---
title: Just Product 2026, Main Stage takeaways
conference: Just Product 2026, Munich, 16 October 2026, ARRI Cinema
organiser: Product Masterclass (Sebastian Borggrewe, Thomas Hartmann)
talks: 8
language: en
---

# Just Product 2026: Main Stage takeaways

Just Product 2026 ran on 16 October at the ARRI Cinema in Munich. The Main Stage carried eight talks under three themes: AI Mindset Shift in the morning, AI End-2-End PM Workflows after lunch, and AI Transformation and Leadership in the afternoon, with Martin Eriksson opening and Simina Pasat closing. The AI Workshop Stage ran four hands-on sessions in parallel. This page documents the Main Stage: every talk as video with chapters, the highlight moments as short clips, a written summary, the key takeaways, the slides as PDF, and everything as Markdown you can hand to your own agents.

Read across the eight talks and a shared argument appears. Nobody on stage was selling AI as a feature. Everyone had already lived through the moment when engineering got four or ten times faster, and everyone was reporting what broke next. The patterns below are the editorial summary of that day. Each one links to the talks that made the point.

## The five patterns of the day

### Building got cheap. Judgment did not.

Martin Eriksson opened with the question that replaced 'can we build it': should we build it. Simina Pasat closed with its mirror image: it has never been easier to build something nobody wants, because the old cost of engineering used to force the boring questions early. In between, Dave Killeen admitted that taste is no longer the moat but the bottleneck, Faith Forster showed six ways teams skip the thinking because the build is so fast, and Felix Geelhaar warned that a roadmap that empties faster also ships every bad idea still on it. The job of product did not shrink. It moved up the stack.

Talks: Martin Eriksson, Simina Pasat, Dave Killeen, Faith Forster, Felix Geelhaar

### The bottleneck moved. It is now product.

Verena Graber told the story most concretely: medflex adopted agentic engineering, the backlog that had been five kilometres long was empty by Friday, and her chat filled with developers asking what to do next. Felix Geelhaar generalised it: AI speeds up the wiring an organisation already has, so 42 issues become 420 at review and the natural reflex is an agent that stamps 'looks good to me'. Dave Killeen called it the lemming problem, remove one constraint and the next one pops up, last week at CI. Simina Pasat saw the same inside GitHub, where one fast team stress tested release processes, reviews and a comms team built for a slower pace. The practical answer from all four: find where work piles up now, redesign around it, then look again.

Talks: Verena Graber, Felix Geelhaar, Dave Killeen, Simina Pasat

### Context is the new work.

Gopika Mohanan put it as a hiring scene: five new colleagues arrive on Monday, fast, tireless, confident and sometimes wrong, and the question is not how autonomous they should be but what context, roles and decision rights they get. Faith Forster's 'abandoned junior' is the same point from the failure side: hand an agent work without the rationale and conclude that AI does not work. Thomas Hartmann showed the operating system Product Masterclass built from exactly that insight, tools connected through APIs, business context, and about 46 skills that encode the processes. Dave Killeen writes down his 'truths' about the next twelve months and gives them to every team and agent. Felix Geelhaar wants that context in a queryable company layer, not in markdown files piling up in repos.

Talks: Gopika Mohanan, Faith Forster, Thomas Hartmann, Dave Killeen, Felix Geelhaar

### Measure signal, not activity.

Faith Forster's first trap was tokenmaxxing: Uber ranked teams by AI usage, burned the year's budget by April, and could not name what customers got. Felix Geelhaar proposed replacing time to market with time to signal, how fast a customer reaction reaches the team that built the feature, against a survey finding that only 4 percent of teams can trace a roadmap decision to customer evidence. Simina Pasat traced how value moved from access to usage to impact and pricing followed, from per seat to per token to Intercom's 99 cents per resolved ticket, the same Intercom story Martin Eriksson used as the example of a company that scrapped its roadmap over a weekend. Dave Killeen was honest that he has no idea who uses Dex, and that love is not data.

Talks: Faith Forster, Felix Geelhaar, Simina Pasat, Martin Eriksson, Dave Killeen

### Leadership became a junior skill. Reviewing became the job.

Gopika Mohanan's core claim: working with agents forces every product manager, including the most junior, to practice delegation, decision boundaries, judgment and teaching from day one, skills people used to wait years to need. Thomas Hartmann described the new job as a three star chef at the pass, inspecting every plate the agents cook and sending back anything with a hair in the soup. Dave Killeen moved from a kanban of agents he had to shepherd to a single calm chat that brings back only the decisions he needs to make. Simina Pasat runs a Friday retro with Claude on what pushed the limits that week. The common thread: direction, review and the quality bar are the parts that stay human, and they need to be practised deliberately.

