very helpful. **Host (Andreea Lungulescu):** I should monetize it. **Ravi:** Yes, you should turn it into a dating product. You'd make a lot of money. Maybe become the number one monetizing app. What else? What other unmet customer problems could be solved with AI? **Participant (Parloa):** I work at Parloa, which is an AI agent management platform, and one of the use cases is customer support. One problem we see is this: say you're a recurring customer calling about an order you purchased two months ago and never received. One issue is context: you don't want to be repeating yourself all the time. Nowadays we can actually solve that problem with AI, because we're cracking context and agent memory. As you said, having so much data is powerful for solving the problem, and it makes loyalty a result: customers come back because they were super happy, because we know everything about them. That was more challenging to resolve in the past, and now, thanks to AI, it's not that hard. I would say it's still hard, but we're getting there. **Ravi:** That's a great example. Customer loyalty is key, and so many companies now take it for granted. Customers have a history with your company. If they've spent years using your product or services, they feel a sense of loyalty; they've invested in your company. But so often you call customer service and there's no record of that, and all that loyalty seems for naught. **Participant:** Something we've gotten really good at is identifying high-impact, low-hanging fruit. It might be a small paper cut, an inefficiency or a frustration that would never reach the quarterly planning cycle of a product or engineering team. It's a relatively small thing, but it has a lot of signal behind it. Sometimes it's just a handful of lines that need changing, or a small fix or tiny improvement that makes a big difference to users' lives. **Ravi:** That's super important. When I was at Tinder, I worked with a product leader from Spotify who said something I never forgot: at Spotify, quality was the most important feature. We ship that first, and then we ship everything else. I think we have more ability to do that than ever before, because we can find those little things that annoy customers and erode their experience, that are easy for us to fix but may not tie to our OKRs. But what is our OKR, ultimately? It's about serving the customer well. Awesome, thank you. One of the examples I really love of using AI to solve customer problems is Descript. This is the palette of AI features available within Descript, and all of them are designed around the creator's use case. The creator's need is: I want to look and sound my best for my audience. One cool feature is called Eye Contact. As a creator recording a YouTube video, I'm probably looking down at my notes, speaking into the camera, looking back down at my notes, speaking into the camera. That makes it very difficult to edit a continuous cut without my eyes darting up and down. So they use computer vision and generative AI to update the footage so that your eyes are always in contact with the camera, instantly solving one of the hardest and most time-consuming editing problems for creators. Every one of these features uses AI, and yet the term AI appears nowhere on here. In fact, I had the CEO of Descript in our AI strategy program, and she said she got an angry call from a customer asking, "Why are you charging me for all these AI features I don't use?" She said, "Well, are you using the product?" "Yes, we use it all the time, we love it." "Well, everything uses AI. We just don't talk about it that way." Customers don't want AI. In fact, I think it's actually a negative: if you put AI on a feature today, customers immediately lean back, because there's a lot of skepticism. But if you say, "Here is the customer problem we're going to solve for you," they don't care how you solve it, as long as you solve it for them and are responsible with their data. I think this is one of the best examples of an incredibly AI-native product that solves customer problems rather than just implementing technology. If we're going to apply this to our own products, the first step is to take an inventory of all the AI capabilities that could solve the unmet problems customers have and figure out the match between the two. Once we know which AI capability or capabilities we want to use for a particular customer feature, we need to do a careful audit of our data. There are certain classes of data that are proprietary to you and, by definition, not in the models' training data. The first is real-time data: traffic data, weather data, anything like that, especially if it isn't easily searchable now that web search is in all the models. That is proprietary, and especially if it's real time and has to be delivered quickly, you have the opportunity to create something that can't be done with the models alone. There's user-specific data: whenever the user has created information on your platform that is personalized to them and not crawlable, that's another element of data you can lean into. There's domain-specific data: you may have built up expertise. One of the disadvantages of ChatGPT being so horizontal is that it competes with everyone. One of our advantages is vertical expertise that we can use as context, so that when we use the model, it does a better job for customers in that specific domain than it does without that context. One of the most interesting ones is human judgment data. You can take a data set, have people evaluate it, and come up with an augmented data set that is much more valuable than the original. A great example is Midjourney. In the early days, image generation models were trained on all the images on the internet, and it turns out that most images on the internet suck, so you get really bad images out. What happens when you curate the images and say, "I only want to include the ones that are aesthetically beautiful"? Midjourney did exactly that. By augmenting the data available to everyone and identifying the most aesthetic images, they came up with an image generation model that was so much better when it launched, and it has maintained that advantage. The last one is reinforcement data. If, as part of delivering the product, we can get feedback from customers on what worked and what didn't, both explicit feedback like a thumbs up or thumbs down and implicit feedback like whether they copied the result, saved it or shared it, those are all signals we can use to figure out whether they liked what was generated. If so, amazing, we reinforce that. If not, we figure out what went wrong and reinforce with that. The last piece is your functionality: things specific to you, hard problems you've spent years solving. These could be specialized workflows; there are some really interesting examples of this at Harvey, the legal AI company. There might be unique algorithms you've built. The TikTok algorithm is the most valuable algorithm in the world: they've accumulated so much information that they can identify, better than any other company, the next piece of content a given user wants to watch. You likely have valuable algorithms in your existing businesses around search, discovery or other elements of how you deliver your product. Then business rules. Even though business rules sound boring and trivial, if you're working with customers on a complicated use case and you've built up business rules, particularly ones specific to and personalized for particular companies, those become superpowers available to your AI that others don't have. The last one is integrations. Say you run an attraction or tour-supply marketplace and you integrate with a bunch of merchant services or suppliers, booking with tour guides who use hundreds of different systems or maybe even write things down on paper. That's an advantage that's hard for anyone else to replicate. So, next question for everyone: what proprietary data or functionality can you leverage to gain an unfair advantage? You have data available to you, and functionality your product has built up over the years. Think about those in terms of an unfair advantage. What can you use, and what will that unfair advantage be? Go ahead. **Participant:** Thank you. Could I introduce a fourth Lego brick? **Ravi:** Sure. **Participant:** I'm biased: I've spent 30 years in open source, so this won't apply to everybody. But with open source, if you're an open source project or product, all of this stuff is baked into the model, into the training data, and that can actually give you an unfair advantage over proprietary or closed platforms. Yes, the model has web search and all these other things, but if you've got an open-source-based product, solution or service, the model has all this rich context about how other people use it and the problems they've solved with it, and that can give you a very, very strong advantage. **Ravi:** Interesting. This came up during the leaders day, actually. People now want software that's molded for them, and if you have something that's open source, you can fork it, use a vibe coding tool, mold it to your customers and create something that really stands out. So that's a great example. And to your point about a fourth Lego block: absolutely, add a fourth block. If you find something here doesn't work for you, that's the point of all the frameworks we share at Reforge. Brian likes to say Reforge is a tool, not a rule. All of this is meant to help you make these distinctions and identify what might be missing that you could build on. Awesome, thank you. **Participant (Leapsome):** Hi, I work at Leapsome. We're originally a people enablement platform, so surveys, performance reviews and so on, and we now also have an HRIS, which is really good as well. That means that, compared to a lot of our competition, we have data on how people are working and performing, as well as core data like manager changes and performance reviews mixed with compensation updates. That gives us a really big moat for making useful AI suggestions instead of hallucinations, to put it mildly. **Ravi:** Perfect, that's a great example. HR data is typically confidential. It's not crawlable and it's not in the training data, and if you've accumulated it over the years, that's an advantage you can really lean into. Let's do one more. Maybe someone in the back or the front. There we go. **Participant (Semrush):** Hi. I work at Semrush, the company that provides data for SEO specialists. We have data about keywords, and it's a great advantage for estimating prompt volumes for AI visibility, using ML to approximate those numbers, which most startups don't have right now. **Ravi:** Awesome, that's a good example. A lot of data has accumulated in SEO, and that data is good not only for SEO but also for AEO, so you can build a really good AEO product from it. Thanks, everyone, for jumping in. I want to go through one last example before we wrap up and do some Q&A: what does it look like when we assemble all the pieces? I want to use one of the fastest-growing companies in AI. Granola is the AI note-taking tool. They recently raised $125 million at a $1.5 billion valuation, and they're one of the fastest-growing AI platforms. The problem is that note-taking is actually a red ocean market for AI. There were already a bunch of players when Granola launched: Fireflies, Fathom and Otter, all AI-native players whose valuations had increased massively. Zoom was getting into the space, Teams is getting into the space, and there are hardware products. All of these products are solving AI note-taking. So how could Granola create a wedge? How could they take a red ocean market and enter it in a blue ocean way? What they noticed was that all of these products solve one particular use case: "I want someone else to take my notes." They solve an autonomy use case, a delegation use case. Granola said there's a set of professional note-takers. Many of us in this room are professional note-takers; we live and die by our notes. We don't want to delegate note-taking to someone else. We don't want to sit in a meeting and then get an email afterwards saying, "Here are the notes from the meeting." We need our notepad up and we need to be taking notes. But if you can make those notes better, that's really interesting. So Granola created an AI note-taking product for people who want to take better notes. Instead of solving an autonomy use case, they solved an amplification use case. That's an interesting spectrum to use any time you're building something with AI: do you want someone to delegate to the AI, or do you want them to use the AI to make themselves more powerful? They built one of the fastest-growing companies in AI without doing anything special at the AI layer. They use completely off-the-shelf models: Deepgram for transcription, which is available to all of us right now, and Anthropic and OpenAI for LLM features like summarization and chat. What they did do was think about how to use those models in a really interesting and proprietary way. The first was context. When you use Granola, a blank page comes up and you can enter notes. It then uses both the notes you took and the transcript of the meeting, combining them into a much better note than a transcription alone, because it knows exactly what you paid attention to and what you wanted to take away: the action items you cited, the specific things you noted down. Because it uses both the context you provided and the transcription, it outputs much better notes. So you now have a repository not just of transcripts but of your notes made better, which becomes part of a flywheel Granola can build over time. That makes them really sticky with an individual user, and also with teams: they can take all the notes people in a meeting took. I might take one set of notes and someone else another, and those can be melded into a team knowledge base that's much more powerful than just a transcription of the meeting. They made a number of innovations on the orchestration side. Most existing note-taking products are web-based agents that log into your meeting as a participant. Granola said, "Instead, we're going to orchestrate the AI within a native Mac app." By doing that, they have access to system-level sound, and they can tell when a meeting starts, no matter what tool it's in, and put up a reminder to start taking notes. Because of this system-level integration, they were able to model the amplification use case much better for their users. They also use a bunch of tools within their product: integration with your calendar and with metadata. It can look at your participants, who you met with and what time you met, all to make what's created more accurate and more useful. So that's one of the fastest-growing AI products, one of the darlings of Silicon Valley, and they got there not by innovating at the AI layer but by innovating at the data and functionality layer. To win with AI, I think it's important to ask four critical questions. What unmet customer problems need to be solved? What AI capabilities can solve those problems in novel ways? What proprietary data can power those solutions and shape those AI capabilities? And what superpowers can your product give to AI? A hundred years ago, when Jean-Marc Côté, that artist, imagined the future, he didn't get everything right, but he did predict that we would all be trapped on Zoom calls. It just looks a little different here. What I want you to take away is that winning with AI isn't just about AI tech. It's about turning what is uniquely yours into a competitive advantage. Thank you, everyone. So glad to have you here. We're going to spend some time on Q&A, and I'd love your feedback on the talk as well; you can scan the QR code. Let me double-check the time. We've got about ten minutes, so time for a few questions. Thank you, everyone. #### Q&A **Ravi:** Go ahead. **Participant:** So I guess it boils down to how uniquely and creatively you can look at a problem and frame it very differently, like in the Granola example? **Ravi:** Yes, and it comes down to really taking stock of what your company can do, as well as what your company understands about the customer. Too often companies look at what's happening outside and say, "The competitor did this." If the competitor did that, chances are they're really good at it, and that may not be the area where you want to compete with them. If instead you can look internally and say, "Here are the customers we have, here's what we know about them, here's how we can translate that into something from an AI perspective that only we can do," that's where you get capabilities that are very hard for other companies to match. **Participant:** Thanks for the talk, Ravi. A question on the flywheel: at what point does this potentially degenerate into a sort of slop inception, and how do you prevent that from happening? **Ravi:** It's a great question: at what point does this data flywheel degenerate into slop? I think it's helpful to define slop as the accumulation of unexamined defaults. Without any context, the AI is just going to default to what it's been trained to do. Whatever it's been trained to do will get better over time, but we as humans will also get more bored of it. So no matter how good it gets, we're still going to read the average from AI as slop. The slop today is prettier than the slop from yesterday, but it's still slop. So the question is how we get the AI not to do its default, and the way we do that is with context engineering. That's why context is so important. It doesn't have to mean a lot of context. Really sharp, very clear context is how you get an AI to do something that doesn't read as slop, because it doesn't read as the average; it reads as something opinionated and exceptional. In fact, less context is actually better than more in most cases, because of something called context rot. Even though context windows today are much larger than a couple of years ago, and you can have a million tokens or more in the context window, once you have all those tokens the model gets confused about what it should pay attention to and has a much harder time retrieving relevant information. Then it starts to create things that look a lot like slop. So I think good context engineering follows a bell curve. You don't want so little context that it generates slop because it's too vague, and you don't want so much that it generates slop because it doesn't know what actually matters. The right context is somewhere in the middle. **Participant:** Thanks. There's a great article you've probably heard of, "The Next Great Distribution Shift" by Brian Balfour. **Ravi:** Yes, I love it. **Participant:** I've read it four times. Brian argues that every distribution shift has stages. There's the generosity stage, when platforms invite you to use them because they need to gain a mass of users and become the standard. Later there's the phase called "walls up," when the conditions for using the platform change rapidly, and in most cases it becomes more expensive or more challenging to use. From your presentation, the central piece of the whole puzzle is AI capabilities, and there are just a couple of vendors providing those capabilities. From a product strategy perspective, how can you structure that risk so that when the walls-up stage arrives, you don't wake up the next day with a huge problem for your product? **Ravi:** Absolutely, it's a great question. Here I'm really trying to answer the question of how you create a product that has fit with your customer. I think distribution is as important a problem today, if not more so, and all of this needs to be done in the context of your distribution strategy as well. Brian talked about the fact that a number of platforms today are moving out of the generosity phase: away from giving free traffic and towards locking things down. LinkedIn is a really good example. It used to be that if you had a good LinkedIn post, you'd get a ton of exposure. Now they really want you to boost the post to get more traction. The AI platforms, Claude and ChatGPT, are still in the generosity phase. ChatGPT a little less so, because they have advertising, so you can pay to play or show up organically. The challenge, though, is that it's not really clear how to meaningfully and directly leverage them as distribution channels right now. So I think we're in a tough phase for AI where distribution is hard. A lot of the traditional cheap distribution has been closed off, and a lot of the free distribution that's supposed to come online as attention moves to ChatGPT, Claude and other places isn't as addressable as we'd like. It means understanding where you can lean in and get through some of those challenges. Some people are being really successful with paid campaigns; others are using influencer campaigns, or astroturfing influencers. Just like you figure out how to create a product competitors can't match, it's important to figure out a distribution strategy that competitors can't compete with or interrupt, and that will scale with your business. Did I answer your question? Okay, perfect. **Host (Andreea Lungulescu):** I'm pretty sure there are plenty of other questions, even if people are probably quite tired, which is amazing. You should do a Slido for questions only. Actually, don't, because you'll be inundated. For those of you who still have questions, because we are at time, please find Ravi a little later. You'll be around, right? **Ravi:** I'll be around. I'm also on LinkedIn, so feel free to connect with me. **Host (Andreea Lungulescu):** And you don't have to boost your posts. **Ravi:** I'm pretty sure I do. I don't boost them, but I probably should. **Host (Andreea Lungulescu):** No, you shouldn't. Before we end: before Ravi came on stage, I asked him to tell me something that would make people go "oh, cute" when they hear about him, and then I forgot to tell you. So I'm telling you now, before he leaves. If I say the word "Kava," what do you think of? *[Audience guesses sparkling wine: cava, prosecco.]* **Host (Andreea Lungulescu):** Kava is the most adorable rescue pup, and the pup is on his home screen. It's his dog. **Ravi:** Not my daughter's dog. My dog. **Host (Andreea Lungulescu):** Now you know what to think of when you hear the word Kava. Ravi, thank you so much for the wonderful presentation, and thank you, everyone, for being here. --- ## Doctolib Stage Deep dives · in person only (not streamed) · recorded with Granola. Hosts: Andreea Lungulescu (morning), Veronika Morozová (afternoon). ### 10:30 · The Interface Is Not Your Product. Now What? — Dajana Aleksic - **Speakers:** - [Dajana Aleksic](https://www.linkedin.com/in/dajana-aleksic/) — Head of Product & CX, Mercedes-Benz. Leads the digital self-service products for Mercedes-Benz financing and leasing customers globally (1M+ app downloads, $1B+ in payments a year). - **Stage:** Doctolib Stage · CIC Berlin · 17 Sept 2026, 10:30 CEST - **Format:** Talk · Track: Build & Craft - **Host:** [Andreea Lungulescu](https://www.linkedin.com/in/alungulescu/) (The Principal Recruiter) - **Official abstract:** The interface is becoming optional. Remove it — and what's left tells you everything about your real strategy. **Summary.** Dajana Aleksic, Head of Product and Customer Experience for Mercedes-Benz's digital self-service products for financing and leasing, uses the concept of "Zero UI" to ask where complexity goes when interfaces become simpler. Prompted by a reorganisation that shifted ownership of the front end to another team, she walks through a diagnosis-first strategy approach — gather signals, map them to underlying problems, name the challenge — with interactive exercises for the audience to apply to their own products. **Key takeaways** - Zero UI (a term coined by designer Andy Goodman in 2015) is not "no screen" — it's machines understanding human language and intent. There is usually still a minimal interface for confirmation. - Customers increasingly expect systems to anticipate their needs; the future customer may be the user's own agent — for example, a finance agent that moves a payment before the user even notices. - When interfaces get simpler, complexity doesn't disappear — it moves elsewhere. In her case, it moved into ownership across teams and the customer journey, not into technology. - Don't skip the diagnosis: gather signals from data, the team and the organisational context, ask "why" to find root causes, and name the challenge as one simple, sayable sentence. - "I may be wrong, but at least I'm not confused" — pick a challenge, define a strategy, move, then adjust as new signals arrive. *The full transcript of this session is not published, at the speaker's request.* ### 12:05 · From Feature Factory to Economic Ownership — Florian Bonnet - **Speakers:** - [Florian Bonnet](https://www.linkedin.com/in/bonnetflorian) — VP Product Solutions & Growth, Veriff. A decade in product and growth at HelloFresh, Typeform, Pleo and Veriff; PhD in theoretical physics, MSc in behavioral science; author of The Power of Analytics. - **Stage:** Doctolib Stage · CIC Berlin · 17 Sept 2026, 12:05 CEST - **Format:** Talk · Track: Strategy - **Host:** [Andreea Lungulescu](https://www.linkedin.com/in/alungulescu/) (The Principal Recruiter) - **Official abstract:** Product teams ship features. Businesses need economic outcomes. I introduce Economic Architecture, a framework to turn product from a feature factory into a system that drives revenue and compounding growth. **Summary.** Florian Bonnet argues that product leadership must shift from stewardship of features to economic ownership, because shipping velocity does not correlate with revenue. He introduces "Economic Architecture", a five-pillar framework (market thesis, value capture design, capital allocation, commercial integration, compounding loops), illustrated with examples from Typeform, Veeva, Stripe, Datadog and Veriff, followed by questions on focus in volatile markets and on expanding into new markets. **Key takeaways** - In a survey he cites, 79% of leaders said their performance is evaluated on revenue, while only 8% use velocity as a success metric; "busyness is not business". - Market thesis comes first and must be narrow: "mid-market SMB" is too vague. Typeform found that marketers doing lead generation for events were about 1% of sign-ups but converted best; Veeva won pharma CRM by specialising. - Design value capture deliberately and early; net revenue retention below 110% signals weak monetisation. With AI agents, per-seat pricing breaks down, and products may compete for headcount or services budgets instead. - Allocate capital across core, growth and explore portfolios (like Horizons 1–3), state explicitly what is deprioritised, and resist moving things from explore to core too quickly. - Commercial integration: before building, know which customers you will sell to, how, and through which distribution channel, so time to revenue is short. - Build compounding loops (e.g. Veriff's data loop, Typeform's "made with Typeform" referral loop), and re-check that all five pillars fit together whenever you add a segment. #### Transcript **Andreea (moderator):** The first time I introduced a speaker today, the speaker probably did better moderation than I did, so I'm sorry. I'm also sorry for those of you who were here for the previous talk in the same room — you'll have to hear me twice. I don't know how many of you there are; is there anyone left in any other room? Florian, you have to deliver. My name is Andreea. I'm probably one of the very few people at this conference who is not from product. However, I do have something to do with product. First of all, I bribed Daniele. No, I didn't. I work in talent acquisition and have done so for 16 years. I've worked in consulting, with Zalando, with Wayfair, with tech. Then I went to the dark side: I built Germany's largest talent community, which is a product — now it's many products. And I also run my own consulting company for talent acquisition, for companies and for individuals. So somehow I'm still there. Most of my recruitment life was with people in tech and product, so some of you would have been my candidates, some of you might have been my hiring managers. If I rejected you, I'm really sorry. I'm actually a gentle rejecter, I promise; I give feedback. Anyway, it's not about me, but you should know who I am and why I'm here. I'm going to sit over there; if you need anything, tell me. Florian — you obviously know him, he doesn't need an introduction. But I will do one, because I did my research and my stalking, as is my responsibility. **Florian:** Very efficient. I was surprised. **Andreea (moderator):** LinkedIn and more. Maybe you know him, maybe you don't, but here's something you may not have found unless you looked deep enough into his past: before he wrote a single product roadmap, he wrote research papers on neutrinos — tiny subatomic particles. My question to him was why; he will explain at some point. He got a PhD in theoretical physics from Paris Diderot, and one of the papers he co-authored has been cited more than 450 times. So, a celebrity. Later he did an executive master's in behavioural science, because why not. So now he can explain how particles behave and how consumers behave, and I think that's what we'll be looking into today. From physics he moved to BCG, then product leadership at HelloFresh and Typeform. Today he leads product for the enablement platform at Veriff — a very cool company I just found out about — which covers verification tasks, customer back office and the billing engine. He also teaches introduction to product management at ESADE in Barcelona. Anyone from Barcelona here? Well, you should move, and then you can meet him every day. **Florian:** You should move. **Andreea (moderator):** He sits on the Productlab board — he helped build this, that's why he's here. He writes a Substack called *The Monday Cake*, which is very good and illustrated, and he wrote his own book, *The Power of Analytics*. And today: companies win by building economic systems that compound, and product teams can design those systems on purpose. He calls this economic architecture, and that's what he will walk us through. Correct? Cool. Can I please get a very loud welcome? **Florian:** Today I want to talk about the fact that product leadership, I think, is shifting. It used to be the stewardship of features, and I