Talks: Gopika Mohanan, Thomas Hartmann, Dave Killeen, Simina Pasat


---

## How AI Broke Your Strategy

**Martin Eriksson**, Author of The Decision Stack, Co-Founder of Mind the Product. Opening Keynote, 09:45, Just Product 2026.

> AI broke an assumption under your strategy; find which one, go up the stack until one holds, and rebuild from there.

### Summary

Martin opens with a prediction that failed. In 2016 Geoffrey Hinton said AI would replace radiologists within five to ten years. AI did get faster, cheaper and more accurate at reading images, yet the US has 16 percent more radiologists than a decade ago and they earn 44 percent more. Cheaper imaging meant preventative screening at scale, more findings, more follow up, more demand. That is the Jevons paradox, first observed with coal and steam engines in 1865. Hinton did not get the technology wrong. He assumed demand was finite. Martin's claim for the room: AI has broken at least one assumption in every company present, and most people do not know which one.

To find it you need to know where it hides. After 30 years in product, co-authoring Product Leadership and advising about 150 companies, Martin boils it down to five questions: where are we going, how do we get there, what matters now, which actions do we take, how do we choose between them. He answers them with vision, strategy, objectives, opportunities and principles, the Decision Stack, but any framework works as long as the layers connect top down through how and bottom up through why. Every layer is a decision, and every decision rests on assumptions. Most companies have a vision on a lobby poster, a strategy in a PowerPoint on a file server, OKRs in a forgotten spreadsheet, and now sprinkle AI strategy on top.

Then come the failure modes. Most companies look only at the opportunities layer and treat AI as a feature, which assumes the customer's problem is a lack of AI. Four in five people on earth have never used generative AI, and a Washington State University study found the label artificial intelligence reduces purchase intent. Chegg built CheggMate with OpenAI while the broken assumption sat higher up: students would not keep paying for answers ChatGPT gives away. The stock halved during the earnings call and revenue fell 75 percent. The second failure is AI as cost cutting. IKEA's agent resolved half of all tickets, the rest were design questions, so IKEA retrained support staff as remote interior designers and booked 1.4 billion dollars in new revenue in year one.

Intercom went furthest. Des Traynor's team saw that per seat pricing and humans staffing support were both broken, kept only the vision that businesses want to support their customers, scrapped the roadmap over a weekend and built Fin. At 99 cents per resolution, Fin passed 100 million ARR in year one, and Martin reports a 3.6 billion sale to Salesforce after a 1.3 billion valuation. His method: name your assumptions by asking what would have to be true, go up the stack until one holds, test it against tokens at zero and tokens at ten times the price, then bet at 70 percent confidence and write down the kill criteria. We are in the DOS era of AI; the best way to predict the future is to invent it.

### Key takeaways

1. Write down the assumptions under each layer of your stack using Roger Martin's question: what would have to be true?
2. Drop the AI as a feature reflex: four in five people have never used generative AI and the label lowers purchase intent.
3. Ask the Jevons question: what would customers do with your product if it were ten times cheaper, faster or more accessible?
4. When an AI agent absorbs half the workload, study what the remaining half reveals before cutting headcount, as IKEA did.
5. Check whether your business model survives AI; outcome based pricing like Intercom's 99 cents per resolution ties revenue to the product working.
6. Plan for two futures, tokens at zero and tokens at ten times today's cost, and keep only assumptions that survive both.
7. Make the bet at 70 percent confidence, write it down, and define the kill criteria that would prove you wrong.
8. Challenge upward by naming assumptions and bringing customer evidence, and ask each leader separately what the strategy is.

### Chapters (timestamps in the talk video)

- 0:00 Host introduction
- 1:12 Hinton, radiologists and the Jevons paradox
- 5:10 The Decision Stack in five questions
- 7:46 Every layer sits on assumptions
- 10:25 AI as a feature: Chegg
- 14:14 AI as cost cutting: IKEA
- 17:51 Going up the stack: Intercom
- 22:45 How to find your broken assumption
- 26:17 Make a bet, write it down