think it really needs to move towards what I call economic ownership. To do that, we will go through what I call economic architecture. I'll start with two numbers — I like math a lot. A survey from last year showed that 79% of the leaders surveyed said their performance is evaluated on revenue. And surprisingly, only 8% said they use velocity as a metric to measure success in their organisation. I was surprised, because everything we keep hearing with AI is that speed is the new moat: you can ship faster than your competition, you can ship product faster to your consumers, and that's what will make you win. But if everybody ships fast because AI is accessible to everybody, is that really the new moat? Maybe not. So it was reassuring to see that revenue is the metric product leaders, and actually every PM, are evaluated on. But the main question is: can we deliver that? What we've seen at least this year and last year is that everybody is trying to increase the shipping of features towards their customers, thinking that's how they'll generate revenue. But if you chart velocity against economic impact, they don't correlate. And I thank Csaba for making me think about this this morning: this is busyness — we are busy — and this is business — we are generating revenue. The two are not the same. You can ship a lot of things, but if you don't structurally engineer how they will make money for your company, your users may have a lot of features in front of them, but it won't generate money. Maybe they are not willing to pay for it. So today I'd like to help you answer a question I think every CEO or CFO asks their product leaders: can you explain how your product makes money? That's what matters most for your CEO, for your CPO, probably for your CTO as well, and so it's what should matter most for you. Who can answer this question if I ask it? You want to tell us? Sorry to put you on the spot. **Participant:** We charge for the main product, and every meaningful feature gets a price tag. We sell a bunch of SKUs. Some of the things we launch we include in the core product for free, but the majority of the really meaningful big rocks that we ship get a price tag and a clear value proposition. So yes, I can explain it. **Florian:** Good, you can answer this question. We'll come back to you after the talk and you can let me know. It's not a bad answer, but I think there are some things missing. That's why I'm here. What happens in most cases I see is that there is a translation gap between product management activity — shipping features; we put stuff in front of our customers and we're very happy because they use it — and actual economic outcomes for the company: is revenue growing? Can you tie the usage of your feature to revenue growth, to retention growth, to churn reduction? This happens at every level in every company, small or big. Sometimes you equate what you ship with economic revenue, but it's not the case. I worked at Typeform. Typeform grew organically very nicely for many years, just because people liked the product, the way it looked, and there was no alternative. Then the environment became denser, growth suddenly stalled, and people realised they didn't understand what they were doing. They kept shipping features the same way, and revenue was not growing, because they didn't understand the fundamentals of generating money with their product. The issue, I think, is not a talent issue. We have good PMs everywhere, and yet this gap exists. It's not a prioritisation issue — sometimes it is, but in many companies you still prioritise what you believe is most important for your product, your customers and your company. It's not a communication issue, even though everybody has communication issues with their company or customers. In 2026 most of us are doing a new job, but this is a structural issue: understanding what product to build and how to translate it into economic outcomes. It's not something you invent from one day to the next. You need to think about it deeply and put a lot of elements in place to generate that at scale, all the time. That's another way to present my CV; I won't go back to it. What I want to say is that my career is quite diverse: I've worked in B2B and B2C, in small and big companies, in marketplaces, SaaS and fintech. So what I'm presenting today is something I believe is quite universal. The language I use may feel B2B, but for B2C you replace sales with marketing, and the principles are more or less the same. So don't feel excluded. The takeaway is that what I'm presenting should work in all your companies. A bit of self-promotion: I wrote a book on product analytics — the QR code is here and at the end — if you're interested in understanding your customers better, setting up the right analytics frameworks in your company and asking the right questions. And since the beginning of the year I've been writing a newsletter, *The Monday Cake*, where at the moment I'm going very deep into economic ownership and architecture. We only have a short time together, so if you want more details and examples, subscribe; it's free. So the problem is execution: translating product management activity into outcomes is what I call economic ownership. Generating revenue is too often left to the marketing team or the sales team. Everything okay? Yes. We leave this to other people in the organisation, when actually, as product people, we should be the ones designing how the company is going to generate money. That's what I want to bring to you, because you can only generate revenue if you build the right product for the right people, and it's not something you should only think about once you've shipped something to your customers. To do that, I propose a framework I call economic architecture, which has five important pillars. We'll go through each in detail, but quickly. The first is the market thesis. As you can see, it's a bigger part of the circle, on purpose, because it's very important: if you don't understand which customer segment you concentrate your efforts on, you will not be successful. Defining this determines what you do in the rest of the framework. The second is value capture design. Once you understand who your customers are and you build a product that answers their problem, how do you extract monetary value from it? What you told us before was this one: you nailed the value capture design. But there are other things we also need to think about. The third is capital allocation. You build a product, you get value out of it, and you continue to develop features over and over. How do you ensure you keep working in the right direction and don't drift out of your market thesis? How do you ensure you still capture the right value and put your resources towards the most important things? Then there is commercial integration. How do you connect to your commercial organisation? How does the market give you the right signals, and how do you integrate them into your practice, so you keep generating money and don't become obsolete? And the last one is compounding loops: the system that strengthens itself over time, so that you generate value more easily every quarter, every year, and every customer acquisition becomes easier with the next customer. In the next — I don't know how many minutes we have, let's say 25 — we'll go into each pillar with some examples. If you have questions, raise your hand now; don't wait for the end. I think it needs to be interactive. The first one is the market thesis: where do we concentrate our economic efforts? The failure mode of the market thesis is being horizontal: you have a large set of customers, you build features for everyone, and you don't focus. I'll take Typeform as an example of the failure mode. Typeform was a form builder. You could use it for your team event, put a form on your website to collect feedback or churn reasons, use it for lead segmentation — for everything. Because everybody can use it, you end up with a tool that is good at a bit of everything, but not specialised and not answering an end-to-end use case. At the beginning it was great; the company grew to roughly 70 million like this, which is not bad. But then specialised solutions appeared, and suddenly people thought that a product doing a bit of something but not everything was too expensive compared to specialised solutions that were maybe the same price or a bit more expensive but covered the end-to-end use case. That's when it stalled. So the most important thing with a market thesis is to be very precise. The one-sentence test for whether you have defined your market thesis properly is: "We concentrate our effort on a specific vertical or sub-segment, because we solve a very specific economic problem, and winning one customer in this sub-segment makes the next one easier because…" When we build products, we sometimes think about our customers' problems in a functional way. But at the end of the day, people want to solve a functional problem because they want to generate revenue or economic outcomes for themselves. And then there is the compounding mechanism. In B2B, for example, it would be references: if you work in ride-hailing and you sign Uber, signing Lyft will be easier because they know Uber uses your product. The main issue I see is that if I ask many people what their market thesis is, they say "mid-market SMB". Mid-market SMB is too vague. What you want is something very specific. For example, at Typeform, I was working on growth at the time, and when we looked at how to generate more revenue, we said we need to be very precise. We looked at the good customers we had that generated more revenue, and said maybe we should get more of them. We said marketers are a good segment. But "marketer" means a lot of things: you can be in CRM, in paid social. So we looked at the use cases, went very narrow, and said we need to focus on marketers doing lead generation for events — people who want to attract people to webinars or events like the conference today — because their use case is very specific and they have specific needs. These people represented about 1% of sign-ups at Typeform, but their conversion rate was the highest of all. Why is it important to focus? If, for this very specific segment, you're not able to tell a story compelling enough for them to buy your product, stay with you, expand their usage and buy more of your products, you cannot convince anyone else. The wider your segment, the more diluted your story becomes. You end up with a wishy-washy story that kind of resonates with people, but not fully, and it's not compelling on the use case, the problem resolution or the feature usage. Sorry — my laptop shuts down very quickly, so I'll stay here. One example here is Veeva versus Salesforce. Who knows Salesforce? Who knows Veeva? One person. I'll tell you the story. Salesforce is a very horizontal CRM and a very successful company. It started with some products and then acquired others; now it's a giant that does a bit of everything. But then somebody realised that selling Salesforce to pharmaceutical companies is very complex, because pharmaceutical companies are very regulated, the way they interact requires certifications, and it's really hard to tune Salesforce to their needs. So pharma companies bought Salesforce and then tried to hack it to make it work for themselves, which was really complex. Somebody said: maybe we can do a better job. They went into a very narrow segment. And the economic problem they solved was not "pharmaceutical companies need a CRM"; the actual problem was that pharmaceutical companies were paying a huge amount of money to tune horizontal CRMs to their own problems and details. Veeva focused on this sub-segment of customers they could sell a CRM tool to, and now they hold 80% of that market. They're a bit like Salesforce in their own segment; they make a lot of money, and they could only do that by specialising. If you want to attack the CRM market today, you have a lot of competition. But if you attack it for a very specific segment, with problems that are specific to them and not easily solved by the major tools, you have a shot at building a consequential company. So now you have your market thesis and your customers, and you need to extract value from them. The failure mode in value capture is: people are using our product and our new features, but monetisation is weak. That's one of the hardest problems: how much are people willing to pay for your product? We always tend to feel that if we make people pay, they won't use the product. But if you solve a sharp problem for them, they will be willing to pay for your solution. The sentence test here is: "We capture value through a specific pricing and packaging model" — are you pricing per usage, per seat, per monthly subscription? — "and we expand as customers do specific things." So you don't only have a pricing and packaging model; you also understand how consumption of your product will deliver more value over time. That's something you need to design. You cannot just put things in front of customers, see what happens, and then design your pricing and packaging, because the later you do it, the harder it is to change. Once people are used to a specific pricing, changing the price point or the packages is really hard, especially for existing customers. How can you diagnose that you don't have the right value capture? If your NRR, net revenue retention, is below 110%. That means that for your existing cohort, in a year's time, the money they generate for you has grown less than 10% compared to today. NRR is not really a customer retention metric; it's how much money I generate with the same pool of customers over a year. Stripe does a very good job at this. Stripe decided to focus on one segment: developers. I'm a developer, I need to add payments for my company, I just want Stripe. I have all the API calls I need, I don't need to talk to anybody, I implement Stripe, and I have a very simple way to get paid. I put in my credit card and then pay based on payments processed. As I process more and more payments because my company becomes successful, Stripe makes more and more money from me through its fee. Their fee structure is a bit complex, with a fixed fee and a percentage fee, but the more transactions you have, the more money they make, and if transactions are larger, they make more money, because they take a percentage of every transaction. And you never need to talk to a salesperson along the way. So they really tailored it to one segment and engineered the pricing and packaging through the lens of developers. Very few people like talking to salespeople, and developers even less. They knew they had to go to developers, to engineers, with pricing and packaging built for that: very simple to understand, very simple to implement, no hassle to renew or to grow usage. I won't tell you the whole Stripe success story, but that was part of the initial success: they really focused on one customer segment. A little interlude on AI, because I think with AI, the way we capture value from our products is going to change, and there are still a lot of open questions. One person delivers a certain amount of work. With SaaS tools, I can do a bit more work than if I did tasks manually; there's a bit of automation, a bit of help. So SaaS was built and priced per seat most of the time. As a company, I get value by giving a SaaS tool to most of my collaborators, because each collaborator who uses it generates more output than without it. So the value tracks headcount for the SaaS company. Now, if you have an AI agent, it can do the work of several people. If you price per seat, the agent uses one seat, but with that one seat it does the work of five people. Suddenly you've decreased the value of your tool, because instead of a customer paying for five seats, they have one seat with one agent. Your revenue goes down, and you can't just increase the price, saying "your agents are on the seats", because some people will use humans, some will use agents, and you may not be able to control whether there's a human or an agent in front of the tool. So maybe what we need to understand is that SaaS is software spend: you compete in the software budget. With agents, maybe you compete for headcount, in the consulting or BPO space, which are much larger budgets. So what you should price is not per seat, but by how many people you replace in the organisation. Instead of $50 or $100 per seat per month, you charge a fraction of an employee's salary per month. If somebody costs €4,000 a month and you charge 20–25% of that, you're at about $1,000 a month. Suddenly you extract much more value from your tool, just because you understand that the consumption is different from the usage. I think a lot of questions will arise in the next 12 to 24 months on how we price AI, but it's something you need to actively think about before you just allow AI agents to use your tool while charging per seat. Capital allocation: how do we direct product and engineering resources? I did a workshop last year on this; we could speak for eight hours about it. The failure mode here is a roadmap driven by demand, by the loudest voices. Protecting ROI was a big theme last year and it's still big. We tend to react to market signals. But you can build a lot of things the market wants, and if they're not willing to pay and you cannot tie it to revenue, at the end of the day you're spending a lot of very expensive people building things that don't create value for the company. At some point you're EBITDA negative, your cash is burning, people say "we need to raise money again", or you do layoffs to extend the runway. So the simple sentence here is: state very clearly to your company where you allocate your capital — which categories of bets you make. They need to be linked to your market thesis and to the way you want to capture money. And also state explicitly what is deprioritised. I think all of us fail here: we say very clearly "we're going to do this", but because we don't say "we're not going to do these other things", those things come back into the roadmap, and that's where the drift happens. You can think about your investments as three portfolios. The first is core: you protect what works. That's your core product today; it makes money, people use it, and you need to keep investing and innovating so you don't become irrelevant compared to the competition. Then there is growth: you have a customer segment with some needs. What is the next adjacent segment, or the next adjacent need you can address for these people? That's growth. **Participant:** Quick question, more of a comment. You mentioned, in one sentence, the differentiation between prioritised and deprioritised. How do you think about the reality that a lot of work is actually on deprioritised things? People say "yes, deprioritised", but somehow we still have to do it anyway. **Florian:** It's not easy. If I could just say it and it worked, it would be easy. For example, we are B2B with very large customers, and of course some of them represent a very large share of revenue. When they come with a demand, it may not be aligned with where we should spend the money based on this framework. So I think it takes discipline at the leadership level first. But especially in B2B, you will never escape building some things that generate revenue in the short term. You'll do it. As product leaders, what you need to ensure is: if we build this for this customer, is there a path for it to become something that serves a larger customer segment, close enough to what we do today? If that customer is in your domain, they should be very close to your market thesis. And if they're not, it's a good moment to ask: should they even be our customers? Maybe we should stop serving them, and maybe at some point they churn. But if we focus on our core segment, we should capture more revenue than we lose. It's not easy, and I don't have the magic answer; I wish I did. But at least at the leadership level, restate these things every quarter. So: core, growth, and then explore — optionality. If you know the Horizons framework, it's similar: Horizon 1, Horizon 2, Horizon 3. Revenue today, revenue tomorrow, revenue at some point in the future. A good example is Datadog. Datadog tailored itself to DevOps, doing observability and infrastructure monitoring, and they chose to capture value through usage-based pricing, because they handle a lot of events. For them it made sense: as companies grow, they monitor more events, so it was a natural way to grow and it expanded naturally with customers. If the company grows, or the customer puts the system on a bigger share of their infrastructure, Datadog grows with them. We call this land and expand: you land in a subset of the business and then you're deployed everywhere. They were very successful; at IPO filing their NRR was above 146%, which I think everybody would like to have. What's interesting is how they grew their product portfolio. They started with infrastructure monitoring, the core of their product. Then they went into log management, the adjacent use case. Then APM and RUM, then security, and now of course AI and LLM observability. So they solve one problem very well, generate money from it, and once it's become part of the core, they go to the next. They focus a lot of their resources there, and meanwhile dedicate very few resources to explore, to try to anticipate what's coming. This is the hardest part of product, I think: understanding how much to dedicate to each, and how to move things from explore to growth and from growth to core. Some people move things too fast from explore to core. Then they dedicate a lot of resources, because it's in production, you have to maintain it, and it becomes an expectation from your customers — but maybe you weren't ready, or the market wasn't ready. Commercial integration. This will feel very B2B, but in B2C it's the same problem. Many times we build a product, and after the fact we go to the people generating revenue and say: "Hey, we have a new product, we need to sell it." That's how you fail, because selling a product needs to be thought through completely: what's the value proposition, how will it be used? The one-sentence test here is: "We take in commercial signals through a specific mechanism." How do you get feedback from the market, and how do you treat it? In B2B it probably comes from your sales team or account management team. How do you do that at scale? How do you actually treat the signal? Because the sales team will always come and say, "This customer needs this, it's very important, it's $10 million in revenue and everybody wants it." Now give me concrete signals about that. You need to be very specific, otherwise you'll start building things that aren't important. Then you need a process to treat the signals and a process to prioritise, to make sure what you build is aligned with a demand you can capture. What I keep telling the PMs I manage or work with is: when you start building something, you should already know and state very clearly the list of customers you will go after, how you will go after them, and how you know they'll generate revenue, so that when it ships, the time to revenue is very short, at least at the beginning. Then you can expand to the rest of the segment you want to target, but you should have a first share of that segment identified, know how to target them, and they're just waiting for this feature. Then the distribution channel is also very important. How you deliver your product matters most. For example, Stripe chose to deliver their product via API and target developers. They didn't start by developing fancy interfaces or an app, because it didn't make sense for their product. At Veriff, for example, we deliver our product through sales teams, and we also have an SMB self-serve segment, like SaaS, with pricing plans and so on. We decided to stop caring about that side of the business, because we really want to focus on enterprise customers first. So the choices we make are: we don't need to explain every detail of all our products on our website, because we know there will be human interaction; people will talk to a sales team. What we need to do is equip the sales team with the right information and assets to share with enterprise customers. We also understand that in our segment, the person who buys, the person who implements and the person who uses are different, so we need to decouple that content, and understand who uses the product at each stage to make it successful. These questions are very important, because you can build the best product, but if it goes through a distribution channel it wasn't built for, you'll be unsuccessful. The example I give here is Datadog again. Sorry, I don't even remember my password — this is before launch, that's why. Datadog is a bit like Stripe: they went through DevOps self-serve. You install Datadog and use it, and then the product signals to the sales team: here is a company that's growing and starting to use a lot of our product. Can we sell them more services, better customer support, consulting services? Do they need bigger packages? Because now there's a product portfolio to sell and they only use a share of it. Then the sales team goes to the customer's procurement team, and that's the next cycle. Most new customers come in that way and then over time grow into a much larger customer base; and of course, as Datadog grew, the sales team also went directly to some customers. But if you don't understand this loop — where your product will surface, how it will be purchased, what the mechanisms for installation are — you can have the most beautiful product and nobody will see it or use it. And the last part is compounding loops. The loop framework has been used a lot before and was popularised a lot. But if you don't build compounding loops, your revenue will grow linearly with all the effort you put in. Especially nowadays, if you want to be successful, you want every effort you put into your product to compound with everything you did before. So you want to understand your system over time and be very specific about designing these loops. You need to state them, make sure they work, and understand what you need to do to generate them for your customers or through your internal systems. I've given two loops here. The first one is Veriff. At Veriff we do identity verification: if you open a bank account and have to scan your face and your ID, that's the type of solution we provide, with biometric authentication and other things. To do that, we analyse a large set of data with machine learning models and identify what is good, what is not, what is fraud. So the more customers we have, the more use cases we see, the more verifications we do, and the more good and bad data we have access to. That means we can train our models much better and increase our detection accuracy. If we increase accuracy, we can reduce friction for the end user, because we need fewer signals to make a decision. At the beginning we asked for a lot of signals; then we understood that maybe we only need two or three to make the same decision that previously needed ten. So we reduce friction for the end user, which means higher conversion for our customers, because we usually operate in onboarding: if it takes long, the end user might drop, which means less conversion for our customers. With higher conversion, we get more customers, because we can tell them: with the competition you will lose business, and with us you'll protect it. And then — sorry, I have fat fingers — the more customers we have, we go back into the loop. That's the loop we are actively designing, and with every new product we build we ask how it fits into it. For some new products we say it doesn't fit this loop: should we build this product, or how can we tie it to our existing products so it gives us signals that reinforce the loop? At Typeform, the referral loop was very important. If you've ever used Typeform or any form builder: the more customers you have, the more use cases you're exposed to and the more forms are out there. The more forms are out there, the more respondents you have — people answering the forms. And the trick was to put a "made with Typeform" button on them. So the more people used the product, the more people clicked on "made with Typeform", saying "What is this? I'm interested." So you get more sign-ups, and more conversion, because you have more use cases covered by your templates and features, so you can serve more of your customers' needs. That's how Typeform grew. At some point this loop became obsolete, because the people coming in from that channel were not people who could become customers. But at the beginning it helped the company book something like 50 to 70 million dollars. So these are examples of loops. There are others you can see here: retention loops — can you decrease the cost to serve with every new customer? Do you have a data advantage? Veriff is a data advantage: the more customers I have, the more data I have, and with this data I can build a better product. That's not only a loop, it's also a moat, because my competition will have a hard time replicating it. And then reference density: referral loops in B2C are references in the B2B world. Again, if I have Uber, getting Lyft will be much, much easier. So this is very important. **Organiser:** Just a quick note: apparently this is a really successful deep dive. We're sorry — this room is meant for 30 or 40 people and it's overbooked. Florian, will you share the slides later? **Florian:** Yes. **Organiser:** Florian is a good friend and on the board, so if you need the slides, we'll share them. And if you'd like, the other stage is starting soon as well. We're so sorry for the inconvenience that many of you are standing. Apparently next year we book a cinema for Florian. But in general I think it's just a sign that it's successful. Florian will for sure share the slides, and I recommend the book; last year we gave it out at the conference. Thank you. **Florian:** Two more minutes. We went very fast through all the sections, and what's important is that we talked about every step a bit in isolation, but they compound on each other. So once you do the exercise, my recommendation is to do a coherence step: look at all the pillars and ask whether they make sense together. If you see one that doesn't fit with what comes before or after, go back to that section and understand what you need to change. Or maybe you need to change the step before it; maybe you need to go back to the market thesis. That's very important. It's an exercise you need to do over time, because you'll focus on a first segment of customers, then hopefully become larger and start introducing more segments, and for each of them you need to redo the exercise. So we're finished. If you want to know more about all of this in detail, follow my newsletter; you get it every Monday. And please give feedback for the conference; they try to collect it for all the sessions. Questions — one in the back there, and then over there. #### Q&A **Participant:** I have a question about the market. Right now the market is shifting completely, there's disruption. I'm in a commerce space for digital creators, and when I look at the market overall, we have geopolitical things. We see GoPro just being acquired by a military company because it has thousands of patents. We see Airtable, Miro, Meetup.com, Eventbrite — they have been acquired, but back in the day they had insane valuations. And we see OpenAI, which was super bullish and horizontal: once they acquired Jony Ive's company for 2 billion, they wanted to build new devices, and now they're becoming more vertical. From your personal experience, how do you as a product person navigate this wildness of the market, from vertical to horizontal? What keeps you in a sane mood and not going crazy, jumping from one initiative to another? Personally, I still like doing classic research on what competitors are doing — for me Shopify, because they started building AI-native things, or Intercom, which started pricing based on value. What's your opinion right now on going vertical? **Florian:** I think focus is the answer. I showed this because it's the same question as the market thesis. When you start to build a product and get a bit of traction, you suddenly feel like a genius, and you say, "We could also solve all these other problems," and you spread out. The issue is that you'll grow a lot, maybe for a long time, because you have a first segment that generates revenue. Then you start tailoring a bit to another segment, and you have early adopters, the curious ones, who pick up the product and keep using it. So revenue starts to compound and you feel like you're off to the races. VCs look at this, see very nice growth, fund you, and you get a crazy valuation. At some point the growth stops, because you don't tailor precisely to the needs of certain segments, and you run out of early adopters. Then you need to develop a product that really solves end-to-end use cases. The only way to solve this is — well, if you have founders, don't always listen to them. But it's very hard to compete with "we have this segment where we can make 10 million tomorrow" versus "this other segment where maybe we make 5 million tomorrow, but we can make 50 million in three or four years". The time preference unfortunately skews short. That's why we need to do this exercise and show people: based on what you're telling us, if we go into this next segment and want to do it properly, we need to build a very different product and put our resources into a very different space, and that's not manageable for the company. But it's really hard. It's very tempting to add one more customer segment, because you feel you'll get all of them and two is better than one. That's the mistake most companies make. The problem is that it takes time to realise you made that mistake. Maybe with AI that will shorten, but it's still two or three years before you realise you should reinvest in the first segment and maybe let go of another one. **Participant:** But would you still use something like the Horizons framework, and still believe in diversifying yourself with something to explore, something to innovate? **Florian:** I think the loop is becoming shorter. Horizon 3 used to be maybe 15 years in the old world; for sure now it's less. It depends on how much disruption the technology can bring to your market. You can still do core work with AI; it's still the bulk of your product. So I think you have to separate what enhances your current product from what will really be a new business line. A new business line is Horizon 3, explore, and it should take time, because you need to redo the work of understanding the customers and their problems. You can go faster now, but still, if somebody says "I start Horizon 3 today and ship in six months" — yeah. **Participant:** Thanks for the presentation. My core product is going to be the same, because it's an asset-based product. As a growth product manager, how should I tune my approach for mature markets, markets that are yet to mature, and new markets? **Florian:** It depends on your context. When you say assets, what do you mean — features? **Participant:** I can give you details. I'm in a growth team for car sharing, so the asset, the value, is the same across markets. **Florian:** So for your core market, how do you increase the value? How do you expand the value across all markets? How do you ensure that what you build is sustainable for the first product you have? Say: core for your core market, and growth and explore for emerging markets. At the end of the day, is the playbook you had for this emerging market the same one you'll use for the next market after it? Then you don't need to do anything; it's just about understanding when to deploy your resources from a go-to-market point of view. If it's a different market, you need to do the whole exercise once again and understand: do we have the resources? Is it even worth it? Because every resource you deploy to a new market should also get more value out of the product you already have. **Andreea (moderator):** Do you want to stay in Berlin for the next two or three months? **Florian:** For the conference? I'll let you talk to my wife, my daughter and my cats, and I think all of them will disagree. **Andreea (moderator):** Are you staying today, though? **Florian:** I'm staying today. **Andreea (moderator):** So this is your cue to bombard him with questions. And believe it or not, you actually still have five minutes. You were super fast; you told me to time you and you didn't even look at me. Okay folks, thank you so much. I think we're hungry. Is anyone else hungry, or is it just me? Thank God. **Florian:** Thanks, everybody. ### 15:00 · For product builders: Let('s) build. — Vito Bica - **Speakers:** - [Vito Bica](https://www.linkedin.com/in/bica/) — Founder & CPTO, Arcate. 20 years of B2B product consulting (Endress+Hauser, KSB, Klenico, ZAGENO); built Arcate, an AI engine that ranks product bets by revenue at risk. - **Stage:** Doctolib Stage · CIC Berlin · 17 Sept 2026, 15:00 CEST - **Format:** Talk · Track: Build & Craft - **Host:** [Veronika Morozová](https://www.linkedin.com/in/veronika-morozov%C3%A1/) (Contentsquare) - **Official abstract:** Most people are talking about the product builder. Vito has been doing it for 20 years. This is a field report from someone who removed his own training wheels and shipped a full SaaS product alone with AI — and is still falling at parts of it. **Summary.** Vito Bica, an independent product builder, presents "product builder" as a mindset and an operating model rather than a role: agency plus learning the craft of different roles, a cross-functional operating scope, and a decision chain from signal to decision. He illustrates it with three personal case studies spanning hardware (a mosquito trap), a clinical mental-health diagnostic product, and his own AI-based discovery tool, followed by questions on automation, go-to-market and scale. **Key takeaways** - AI expands what one person can execute, but it does not remove the need to make assumptions visible or to assign decision ownership; a tool doesn't make you a product builder. - His model has two parts: a decision chain (signal → evidence → inference → assumption → test → decision, later extended with outcome, learning and scale) and an operating scope that spans domain, user context, engineering, data and commercial. - Learning the domain deeply comes first: mosquito wingbeat frequencies and life tables, ICD-10 diagnostic criteria, then encode the domain's constraints into the build and validate against reality. - In the clinical case, replacing a three-hour ICD-10 assessment with a weighted 20-question screener in plain language cut it to about 20 minutes and reduced patient burden. - Agency means taking responsibility for the whole chain even when a project fails or never reaches production; finish, validate, then keep or kill. - He does not run a fully automated pipeline: he doesn't trust first outputs and treats coding as part of knowledge building; the approach needs at least 10–20 customer conversations as input. #### Transcript **Host (Veronika Morozová):** The first session on our Doctolib stage is about building AI products, and I have to say Vito is a real builder. When I was preparing, I wrote to him on LinkedIn and asked whether there was any story about him I could use to welcome him on stage. He said that when he was nine years old, his father ordered an '80s hi-fi stereo system with seven modules and a VHS recorder, and went to work, because he wanted to assemble it himself afterwards, following the manual. But when his father got home from work, everything was built and Vito was watching his favourite cartoons. His father asked, "Did you at least read the manual?" And he said, "No." That describes Vito. So welcome to the stage. **Vito:** Thank you very much. I'm a bit overwhelmed, so many people here with me. I think it's my first speaker gig in a long time. I brought my daughter's kids' bike here, the little one, to remind myself that every day for me is day one. Day one means that being a product builder, for me, is in particular about delving into new domains, new industries, new topics every couple of months. The reason I function this way is that I'm super curious; I'm driven by curiosity. When I was a kid, every day there was something new to explore, and I just tried to figure out how things work. That's why I brought the kids' bike today. I think the clicker should work. The first two screens are a bit of reflection, because I threw away 60 slides before I got here. I thought: if I talk about product building today, I won't talk that much about AI product building, because being a product builder is a mindset. If I came here and talked about building products with AI, you could just run a search on Perplexity, learn what it means to be a product builder and try to mimic it. But you wouldn't really learn it, because if you run a couple of prompts, it will say the same thing every time: get to know the customer problem, form some hypotheses, do some vibe coding and hand something over to engineering. That's not being a product builder. Being a product builder means much more, at least for me. Then, in August, there was a description of my deep dive on the web page, and I was forwarded some feedback on it. I was in Sicily, sitting there, and thought: maybe this guy doesn't know me. What really caught me was this line: "It feels like a very self-centred presentation." That made me think. If I do a self-centred presentation, you'll think, "Ah, he's talking about his portfolio, about his experience." But I think it's more about something else. At the same time in August — my second passion is cooking. This is a recipe from Palermo, *tenerumi*, the young shoots of the snake gourd, a gourd from Sicily that looks like a snake. It's an old recipe, and I've worked on it for about 15 years until it got to this, let's say, Michelin-star presentation quality. And my father said: "Good balance." And I thought: yes, it's the same as what that person said about self-centred. Here there is heritage, legacy, tradition, being curious about building recipes and making them better all the time. It takes a long time until a recipe gets to the point where you say, "Okay, now it makes sense." That's why I built my presentation a bit more self-centred: because it's my way of thinking, the way I grew up as a child. In 1998 — around the same time as the hi-fi — I asked my father's chef for help with electronics. Downstairs there was a restaurant, and upstairs we lived. He asked, "What do you want to do?" I said I wanted to build a one-way FM transmitter, because of this radio from the '80s, a Telefunken Partner universal radio, that was standing in the kitchen. So I built this one-way FM transmitter with him, and I tuned the radio to its frequency. I waited for my mom to come into the kitchen while I was sitting in my little room, and I tried to scare my mom through the radio. It didn't work, because she heard it was coming through the radio and just turned it off. The job to be done was clear. The customer problem was also clear. It was a one-shot product; it never went into production. Being a product builder was always possible for some people, and today it's possible to be this type of unicorn because AI changes accessibility. You can run everything through AI and think you can be a product builder. But it's not a tool that will make you a product builder; it's the way you think. AI can expand what one person can execute, but it does not remove the need to make assumptions visible or to assign decision ownership. I'm not advocating for the role, because maybe it's not for everyone. From my experience with companies, I seldom work in squads or big teams. I get my projects, even from companies that make $4 billion a year, and I steer them myself: no distractions, no big meetings. I pick the people I need to talk to, but I run it myself. I'm here today to describe the operating model behind that as exactly as I can. I had a talk with Andreea, who was moderating here earlier. The fun thing about preparing this deep dive was figuring out what feels natural to me about being a product builder; it took me a long time to understand my frameworks and how I organise my work. It takes two components. There is a decision chain that logically passes from signal to decision — basically my product management model to get from signal to decision. And then there is an operating scope, which means the cross-functional part, what you work across today. So it's a different way to see your own work: not three roles or multiple roles, but one mode — not a role, sorry, a working model. I'll show three case studies to prove it holds up across hardware, software and AI. And let's be honest: what you do with it is up to you. I can only tell you how it works for me; maybe it doesn't fit you. The foundation is maybe not only agency, but agency is one of it: the ability to direct your own actions towards goals and take responsibility for the outcome. The second is that if you can understand and learn the craft of different roles, you can delve into any new domain and any new topic with little effort. It makes you move very fast into new areas; I'll explain a bit later. This is from the movie, but I like it very much because the title says everything: *Everything Everywhere All at Once*. That's the operating model. Because you cannot delegate work to others — you can, but most of the time you're organising the context of what you're doing. That's where AI comes into context. You try to organise everything that needs to be done to move forward, but you have to do the organising. You have to know which step is necessary now, what you learned and how to apply it, and then move to the next area. So it's a lot of back and forth. The operating scope is that I have fields, and then you have loops, and you have the trajectory from discover to test. And you have to cover the domain, user context, engineering and tech, data logic, and even commercial. In particular, something I love very much is the domain: learn the domain's logic — it could be clinical criteria or an industrial process — then translate it into domain rules, encode the domain's constraints directly into the build, and validate against the domain's reality. This is something I really love to do at the beginning, because here you create all your distance from the others. Then you can slowly move forward and add new things. And then the decision chain is the model I use in product management. You have raw inputs: customer, market data, usage. Then you have graded evidence, with its source, context and commercial relevance. Then inference, which means what we think the evidence means. Then the assumption, and a score for the business. Then you have to test it, and then you have to take a decision and own the decision. Now it may get a bit nerdy. I will not show the super success stories, but rather projects that maybe didn't go into production, because I don't want to make a hero out of the product. I think it was 2017, I was in Sicily, and mosquitoes love me very much, but I don't love them. They have been attacking me since I was a child. At that point I thought I have to do something against it. So I started researching — pre-AI, on the phone — what systems exist to act against mosquitoes at home, what tools or traps. I saw there are a lot of traps out there, but they don't really work. It's just blinking lights, not based on science. So I started diving into the topic, and I had to learn how mosquitoes work. For example, to monitor the type of mosquito: they have very specific wingbeat frequencies that differ from species to species. So I built an Arduino sensor system that could identify the species. This one was just a simple *Culex* — *Culex* is the common European mosquito. Then I started building mosquito life tables to understand how mosquitoes work. It's important because it's taxonomy: the whole taxonomy is in there, the names, but also the wingbeat frequencies. Then how they feed: are they daytime biters, twilight biters, night biters? Daytime biters are active only in the morning; twilight biters are in the shade at twilight. Then the circadian dynamics. And also how long it takes for an egg to hatch and how much warmth energy they need from winter to summer for the eggs to hatch. Sometimes they need about 20 degrees over a period of 30 days, so the warmth energy must accumulate over that time and then they hatch. So I could build this table, match it against weather data and forecast when the season begins. That's all about learning the domain. To test whether there was willingness to pay: I knew Miami has a huge issue with mosquitoes, and I knew when the season would start. So two weeks before, I set up a Facebook campaign to promote the product. I had a landing page, and within a couple of hours I got 50 requests to buy it. Then I started ideating. This is a trap concept from the CDC in the US from the '60s. It's like a vacuum: they get sucked into it. And there's an attractant called octenol, which comes from our feet and our breath, and mosquitoes are attracted by it. There's octenol here at the bottom; it emanates because of the suction, which spreads the octenol. The mosquito comes, sees a black shape, goes in. And here is the microphone to detect the species. Then we went into hardware prototyping. This is the 3D model we developed with 3D printing, five iterations; we spent a lot of time building a setup that could potentially work at scale. And this is a 3D marketing asset I built. But the thing never got into production, because the investors preferred to invest in, I think, quick commerce. So agency doesn't mean you're always right. Agency means taking responsibility for the entire chain, even if it doesn't work, even if it breaks. At least you went from A to Z and could validate the idea. And this is basically what I do all the time: from discover to test, I have my counter-forces, and I try to move upwards. For example, regulation and compliance is the next case. This is a clinical company, a spin-off from ETH, that built mental health diagnostics. It was a clinical product, a medical device with a certificate. That's why regulation is a huge topic here: not only monetisation, but keeping it stable for the licence, for the certificate it has. I had to learn the ICD-10 psychiatric diagnostic criteria. I knew them a bit, because I had depression while I was building the mosquito trap. Then I had to rewrite the clinical questions for patient comprehension, encode the redesigned logic into the assessment engine, validate the automatic reports against clinical ICD-10 standards, and so on. The user context was that the whole assessment lasted three hours. Three hours — sorry, I get a bit emotional — because this is the ICD-10 library; it's very complex, many symptoms have to match, so the assessment is very complicated. I had to learn the whole catalogue and understand how the assessment should be built. And — sorry, emotional — I had the most harrowing interviews. There were people who were really destroyed. I said to the founders: it's not possible that someone with serious mental health issues has to do a three-hour assessment. This was the original clinical model: you had to go from F20 to F99 and answer every question, and it took three hours to complete. An immense cognitive load, a massive list of drop-offs. It was precise, but it was heavy. So in the new model — I love data, I love building context and making it better — we built a screener with 20 questions. It's weighted, it has factoring in it, and with 20 questions you could identify the issues, from depression to schizophrenia to everything else, and give a hint whether there are comorbidities or more issues, so that after the screener people could have a personal one-to-one session. We cut the assessment from three hours to 20 minutes. Also the language: the assessment was written in clinical language. In German there's also *Leichte Sprache*, simple language. This was pre-AI, so we used a simple-German language tool and rebuilt all the *Aussagen*, the statements. Instead of asking questions, we gave statements, and we introduced a semantic differential scale, which is very easy to answer. When we were ready, the clinical product was done, and we thought: how can we test it? Sorry, because Claudia was talking about Python — we built the entire MVP in Python, and it worked. Then we attached a payment step to it to see if it works for B2C. They have since pivoted from B2C; now it's B2B2C. The product is still the same. Now case study number three is Arcate. It's my own product, because I did a lot of discovery in the last couple of years. One of the biggest projects was for Endress+Hauser, sensor systems for industrial manufacturing. It was a discovery project that lasted six months. I was working alone, running everything from the customer interviews to the high-fidelity user testing at the end. It was the first discovery project I completed powered by AI; only the scoring was still like RICE and those frameworks, with the team, to identify reach, impact and so on. After a couple of months — because it was very successful: we lowered the customer effort score by 2.5 points on a five-point Likert scale, and the product outcome, increasing the self-service rate by 30%, was also reached, at least in the high-fidelity test — I thought: what if I can build the whole model for doing discovery, from signal to decision, the same way? So here too I started with the decision chain. Product needs, containers, insightful roadmaps, and also discovery tools. Sales holds the signals, product holds the roadmap, and revenue connects neither. Priorities must be set by measurable ARR at risk, in that case, because I deliver only for B2B companies. The assumption is that an AI model can continuously rank customer signals by ARR at risk as a correct authority. As a side note, not super correct: the ranking happens via AI, but the scoring happens via a data model — a very complex Fermi formula. I did many simulations on data I had; I asked companies to give me anonymised data, and the Fermi formula is still beating senior PM judgement. And it's a pure formula. The test was simulations across several hundred signals, comparing its ranking against a senior PM using Kendall's tau. The target was above 0.85, the benchmark was the senior PM, and now it reached 0.94, which is basically the same. Ah, no — 0.24, sorry. So I committed to building Arcate. Yesterday, during the workshop, I introduced a model for getting from signal to decision, or to MVP, and I included a simplified version of this case, only for the scoring. I was happy that the people joining the workshop somehow validated the scoring. So no pitch today. But finally, here's the AI case. I didn't want to pitch my product, only to explain that everything I did before AI as a product builder now also works with AI. I built my whole context: my research, the domain knowledge, the positioning, the GTM strategy, the pricing models, all the interviews, the synthesis — everything now runs in my file system for this project. I build up the knowledge and move forward on the chessboard one step at a time, always thinking about everything that's interconnected. After three months I started building a UI-based SaaS. Then, after a few months, I realised I was always working in my IDE, so I stopped working on the UI, and now the whole software runs as MCP tooling. So yes, the correlation with expert ranking is exactly the result. This is my nerdiness about building technical, data-driven products. But the business outcome still needs to be proven. I'm now starting to see whether people will buy the product. Otherwise I'll just give it away. It was an experience. My grandmother had a lot of sayings, and one is: learn a craft. You never know when you'll need it, but learn it and put it in your pocket. That's being a product builder: learn the craft, put it in your pocket, and you'll have it when you need it someday. These three cases don't prove a universal method, because they're just a snapshot. But they show that the same pattern and the same operating logic can work across different contexts. So the agency rules apply. Never wait to be asked: hardware and field prototypes, a reworked assessment model, a data product, not just an analysis. See problems before they're obvious: connecting detection with forecasting in the physical trap; early on seeing patient burden, not physicians, as the bottleneck; seeing evidence gaps behind unprioritised roadmaps. Guard your attention — these are my rules as a product builder — so don't get distracted; focus on what's needed. Health, biology, hardware — put the product together; clinical logic, patient persona, language, assessments — put together. Every topic needed to build a product, I keep in my mind or in my context. Stay curious: there are new people, new topics and new things all the time. Finish things: even if it doesn't work yet, bring it to the end so it gets validated, and then you can keep it or kill it. And the complete decision chain: I started earlier with signal, evidence, inference, assumption, test and decision. But later — much later, in 2020 — I learned that outcome, learning and scale are part of it. Before, I was sometimes just building for the sake of it, because it was just brilliant. The popular narrative says the faster we can build, the more expensive judgement becomes. And it also says a product builder is a full-stack product manager. No: a product builder holds the complete context of a product and is driven by the desire to act right now, without waiting for someone else to move first. Nothing left to say. No playbook. Now you can shoot your questions — we have half an hour. #### Q&A **Host (Veronika Morozová):** If you have questions, raise your hand. **Participant:** Are you thinking about a next step where this feedback goes automatically into coding on its own? Do you see the technical possibilities already there to automate end-to-end discovery and delivery? **Vito:** Pre-AI, I was coding from day one. I had data, I built up data and started doing little tests: a function, something I wanted to realise. Then I added engineering chunks to the rapid prototype — back in the day it was called rapid prototyping. I created small chunks of engineering, coding, tested something and put it aside. Nowadays, yes: I think about something, I get new insights, and then I generate code. But it's not automated yet. I don't trust the first output, so I iterate on a lot of things. I don't have an agentic pipeline for my work, because I need time to reflect. But I build up knowledge, and engineering and coding are part of knowledge building, because they prove certain things. **Host (Veronika Morozová):** Another question. **Participant:** We talk a lot about building. It's amazing, accelerated; we all get smarter and get to do other people's jobs very nicely. But I think building was not always the biggest blocker. You're a very independent, autonomous person who builds the whole thing. How does the other side change — going to market, convincing customers, or learning and finding things out with customers, and maybe investors? **Vito:** I don't know. But what I do is build the knowledge. I talk to customers, I do the discovery calls, and I get the transcripts automatically now: from Demodesk I get the transcripts via MCP, and they go into my context. From there I do the synthesis, because I have the questions and know how they should be synthesised. Then I try to understand what the customers need, why they're willing to pay, and all that. And then I hand over the knowledge to the respective departments — what they need to know from me. **Participant:** At what scale does this work? **Vito:** I didn't include it here, but it scaled at an enterprise company doing four and a half billion dollars a year. **Participant:** I mean the other end: where does it start to be efficient? When you have a couple of dozen inputs or feedback from different angles, or hundreds? Just to get an idea of where this could be applied. **Vito:** It doesn't work if you don't have any customer data. You need to have talked to at least 10 or 20 customers. **Participant:** Correct, so that's the scale. **Vito:** Yes. For my own product, Arcate, it started before I talked to customers, but I was eating my own dog food. I was the customer, and I knew exactly what really annoys me. That was my motivation. But I had no customer data. **Participant:** Super, thank you. **Host (Veronika Morozová):** Another question? Okay, then you have more time for coffee. Thank you for joining, and we'll see each other here at 4:40 if you want to join the next session. Thank you very much. ### 16:40 · Start the Hard Talk — Alexandru Bleau - **Speakers:** - [Alexandru Bleau](https://www.linkedin.com/in/alexandrubleau/) — Senior Innovation Manager, E.ON Group Innovation. 