### Highlight moments

- 2:57 to 4:09: **Hinton gave radiologists five years**. The opening story in full: AI got better at imaging, radiologists grew 16 percent and earn 44 percent more.
- 6:53 to 7:46: **Every level of the stack is a decision**. Defines the Decision Stack and lands the poster, PowerPoint and spreadsheet punchline about sprinkled AI strategy.
- 12:58 to 14:14: **Chegg never asked which assumption changed**. A complete cautionary tale with the numbers: stock halved, revenue down 75 percent, students just wanted homework done.
- 15:06 to 16:35: **IKEA retrained support as interior designers**. The strongest counterexample to AI as cost cutting, ending on 1.4 billion dollars of new revenue from the same humans.
- 20:50 to 22:18: **Intercom scrapped its roadmap over a weekend**. Shows what a strategy level rethink looks like, from per seat pricing to 99 cents per resolution and the Fin outcome.
- 24:34 to 25:31: **We pass five year strategies on vibes**. The most practical challenge to product people: take assumption testing up the stack, because you cannot A/B test 2028.

### Quotes

- "Today's AI is the worst AI you are ever going to work with." (8:31)
- "Turns out Chegg's customers didn't really want Chegg. They definitely didn't want AI. They just wanted their homework done." (14:01)
- "To me, that's using once in a generation technology to shrink yourself." (14:35)
- "AI is the biggest growth lever your business has ever had, and treating it as a cost lever is a failure of imagination." (16:35)
- "We obsessively test these relatively cheap decisions, and we pass five-year strategies on vibes." (24:56)
- "You can't A/B test 2028." (25:27)

### Q&A (not in the video, summarised from the stream)

**Q: Which assumption do you hear most often that is simply wrong?**

That none of their assumptions have changed. Companies react to the speed of AI but skip going back to the drawing board to check whether their objectives, strategy or vision still hold. In his workshop the day before, that check surfaced broken assumptions quickly.

**Q: Can we use AI to explore and forecast the upper layers of the stack, or are those human only?**

AI is a strong thought partner for research, options and checking the consistency of an argument. The leap of faith at strategy level and connecting signals across customers stays human, because AI is not happy making bets. Validate its output with real people, in the spirit of Lean Startup and Steve Blank's customer development.

**Q: How do you challenge vision and strategy in a 20 year old founder led company that resists change but leans into the AI hype?**

Challenge the assumptions, not the person: say the strategy rests on enterprise customers needing X, ten interviews show they no longer do, what changes? Always bring evidence, or the highest paid person's opinion wins. His favourite trick is asking each leader individually what the strategy is and playing the different answers back to them.

**Q: Is a human in the loop a real requirement or just a limitation we accept because someone must be accountable?**

It depends on the level. Agents handling tickets or code can run without a human once evals and testing are in place. For objectives, strategy and vision a human stays in the loop, not only for accountability but because industry judgment is needed to filter the stupid ideas from the great ones.

### Mentioned

- Geoffrey Hinton, the 2016 prediction that AI would replace radiologists within five to ten years
- Jevons paradox, William Jevons on coal consumption rising after James Watt's steam engine, 1865
- Product Leadership, the book Martin co-authored to interview hundreds of product leaders
- The Decision Stack, Martin's book and framework (vision, strategy, objectives, opportunities, principles), first two chapters free via QR code at thedecisionstack.com
- Mind the Product and ProductTank, the community Martin co-founded, credited by host Sebastian for his own move into product
- Gmail launch in 2004 with 1 GB of storage, Google's bet on falling storage costs
- Washington State University study, the term artificial intelligence in product descriptions reduces purchase intent
- April Dunford, customers do not buy AI, they buy the elimination of a pain point
- Chegg and CheggMate, CEO Dan Rosensweig's May 2023 earnings call, homework help business hit by ChatGPT
- IKEA, support agent resolved 50 percent of tickets, staff retrained as remote interior designers, quote from Ulrika Biesert
- Rita McGrath, strategy is more critical in a fast moving world because organisations make many more decisions
- McKinsey report, a third of companies trying to replace at least one SaaS product
- MIT study, 66 percent of in-house software builds fail
- Anthropic, still uses Slack, Workday and Figma rather than rebuilding them
- Intercom, Des Traynor, Fin AI agent at 99 cents per resolution, company renamed Fin, reported sale to Salesforce
- IBM manual from 1979, a computer can never be held accountable so must never make a management decision
- Roger Martin, the question what would have to be true
- Annie Duke, Thinking in Bets, calling it a bet and writing down kill criteria
- Margaret Heffernan, just because we cannot predict the future does not mean we cannot prepare for it
- Xerox PARC, Alan Kay and Adele Goldberg, the graphical user interface and the line that the best way to predict the future is to invent it
- Lean Startup and Steve Blank's customer development, cited in the Q&A on validating AI output with humans

### About the speaker

Martin Eriksson has spent three decades helping organisations figure out why they are stuck and what to do about it. He has built digital products since 1994, worked at the Financial Times, Monster, Huddle, Covestor, Cazoo and EQT, and co-founded Mind the Product, the world's largest product community. His book The Decision Stack distils what it takes to align an organisation from vision to execution.