15+ years building digital and physical products in gaming, travel, advertising, blockchain and energy; facilitates hands-on LEGO® Serious Play® sessions. - **Stage:** Doctolib Stage · CIC Berlin · 17 Sept 2026, 16:40 CEST - **Format:** Workshop (LEGO® Serious Play®) · Track: Leadership - **Host:** [Veronika Morozová](https://www.linkedin.com/in/veronika-morozov%C3%A1/) (Contentsquare) - **Official abstract:** Everyone avoids a conversation. In this hands-on LEGO® Serious Play® session, you’ll build it, uncover what’s underneath, and decide what to do. No slides. All action. **Summary.** Alexandru Bleau, Senior Innovation Manager at E.ON Innovation and a product coach and LEGO® Serious Play® facilitator, runs a hands-on workshop on the conversations we avoid. Participants build models of the conversations they are avoiding, reflect in writing, share at their tables with curiosity questions, and finish by committing to a concrete next action. **Key takeaways** - Making your thinking explicit, by building it and looking at it from a third-person view, helps you see a problem differently than reflecting only in your head. - The workshop sequence: build three to five avoided conversations, choose the one with the most charge, summarise it in one sentence (who, why now), then model the conversation itself with its people, emotions, blockers and what you gain if it goes well. - Sharing with one "pure curiosity" question per listener surfaces new perspectives and assumptions. - Participants noticed that the hardest conversations are often the ones with ourselves, and that we tend to list only the feared outcome and leave little room for good possibilities. - End with a next action: what you will do, by when (a time frame or condition), and how you'll hold yourself accountable. A journey of a thousand miles starts with a single step. #### Transcript **Host (Veronika Morozová):** This session will be more interactive. I'm curious myself to see what we're going to do today, because Alexandru is a product manager at E.ON in his day-to-day life, but he is also a LEGO facilitator. He has been a LEGO fan since childhood. When he was a little boy, he always wanted the fanciest LEGO on the market, but he couldn't afford it, so he and his brother were always building new pieces themselves. That became his mantra for life and for product management: be focused on what you want and then build upon it. Today he's going to teach you about having difficult conversations, by building LEGO. Alexandru, the stage is yours. **Alexandru:** Thanks, everyone. I'm excited to be here because I get to have fun and call it something serious. There will be a lot of serious things you choose to go through in this session: it's about building the conversation we're avoiding. It's a combination of LEGO Serious Play, the method offered by LEGO, and coaching, as I also coach product managers and product people, and I thought it makes sense to combine them. It also ties into what Daniele mentioned this morning, that we're trying analog things, so you could consider this the offline Miro board. There's something very helpful about making our thinking explicit. Usually we reflect and think in our heads, but there's something very powerful about looking at our thinking from a third-person view. That's what we'll try to do today. I want you to leave with something tangible, which is why you'll find reflection papers on your table, not just the LEGO. I encourage you to use them, write things down, and share what you're comfortable with during the sharing moments. If you're alone at a table, please join another table, as we'll have exchanges within the tables. There are some slides. Initially I said there'd be no slides; the slides will just repeat the questions and tasks visually. A bit about myself. My official title is Senior Innovation Manager at E.ON Innovation, which is part of the E.ON Group. We try to move the needle in a quite tricky environment in terms of politics, regulation and priorities. E.ON might look like a single company from the outside, but it's actually a huge group of over 70,000 people with a lot of companies. If you know Bayernwerk, Westnetz or Westenergie — they have different names, but they're all part of the E.ON Group, and they all need innovation and improvement to better serve their own customers. I also do coaching and training under the brand Product Sparks, which I've also been doing in Romania. I'm originally from Romania, despite my name sounding French; I was born with it, so nothing to change there. Mainly on product outcomes, communication and facilitation with LEGO Serious Play. And I make my own bow ties — actually I sell the straps. Buy one strap and you can build anything else; we just made a bow tie out of LEGO here. That's about me. And I wish people a crazy day, because why not? We should have a good kind of crazy. So what are we going to do today? You have your canvas, the board you'll build on, and you have the reflection paper. I would kindly ask that you don't steal LEGO parts at the end. I know it's tempting — even for me, and it's my LEGO, I'm still tempted to sneak something into my pocket. If you have questions about how to build, I'm happy to answer, and happy to address any other thoughts or comments. But mostly this is space for you to reflect and think, and time for you to build and put those thoughts out. We'll switch between building and reflecting, between the models and the paper; you'll see it on the slides. I encourage you to take photos of your models, and I'd like to ask whether anyone objects to me taking photos of you while we build. If anyone disagrees, I won't take any photos, and there's absolutely no shame or harm done. But I think you'll have awesome builds, and I'd love to take some photos. All right, great. With that, let's kick it off and jump straight into building. I want you to build a model with three to five things you've been avoiding discussing with someone. I don't know who that someone is for you; it could be yourself. It might be a raise, an issue with one of your co-workers, something with one of your stakeholders, or with yourself — I don't know, quitting that donut at midnight, or taking a step, or sleep. Put your thoughts on the table. You can split your plate space in any way you want. And "quantity" refers not to the number of conversations, but to the number of bricks: just use the bricks to express yourself. The timer on screen won't work, so I'll set a timer on my phone. *[Building round 1: three to five avoided conversations.]* All right, time is up. That was some quick thinking; six or seven minutes might not be enough, but we have our topics. Now we jump into reflection with that model. Choose the one with the most charge for you and use the sheet. Reflection one has three questions. If you have several topics, choose one; if you have one, that part is simple. Answer those questions for yourself. What's the conversation about, in one sentence? That's actually important: try to synthesise for yourself what the issue is with that conversation. Who is it with? And why exactly this one, why now — why does your mind say this topic, at this point in time? You have four minutes for this one; I can give you another minute if you need. *[Reflection 1.]* Two minutes remaining. Who needs more time? Everyone ready? Then we jump again. There will be more reflection, but now we're building again. You have up to 12 minutes to build a model of the conversation you chose. You might have chosen the topic, or you might already have gone into the conversation itself. What I want you to do now is think of the conversation as if you were having it right now, and build a model of that conversation. Put the people inside, put your emotions inside, put the blockages inside — everything you think is part of that conversation, particularly what's blocking you, but also what you get if you have it. You can remove all the parts or just modify them. *[Building round 2: modelling the chosen conversation.]* Oh, you're done? That was the conversation — a full conversation with all the blockers, what's stopping you, and what you get if you have it successfully. Two more minutes. While we wait, one more question: when did I change my bow tie? Some time before this session — almost half a day before. I once moderated a whole event for my company, and I switched bow ties five or six times, sometimes during the moderation itself; I just turned around. So weird can also be fun. All right, we have our conversations, with our emotions, potential successes or failures, whatever surfaced when looking at it. Now I'd like you to share. This is also a powerful step. You don't need to use names; share only what you want to share. We'll share at the tables with each other. Since we finished the building part quite early, we have some extra time, so maybe it's not just one minute per person. Everybody shares their model, and then somebody from the table asks one question. Try to ask a question of pure curiosity. You're exploring new land: this conversation somebody else is struggling with. So one question, and then you switch. Once everyone has gone through this ritual — I present my model, somebody asks one question — if there's time, you can do another round with open questions. Somebody from each table goes first, or I'll nominate. **Participant:** So we're talking about the model, or…? **Alexandru:** Yes — the conversation. You share your model of the conversation. **Participant:** But are we explaining the model or the conversation? I could talk about the model, the theory, or I could just talk about the conversation. **Alexandru:** You're talking about the conversation, using your model. Thanks for clarifying. So again: you share your model, very short, half a minute to a minute. Present your conversation through the model, then someone asks a question, and you switch to the next person. Is that clear for everyone? **Participants:** Yes. *[Table sharing rounds.]* **Alexandru:** Can I have your attention? If I raise my hand, everyone raises their hand — that way we get quiet. Awesome stuff. Lots of discussion, and clearly a minute and a half per person isn't enough, which is great. It's always a session that runs out of time, and I kind of always want it to. Now we switch to reflection. We've discussed a lot and got questions that might have shifted your perspective in new or interesting ways, or something someone else asked or presented might have given you an aha moment for your own conversation. Write that down. You don't necessarily need to answer all the questions; they're there to help you put into words what's stopping you and what assumptions you're making. Not everything we assume is actually true or likely, but it's good to write it down. Three minutes to write things down. *[Reflection 2.]* Two minutes remaining. Thirty seconds. All right, we know what's showing up in the room. Can anyone share an aha, something that came up for you? Yes. **Participant:** At our table, the main topic was uncertainty, independent of the question: career uncertainty, uncertain choices. Do I go here, do I go there? **Participant:** I have two insights. One for myself: at the very beginning, when you said we should think about conversations we've been avoiding, I had a very hard time thinking of any. I thought either I'm very clear and ask all the questions, or maybe there's some self-reflection needed. And I realised that these are actually conversations I need to have with myself. With others, I don't have issues bringing up what I have to say, but the difficult ones are the conversations with ourselves. And it seems to be that way for other people too, at least in a lot of cases. So, just an insight. **Alexandru:** Conversations with ourselves tend to be the hardest, probably also because we spend a lot of time with ourselves, we ask ourselves a lot of questions, but we also leave a lot of questions unanswered. Wonderful, thank you. Anyone else? Yes. **Participant:** I noticed something about how I answered the questions in sequence. The second question is "What are you assuming will happen?", and I only listed one thing. Then at the third question, "What are you afraid will happen?", I noticed I had already put the one thing I was afraid of under the second question. There could have been other assumptions, but I directly put the only bad outcome as my assumption, and that stopped me from even thinking of other possibilities. That was a pattern I noticed while answering one by one. **Alexandru:** Maybe one more. Anyone want to share? **Participant:** I think mine is similar. My brain comes up first with the not-so-good options, the ones I'd need to deal with afterwards, because if there's a problem, I need to solve it, so I need to know up front. My brain is mostly full of that part. But maybe the other options will happen as well. I'm not giving equal space to the good outcomes and to the things I'm afraid of, as options. **Alexandru:** We humans are very fickle beings. One thing that's very amusing is that we usually talk about bad things five times more than about good things, when in theory it should be equal. We even have a word for talking about bad things: gossip. What's the word for the good things? We don't have one. That's a lot about our culture. **Participant:** One last sentence — I recently read: "If you overthink, overthink the really good things." **Alexandru:** I agree. All right, the last part, but certainly not the least: your next action. I don't want you to just walk away with the paper; let's do something about it. You don't need to overwhelm yourself. The journey of a thousand miles starts with a single step. So think about what you want to do next. What will you do? By when will you do it — set a time frame or a condition. I heard of the postman who does 20 miles a day, every day, regardless of the conditions — whether it's snowing or the sun is shining, rain or sleet or snow, the man's got to go. And very important: how will you hold yourself accountable? Usually we're not strong on self-motivation all the time. Two minutes for this one. And if you haven't yet, don't forget to take photos of your models. *[Individual next-action planning.]* All right, our time is unfortunately up for today. As I mentioned at the start, we wanted to use this as an in-person Miro: actually put thoughts onto the table, talk a bit about them, and get helpful perspectives from others at the table, using LEGO Serious Play and coaching. So I want you to think not just about the conversation you chose, but also about how this method helped you. In the end, even if you don't do LEGO Serious Play, LEGO is everywhere, and other kinds of bricks too — choose your guilty pleasure. You can do this for yourself and try it out with your thoughts on paper. So with that, thank you. Please leave the LEGO. And your feedback is really appreciated: if there's a way we can improve the session — or if it was horrible, "LEGO was great, but the facilitator…" — let us know. --- *Productlab Conference 2026 · productlab.app · Transcripts edited from machine captions, September 2026.* --- # Productlab Conference 2025: talks and learnings The first Productlab Conference took place in Berlin in September 2025. Below: an overview recap of the day, then one write-up per talk from the Productlab blog (https://productlab.app/blog). ## Productlab 2025 Conference Redux (overview) Recap by: Dave Feldman (Freelancer) Read online: https://productlab.app/blog/2025/conference-redux Dave Feldman attended the first Productlab Conf in Berlin and wrote the recap that did not fit on LinkedIn: every talk, and what stuck. Again, thanks and kudos to Daniele and the Productlab team for putting together a fantastic day—it’s not an easy feat! Note also that there were three parallel tracks: the sessions I attended and have summarized here don’t represent the whole conference. ### High-Level Thoughts AI and vibe-coding are allowing more people to build higher-fidelity prototypes. On a high-trust, psychologically safe team this can democratize idea-generation and spark great conversations. AI seems to be the catalyst for a lot of teams reevaluating their processes in ways that aren’t specific to AI. Processes have calcified and become overly “waterfall” and pedantic; this hits PMs especially hard because great product-development requires ambiguity and flexibility, which is in conflict with predictable process. For example: I’m still a believer in understanding the customer problem separately from testing solutions, which is a bit of process/framework…but those don’t have to happen in series and sometimes it’s fine for them to bleed into each other. We’re starting to get real about AI. I see this in the news and on LinkedIn, too. That’s great: all the “AI will take all our jobs and solve all our problems and destroy the world” talk is exhausting and makes us feel gaslit. Whereas, “Hey, AI can do some cool stuff and we’re figuring it out,” is energizing. Vibe-coding is not production coding — I’ve always felt that way and it was good to hear it from multiple speakers. Humans are great at stuff that AIs aren’t, and some of that stuff is needed in product management. Find things that are “un-LLM-able” and do them. ### Stefan Ostwald: The Intelligence Age Stefan talked about his vision for AI and agents in particular. He laid out some useful principles when thinking about AI product design: Design for reference, not description. I’ve talked about this as “direct manipulation” — basically people will want to manipulate artifacts or parts of artifacts with AI. Not, “Change the blue one second from the right,” but, “Here, change this one.” This is also a critical component of the project I’m doing with. Design for review, not perfection. AI makes mistakes, so build products that expect it to. Design for behavior validation, not configuration. This is in part about evals and automatic prompt optimization. ### Fabian Strunden: AI from Discovery to Delivery In this roundtable, Fabian laid out the first compelling case I’ve heard for PMs vibe-coding prototypes. His team uses them as a communication tool, acknowledging that they’re not “real” designs and simply serve to communicate an idea — it’s expected that designers and others will push back and it’s really just about the conversation. They use the hashtag #badidea to indicate these moments. This makes a lot more sense to me than what I see on LinkedIn — “PMs can stop doing everything else and just vibe-code prototypes!” — since it doesn’t imply that such prototype are sufficient…just useful. A couple things occurred to me as a result of this roundtable: We see a lot more talk about AI accelerating solution discovery than problem discovery…but in fact it’s useful for both, e.g., automating literature reviews, coding user research transcripts. (I’ve been doing a bit of this in my productivity-app research.) For prototypes to work as conversation, you need a high degree of trust and psychological safety within a team. ### Pranav Pathak: Scale Experiments with AI Pranav got into specific frameworks and examples for how to prioritize and evaluate AI product experiments. He explained the PROVE-IT prioritization framework — Pain, Reach, Own data, Verifiable, Execution loop, Integration time, Total cost. He then took us through two real-world examples from Booking explaining how the framework applied, and how each of the projects turned out. In a world of vague hype it was great to see such concrete examples and analysis. I also liked what he had to say about metrics: track the right metrics. A good metric is clear, reflects the actual product goal, and is iterable. Business metrics are problematic because they’re also sensitive to outside factors like the global economy; engagement metrics are too easy to game. ### Lucie McLean: Navigating Organizational Change Lucie led this roundtable discussion of organizational change — a topic that’s always on people’s minds and especially these days with layoffs and the spectre of AI. As is so often the case, I was impressed and humbled by people’s willingness to ask for and offer advice with thoughtfulness and vulnerability. ### Elena Leonova: The Brutal Truth About AI Product Success A lot of AI strategy is hype. Key point here: Your moat can’t just be the AI model because everybody has the model. What meaningful differentiator are you delivering beyond a prompt? Potential moats include your data, your behavioral understanding, or your workflow (i.e., does it match what your users need and expect?) ### Asya Kuznetsova: Escaping the Paid Growth Trap Enjoyed this not only because it wasn’t about AI, but because Growth isn’t my core skillset. Asys says beware paid growth: its techniques keep changing, and it can be expensive and unsustainable. She then went into detail about techniques and perspectives that work: Word-of-mouth is powerful and is something you can drive. Organic recommendations are powerful: encourage customers to notify other people about your product as part of the actual functionality. Ex: splitting the bill in Revolut, requesting money from non-user contacts in Wise. (I recall emails saying, “You’ve got cash!” in the early days of PayPal, too.) Ask yourself: How can customers benefit from having others use your product? Incentivized recommendations (e.g., referrals) work well, too. Things like rewards or discounts for referring friends & colleagues. The right incentives vary not only by product but also culturally. Counterintuitively, you want to nudge customers to refer early in the journey: after the first food order, etc. This is when delight is highest. In a noisy world, people still trust people—a sentiment echoed later in Francesca Cortesi’s closing keynote. ### Simone Basso: The Community Bet Simone talked about a new product he and his team built at WeRoad. They started with an idea and, with a team of 5–6, went from zero to MVP in a couple months. Two interesting tidbits here: Yes, they vibe-coded a prototype in a couple days…and then redid it from scratch to create the actual MVP. What they learned from the MVP? They hadn’t fully understood the customer problem. I asked Simone about this after the talk: if he did it over, would he pause and do a little research? He said no: this was the fastest way to learn. Reminder that discovery and implementation are different activities using some of the same methods. ### Francesca Cortesi: AI is here — Now What?! There’s so much uncertainty, ambiguity, and rapid change around AI. What’s happening? What’s going to happen next? Cortesi doesn’t know. Most of us don’t know. And that’s OK. Instead of reacting with fear, let’s react with curiosity. AI is machines: when you play the speed game with machines, you lose. But humans are good at things that AI is not — things that are “un-LLM-able.” Find the un-LLM-able and focus on that instead. Curiosity, learning and unlearning. Thinking creatively. LLMs offer the illusion of thinking but can’t think outside the box. Trust: LLMs hallucinate and that makes trust an un-LLM-able skill. “Trust is your most important currency.” Don’t react: strategize. That’s un-LLM-able too. You can go fast…or you can go intentionally in the right direction. (I’m reminded of the old saying, “Don’t mistake motion for progress,” which has always been an issue.) Here are Francesca’s slides, shared with her permission. #### About the Author Dave Feldman is a versatile product leader with 25 years of experience spanning Product, Design, and Engineering. A two-time founder with exits to Google, he’s held leadership roles at Google, Facebook, Heap, and Yahoo. Now based in Berlin, Dave advises the Productlab community on industry trends and maintains strong connections to the US product scene. You can follow him on LinkedIn. ## How SumUp saved 1,700 engineering hours by promoting a customer support agent Speaker: Liza Mello Read online: https://productlab.app/blog/2025/liza-mello Video: https://www.youtube.com/watch?v=3URbh4AJVGw A counterintuitive move that turned chaos into clarity. Liza Mello on the hire that gave SumUp's product team its focus back. TL;DR: Liza Mello, Product Lead at SumUp, faced a common problem: her engineering team was drowning in bug reports (many of which weren’t actually bugs). Instead of hiring another PM or engineer, she promoted a senior customer support agent to a new role—part-time. The result? 1,700 engineering hours saved over three years, fewer bugs, and a scalable system that now handles merchant feedback, regulatory research, and user outreach. ### The problem: Bugs everywhere, progress nowhere Three years ago, Liza’s product team at SumUp was stuck in firefighting mode. Developers were interrupted by customer support agents at least once a day. Bug tickets got lost in JIRA. Urgent issues slipped through the cracks. And worst of all? Many of the reported “bugs” weren’t bugs at all—they were misunderstandings or feature requests dressed up as technical issues. The team wanted to move faster and deliver more value, but they couldn’t focus. As one PM put it: “Yes, I agree we need to go faster, but my team is constantly distracted by bugs coming our way—and many aren’t even bugs. It’s wasted time.” Classic catch-22: Product wanted engineering to ship faster. Engineering wanted Product to shield them from noise. ### The unconventional solution: Hire from within Instead of adding headcount to Product or Engineering, Liza looked inward—to Customer Support. She recruited Alex, a senior CS agent with three critical qualities: Deep product knowledge (could distinguish bugs from correct behavior instantly) Attention to detail (tech-savvy, interested in engineering tools) Trust from the support team (crucial for adoption) The role started part-time. No budget battles. No lengthy hiring process. Just an experiment: What if Alex became a “Product Specialist” who triaged all bug reports before they reached the engineering team? ### What Alex actually did (and why it worked) Alex’s initial scope was simple but powerful: Bug Triage: Troubleshoot incoming issues Enrich tickets with context (error logs, impact assessment via Grafana) Filter out non-bugs and duplicates Pass only validated, detailed bugs to engineering The impact was immediate. Because Alex came from customer support, he could: Train CS agents to troubleshoot basic issues themselves (reducing deflected bugs) Communicate changes back to the support team organically Distinguish real bugs from user confusion with expertise only a CS veteran could have The numbers speak: Over three years, SumUp saw fewer bugs (both real and deflected) and saved an estimated 1,700 engineering hours—nearly one full year of an engineer’s time. ### The evolution: From shield to enabler After three months of success, Liza realized something important: This role could do more than just shield the team. It could enable them. Alex’s scope expanded to include: #### 1. Merchant Feedback Insights Alex became the team’s voice-of-customer engine, synthesizing feedback from: Support tickets Cancellation surveys Trustpilot and app store reviews Every month, Alex presents categorized insights to Product, Design, and Marketing. They discuss what’s worth pursuing immediately vs. later. (And yes, AI helps automate parts of this now, but Alex still adds the human layer of context and follow-up.) #### 2. Contacting Merchants When the team gets intriguing feedback but doesn’t have time to investigate, Alex reaches out to merchants directly. He’s also supported proof-of-concept testing where part of the flow required phone calls—something PMs rarely have bandwidth for. As Liza notes: “These won’t be as qualified as user interviews. Choose which ones you’d rather do yourself and which you don’t.” #### 3. Regulatory Research SumUp operates in heavily regulated markets (invoicing, accounting). Alex monitors regulatory changes—like new VAT rates or electronic invoicing requirements—and explores potential market expansions. He does the preliminary legwork so Product can make informed decisions faster. ### The AI question: Could it replace Alex? Liza addresses the elephant in the room head-on: Could AI do all this? Her answer: “AI definitely supports this role, but couldn’t yet automate it completely.” Here’s why: AI can help analyze feedback at scale, but Alex adds context, asks follow-up questions, and knows the CS team intimately Bug triage requires judgment calls that tools can’t always make Human relationships matter—both with the CS team (who trust Alex) and with merchants (who need a familiar voice) That said, AI has made Alex more efficient. What used to take manual effort now happens faster, meaning one Alex can support more teams. SumUp started with a 3:1 team-to-Alex ratio and believes it can scale further. ### Key lessons from Liza’s experiment 1. Start small, prove value, then expand The part-time approach avoided budget battles and allowed the team to test without commitment. Once the ROI was clear, the role became permanent. 2. Hire from Customer Support—seriously CS agents have an “untapped potential” that most companies ignore. They know the product deeply, understand user pain points, and can bridge Product and Support in ways external hires can’t. 3. Evolution over perfection The role changed dramatically over three years. What started as bug triage expanded to feedback synthesis, user research, and regulatory monitoring. Don’t be afraid to iterate. 4. Include them in rituals Alex attends design reviews and sprint planning. Why? Because he knows users better than anyone. His insights prevent blind spots before they ship. 5. AI is a multiplier, not a replacement Tools help scale the work, but the human expertise—especially from someone who’s been in the trenches—remains irreplaceable. ### Is this scalable? Is it sustainable? A great question from the audience: “Won’t you just need more roles as problems grow?” Liza’s take: Yes, this is a permanent role. But it’s not a linear scaling problem. Better processes, automated testing, and AI tools mean Alex can handle more over time. The key insight? This role prevents problems from ballooning in the first place. As one attendee reframed it: Alex isn’t a luxury hire. He’s Product Operations—the connective tissue that lets Product and Engineering focus on what they do best. ### The bottom line If your team is drowning in noise—bug reports, user feedback, regulatory changes—before you hire another PM or engineer, look inward. Your Customer Support team might have the answer sitting right there. Liza’s story proves that sometimes the best hire isn’t a hire at all. It’s a promotion. And the best solution isn’t adding more resources—it’s leveraging the expertise you already have in smarter ways. What SumUp got: 1,700 engineering hours saved Faster feedback loops Stronger Product-Support collaboration A scalable model that AI enhances (but doesn’t replace) What it cost: One part-time role (initially) The courage to try something unconventional Not a bad trade-off. Liza Mello is a Product Lead at SumUp with 14 years in product management. She leads teams across invoicing, accounting, and point-of-sale products. Before moving to Denmark, she built her career in Brazil, bringing a uniquely resourceful approach to product operations. Connect with her on LinkedIn. ## Stop Building Better Hammers. Start Building Carpenters. Speaker: Stefan Ostwald Read online: https://productlab.app/blog/2025/stefan-ostwald Video: https://www.youtube.com/watch?v=Kqm6uCrL3u4 What Stefan Ostwald taught 250 product managers about the AI revolution happening right now. I’m sitting in my Gesundbrunnen apartment right now, still processing what happened at Productlab Conf in September. Gesundbrunnen—yeah, not exactly the Berlin you see on Instagram. But here’s the thing: while everyone’s fighting for overpriced flats in Kreuzberg, this neighborhood near [Nord]bahnhof is quietly transforming. New bars opening up. Prenzlauer Berg parents discovering they can actually afford space here. Tech people realizing the S-Bahn and FEX connection is unbeatable. Like AI, sometimes the smartest move is the path no one’s talking about yet. Which brings me to Productlab. CIC Berlin (the space previously known as The Factory) was buzzing with energy—250 product managers, leaders from across Europe, all trying to figure out what the hell we’re supposed to do with AI. Then Stefan Ostwald walked on stage. And honestly? Everything shifted. ### The moment that changed everything Stefan started with something deeply personal. Since he was a teenager, he had one obsession: Can I build something that works FOR me, not WITH me? Failed robot bachelor thesis. Failed carpooling startup. Failure after failure. But here’s what got me: he never stopped trying. Because there’s something magical—his word—about systems that work for you. Systems that help you even when you walk away. And with Paloa, he finally cracked it. Not through robotics. Not through clever algorithms. Through large language models that can reason and act. ### The question that breaks everything open Stefan asked us something that’s been haunting me ever since: “What if, instead of building a better hammer, we built the carpenter?” Let that sit for a second. We’re all sitting here in our product teams, building better search tools. Better analytics dashboards. Better checkout flows. Even our “AI features” are just… better tools. But what if we’re thinking about this completely wrong? ### Tools vs. Workers: The test that matters Here’s the framework Stefan gave us, and it’s so simple it hurts: Can you walk away and it keeps working? If NO → You built a tool If YES → You built a worker That’s it. That’s the difference between the past and the future. He painted this picture of Jesse, a consultant with a travel agent worker. Her flight gets delayed. But before she even panics, her agent has already: Rebooked her on the next flight > 2. Informed her hotel > 3. Extended her car rental > 4. Got her a voucher for a free meal > 5. Filed €400 compensation paperwork Jesse barely had anxiety. The worker handled everything. Six months ago? That day would’ve been hell. Now? Few Seconds. These experiences are currently only available to CEOs with amazing assistants. In the future, they’ll be available to everybody. ### This is happening across EVERY industry Stefan didn’t just talk theory. He showed the transformation happening everywhere: CRM: Not dashboards—workers qualifying leads and following upAnalytics: Not dashboards—analysts digging into data and alerting youCompetitive analysis: Someone watching the market for you 24/7Customer service: Paloa’s client Minia increased satisfaction by 38% The models are ready. They’re good enough right now. What’s missing is the applications on top. That’s what WE need to build. ### The hard truth about transformation Here’s where Stefan got real with us. Like, uncomfortably real. At Parloa, they had a successful rule-based automation product. Tripling revenue every year. Everything was working. And they made the shift anyway. He didn’t sugarcoat it: “This was really really intense. The commercial team was shielding pressure from competitors who were advancing faster on the old pattern. Technology was hustling hard to get the new product running.” But here’s his point: What choice do you have? #### Two paths: Path 1: Keep going with what works→ Short-term success + low risk→ Long-term: Someone else will rethink the pattern and disrupt you Path 2: Transform now while you’re successful→ Hard now + significant risk now→ Realistic chance of long-term success The question isn’t “will industries be disrupted?” The question is: Will YOU be the one disrupting it? ### What to do Monday morning Stefan gave us homework. Real, practical homework: Ask yourself: What will be your WhatsApp moment? (Context: WhatsApp completely rethought communication when SMS was dominant) Then, when you go back to your team: What problems do your customers always have? What do they always need to look at? What do they always need to drill down into? Then ask a different question: Not “how can we improve our tool?” but “Can we build something that works FOR them?” Once you ideate, test it: Can you walk away and it keeps working? If yes, you’re onto something. ### Four principles for building AI workers Stefan gave us his battle-tested framework from building at Paloa. This isn’t theory—this is what actually works: #### 1. Recognize the spectrum It’s not binary (tools vs. workers). There’s everything in between—like co-pilots in coding where AI does work but humans review before deployment. Consciously decide where on the spectrum you should be for each problem. #### 2. Design for references, not descriptions Imagine trying to describe “the blue pants in the upper left” to an agent. Exhausting, right? Instead, make things explicitly referenceable. Like Notion’s @ symbol. Make implicit references explicit = faster velocity + better quality. #### 3. Design for review, not perfection AI makes errors. It will continue to make errors. The job isn’t to make it perfect—it’s to make working with an imperfect AI easy. Look at Cursor in coding: smooth iteration cycles, easy accept/decline, simple rollback when AI goes off rails. Design for the error state, not the perfect state. #### 4. Design for behavior validation, not configuration Traditional systems were deterministic—add a rule, it follows the rule. AI agents? They’re like humans but worse. Just because you tell them something doesn’t mean they’ll do it. The truth is what actually arrives at the user, not what you put into the model. Stefan’s team simulates hundreds of scenarios, then uses another AI agent to evaluate them automatically. Build the validation loop, and you can even do automatic prompt optimization. ### Why this hit different Look, I’ve been to a lot of conferences. Heard a lot of talks about AI. But Stefan’s talk hit different because it wasn’t about “AI strategy” or “digital transformation” or any of that consultant-speak bullshit. It was about a guy who’s been obsessed with building autonomous systems since he was a teenager. Who failed multiple times. Who finally figured it out. And who’s willing to tell us—not from theory, but from the trenches—exactly what it takes. He stood up there and said: The age of AI is here right now. If you don’t do this, somebody else is out there doing it. And you know what? He’s right. ### My take from Berlin I’m writing this from Gesundbrunnen—not the coolest neighborhood, I’ll admit. But it’s near Hauptbahnhof, rent is actually affordable, and something interesting is happening here. Cozy bars are popping up. Mummies are migrating from Prenzlauer Berg. Tech people are discovering what locals already knew. It reminds me of Stefan’s talk, actually. Everyone’s chasing the obvious paths—Kreuzberg, Mitte, Prenzlauer Berg. But sometimes the smartest move is to take the uncovered path. To see what others are missing. To bet on transformation before it’s obvious. Berlin has always been about reinvention—tearing down walls, building new things from the rubble, embracing change even when it’s uncomfortable. Stefan’s message feels very Berlin to me: Don’t wait for the perfect moment. Transform while you’re successful. Disrupt yourself before someone else does. CIC Berlin, where we hosted Productlab (previously known as The Factory)? Stefan mentioned they started there years ago. Full circle. From startup struggling to make autonomous systems work, to standing on that stage showing 250 product managers exactly how to do it. That’s the journey we’re all on right now. The models are ready. The technology is there. The question is: Are we bold enough to stop building better tools and start building workers? ### What’s your WhatsApp moment? I want to hear from you. Seriously. What’s the thing in your product where users are constantly drilling down, constantly looking at dashboards, constantly doing repetitive work? What would it look like if that worked FOR them instead of WITH them? Reply to this email. Let’s figure this out together. Because Stefan’s right about one thing: Someone’s going to rethink your industry. Might as well be us. —Daniele P.S. Productlab 2026 dates coming in December. If this year taught me anything, it’s that we need to keep having these conversations. The revolution is here. Let’s architect it together. DM me on LinkedIn! #### About the Speaker Stefan Ostwald, Chief AI and co-founder of Parloa, is a leader in conversational AI with nearly a decade of experience making customer communications easier and more secure through AI. Connect with him on LinkedIn. ## How to build un-AI-able skills (and why "I don't know" is your superpower) Speaker: Francesca Cortesi Read online: https://productlab.app/blog/2025/francesca-cortesi Video: https://www.youtube.com/watch?v=qGuT4Au2n-I The closing keynote everyone needed to hear at Productlab Conf 2025 Six months of research. 100+ slides. Dozens of books, podcasts, and expert conversations. And Francesca Cortesi stood on stage at Productlab Conf 2025 and said the three words no keynote speaker is supposed to say: “I don’t know.” Not once, but repeatedly. In front of 300 product managers who came looking for answers about surviving the AI revolution. And it was exactly what we needed to hear. ### The panic we’re all feeling When Daniele Ronca (Conference’s founder) asked Francesca to close the conference with a talk about AI, she had 30 seconds of panic. Then clarity: she didn’t want to talk about the tech. She wanted to talk about how we should relate to the tech. Because here’s what we’re all dealing with: emotions ranging from excitement about AI’s possibilities to anxiety about mass layoffs. The tech is here to stay. We can’t control that. But we can control how we respond to it. Francesca dove deep into research, expecting to emerge with all the answers. Instead, she found something more valuable: a framework for navigating uncertainty when nobody actually knows what’s coming. ### The skill that lasts 18 months (not 5 years) Remember when the lifetime of a professional skill was 5 years? That felt manageable. You could invest time, build expertise, and reasonably expect it to pay off. Now? 18 months. Don’t believe it? Remember prompt engineers? Those $200K+ jobs that were supposed to be the golden ticket? They disappeared almost as fast as they appeared. Not because the skill became less important, but because AI companies optimized for mass adoption and made prompting dramatically easier. This leads to Francesca’s first major insight: The real skill is learning and unlearning. Not what you know today, but how quickly you can learn tomorrow. ### The framework: Curious intentionality Instead of running frantically trying to learn everything (the image of overwhelmed people running resonated with everyone), Francesca advocates for strategic curiosity. She calls it “curious intentionality” and breaks it down into five questions: 1. How does this actually work? LLMs are pre-trained on existing data. They can’t “think” – they give us the illusion of thinking. They can’t think outside the box because they only use the data they have. Understanding these limitations is power. 2. What are they optimizing for? Big AI companies are optimizing for mass consumption, not perfection. This is why hallucinations aren’t going away. They don’t care about fixing them – they care about usage. 3. How is it evolving? Watch the trajectory, not just the current state. Remember AI-generated videos six months ago versus now? The improvement curve tells you where to place your bets. 4. Where are the gaps? This is where your opportunity lives. What can’t AI do? What won’t it be able to do? 5. What should I focus on? Francesca’s advice mirrors product strategy: don’t spread yourself thin. Pick your eggs, choose your baskets intentionally. Skip an innovation wave? You’ll be fine if you’re paying attention. Don’t let FOMO drive you – let intentional curiosity guide you. ### What is “un-LLM-able”? This question became Francesca’s north star. In product strategy, you look for things that are hard to copy – sustainable competitive advantages. The same principle applies to your career. Her answer centers on one word: Trust. And here’s why it matters: AI hallucinations aren’t getting fixed. The big companies aren’t even trying. They want adoption, not accuracy. As fake images multiply and skepticism grows, the human who can be trusted becomes invaluable. Francesca breaks trust into three components: #### 1. Credibility Deliver what you say you’ll deliver. Simple, but foundational. #### 2. Value consistency This is where Francesca showed the Jason Knight meme that hit hard: teams rushing to build AI features while basic customer needs go unmet. The customers waiting? They won’t wait forever. Either they’ll build it themselves, or someone else will build it for them. The question isn’t “Can we add AI to this?” It’s “Are we creating actual value for our customers?” #### 3. Collaboration Here’s where Francesca shared a hard-won lesson. Six or seven years ago, when stakeholders wanted to build the wrong AI feature, she’d try to prove them wrong. It never worked. Too much ego, too much reputation on the line. Her new approach: “This is my thinking. What am I missing?” Instead of collecting data to convince people, she collects data, presents it transparently, and invites collaboration. She involves stakeholders in the thinking process rather than trying to change their minds. ### The process that works Francesca shared her framework for navigating uncertainty (which, let’s be honest, is all of product management): Study → Frame the question → Explore possibilities → Test → Converge And critically: Involve and learn at every stage. Then repeat with AI: Study AI → Frame choices around AI → Explore AI possibilities → Test with AI → Converge Communication and execution happen throughout, but the key is continuous learning and stakeholder involvement. Just to let you know, this is taking time. ### Why “I don’t know” is the answer Standing on stage, admitting uncertainty could have felt like weakness. Instead, it was the most powerful moment of the conference. Because Francesca normalized what we’re all feeling: We’re ALL learning. We’re ALL on a learning curve. Her presentation wasn’t a set of answers. It was a process she’d gone through – and is still going through. When she realized she didn’t know and still had to speak about it, her “product brain” kicked in. What do you do when you don’t know? You dig into the question. ### The takeaway As AI rapidly evolves and skills become obsolete every 18 months, your competitive advantage isn’t what you know. It’s: How quickly you can learn and unlearn How intentionally curious you are (not just reactive) How much you can be trusted (credibility, value focus, collaboration) How well you involve others instead of trying to convince them The tech is moving fast. Faster than any of us can keep up with completely. But if you’re curious, intentional, focused on creating real value, and building trust through collaboration, you’ll pick up the signals you need. You’ll get there. And if you skip an innovation wave? Remember the prompt engineers. You’ll be fine. ### One more thing At the end of her talk, Francesca said: “I hope I gave you some food for thought.” She did more than that. She gave us permission to not have all the answers while showing us exactly how to move forward anyway. In an age where AI is changing everything, the most human skill – admitting “I don’t know” and genuinely seeking to learn – might be the most valuable one of all. Francesca Cortesi delivered this closing keynote at Productlab Conf 2025. You can learn more about her on francescacortesi.com or connect on LinkedIn. ## When 90% of AI Agents Fail: What Zendesk’s Team Learned Building the Real Thing Speaker: Mirza Beširović Read online: https://productlab.app/blog/2025/mirza-besirovic Video: https://www.youtube.com/watch?v=73_Uov3zmK4 Based on a talk by Mirza Beširović at Productlab Conf 2025 At Productlab Conf 2025, Mirza Beširović, AI Product Leader at Zendesk, shared what their team has learned building and scaling AI agents in real production environments — where systems need to withstand high interaction volumes, unpredictable customer behavior, and real business stakes. This post summarizes Mirza’s talk. Quotes are from the session, but the wording and commentary are mine. Mirza stood on stage and asked a simple question: “Raise your hand if you’ve built an AI agent in the past six months.” About half the room raised their hands. “Keep it up if it actually worked the way you wanted it to work.” Half the hands went down. “Keep it up if you’re still using it six months later.” Only about 10% remained. The room got quiet. That uncomfortable kind of quiet when someone just said the thing everyone was thinking but nobody wanted to admit. Mirza leads Zendesk’s AI agents platform. They process millions of customer service interactions, dealing with handoffs that fail, latency that kills trust, agents that break. But they’re figuring it out at scale. What Mirza shared wasn’t theory. It was battle scars. “You’re not alone. We’re all still learning and experimenting. Most of us are still figuring out how this is working.” That honesty? That’s what made the room lean in. ### The Job Has Already Changed (We Just Don’t Know It Yet) “Most PMs don’t realize that the product management job has already changed drastically and fundamentally. We’re all still kind of catching up with it.” That’s what Mirza opened with. And I felt it land in the room. We’re still writing PRDs for probabilistic systems. We’re still running sprints for research that takes either 6 days or 6 months. We’re still planning features when we should be orchestrating capabilities. The tools broke. The methods broke. And we’re all pretending they still work. ### Three Tensions Breaking Everything We Know Mirza identified three fundamental tensions emerging that are breaking traditional product management: #### 1. The Control Paradox Users tell you they want control. They want to “train the bot.” They want to influence how the system behaves. But they absolutely don’t want any complexity. They don’t actually want to know how it works. Real example from Zendesk: A customer writes an instruction: “When a customer says X, always escalate to a human agent.” The very next instruction they write? “But resolve everything autonomously.” This isn’t stupidity. This is humans being human. “Most users want the illusion of control, not actual system effectiveness.” We want the illusion of control, not actual system effectiveness. The solution isn’t more control panels or more settings. It’s designing touch points that preserve user agency without breaking the underlying system. What this looks like in practice: Binary signals where human judgment actually matters (thumbs up/down, but contextualized) Evaluation layers that aggregate feedback before modifying behavior Three-layered architecture: big picture analytics, drill-down capability, and qualitative feedback with conflict detection Here’s the insight that stuck with me: Users don’t want control—they want transparency. They want to understand how the system is performing and why it’s doing something. It’s a trust-building exercise, not a configuration challenge. #### 2. The Orchestration Trap Specialized agents perform better. A focused agent that handles order cancellations will outperform a general chatbot every time. But here’s the trap: More specialization = more agents = more handoffs = more latency = more failure points. When Zendesk handles a simple customer query like “cancel my order,” it’s not one agent. It’s a choreographed dance: Agent detects intent Agent checks knowledge base Agent makes API call to retrieve data Agents validate context handoff Agent generates response Agent evaluates response quality LLM judges overall conversation success Seven agents. One query. And users expect sub-second responses because that’s how humans communicate. “Humans have very low latency. AI agents don’t. The number one problem we all have is latency, latency, latency.” That gap? That’s what breaks trust. This is the shift: You’re not building product interfaces anymore. You’re creating digital workspaces where AI agents collaborate to solve human problems. “We’re creating digital workspaces for AI agents to collaborate with one another. Think Second Life, but agentic.” Mirza called it “Second Life, but agentic”—specialized agents with clear roles, capabilities, and purposes, passing context between each other. We’re designing the space where agents work, not just the output they produce. #### 3. The Evolution Crisis Traditional planning methods are failing for AI products. AI research cycles don’t fit into deterministic PRDs. Things can take 6 days or 6 months—it’s the nature of probabilistic systems. “You can’t just write an upfront spec. I mean, you never should have been writing an upfront spec anyway, but now you really can’t with an AI system.” You can’t write an upfront spec. (And honestly, you never should have been doing that anyway.) This is causing teams to fragment along “AI” vs “traditional” lines. And it raises an uncomfortable question: Should teams building AI products stop building user-facing features? Should autonomous agents be building the features instead? Mirza didn’t answer that. He just asked it. And it hung in the air. Because that’s the future we’re walking into, whether we’re ready or not. ### What’s Actually Changing in the PM Toolkit The evolution from feature builder to orchestrator is happening now: PRDs → Prototypes (show, don’t tell)Feature specs → Context architecture (prompts, agent instructions, context engineering)User stories → Agent instructions (what should the agent do in this system?)Product planning → Research cycles (hypothesis testing, not roadmap commitments)Launch planning → Continuous orchestration (how agents interact and evolve)Experience teams → Capability teams (building what agents can do) And then there’s the new skill that will define PM interviews in 2026: Evals. Frameworks that measure AI system performance—metrics, scorers, evaluation datasets. “The one skill I promise you everyone will be testing for in job interviews in the next year or so will be about evals. So learn evals.” Mirza’s prediction: If you’re not learning this now, you’re already behind. ### The New PM: Builder + Orchestrator The PMs who will thrive aren’t just strategists. They’re: Builders: Creating agents for themselves (personal experiments) Prototyping relentlessly (not writing specs) Designing agent interaction patterns Understanding what breaks and why Orchestrators: Architecting how humans and machines collaborate Deploying agents to solve business outcomes Managing context flow between systems Creating transparency and explainability Does everyone need to become a coder? No. Strategy and market understanding remain critical. But the ability to prototype, to experiment, to show instead of describe—that’s the new table stakes. ### Your Homework This Week Mirza ended with a challenge: “The question isn’t whether you’ll become an orchestrator, it’s whether you’ll do it intentionally.” Here’s what he wants you to do: Build one agent. Use Zapier, ChatGPT, whatever. Map one of your workflows (personal or work) as an agent interaction. Identify one human touchpoint where you could simplify the experience. Prototype instead of writing a PRD. Try running a research cycle instead of a sprint. Share your learnings publicly. Create social proof. Document what breaks and why. This is how you break into an AI PM role even if you don’t have one yet. ### What This Means For Us I’m sitting here at The Social Hub Berlin on a rainy Monday, and I keep thinking about that 10% stat. Only 10% of agents are still being used six months later. That’s not because the technology isn’t ready. It’s because we’re still applying old methods to a new paradigm. We’re trying to plan the unplannable. Control the uncontrollable. Simplify the irreducibly complex. And the teams that figure this out first? They’re not writing better PRDs. They’re not running tighter sprints. They’re prototyping. They’re building evals. They’re designing for orchestration, not features. They’re becoming the orchestrators Mirza described. ### The Uncomfortable Truth Here’s what Mirza said that I keep coming back to: “None of us have all this figured out. But community is the competitive advantage that we as product people have.” This is why Productlab exists. These conversations—honest ones about what’s breaking and what’s working—are how we collectively figure this out. Because the PM job has already changed. The question is whether you’ll lead that change or react to it. —Daniele P.S. Mirza runs a newsletter for aspiring product leaders at . If you’re serious about breaking into AI PM, subscribe. He’s documenting the journey in real-time. Connect with him on LinkedIn. Check the Slides ## How WeRoad Turned 1,200 Solo Travelers Into a Year-Round Community Movement Speaker: Simone Basso Read online: https://productlab.app/blog/2025/simone-basso Video: https://www.youtube.com/watch?v=0Ug1Q8wdDMU WeRoad had thousands of travelers who loved their trips. The problem? They only traveled twice a year. When Simone Basso and the team launched WeRoad Events in 2024, they faced a problem most travel companies would kill for: they had thousands of customers who loved their group trips—but only traveled 1-2 times per year. The Bet? Build a year-round community of local events across 18 European cities. Running clubs. Dinner with strangers. After-work drinks. In just one coordinated evening across Europe, they got 1,200 people to show up. At Productlab Conf 2025 in Berlin, Simone shared the brutal truth about building community products—the missteps, the data, and the principles that turned WeRoad Events from “a bit of a shit show” into a referral-generating growth engine. “Speed of execution is much more important than perfection. We made a lot of errors on the way. It doesn’t really matter as long as you correct them super fast.”— Simone Basso, CTPO at WeRoad ### Understanding the Flywheel WeRoad isn’t just a travel company where you book and go. It’s built on a community flywheel that most PMs miss. Here’s how it works: Travelers join through social channels, do a weekend group trip with 15 strangers aged 25-49, then some become coordinators (paid guides who travel for free). These coordinators organize more events and trips, attracting more travelers. In the center? A community of people passionate about solo travel and real-life connections. The insight: 60% of people who book a trip join the Facebook group (Solo Travelers United) just before or after booking. They’re not just buying a trip—they’re buying into a community. “Fundamentally, it is an evolution of dating apps. Tinder and the likes are all going down, but people do want in real life connections. Doing it in a group setting is much easier than in a dating setting.”— Simone Basso ### 5 Principles for Building Community Products That Actually Scale Simone distilled WeRoad’s journey into 5 hard-won principles. This isn’t theory—it’s what actually worked when they had to fix their “shit show.” #### 1. Ship Fast, Fix Faster—Perfection Is the Enemy WeRoad launched their events app in 6 weeks. No monetization. No payment system. Just search, details, map, and RSVP. They cut scope aggressively. Reused the design system from WeRoad. Focused only on what users absolutely needed to find and attend an event. The result? They launched on time—but engagement dropped after the initial spike. Daily active users weren’t growing. Monthly active users weren’t growing. “We made a lot of errors on the way. It does not really matter as long as you correct them super fast. It will be much worse to try and get everything perfect—the supply perfect, the message perfect. We would still be building.”— Simone Basso Action item: If you’re building a community product, launch your MVP in under 8 weeks. One core feature that works is better than three half-built features. #### 2. Clear Messaging Beats Clever Complexity Their original landing page was a mess. QR codes on desktop. Unclear value prop. Too many event types across 30 cities in 5 countries. Compare that to a competitor (Timeleft): “Stranger dinner Wednesday. Boom. Done.” WeRoad’s fix? They narrowed to 3 event types: Running clubs, Dinners with new friends, We Work After Work. Same message across all 18 cities. One coordinated night across Europe. The clarity unlocked growth. Last coordinated event: 1,200 attendees in one night. Action item: If your product needs more than 10 words to explain, you’re confusing users. Cut your event types, your messaging, your supply—until a stranger gets it in 5 seconds. #### 3. Community Products Need Identity, Not Just Supply WeRoad initially tried to be everything: bottom-up events, top-down organized dinners, 30 different cities, diverse event types from London to Bari. It became “a completely unscalable problem.” The mistake: treating WeRoad Events like WeRoad. But events aren’t trips—they’re a different market, different people, different identity. “Having a clear identity and a clear supply is absolutely necessary to make it work and to grow our community. You cannot just pretend you’re good at WeRoad Events—it’s a completely different thing.”— Simone Basso Action item: Define your community’s identity in one sentence. If you can’t, your users definitely can’t. #### 4. Build for Repeat, Not Just Acquisition The real metric that matters? Cohort retention by month. 21% of people who attend an event come back the next month. 8% come back two months later. About 20% become regular community members. More revealing: 6-7% of event attendees had already done a WeRoad trip. 2% convert to become WeRoad customers after attending events. The network effect is working. Events drive trips. Trips drive more events. Community drives both. But here’s the gap: WeRoad doesn’t yet nudge people enough. Someone does a trip, makes friends—then those friends do a running club in Berlin next week. WeRoad should connect those dots. Action item: Map your retention by cohort month-over-month. If 20%+ aren’t coming back, you don’t have community—you have transactions. #### 5. Own the Full Community Lifecycle WeRoad replaced Meetup with their own platform. Next target? The Facebook group. Why? They can’t control who people talk to, what they ask, or who responds. The pre-trip research phase—where fear and uncertainty kill conversions—happens entirely outside WeRoad’s control. One clever solution they built: post-booking, travelers can share which flights they’ve booked. Other group members see it and coordinate to fly together. Small feature, massive reduction in “fear of traveling alone.” “Getting ownership of the full cycle of the digital community is as important as making sure they have a great community experience. We’re never going to replace WhatsApp or Instagram for how people talk—but the upper funnel of community is crucial for growth.”— Simone Basso Action item: Audit where your community lives. If critical conversations happen off-platform, you’re blind to what drives or blocks conversions. ### The Numbers That Matter When asked about growth mechanics, Simone shared specific data: 900 referral friends in 4-5 months—almost €1 million in referred revenue. The incentive structure: 5% discount immediately to the referred friend, 5% discount on your next trip (up to 50% off after 10 referrals). But here’s the opportunity cost: they only talk about referrals to existing WeRoaders. They don’t nudge new customers pre-trip or post-trip enough. “We don’t do enough nudging of new customers who could have other friends and bring them into the community even before they go on a trip.”— Simone Basso The user activity spike after refocusing the platform? It “fundamentally picked up”—exact numbers still climbing weeks after the messaging fix. ### What WeRoad Got Wrong Simone was brutally honest about their mistakes: Mistake 1: They tried to stop paying coordinators to organize events, thinking community loyalty would sustain it. Coordinators refused. Supply collapsed. Mistake 2: They spent weeks debating monetization—membership model vs. pay-per-use—before shipping the MVP. Premature strategy discussion that delayed value. Mistake 3: They lost sight of the community during the platform migration. Changing the supply model and the platform simultaneously confused everyone. “We had to rush to organize all these different types of events in 5 countries, 30 cities. From London to Berlin to Milan to Bari—very complicated. It became completely unscalable. We fundamentally screwed up.”— Simone Basso The fix came fast: focus on 3 event types, 18 cities, consistent messaging, and rebuilt relationships with coordinators. ### The Conviction: Community Is the Moat Simone’s bet is simple: community compounds. Travelers who attend events repeat more. They book more trips. They stay engaged year-round instead of 1-2 weekends annually. The next phase? Replace Facebook groups with their own platform. Control the pre-trip conversations where fear and uncertainty live. Long-term vision: WeRoad becomes the platform where solo travelers discover trips, meet locals, attend events, and build friendships—all in one ecosystem. “Running clubs are the new dating. People want in-real-life connections in group settings—it’s much easier than one-on-one dating.”— Simone Basso The timeline: by end of 2025, own the full digital community experience. ### Your Move Building community products isn’t magic. It’s discipline. Simone and the WeRoad team proved that with the right framework—ship fast, clarify your identity, design for repeat, own the lifecycle—you can turn one-time travelers into year-round community members. Start with one thing tomorrow: Cut your event types to 3 or fewer. Make them crystal clear. Then add coordinated launch days across all your locations. The community product that wins isn’t the one with the most features. It’s the one people understand in 5 seconds and come back to every month. Connect with Simone: LinkedIn: Simone Basso Medium: Simone Basso X: @simone_basso ## Escaping the Paid Growth Trap: How Product-Led Growth Beats Paid Ads 10:1 Speaker: Asya Kuznetsova Read online: https://productlab.app/blog/2025/asya-kuznetsova Video: https://www.youtube.com/watch?v=4mJT1UMGmag The growth strategy that gets stronger as marketing costs rise The bottom line: Most companies burn money on paid ads while their competitors grow organically at 10x lower cost. The secret isn’t more budget—it’s building products people want to share. When Asya Kuznetsova was first tasked with “grow through word of mouth,” she was puzzled. How do you control what someone somewhere tells someone about your product? How do you measure dark matter? Today, as Sr Group Product Lead at Kraken, Asya has cracked the code. Her teams have consistently achieved 70-95% organic growth by treating word-of-mouth not as luck, but as a buildable system. At Productlab Conf Berlin 2025, she shared the exact playbook that helped fintech companies scale from millions to hundreds of millions of users—without burning through ad budgets. ### Why Word-of-Mouth Matters More Than Ever In a world drowning in paid ads and AI-generated noise, people still trust people. Consider these cases: → Notion grew from 4M to 20M users in one year. 95% came through shared templates—essentially free word-of-mouth built into the product. → Spotify hit 650M users largely through social features. When you click “create” in Spotify, 3 out of 4 options involve friends. → Multiple fintech companies achieved 70%+ organic acquisition by building sharing directly into their core user journeys. The companies you still use today? They’re probably the ones you discovered through word-of-mouth—because they were designed for it. ### The Three Pillars of Word-of-Mouth Growth Most companies hope for word-of-mouth. Smart companies design for it. Asya’s framework has three components: 1. Organic Recommendations These happen when customers notify others about your product simply by using it. Examples: → When Spotify users share songs on Instagram, non-users see that their friends prefer Spotify → When fintech users send money, recipients get branded notifications showing benefits and savings → When Notion users share templates, they’re distributing free marketing Think about your product: What features would automatically notify others when customers use them? 2. Network Effect Features Build features that make your product more useful when others join. Real examples from fintech: → Revolut’s “split the bill” feature—pay at restaurants, split in-app, friends pay you back instantly → Monzo’s group expense management—roommates pool expenses, easily settle up → Payment request features—one click to sync contacts, one click to request money Customers invite friends not for rewards, but because the product becomes more valuable with them on it. 3. Incentivized Referrals This is where most companies fail—they just throw money at customers hoping they’ll invite friends. The key? Connect referrals to customer journeys. Asya’s research revealed customers in fintech primarily recommended products for travel. So the entire referral program pivoted: “Invite friends to travel, give them a free travel card, earn cash rewards for travel.” Result: Referral conversion improved several times over generic “invite a friend” programs. ### The Referral Multiplication Effect Here’s why referrals are the fastest path to impact: John loves your product. You incentivize him to invite friends. He invites 5 people and gets rewarded. John feels appreciated and becomes more likely to recommend again. His friends receive incentives to try. They love the product because it came from someone trusted. They become loyal customers. John’s friends are now more likely to recommend because they came through recommendation themselves. John’s 5 friends bring 25 more. Those 25 bring 125. Exponential growth. Dropbox proved this works at scale: they grew 40x in 15 months (100K to 4M customers) through a simple referral program. ### The Recipe for Successful Referrals Most referral programs fail because they’re disconnected from customer journeys. Here’s Asya’s tested recipe: Step 1: Match customers in their first weeks Send referral prompts when customers are in their “aha moment”—when they’ve just experienced core value and want to share. Step 2: Give clear direction Instead of generic “invite friends,” be specific: → “Invite friends to [specific use case]” → “Give them [specific benefit they’ll understand]” → “Earn [clear reward]” Context removes the main blocker and improves conversions several times. Step 3: Connect referrals to other social features When you share a payment link, attach a referral link. When you split a bill, attach a referral link. When you connect all word-of-mouth features, they multiply each other’s growth. Referrals are never standalone—they’re part of your entire social ecosystem. Step 4: Maintain healthy economics Manage referrals by payback period: how fast do you earn revenue from invited customers compared to reward costs? Two levers to pull: → Decrease reward cost (variable rewards: “up to €50” where only small percentage gets top amount) → Increase invited customer revenue (require actions that correlate with retention to claim rewards) For example, one fintech company found customers who spent €200 had healthy retention. To claim referral rewards, invited customers must spend €200—aligning incentives with business health. ### The Cultural Challenge One audience member raised an important point: cultural differences matter. In Brazil, people share referral links freely across any channel. In Germany, people feel uncomfortable referring others for personal benefit. Asya’s insight: When you talk to customers who don’t share referral links, they’re often motivated by helping others—but the money component feels uncomfortable. The solution? Build referral mechanics that feel like genuine favors, not ads. Frame benefits around helping friends discover value, not earning bounties. ### The Retention Advantage A critical question from the audience: Do referred customers actually stick around, or is it just bounty hunting? Asya’s data across multiple companies: Retention for invited customers is consistently higher than for customers from other channels. Why? They came through trusted recommendations. They had better onboarding experiences. They joined with context about the product’s value. ### B2B Referral Strategy Growing B2B through referrals is harder—businesses have more barriers, they fear reputation damage. Asya’s workaround: Target third parties who work closely with your customers but aren’t direct clients. Examples: → Accounting firms that work with your business customers → Real estate agents who interact with your user base → Industry consultants who make regular recommendations These intermediaries will recommend your product for incentives with much lower risk than asking your customers to become evangelists. ### The New Growth Playbook Every great company Asya knows didn’t rise to the top by outspending competitors on ads. They succeeded because customers spread the word. The smartest ones didn’t wait for it to happen—they designed for it. If you want exponential growth: Build something customers want to talk about Give them the tools to do it Connect those tools to natural user journeys Make sharing feel like helping, not selling This approach cuts acquisition costs 10-fold while building defensible growth. In rising competition, marketing costs skyrocket. Companies burn through money just keeping up with competitors. But in a world full of noise and paid ads, people still trust people. And now with AI shifting standards, authentic word-of-mouth becomes more vital than ever. Connect with Asya: Follow her on LinkedIn where she regularly shares growth insights. She’s currently hiring a Senior PM at Kraken to work on exactly these topics. ## From 12 Months to 3 Months: How Booking.com Cracked the Code on AI Experimentation Speaker: Pranav Pathak Read online: https://productlab.app/blog/2025/pranav-pathak The Behind-the-Scenes from Productlab Conf Berlin Hey there, After Productlab Conf, we asked you which talks resonated most—and Pranav’s session on scaling AI experimentation came out as one of the absolute favorites. So many of you reached out asking for a deeper recap to share with your teams. Here it is. For the community. You know that moment when a speaker starts their talk and you can just feel the energy shift in the room? That happened when Pranav took the stage at Productlab Conf on September 18th. The lights were bright, the Productlab Stage was packed, and Pranav—who leads Product Development for AI at Booking.com—opened with a confession that made everyone lean in: “I can’t see much because of the lights, but wow, we have a full house!” What followed was one of the most honest, actionable talks on AI experimentation I’ve seen. No fluff. No AI hype. Just real lessons from someone who’s been doing AI for 10 years (yes, before it was cool) and has the battle scars to prove it. ### The Failure That Changed Everything Let’s start with the uncomfortable truth Pranav shared: his first chatbot was a complete disaster. Back in 2017, his team spent 12 months. They involved three separate teams. They built a “virtual post-booking customer service agent in training” using basic natural language processing (remember, no GenAI back then). The result? Customer satisfaction didn’t move. At all. Here’s what hit me: It’s not just that it failed. It’s that it took a YEAR to fail. But here’s where the story gets interesting… ### The Transformation: Tripling Their AI Success Rate Fast forward to today. Booking.com’s AI experimentation success rate is now more than triple their standard product launch rate. Let me repeat that: Their AI experiments succeed at 3x the rate of traditional features. That’s not luck. That’s a system. And Pranav broke down exactly how they did it. ### The Four Principles That Actually Work #### 1. Spot the Right Opportunity Pranav shared their prioritization framework—and it’s beautifully simple: Pain: How much does this problem hurt? Reach: How many users does it affect? Data: Do we have what we need? Proof: Can we validate this works? But here’s the nuance he added during Q&A: “It’s not just about the numbers. If something’s going to generate €20 million more for the business, I will take that over everything else.” They’re doing 200 BILLION predictions per day at Booking. The scale is wild. But they still ask: What moves the metric the most? #### 2. Track the Right Metrics This one really resonated with the PMs in the room. Pranav was crystal clear: don’t track vanity metrics. At Booking, they don’t obsess over weekly active users. They track conversion. Why? Because: It overlaps business and user value You can isolate what caused it to move You can continuously iterate on it It’s actionable “If you make a change to the product, you set up an A/B test, it moves up or it moves down. You’re able to isolate the cause of that movement quite easily.” #### 3. Democratize AI Across Your Organization This might be the most underrated insight from the talk. When Booking first asked teams “What do you want to build with AI?”, they got 1,000 submissions. Only 5 were viable. Today? They get about 100 submissions per year. 90 of them are viable. What changed? They trained their org. Not on transformer architectures or how LLMs work under the hood. They taught six practical skills that AI is really good at. “If the AI skills in the organization are concentrated towards one team, no one else is ideating,” Pranav explained. And without ideas, you can’t scale. #### 4. Learn the Tools, Experiment, Ship Fast Remember that first chatbot that took 12 months? Their latest version took 3 months with just two teams. Same company. Same complexity. Different approach. They now: Centralize the tech that drives experimentation Automate the iteration cycle (prompts, user feedback, model responses) Make it easier for teams after their first success “The first time that the team experiments is always much harder,” Pranav admitted. “But then the second and third and fourth iteration pick up quite naturally, because the team now knows where the success comes from.” ### The Real-World Examples That Prove It Works Pranav showed us three live products that exemplify this approach: 1. AI-Assisted Host RepliesHotel partners can now auto-generate responses to guest messages. It’s been wildly successful—partners love it because it saves time, guests love it because they get faster replies. 2. Personalized Property DescriptionsInstead of showing the same generic hotel description to everyone, Booking now uses AI to tailor descriptions to user profiles. A family sees “five bedrooms, five bathrooms, ample space for families.” A solo traveler sees something completely different. Same property. Different story. 3. The Chatbot That Actually WorksThree months to build. Two teams. Good customer engagement. Good task completion. Good conversion rates. It’s the redemption arc of that 2017 failure. ### The Uncomfortable Questions (and Honest Answers) The Q&A session was gold. Someone asked about hallucinations and inconsistency—the classic LLM problems. Pranav’s response? They don’t lower temperature (that kills the magic users expect). Instead: They use evaluation models (evals) and train their own They demand structured outputs (JSON responses) They send a percentage to human evaluators They use deterministic models for critical paths (like property recommendations) “We have very low tolerance for factual accuracy, actually, because big business, lots of risk.” Another question was about scaling after experiments. How do you shorten cycle time? The answer: centralize the tech. If five experiments all use question-answering, build it as a shared service. Iterate across all five together. ### What We’re Taking Away Here’s what this talk taught us about scaling AI experimentation: Start with real problems, not AI solutions. Pranav emphasized: “Our customer problems are not changing. Travel still has the same complications. We just now have a tool that’s really good at taking some of those problems away.” Embrace failure as data. That 12-month disaster in 2017 taught them more than any success. Now they’ve streamlined legal compliance assessments from 3-4 months down to 3-4 days. Make AI accessible to everyone. When your whole org can think in AI terms, you get better ideas. It’s that simple. Ship fast, learn faster. They have 200+ GenAI implementations prioritized and dozens of successful product launches already adding meaningful revenue. ### The Energy in the Room Standing at the back of the Productlab Stage, watching PMs frantically take notes while Pranav fielded questions, I was reminded why we do Productlab Conf. It’s not just about the content (though Pranav’s was exceptional). It’s about that moment when someone shares the real story—the failures, the breakthroughs, the behind-the-scenes details—and you realize you’re not alone in struggling with these problems. By the way, Pranav offered a beer to anyone who could guess how many predictions Booking makes daily. Someone shouted “5 billion!” They were off by 195 billion. ## How to build AI products that don't get killed by OpenAI Speaker: Henrique Cruz Read online: https://productlab.app/blog/2025/henrique-cruz Video: https://www.youtube.com/watch?v=b8rrLvNtywE What going from 100K to 1M users in 6 months taught Rows.com about building products that improve every day + Slide to Download When and his team at Rows.com launched their AI Analyst feature, they did something most startups dream about: they turned a $6,000 influencer campaign into 30 million impressions, 1 million sign-ups, and thousands of backlinks. But here’s the real story—it took them 5 years to reach their first 100,000 users, then just 6 months to hit 1 million. At Productlab Conf 2025 in Berlin, Henrique shared the hard-won lessons from building what they claim is the world’s smartest AI spreadsheet. Unlike most AI product talks filled with theory, this was pure tactical gold from the trenches. “In maybe 8 or 10 years of building and working in product, this is the first time that I can tell you that there’s a metric that it’s actually helpful to look at every day.” — Henrique Cruz ### The Speed of Now Tech adoption isn’t just fast anymore—it’s exponential. What once took decades now happens in months. The next breakout app could reach a million users in hours. Henrique’s insight? Timing matters more than perfection. They launched when the market was ready, even though the product wasn’t perfect. The result? They were “in the right place at the right time”—but that place was earned through consistent execution. “It was the right(ish) product, at the right time. A $6k influencer campaign turned into 1 Million sign-ups.” ### 5 Principles for Building AI Products That Improve Daily Henrique distilled years of experimentation into five core principles. This isn’t theory—it’s what actually works when thousands of users are hitting your AI every day. #### 1. Pick Problems AI Can Actually Solve Not every problem needs AI. The sweet spot? Tasks where success is immediately obvious to users. Henrique’s rule: Look for problems where users can instantly tell if the AI nailed it or completely missed. Avoid fuzzy creative work that needs endless back-and-forth iteration. That’s why spreadsheet formulas work beautifully—you either get the right answer or you don’t. Action item: Audit your AI use cases. For each one, ask: “Can the user tell within 5 seconds if this worked?” If not, reconsider. #### 2. Don’t Get “Killed by OpenAI” This was Henrique’s blunt warning: if you’re not 10x more useful than raw ChatGPT, you’re toast. The app layer has 80%+ margins, but only if you’re genuinely more valuable than the base model. His team assumes models will be fast, cheap, and smart. They’re building for a world where AI costs approach zero. Your defensibility can’t come from the model—it must come from your product. Action item: Run this test: Ask ChatGPT your product’s core question. If it gets 70%+ of the way there, you need more defensibility. Add domain expertise, workflows, or data integrations that the base model can’t access. #### 3. Price Where the Puck Is Going Henrique introduced a brilliant pricing framework based on two dimensions: frequency of use and token consumption. The matrix: Low frequency + Few tokens = Free (no-brainer) High frequency + Many tokens = Cannot afford to offer free trials Low frequency + Many tokens = Free to try (onboarding friction is worth it) High frequency + Few tokens = Freemium sweet spot His team watches these metrics daily because token costs are dropping fast. What’s expensive today might be free next quarter. Action item: Map your features onto this matrix. Adjust your pricing tiers accordingly. Review monthly—the AI cost landscape shifts constantly. #### 4. Treat Prompts as Code This might be Henrique’s most important insight: Prompt engineering IS software engineering. “We noticed that our answer rate had dipped from, you know, let’s say 85 to 73 or something and it had dipped because the day before we had made a change on how to process part of the instructions. We had run a few tests, it had passed, but then once hundreds of people are using it at night in the US or whatever, clearly this wasn’t working. And so we basically had to roll back on Saturday.” — Henrique Cruz on why prompts need version control At Rows, prompts go through pull requests, testing environments, and automated evaluations. They version everything. They rollback when things break. One Saturday, they caught their answer rate dropping from 85% to 73% because of a prompt change—and rolled back immediately. The anatomy of their prompt (26,000 characters for their game analysis feature): Context: What the app knows about the user Instructions: The system prompt (broken into objectives, optimal solutions, and rules) Examples: Concrete demonstrations Reflection: When the AI checks its own work Action item: Stop treating prompts like ad-hoc text. Set up version control, A/B testing infrastructure, and automated evals. If you’re not treating prompts like production code, you’re doing it wrong. #### 5. Establish KAIzen Routines Henrique loves the Japanese concept of Kaizen—continuous improvement. He calls it “KAIzen” for AI products (clever, right?). “This is my favorite thing maybe that I do every week, which is actually reviewing the prompts of what the users are asking the AI… You try to get intuition about what are people using the product for. That’s the whole goal.” — Henrique Cruz on weekly prompt reviews Here’s what they review: Daily: Answer rate: Can we parse and execute what the AI returns? (They went from 86% to 95%) Prompts per day: Volume tracking (they’re 5x vs. their baseline) Cost per prompt: Currently under $0.08, but they know it can drop 90% Failed questions: What broke and why? Weekly: User conversations: Read actual prompts to understand use cases (his favorite ritual) New models: Test every new release from OpenAI, Anthropic, Google, Meta Model switching infrastructure: They keep their architecture model-agnostic so they can swap providers in minutes Henrique shared actual user prompts they see: people trolling them (”Who is LeBron James?”), asking for impossible features (”Can you see who doesn’t follow me on Instagram?”), wanting everything in German, and exposing bugs (”Where is it done?”—when the AI said “done” but did nothing). This weekly review gives them intuition about what users actually want versus what they thought they wanted. Action item: Block calendar time NOW—30 minutes daily for metrics, 2 hours weekly for prompt analysis. This isn’t optional surveillance; it’s how your AI product learns. ### The Numbers That Matter When asked about the impact of these routines, Henrique estimated: 80% of improvements came from big architectural changes 20% from daily kaizen tweaks But here’s the key: those daily reviews inform the big architectural decisions. You can’t make smart big bets without ground-level intelligence. “Maybe 80% of the improvements we’ve had over the last five or six weeks was a big architecture change… But these things are really important to get intuition about what’s working and not working.” — Henrique Cruz on the balance of daily tweaks vs. major changes ### The Honest Truth About User Feedback Henrique’s team debated adding thumbs up/down buttons. They decided against it. Why? It creates biased samples. Only motivated users click—either very happy or very angry. “We always said no because it’s going to really be a biased sample. A small percentage of people will click yes or click no and then maybe overindex on the people that said yes.” — Henrique Cruz on why they don’t use thumbs up/down feedback Instead, they look at behavior: Do users come back? Do they keep prompting? What does the conversation history show? Some users hit their AI 70 times in one day and never reply to emails. Clearly it’s working for them. Retention tells the real story. “Sometimes you have people who don’t reply to any email and use it, you know, had to use it like 70 times in one day prompting. Clearly, it worked for something for him or her.” — Henrique Cruz ### Build with Belief, Not Hope Perhaps the most striking moment came during Q&A. Henrique was asked how far ahead they bet on pricing dropping. His answer revealed their conviction: “We’re very close. Within 12-18 months, we’ll run most things on open-source models, and cost won’t be a problem.” “I think we’re very close… we already use an open-source model in the AI analyst for a small task and the open source model is basically free… we think that within the next 12-18 months we’ll be running most things on open source and then the cost is not really a problem.” — Henrique Cruz on betting on the future of AI costs They’re not optimizing for today’s costs. They’re building for a world where smart AI is essentially free. That changes everything about product strategy. ### Your Move Building AI products isn’t magic. It’s discipline. Henrique and the Rows team proved that with the right framework—picking the right problems, treating prompts as code, establishing daily rituals, and pricing for the future—you can build products that genuinely improve every single day. The AI landscape moves fast, but the fundamentals remain: understand your users, measure obsessively, iterate constantly, and always assume the technology will get better and cheaper. Start with one thing tomorrow: pick ONE daily metric to review. Answer rate, prompt volume, user retention—just one. Make it a ritual. Then add the weekly prompt reviews. The AI product that wins isn’t the one with the best model. It’s the one that gets 1% better every day while everyone else ships and forgets. Want to dive deeper? Henrique is active on X at @henrm_cruz, Substack , and he his the one who taught me the most about Growth and PMM on LinkedIn. You can try Rows at rows.com Here is his presentation! ## Why 97% of companies fail at product transformation Speaker: Chris Compston Read online: https://productlab.app/blog/2025/chris-compston Video: https://www.youtube.com/watch?v=wmK_YBQUhdg The brutal truth about going from project to product—and the 4-part framework that actually works Here’s what nobody tells you about product transformation: 97% of organizations trying to shift from project to product-thinking face significant implementation challenges. Even worse? Only 2% have actually operationalized product operating models at scale. I just watched Chris Compston (founder and consultant who helps organizations build great product operating models) break down exactly why this shift is so hard—and more importantly, how actually to make it work. If your organization is stuck shipping features instead of creating outcomes, this is for you. ### The shift everyone talks about (but few achieve) Let’s be honest. We all know the theory: Stop saying: “We shipped new features and met our deadline” Start saying: “Our customers can now achieve something new” Stop saying: “We never have data, so we’re guessing” Start saying: “We’re making evidence-based decisions” Stop saying: “Sales told us to build this” Start saying: “We know what customers value” Sounds simple, right? So why does almost everyone fail? ### The three reasons transformation is brutally hard 1. No actionable guidance Everyone encourages product thinking (books, conferences, podcasts, ChatGPT). But there’s a massive gap between “be more outcome-focused” and actually doing it within your organization’s unique context. 2. The economy is punishing Tech layoffs. Global instability. The AI revolution. You’re being asked to do way more with way less—fewer people, smaller budgets, zero room for “safe to fail” experiments. 3. Organizations are built for certainty Most companies operate on the manufacturing model: “We’ll deliver X things and make Y money.” They’re built for certainty, which makes them reactive rather than responsive to change. In these companies increasing efficiency is often the only lever they can pull to differentiate them from their competitors. ### The stat that should scare you Customer-focused organizations are 60% more profitable. And here’s the kicker: 50% of organizations say that in 5 years, 80% of their work will use product thinking. That means your competitors are making this shift right now. The question isn’t whether to change—it’s whether you’ll be in the 2% who actually succeed. ### The framework that changes everything Chris developed his approach after some significant learnings from his time introducing change Farfetch (a luxury fashion platform where a transformation affecting 1,000 people took 3-4x longer than predicted). Here’s what he learned: #### Part 1: Connect strategic intent to tactical change Don’t just say “let’s make more money” and “let’s be outcome-focused.” Instead: Identify specific organizational capabilities you need to build—things like: Iterative learning Evidence-based decision making Customer discovery Cross-functional collaboration Connect the high-level vision to concrete behavioral changes. #### Part 2: Balance resources and incentives You need BOTH: Readiness: Do people have the right skills, tools, time, and space? Motivation: What’s in it for them? Salary increases or role changes Learning budgets Recognition (internal and external) Growth opportunities Most transformations focus only on readiness and wonder why nothing sticks. #### Part 3: Cultivate long-term commitment Don’t do this: Top-down directives (”Go be outcome-focused!”) Narrow team focus (”Fix iterative learning in your squad”) Vague mandates (”We’re a product company now—figure it out”) Do this instead: Create cross-functional working groups that: Pull from different roles (product, design, data, engineering, finance, marketing) Include different seniority levels Work across multiple teams Focus on one organizational capability at a time (e.g., “improve iterative learning across the entire org”) This is how you get broad organizational impact instead of isolated team improvements. #### Extra: Show, don’t tell (or ask) When you’re stuck with external delivery commitments but trying to shift to product thinking: Don’t: Ask permission to experiment Do: Find capabilities you can shift quietly, prove the value, then show the results Experiment in safer spaces where you can demonstrate impact without high-stakes visibility. ### The real-world challenge (and solution) Someone in the audience asked the perfect question: “We have external commitments through next year. How do we shift to product thinking when we’re locked into deliverables?” Chris’s answer was brilliant: Your CPO (or top product leader) should be putting in the hard work to convince the exec team that outcome-focused work drives MORE business impact. They’re the ones who own the budget and the funding model. As a PM on a team, it’s not your responsibility to make this happen; however, you can support this shift. By starting small, experimenting with more outcome-focused work, changing how you produce discovery, or ensuring iterative learning is taking place. You can’t always predict how much revenue is going to be generated in a year’s time. You can, however, be more confident with that estimation if you’re using a funding model that allows you to build products and learn on a more regular basis. ### What to do right now Audit your organizational capabilities: Which ones need work? Where are the gaps between your strategic vision and tactical reality? Balance readiness and motivation: Don’t just focus on skills and tools. What’s actually motivating people to change? Start small and cross-functional: Pick ONE capability. Create a working group across teams and functions. Make visible progress. Show value before scaling: Find safe spaces to experiment. Prove the model works. Then expand. ### The bottom line The product shift isn’t optional if you want to lead. Customer-focused organizations are 60% more profitable. Your competitors are already moving. But transformation isn’t about reading the right books or attending the right conferences. It’s about connecting strategic intent to tactical change, balancing resources with incentives, and cultivating genuine commitment across your organization. Most importantly: It’s about being in the 2% who actually make it work. P.S. Chris is writing a book on this framework with co-authors Esra Yetis and Hugo Froes. You can learn more at maketheproductshift.com or connect with Chris on Substack: or LinkedIn. Check the Slides ## How Product Managers Survive (and Thrive) When Growth Isn't Enough Anymore Speaker: Bandan Singh Read online: https://productlab.app/blog/2025/bandan-singh Video: https://www.youtube.com/watch?v=EVtpwnDWpAQ How to survive as a PM when your company shifts from growth to profitability. When Bandan Singh moved from Gojek to Booking.com to Rivety, he thought product management was all about user growth and conversion optimization. MAUs, DAUs, experiment after experiment—that was the game. Then Rivety handed him full P&L ownership. “I told the company hey I’ve never handled a P&L are you sure you want me,” Bandan said at Productlab Conf 2025 in Berlin. They wanted him anyway. Either he’d learn something new or he wouldn’t last long at the company. He survived 3 years—and discovered that product management had been missing half the equation. At Productlab, Bandan shared the hard-won lessons from managing profitability while everyone around him was still obsessing over growth metrics. This wasn’t theory. This was a product leader who learned to kill products, redesign pricing, and make layoff decisions—all while keeping the product organization healthy. “Product management has just been too much about customer delight. Sustainable businesses are also about building a business at the end of the day.” — Bandan Singh ### The Trap Most Product Managers Fall Into Bandan’s career mirrors what’s happening across the industry right now. At Gojek (Southeast Asia’s competitor to Uber and Grab), he was the “cool growth guy.” Small tweaks in the customer journey brought more users. The company was preparing for an IPO. Topline was everything—valuation was key. Everything was about expansion, not efficiency. At Booking.com, he learned experimentation at scale. How a small button change could move conversion numbers. But he never connected his work to the bottom line. He had no idea what his actual contribution was to Booking.com’s profitability. “I would not know what led to true value for the company like what was my contribution in terms of the bottom line.” Then came Rivety, a B2B mobility payments company in Berlin. Full P&L ownership meant looking at parts of the organization where he had no direct authority—but was responsible for outcomes anyway. The learning curve was brutal. Pricing. Killing unprofitable products. Managing costs below revenue. Client-level profitability. And yes, layoffs. “The honest truth about a lot of hard decisions in companies—layoffs being one of them—they all come from a P&L decision.” Action item: Stop measuring your impact solely by user metrics. Start asking: “What’s the actual revenue impact of what I’m building?” ### Why Everything Changed (And Why It Matters for You) Between 2018 and 2025, Bandan watched the entire product management playbook flip. Money is no longer cheap. Companies that went public with “we’ll be profitable one day” promises are now scrambling to actually deliver profitability. Gojek, Grab, and dozens of others—they all said scale would solve everything. It didn’t. “Today if you look at Gojek and it’s true for a lot of companies that did IPOs recently now they’re all looking for profitability.” The macroeconomics are tough. AI is accelerating this shift—it’s about value creation but also about cost-saving. Retention becomes more important because it brings you closer to profitability. Customer lifetime value matters more than vanity metrics. Looking back, Bandan realized what he’d do differently at Gojek: track the cost of acquiring users. “We don’t know at what cost we are acquiring users. I mean you have given me this team, you have given me these engineering resources and UX. I work with them and I deliver it and I get some monthly active users. But I would rather say that can we look at the cost of acquiring those users so that at least at a user level I know the unit economics.” The shift for product managers: If you want to be successful—especially in senior product roles—you need to focus on business viability as much as customer delight. Maybe even overindex on business viability for a while. Action item: Calculate your user acquisition cost. If you’re in a product-led company and it’s “spread across the full product base,” work with your finance counterpart to estimate it. ### The Framework: 4 Levels of Looking at Profitability Bandan broke down the P&L into a simple framework that any product manager can use—even if you don’t have full P&L access. He showed a mock-up of what he was responsible for at Rivety. The key insight: different profitability metrics serve different purposes. #### 1. Revenue ÷ Time (Throughput) This ignores the cost of rails entirely. It’s pure throughput. Most product managers stop here. They celebrate when revenue goes up. But this metric is incomplete—it doesn’t tell you if you’re actually creating value for the business. #### 2. Operating Profit ÷ Time This gets blurred by corporate growth bets and cost allocations. It’s useful for CFOs, less useful for product teams making day-to-day decisions about what to build. #### 3. Net Profit ÷ Time Distorted by financing and tax considerations. Again, not the right metric for product decisions. #### 4. Gross Profit ÷ Time (The One That Matters) “The true speed at which the product engine creates value.” This is where product managers should focus. Gross profit strips away the noise and shows you: is this feature, product, or initiative actually creating value for the business? Bandan’s rule: Take revenue, subtract direct costs (COGS), and you get gross profit. That’s your North Star. “At a given point in time it’s important to know what is more important for the organization—is it gross margin or is it net? Depending on what the company is striving for, that is the more important metric.” Action item: Find your product’s gross profit metric. Work with finance to understand your direct costs. Start tracking gross profit ÷ time as your primary business health metric. ### The Numbers That Actually Matter When asked about how to start thinking about profitability, Bandan shared two calculations that changed how he prioritized work. Profit Velocity: Gross profit per week or per sprint. This tells you how fast your product is generating value. It’s not about the absolute number—it’s about the trend. Is profit velocity increasing or decreasing with each release? Payback Period: Build cost ÷ profit velocity. Let’s say you spend €100,000 over 3 months building a feature. Your product generates €50,000 in gross profit per month. Your payback period is 4 months. “This is a conversation starter on no two features are the same. Even though something might have low effort high impact after they launched, some features might take more time to pay back and some others will take less time.” The insight: not all “low effort, high impact” features are equal. Some have fast payback periods, others don’t. This changes how you prioritize. Action item: Calculate the payback period for your last 3 major features. Share this analysis with your leadership. Start using it in roadmap discussions. ### The Honest Truth About Getting P&L Access Bandan was brutally honest: getting profitability data in your company is hard. You’re not alone if you’re struggling with this. But here’s what he recommended for senior product managers and product leaders: During discovery: Ask not just “Is this valuable for users?” but also “Is this profitable for the business?” In prioritization: Rank initiatives by profit potential per sprint or per week—not just by user impact. Set the culture: Ask your leaders, your peers, for a view into profitability. Read company statements and reports if your company is public. Bandan used to read Booking.com’s income statements even though nobody asked him to. “I used to read Booking.com’s income statements to see how much it cost to build this accommodations business. I also started looking at how much of direct costs are going in, how much of what is driving Booking.com’s profitability, what is not. These are all amazing things that help me today.” At the time, nobody cared that he was doing this. He wasn’t responsible for it. But today, that self-education became his competitive advantage. In leadership conversations: Always connect your roadmap back to P&L levers. Show how your backlog choices affect costs and revenues. When Bandan encounters shared costs (like infrastructure spread across multiple products), he partners with finance and accounting to get answers. “You need good partners in the organization. For me it was not my product leader, they were my finance and accounting counterparts that would help me out with my questions.” One surprising thing: in his direct costs, he had to account for other product teams’ contributions. His product was part of a value chain. To deliver revenue, the full chain mattered—not just his piece. So 10% of another product group’s cost became part of his direct costs. Action item: Schedule time with your finance counterpart this week. Ask: “How is my product viewed in the full product portfolio?” Request a view into your product’s profitability, even if it’s just directional. ### What This Means for Your Career Profitability isn’t a dirty word. It’s resilience. When downturns happen—COVID, AI disruption, whatever’s next—profitable companies survive. They don’t have to cut half the team when VC funding dries up or interest rates rise. “Profitability ensures resilience. When there’s a downturn, when things are bad, you’re able to fight a lot of these downturns as a company.” Product managers used to be able to ignore all of this. Growth solved everything—or seemed to. But the playbook requires a new chapter now. Business viability and customer delight have to go hand in hand. And Bandan suggests overindexing on business viability for a while—especially if you’re moving into senior product roles where you’re closer to stakeholders making P&L decisions. The message isn’t to abandon customer delight. It’s to add the other half of the equation that’s been missing. “Creating sustainable businesses is not always about customer delight. It’s also about building a business at the end of the day.” Start with one thing: know your unit economics. Work backward from there. The product managers who thrive in the next decade won’t just be the ones who can optimize conversion and drive MAUs. They’ll be the ones who can build products that are profitable, resilient, and sustainable. Connect with Bandan Singh: Substack: LinkedIn: Bandan Singh ## How to Build Product Teams That Actually Care (And Keep Them That Way) Speaker: Vivek Juneja Read online: https://productlab.app/blog/2025/vivek-juneja Video: https://www.youtube.com/watch?v=_TY-ieKxL2c What 12 years of building teams at Zalando, On, and GS taught Vivek Juneja about creating outcome-driven organizations When Vivek Juneja walked into On—the Swiss sportswear brand backed by Roger Federer—as one of their first engineering managers, he faced a problem most product leaders know too well. The company was scaling fast, but they needed to build a product and engineering organization from the ground up. Not just any organization—one that could move at startup speed while competing with retail giants like Zalando, where Vivek had spent years leading teams. At Productlab Conf 2025 in Berlin, Vivek shared 12 years of hard-won lessons from building outcome-driven teams at three major retailers. This wasn’t theory from a consultant. This was tactical gold from someone who’s been in the trenches, building teams that ship. “Building these teams is hard,” Vivek told the 250+ product managers in the room. “But keeping them? That’s harder.” Here’s what actually works. ### The Three Pillars: Why Your Teams Need to Care About More Than Just Shipping Features Most product leaders say they want empowered teams. They want people who care about customers and outcomes, not just shipping features. But here’s the reality: saying you want it and actually building it are completely different things. Vivek distilled 12 years of experience into one framework that cuts through all the noise. He calls it the three pillars of care: → Care for your customers — not just features, but the people using them→ Care for the problem space — not the fancy new tool or API, but the actual problem you’re solving→ Care for product economics — how your work affects revenue, retention, and costs “If you don’t get anything from today,” Vivek said, “this slide is enough.” The framework is simple. The execution? That’s where most teams fail. Here’s how to actually build teams that demonstrate this level of care. ### Everyone Should Be Selling (Yes, Even Your Engineers) Here’s a story that Vivek keeps going back to, even years later. During his first month at On, an engineer did something remarkable. Without anyone asking, this engineer set up a lab at home. He bought every On shoe model he could get his hands on using company vouchers. Then he started measuring them—literally taking measurements and creating spreadsheets comparing products. Why? The engineer noticed something. On was labeling certain shoes as “wide fit,” but when he measured them, the data didn’t match up. So he posted his findings in Slack with detailed analysis and datasets. “Who is this person?” Vivek wondered. “Is he part of the physical product development team?” No. He was a web infrastructure engineer. Someone who could have easily said, “Not my job.” But here’s what made it powerful: this engineer cared enough about what customers were buying to flag a problem that the physical product team had missed. When the company was small enough to move fast, they took his feedback, went back to the factory, and fixed the wide-fit issue. Vivek has used this screenshot (with the engineer’s face hidden) in every talk he’s given since. It’s the perfect example of someone who gives a damn. Action item: Rotate your team members through sales for a day. Not to make them better at selling—but to see the customer’s eyes light up when they actually get value from your product. It changes everything. ### The Radical Ownership Toolkit: How to Build DRIs That Actually Own Outcomes Everyone talks about ownership. CEOs say, “I want my teams to own everything.” But what does that actually mean? And more importantly, how do you build it? Vivek introduced a concept from Apple’s playbook: the DRI (Directly Responsible Individual). Ken Kocienda, the designer who built the Apple keyboard, embodied this perfectly. He was the IC (individual contributor) who scheduled meetings with the VP of legal to go over details himself. No middleman. Just pure ownership. But here’s where most companies fail: they talk about ownership while operating in one of three modes: Consensus-driven: Everyone needs to say yes before anything moves Confrontational: Lock people in a room, let them argue, hope they come out aligned Hierarchical: Can’t decide? Push it to the boss “None of these is better than the other,” Vivek clarified. “Just know where you are. Know your operating environment.” Once you know where you are, you can build real ownership using guardrails instead of rules. The Guardrail Framework: → Constraints of time: Give people a deadline. Theory of constraints always works.→ Constraints of budget: Here’s your budget. Play within it.→ Constraints of scope: Here’s your playground. Build whatever you want within these boundaries.→ Philosophy and principles: Document your design principles so teams aren’t fighting about whether a button should be red or blue.→ Sacrificial vs. non-sacrificial beliefs: What can teams sacrifice? What’s sacred? “We care about customers—that’s a sacrificial belief,” Vivek explained. “You can negotiate on execution, but the core principle is non-negotiable.” Action item: Write down your team’s 3 non-sacrificial beliefs this week. What’s sacred? What absolutely cannot be compromised? Share it with your entire organization. ### What Doesn’t Work: The Firefighting Trap That Kills Ownership Here’s where Vivek got brutally honest about what keeps teams from staying outcome-driven. Years ago, a manager told him: “Vivek, you’re too quick at firefighting. Let the spark and the fire be there for some time. Let it burn. We’re all adults. You don’t have to just jump in and throw water at it.” Vivek’s instinct was to fix every problem immediately. No fires on his watch. But his manager was right. By solving every conflict, every decision, every debate, he was preventing his team from developing the muscles they needed to own outcomes. “Avoid the tendency as leaders to resolve conflicts—even decision conflicts about your product,” Vivek advised. “Let it not be resolved quite quickly, especially not through you. Just let it simmer. Let people debate, fight a bit, demonstrate the care, prove you wrong or right.” One team Vivek worked with created a rotating “conflict resolver” role. It was like a hat that moved around the team. If you weren’t the conflict resolver that week, you could fight as hard as you wanted in discussions without worrying about facilitating. The result? People actually started caring about decisions because they had to own them. Action item: For the next 2 weeks, resist the urge to step in and resolve team debates. Let your teams argue (productively) and come to their own conclusions. Watch what happens. ### The Question That Reveals Everything About Your Organization Near the end of the talk, someone asked Vivek the question every product leader struggles with: How do you create an appetite for ownership in a team that’s gotten comfortable? “It’s a good question,” Vivek said. “And I want to approach the answer by saying how I operate—and you would operate completely different, and that’s okay.” Then he broke down two approaches: Approach 1: The Incremental Shift If a team has been comfortable for years, any idea about radical ownership is like a virus to that system. They’ll fight it. So you find a medium-to-low-risk feature or capability—not a quick win, but something that gives people feedback on what ownership actually looks like. You celebrate it. You talk about it. You show people, “This is what ownership looks like.” Approach 2: The Band-Aid Rip Some leaders just say: “From tomorrow, you’re all going to be owning certain things. It’s absolutely okay if you don’t know that yet, but this is how we operate. We’re giving ourselves 4 months to settle in. From now on, everything we build has a single owner. No decisions by committee. If you want advisory, get advisors and move.” “Both work,” Vivek concluded. “It depends on where you are and how much influence power you have within the organization.” ### Keeping Teams: Why Sales Rotations Matter More Than You Think Building outcome-driven teams is one thing. Keeping them as you scale? That’s where most companies fail. Vivek’s advice: Sell, sell, sell. Have periodic rotations where your teams spend time in sales. Let them experience what it’s like to convince a hard customer. Let them see the customer’s jaw drop when they realize your product solves their problem. “One of my managers told me to just spend a day with sales,” Vivek recalled. “The excitement when you’re meeting a customer, when you convince them—it’s a whole different game when they actually get onboarded and start using your product.” And here’s the controversial part: don’t fix all problems. Let some fires burn. Let adults figure it out. Final action item: This quarter, have every product manager, designer, and engineer on your team spend 1 day shadowing sales or customer success. Not observing—actively participating. Watch how it transforms their perspective on what you’re building. ### The Bottom Line Building outcome-driven teams isn’t about copying Google’s playbook or reading another framework doc. It’s about building teams that demonstrate care—for customers, for the problem space, and for product economics. It’s about giving people guardrails, not rules. It’s about letting fires burn so people learn to own their decisions. It’s about rotating engineers through sales so they understand what they’re actually building for. As Vivek put it in his closing: “In the world of agentic systems and AI products, what will not change is building teams that can move fast and build towards outcomes.” The tools change. The fundamentals don’t. ### Connect with Vivek Vivek Juneja has led product and engineering teams at Zalando, On, and GS, building outcome-driven organizations from the ground up. He’s also passionate about education and teaching, creating content to help product leaders build better teams. → LinkedIn: Vivek Juneja→ Substack: ## Why Most Companies Are Asking the Wrong Questions About AI Speaker: Elena Leonova Read online: https://productlab.app/blog/2025/elena-leonova Video: https://www.youtube.com/watch?v=MuLWQtzoUsc The Four Strategic Frameworks That Stop You From Building AI Features Nobody Uses. When Elena Leonova walked onto the Productlab Conf stage in Berlin, she opened with a poll that made the room shift uncomfortably in their seats. “Am I the only one who’s been asked to do AI without a clear product strategy?” Nearly 70% of hands went up. As Chief Product Officer at Spryker (a B2B digital commerce platform serving large enterprises) and coach to 100+ product managers, Elena has spent years watching companies struggle with the same problem: everyone’s being told to “do something with AI,” but nobody knows what that something should be. The result? Teams running around like headless chickens, shipping AI features that get 20% adoption rates, and wondering why their revolutionary technology feels more like expensive window dressing. Elena’s talk wasn’t about prompts, tokens, or the latest model releases. It was about the strategic frameworks that determine whether your AI product becomes essential or gets killed in the next budget review. ### The Real Problem: We’re Not Just Lacking Strategy—We’re Asking the Wrong Questions Here’s what Elena has observed across dozens of companies: product teams are so focused on “Should we do AI?” that they never ask the question that actually matters. The wrong question: “Should we add AI to our product?” The right question: “What game are we playing, and what context demands our next move?” Elena shared four brutal truths that most product leaders don’t want to admit. These aren’t theories—they’re patterns she’s seen repeatedly in established companies trying to evolve in the AI era. ### BRUTAL TRUTH #1: Hype ≠ Strategy The symptom: Your leadership team comes back from a conference buzzing about AI. Within a week, you’ve got a mandate to “AI-ify” everything. No clear outcomes. No defined value. Just AI for AI’s sake. The reality: AI adoption hinges on clear strategic intent, not just excitement. Elena breaks down what real strategic clarity looks like: → Point A: Where are you now? (Current state with hard numbers) → Point B: Where are you going? (Target state with measurable outcomes)→ 3-5 bold bets: Specific hypotheses with customer proof “You need to still have a clear strategy about what kind of change you are trying to create and why,” Elena emphasized. “You need to understand what kind of value you’re trying to create and make sure your customers want that value before you start thinking about AI features.” Most companies skip Point A entirely. They can’t articulate their current baseline, which means they have no way to measure if AI actually improved anything. Action item: Before your next AI roadmap discussion, force your team to complete this sentence: “We will know our AI investment succeeded when [specific metric] moves from [current number] to [target number] because our customers can now [specific outcome].” ### BRUTAL TRUTH #2: AI Labels ≠ Value The symptom: You slap “AI-powered” on existing features and wonder why adoption doesn’t change. The reality: Simply adding “AI” to existing capabilities won’t inherently create value without a defined purpose. Elena introduced a framework that cuts through the noise: The Four Innovation Levers. Every successful product innovation moves the needle on at least one of these: #### 1. Make it Cheaper (Cost ↓ / Margin ↑) → Lower costs for customers or reduce your cost to serve→ Example: AI-powered translation replacing human translators—still involves people for quality checks, but costs drop 60% #### 2. Make it Faster (Cycle Time ↓ / SLA ↑) → Compress time-to-outcome for customers→ Example: Website builder that uses AI for layouts only wins if it actually gets customers to launch faster—not just make prettier pages #### 3. Make it Easier (Friction ↓ / Conversion ↑) → Remove steps, simplify workflows, reduce cognitive load→ Example: Auto-filling forms based on previous context instead of making users type again #### 4. Make it Smarter (Relevance ↑ / AOV ↑) → Personalize experiences, anticipate needs, deliver context→ Example: Recommendations that actually increase average order value, not just look clever Elena ran a live poll asking which levers attendees were targeting with their AI features. The responses were scattered—compliance, adoption, data quality, strategy. That’s the problem. When you don’t know which lever you’re pulling, you can’t measure if it’s working. “When you implement an AI feature, you’re not just throwing an AI label on top of the feature,” Elena explained. “The value itself needs to be absolutely clear to the customer. You’re actually trying to move the needle on one of those four levers.” Action item: Audit every AI feature on your roadmap. Force each one to declare which lever it’s pulling and what the measurable outcome is. If a feature can’t clearly articulate its lever, kill it or redesign it. ### BRUTAL TRUTH #3: Innovation ≠ Always the Right Answer The symptom: Everyone wants to “innovate” with AI, build the revolutionary product, change the game. Meanwhile, your core business is bleeding customers because basic workflows are slow. The reality: Understanding if your goal is evolution (efficiency) or revolution (reinvention) is critical. “We need to innovate” isn’t a strategy. This is where Elena’s framework gets surgical. She introduced a critical distinction most teams never make: #### What Game Are You Playing With AI? EVOLUTION: Optimize the Current Game Outcome focus: Cheaper, Faster, Easier Who: Existing customers, current segments Proof to seek: Repeatable efficiency lift, quality holds Good when: Low readiness, short runway, high compliance risk Risks: Incrementalism, competitors leapfrog you REVOLUTION: Change the Game Outcome focus: New mechanism, “smarter” behavior Who: Early adopters, new segments Proof to seek: 10x task completion, new usage loops Good when: High readiness, long runway, exec sponsor appetite Risks: Adoption risk, longer time-to-value, higher costs “Most of the time companies are lacking a successful strategy for how to actually make bets,” Elena said. “With AI, you need to start making even bolder bets, but at the same time you need to have a strategy behind it.” Elena emphasized that established businesses—companies that have been around for decades or centuries—are in a fundamentally different position than AI-native startups. They have existing products, existing customers, and existing revenue streams. “We’re not talking about the new startup that is starting from scratch. They have existing products with existing customers and that’s where the majority of their money is coming from. So you have to continue to optimize it.” But here’s where it gets interesting. You can’t just pick one game and ignore the other. ### BRUTAL TRUTH #4: Focused Bets > Spray-and-Pray The symptom: Your team is running 15 AI experiments with no clear success criteria, no kill dates, and no portfolio strategy. The reality: Random experimentation without clarity, alignment, and a disciplined portfolio leads to poor outcomes and resource drain. Elena introduced a betting framework adapted from portfolio management theory: The 70-20-10 Model. But here’s the twist—you can’t apply the same percentages to every company. #### Place Your AI Bets Based on Your Context: CORE (Optimize Existing): → Double down on high-performing products→ Systemize marketing campaigns→ Automate repetitive tasks→ Invest in customer retention ADJACENT (Expand Category): → Find adjacent customer needs→ Pilot new versions of your offer→ Enter nearby markets→ Form strategic partnerships TRANSFORMATIVE (Bold Moves): → Ask “If we were to completely change how XYZ is done, what would we do?”→ Ask “What would make us obsolete?”→ Set long horizons (don’t expect immediate adoption) #### But the allocation changes based on your context: Venture-Backed (Chasing a round)? → Bias Revolution for narrative and optionality while managing runway→ 20/30/50 (Core/Adjacent/Transform) Public Company (Margin pressure)? → Bias Evolution—operational efficiency is oxygen for stock performance→ 80/15/5 (Core/Adjacent/Transform) Profitable & Private (Sustainable growth)? → Run a barbell strategy—protect core, fund bold bets→ 60/20/20 (Core/Adjacent/Transform) “The thing is that you cannot apply this model to any company,” Elena warned. “It depends. Context decides.” When asked about the biggest challenges teams face with AI strategy, the Productlab audience responses were telling: compliance, adoption, data quality, capacity, resources, strategy itself. Elena’s point: “There’s no consistent answer that people have. Pretty much across the board, people talk about that yes, everybody has a mandate to go and do AI, but at the same time there’s like hundreds of reasons why—on the surface it seems like other companies are really successful, but most of the time they are not. They’re also struggling to understand what they actually do.” The critical rule: Every bet must have: → Success KPI→ Guardrail KPI→ Kill/Scale date→ Quarterly rebalancing “Don’t ask ‘Should we do AI?’ Ask: What’s the right balance for our context—now? Then start making your decision.” ### From Theory to Action: What to Do Monday Morning Elena closed with a challenge to the audience: change the conversation entirely. Stop asking: Should we do AI? Are our customers ready for AI? What AI features should we build? Start asking: What is the right context for my company? (Venture-backed? Public? Private?) What game are we playing? (Evolution or Revolution?) Which value lever are we pulling? (Cheaper? Faster? Easier? Smarter?) What’s our betting portfolio? (70-20-10? Something else?) “Go with that to your leadership team, which will make you more strategic than anybody else,” Elena said. “And I can guarantee you that if you do that, you’re going to be definitely seen as more strategic, but also you’re going to be speaking the language that business understands more than just if you were talking about a specific product feature.” ### The E-Commerce Reality Check During Q&A, Elena was asked about the biggest AI use case in e-commerce given her Spryker background. Her answer was refreshingly honest: “To be honest, I don’t know what’s going to be the next thing because the world is changing quite a lot.” She explained that large B2B manufacturers (like Siemens Healthcare, which makes medical devices) aren’t easily adopting AI tools due to compliance concerns. But smaller e-commerce companies face a different threat: discovery moving entirely to LLMs. “It’s easy to go to ChatGPT and say, ‘I have these pants, I have these shoes, find me a t-shirt or a top that is going to go with this.’ I don’t need to go and search even on Amazon for that because maybe that product is not there.” The takeaway? The channel itself might change, but the strategic frameworks remain constant. You still need to know: Who your customer is What value you’re creating for them Whether they’re willing to pay for that solution Whether you can execute it “You should not be thinking from the types of product. You need to be thinking about who your customer is. What kind of value do you want to give them? Actually test for their willingness to pay for that solution and then are you actually able to execute for it.” ### Your Move Elena’s frameworks aren’t theoretical constructs—they’re battle-tested approaches from someone who’s been in the trenches building products for established enterprises that can’t afford to chase every shiny AI object. The difference between companies that succeed with AI and those that burn cash on unused features comes down to one thing: strategic clarity. You can have the best models, the smartest engineers, and the biggest budget. But if you don’t know what game you’re playing, which lever you’re pulling, and what context you’re operating in, you’re just another team running around with your hair on fire wondering why “AI” isn’t working. Start with one thing this week: Get brutal clarity on whether you’re playing evolution or revolution. That single decision will cascade into better bets, clearer priorities, and AI products people actually want to use. The brutal truth? Most companies won’t do this work. They’ll keep spray-and-praying, keep slapping “AI” labels on features, and keep wondering why adoption stays stuck at 20%. Don’t be most companies. This article is based on Elena Leonova’s talk “The Brutal Truth About AI Product Success” delivered at Productlab Conf 2025 in Berlin. The conference brings together 250+ senior product managers and leaders to tackle the hardest challenges in product management. Connect with Elena Leonova: LinkedIn - Instagram ## AI agents are coming for your business model Speaker: Hayder Schneider Read online: https://productlab.app/blog/2025/hayder-schneider Video: https://www.youtube.com/watch?v=crGWaOvGwO8 Dr. Hayder Schneider's framework for moving from AI theater to P&L impact—before agents eat your business model When Dr. Hayder Schneider asked the 250+ product managers at Productlab Conf Berlin whether AI would be the biggest opportunity or the biggest bubble of our time, the room split down the middle. That’s the wrong question. AI is getting cheaper, better, and faster every single day. Inference costs have dropped 95% in 18 months. Open-source models like Qwen and DeepSeek rival proprietary ones. The question isn’t whether AI matters—it’s what you’re going to do about it. Hayder, Head of Product at Aleph Alpha (a B2B enterprise GenAI company that raised $500M in 2023), spent years building AI products for industrial-scale customers. His talk cut through the hype with a brutally simple framework: two questions that determine whether your business model survives the agentic era. This isn’t theory. It’s what’s already happening. ### The Gen AI Paradox (And Why 80-95% of AI Projects Fail) Here’s the uncomfortable truth: between 80% and 95% of companies report no measurable impact from their GenAI initiatives. Why? The Gen AI Paradox. Companies face two bad options: Horizontal use cases (ChatGPT, Copilot, Gemini): Scale broadly but generate shallow, unmeasurable value. Nobody can actually quantify how much value Gemini added to their organization. Vertical use cases: High value, function-specific, industry-specific—but they’re stuck in pilot hell. Getting them to production is brutally hard. Agentic AI might be the way out. But first, you need to understand what agents actually are. ### What AI Agents Actually Are (And What They’re Not) Agents are NOT: → A set of prompts → Tool calling or MCP protocols → Just reasoning models → Programmatic loops → Simple process automation An AI agent is a system that uses reasoning and planning to solve complex, multi-step problems by using tools and memory to act and interact with its environment. The foundation? An augmented LLM with three critical capabilities: Retrieval — Access to information Tool calling — Ability to use external systems Memory — Retention across interactions Think of it as upgrading from a calculator to a research assistant who remembers your last conversation, can Google things for you, and makes decisions about what to do next. Now here’s where it gets interesting for business models. ### The Agentic 4 Moves Matrix: Your Strategic Map Hayder’s framework is deceptively simple. Just two questions: 1. WHAT kind of work are you doing? → Optimize (improve existing processes) → Innovate (create new capabilities) 2. WHERE are you doing this work? → Internally (inside your company) → Externally (customer-facing, market-facing) Plot them on a 2x2 matrix, and you get four strategic moves. ### Move 1: Optimize Internally — From AI Pilot to P&L Impact The trap: 90% of GenAI pilots never reach production. Why they fail: The six predictable traps: → 10%-not-10X Trap: Marginal improvements don’t justify the complexity → Unit Economics Trap: Token costs kill profitability at scale → Prototype-to-Production Trap: What works in a demo breaks with real users → Solution-without-a-Problem Trap: Building AI for AI’s sake → Users-don’t-accept-mediocrity Trap: 85% accuracy isn’t good enough when humans expect 99% → Maintenance-and-ongoing-cost Trap: Nobody budgeted for the care and feeding The fix: Don’t optimize for today’s costs. Optimize for where costs will be in 12-18 months. Inference costs are dropping 50%+ annually. Build your business model assuming AI will be nearly free. Hayder’s team at Aleph Alpha already uses open-source models for specific tasks where cost is essentially zero. They’re betting on a world where smart AI costs nothing—and building accordingly. ### Move 2: Innovate Internally — Out-Strategize the Competition This is where things get sneaky. Case study: Anthropic’s multi-agent research system fully automates market research, competitor analysis, investment trends, technology trends, and report generation. What used to take a team of analysts 2 weeks now runs overnight. Continuously. The opportunity: Most companies think of AI as customer-facing. The biggest gains right now are internal—using agents to do work that never had software solutions before. Examples: → Competitive intelligence that updates daily → Automated compliance monitoring across regulations → Supply chain optimization that responds to real-time disruptions The companies winning here are those that treat agents as autonomous team members, not just fancy autocomplete. ### Move 3: Optimize Externally — Separating Signal from Noise in Sales and Marketing The market is shifting from human-first to AI-first customer experience. AI agents are starting to handle end-to-end customer journeys: → Voice and chat interactions → Human-like, personalized experiences → Proactive service (reaching out before problems escalate) The business model shift: If an agent can handle 95% of customer interactions, what happens to your per-seat SaaS pricing? Companies that priced on seats, users, or usage are scrambling. The winners are shifting to outcome-based pricing—but that only works if you can attribute outcomes to actions. If you can’t measure the value your product creates, you can’t price for outcomes. Hayder’s warning: Start experimenting with outcome-based pricing now, even if it’s messy. The companies that figure this out first will have a massive advantage. ### Move 4: Innovate Externally — Building Your Agentic Advantage Here’s where Hayder got provocative: AI agents will disrupt SaaS. Think about his Berlin trip example: Current reality: → Hotel: Open travel site, compare prices manually, book → Train: Switch to DB app, re-enter dates, check schedules, book → Weather: Open weather app → Clothes: Decide outfit, check forecast again → Food: Research restaurants in separate apps → Family: Set reminder to call kids → Wellness: Make sure everything’s ready for tomorrow Agentic future: → “I’m going to Berlin September 18-19 for a conference” → Agent books hotel, train, checks weather, suggests restaurants, sets reminders, handles everything end-to-end You went from 7 apps to 1 conversation. Why this matters for SaaS: The era of selling “engagement” is ending. The era of selling “outcomes” is beginning. Hayder’s framework for the transition: Old model: App/Solution → Engagement New model: Orchestration → Outcome The question every SaaS company must answer: Will you OWN the orchestration layer, or will you become a commodity API that someone else’s agent calls? Google, Microsoft, and Amazon are all racing to own this layer. So are hundreds of startups. The winners will capture the transaction layer—where all the revenue flows. ### The Nuance: Which SaaS Survives? When pressed on his “AI agents will kill SaaS” claim, Hayder got more specific. High risk: → Marketing tools (easy to automate) → Basic customer support platforms → Horizontal productivity tools → Anything where adoption is already high today Lower risk: → Industrial/manufacturing software (physical operations) → Deep vertical software with complex workflows → Systems that integrate with messy, legacy infrastructure → Enterprise tools where compliance and security dominate But even in “safe” categories, the question remains: Are you building for engagement or outcomes? ### 5 Steps to Future-Proof Your Business Model Hayder’s closing advice was tactical: 1. Face reality → Name both the risks and the upside. Don’t pretend agents aren’t coming for your market. 2. Obsess over outcomes → Customers will expect end-to-end results. “We helped you do the thing” beats “We gave you a tool to do the thing.” 3. Build the bench → You need world-class AI engineering + AI-fluent leadership. This isn’t optional anymore. Hayder estimates there are maybe 50-100 people in Germany who can properly run an inference service at scale. Hire them now. 4. Nail economics → Understand token economics and outcome-based pricing. Run experiments today with hybrid pricing models. 5. Think laterally → What OTHER problems do your customers have? If you own a travel app, maybe you should own the weather, restaurant, and reminder layer too. The companies that survive won’t be the ones with the best AI. They’ll be the ones with the best business model for an AI-first world. ### The Hard Questions on Pricing One audience member asked about the challenges of outcome-based pricing for agents that run through dozens of steps. Hayder’s answer was sobering: “I wouldn’t recommend anyone just going for outcome-based pricing because it’s probably the hardest pricing you can nail.” Why? Two reasons: Attribution problem: You need to prove this outcome came from this action. That’s hard when agents are making 50 decisions in a chain. Value quantification problem: Even if you can attribute the outcome, what’s it worth? Not always straightforward. The bigger issue? Agent safety and control. Agents can go rogue—there are documented cases of agents with root access deleting databases. You can’t price by outcome if your agent might accidentally nuke production. The solution isn’t avoiding agents. It’s building the risk management, security, and monitoring infrastructure BEFORE scaling. ### The Dependency Question (And Why Open Source Matters) Another critical question: What about dependency on LLM providers like OpenAI? Hayder’s take: The risk is real. Political risk: An administration could cut off API access to certain countries. You’re done. Economic risk: You have no control over pricing. API costs could 10x overnight. Data risk: You’re sending potentially sensitive data to third parties with unclear processing practices. His recommendation: Own as much of the stack as possible—but recognize this comes at massive cost. Aleph Alpha runs their own models. The infrastructure cost is substantial. The talent cost is even higher. But for enterprise customers in regulated industries (banks, automotive, industrial), it’s the only acceptable path. For everyone else? Build model-agnostic architectures. Be able to swap providers in minutes, not months. Test every new model release from OpenAI, Anthropic, Google, and Meta. Stay flexible. Within 12-18 months, Hayder predicts most companies will run most tasks on open-source models where cost essentially disappears. Build for that world. ### Your Move The agentic AI era isn’t coming—it’s here. The question isn’t whether agents will reshape business models. The question is whether you’ll be ready. Start with the framework: → Map your current business onto the 4 moves matrix → Identify which quadrant represents your biggest opportunity → Identify which quadrant represents your biggest threat → Start experimenting NOW The companies that figure out outcome-based pricing, agent orchestration, and model-agnostic architectures in the next 18 months will have a 5-year head start on everyone else. The era of engagement is ending. The era of outcomes is beginning. Which side of that transition will you be on? ### Connect with Hayder Dr. Hayder Schneider is Head of Product Management at Aleph Alpha, where he’s building enterprise GenAI products at scale. He specializes in deep tech and AI/ML platform products. Connect with him on LinkedIn to follow his insights on agentic AI and business model innovation. ## How HelloFresh and UXCam moved from AI productivity hacks to systemic transformation Speaker: Silvanus Alt & Ebrahim Kargar Read online: https://productlab.app/blog/2025/silvanus-alt-ebrahim-kargar Video: https://www.youtube.com/watch?v=4JlGdq1ZaBo The hidden playbook for turning 40% efficiency gains into organization-wide AI capability Bottom line up front: Most companies are stuck using AI as fancy productivity tools. The real winners are evolving from individual efficiency gains to systemic organizational transformation—and it requires a completely different playbook than you think. At Productlab Conf, Ebrahim Kargar (Staff PM AI/ML at HelloFresh), Silvanus Alt (CEO of UXCam), and Ahmed Wafaey (Sr. Director at Delivery Hero) revealed what actually separates AI adopters from AI transformers. Here’s what they learned the hard way. ### The Silent AI Epidemic: Your Team Is Hiding Their Best Work The most shocking insight? While only 20% of workers admit to using official AI tools, over 40% are secretly using AI and reporting massive performance gains privately. Why the secrecy? Ebrahim identified the fear-based thinking destroying your AI transformation: “If I reveal AI helped me, managers will stop respecting my work” “Productivity gains might lead to layoffs for me or my colleagues” “If they know I’m more productive, they’ll just expect more work” This underground AI usage creates a fundamental barrier to systemic thinking. You can’t transform an organization when everyone’s pretending AI doesn’t exist. The fix: Create psychological safety around AI usage. Leaders must explicitly communicate that AI is a capability enhancer, not a job replacement signal. ### The Real Transformation Starts After Productivity Everyone starts their AI journey the same way: 40% efficiency improvements, faster documentation, time savings. But here’s what separates the winners: HelloFresh’s evolution: They reduced recipe creation from 18 weeks to 1 week using GenAI. But the transformation wasn’t the time savings—it was that this fundamentally changed their entire product development process and what they could now experiment with. The pattern: Individual productivity gains → Changed workflows → New organizational capabilities → Different strategic possibilities As Ebrahim put it: “The biggest piece of cake is focusing on effectiveness. Figuring out what is worth building, testing and validating hypotheses as fast as possible, having insights that none of your competitors has.” ### The AI-First Build Process That Actually Works UXCam’s radical approach: Silvanus shared how they transformed their entire development process: Their Build Formula: Rapid prototyping with AI (Lovable, etc.) → Ship testable versions fast Gather feedback → Iterate in hours, not weeks Hand off full spec → Engineering hardens and ships “Vibe Coding” with AI IDEs → Keep improving agent usage The game-changer: They explicitly document which AI model they used and why (reasoning vs. coding vs. speed) in every spec, PR, and demo. Why this matters: Transparency kills the fear. When everyone sees AI as a tool choice—like picking React vs. Vue—it becomes normal infrastructure, not a secret weapon. ### The Cultural Traps That Kill AI Transformation Trap #1: Over-trusting eloquent LLM answersConfident AI outputs without context = wrong answers. The fix? Human-in-the-loop on high-impact steps, golden data sets for validation. Trap #2: Expecting prod-ready code on day oneDevelopers try agents once, get messy output, give up. The fix? Treat AI agents like junior developers—they need guidance, iteration, and clear constraints. The real transformation: It’s not about better technology—it’s about building organizational capabilities that compound over time. ### The Framework: How to Actually Score AI Initiatives Stop chasing impressive demos. Silvanus shared UXCam’s evaluation framework: Value × Feasibility × Trust Then get razor-specific: One job → What exact problem does this solve? One metric → How do we measure success? Clear tests → What are our kill metrics per stage (Explore → Pilot → Scale)? Critical insight: Embed AI engineers with product teams. Optimize for customer value, not flashy technology demos. As Silvanus said: “Start narrow; scale what works. Choose and benchmark models for the job and be ready to swap.” Resist AGI scope creep. ### What to Actually Measure (Beyond Time Savings) Forget efficiency metrics. Here’s what HelloFresh tracks for AI transformation: Efficiency layer: Time savings on specific tasks Effectiveness layer (where the money is): Speed from insight to action Number of valid AI-generated hypotheses Validated prototypes Prototype-to-build ratio Outcome metrics: Cost savings, customer engagement, ROI The insight: AI’s biggest opportunity isn’t automating existing workflows—it’s figuring out what’s worth building in the first place. ### The 3-Phase Practical Implementation Roadmap Phase 1: Foundation Identify high-impact workflows for immediate productivity gains Build cross-functional AI teams with clear mandates Establish ethical AI guidelines and measurement frameworks Phase 2: Systemic Integration Evolve successful AI implementations into decision-making processes Train leaders to think systemically about AI capabilities Scale proven approaches across different organizational contexts Phase 3: Transformation Embed AI thinking into strategic planning and operations Build organizational learning loops that improve AI implementations over time Develop AI-native approaches to new challenges ### The Leadership Requirements (This Is Non-Negotiable) Ebrahim’s observation: Transformative leaders go beyond saying “AI is important.” What actually works: Paint a clear, compelling vision of what work will look like in the future Revisit expectations for different crafts and seniority levels Create incentives and a safe environment for people to learn and upskill Mandate adoption with resources, not just directives The three pillars of AI transformation: Workforce upskilling → Guidelines, resources, training Infrastructure and tooling → Data foundation, tool integration, clear policies Culture → Constant experimentation to find best use-cases for your org ### The Bottom Line Systemic AI transformation isn’t about implementing better technology. It’s about building organizational capabilities that compound over time. The winners understand three things: Productivity is the appetizer, not the meal → Real transformation changes decision-making and strategic capabilities Culture change is harder than technical implementation → Invest in psychological safety, transparency, and internal champions Measure outcomes, not adoption → Track business impact, not tool usage The gap between AI adopters and AI transformers isn’t technical sophistication—it’s organizational courage to fundamentally rethink how work gets done. What’s your organization’s biggest barrier to AI transformation? Reply with your experience—I’m collecting insights for a follow-up piece on the most common implementation failures and how to avoid them. ### About the Speakers This fireside chat brought together three product leaders at the forefront of AI transformation: Ebrahim Kargar is Staff Product Manager AI/ML at HelloFresh, where he champions and drives GenAI transformation across the organization. He’s leading AI/ML initiatives at scale—from supply chain optimization to personalization—and recently reduced recipe creation time from 18 weeks to 1 week using GenAI. Connect with Ebrahim on LinkedIn → Silvanus Alt, PhD is Founder and CEO of UXCam, an AI-powered analytics platform serving 37,000+ products worldwide. Under his leadership, UXCam transformed into an AI-first organization, implementing “Vibe Coding” practices and launching Tara, their AI Product Analyst that connects product and engineering data to reveal not just the what, but the why behind user behavior. Connect with Silvanus on LinkedIn → Ahmed Wafaey (Moderator) is Senior Director of Product Management for Global Quick Commerce at Delivery Hero and writer at Product Voyagers, where he explores the intersection of product strategy, AI transformation, and operational excellence at scale. His insights on moving beyond productivity wins to systemic transformation set the foundation for this discussion. Connect with Ahmed on LinkedIn → | Read Product Voyagers → --- # Contact and canonical URLs - General contact: team@productlab.app - Company and group packages: team@productlab.app (subject: Custom Company Package) - Website: https://productlab.app - Productlab Conference 2027 (8-10 September 2027, Berlin) pre-registration: https://productlab.app/waitlist - Leaders Studio 1:1 coaching: https://productlab.app/leaders-studio-coaching - Leaders Circles: https://productlab.app/leaders-studio-circles - Blog (talk write-ups): https://productlab.app/blog - Newsletter: https://newsletter.productlab.app - Community events: https://luma.com/productlab-community - LinkedIn: https://linkedin.com/company/productlab3 - YouTube: https://www.youtube.com/@ProductlabCommunityConference - Instagram: https://www.instagram.com/productlab_conf/ - This file: https://productlab.app/llms.txt Last updated: 2026-10-06. For anything not covered here, see https://productlab.app or write to team@productlab.app.