- Website: https://martineriksson.com
- The Decision Stack: https://www.thedecisionstack.com

---

## AI Emptied Our Backlog. Now What?

**Verena Graber**, Product Manager at medflex. AI Mindset Shift, 10:30, Just Product 2026.

> When AI makes delivery nearly instant, product becomes the bottleneck, and the old methods around it have to go.

### Summary

Verena Graber opened with the image Sebastian had teased: a product manager who feels like an octopus with 800 arms. At medflex, a Medtech company that uses AI to support communication and automate processes for healthcare providers, her work ran from research and design to roadmaps, testing, documentation, client onboarding and sales calls. In spring 2026 she took the Product Masterclass cohort on product management in the AI era, which gave her three blocked hours a week to experiment. Instead of chasing what was missing, she started with what the company already had. Sales and support were transcribing most calls, so she sorted those transcripts into two folders, new clients and existing clients, because the AI kept confusing what existed with what was new.

She had AI read the transcripts for recurring complaints, what new customers needed and the most asked questions, and later added the support ticket pipeline to surface bugs that kept coming back. The process still runs today. Every new transcript lands in a folder, gets scanned, and the resulting list is checked against the backlog with almost no human interaction. Challenging the backlog against those insights changed it fast. It had been about five kilometers long, with tickets from years ago nobody remembered. A two year old ticket turned out to still be relevant and went straight to ready to pick, while a ticket from the week before got deleted. Without hours of interviews, the backlog shrank by almost 50 percent, and everything left had been reviewed recently.

Then the tech team introduced agentic engineering and development cycles sped up roughly four times. A week started on Monday with a full list for the developers. By Friday nothing was left, and her Teams chat filled with "what can I do next?" and "is there anything I can do now?". Product had become the bottleneck. Testing, finalizing releases and keeping the team informed left no room for new ideas, so the team looked at its methods. Sprints went first: no point planning a two week block when tasks finish within minutes. Story points went next, which hurt, because those numbers had been her argument with stakeholders for saying no. Scrum Poker rounds filled with coffee pot cards because developers could no longer estimate effort the agent would do anyway.

Groomings stayed, but changed. Instead of comparing tickets, debating sentences and moving two pixels, the sessions now ask what the goal of an idea is and how design and development together can reach it faster. The bottleneck still sat in the methods, so the CEO, Verena and the two lead developers met over a rough prototype the CEO had sketched in about two hours. It held the big ideas the team had never known how to start. They gave the top three ideas to three developers, one each, with only the sketch and one question: what can you do with this using your agents in two weeks? Work planned to take two years at the start of the year took two weeks and now sits on the testing stages.

Her product role now has four arms. Research and insights, where humans still decide what gets built and when. Exploration and design, the most human part, using Claude Design and Figma with faster prototypes and more variants. Grooming without estimates, asking what would make the team faster or the product more lovable, while an agent writes the tickets. And documentation, which agents draft and she reviews, so she can explain what is live today and train support instead of answering the same question again. Her closing advice: do not be afraid to be the bottleneck. Push to that point, then use it to see what you still need, what you can drop and where to work closer together. No perfect method exists yet, so let the role become fluid.

### Key takeaways

1. Start with data you already have, such as sales and support call transcripts, before recruiting interview partners.
2. Separate transcripts of new and existing customers before analysis, otherwise the AI mixes what exists with what is requested.
3. Check every backlog item against live customer insight; expect old tickets to resurface and fresh ones to die.
4. When agents make delivery near instant, drop sprints and story points instead of keeping rituals that no longer fit.
5. Run groomings on the goal of an idea, with design and development in the room, rather than on ticket wording and pixels.
6. Hand big ideas directly to developers with a rough sketch and a two week agent budget, then review what comes back.
7. Let agents draft tickets and documentation; keep human time for deciding what gets built and when.
8. Collect customer feedback during onboarding, when customers are curious anyway, instead of scheduling separate validation calls.

### Chapters (timestamps in the talk video)

- 0:01 Intro: an octopus with 800 arms
- 5:05 Research pipeline from call transcripts
- 6:23 Cleaning a five kilometer backlog
- 7:40 Agentic engineering empties the backlog
- 9:15 Killing sprints, story points, old groomings
- 12:59 Three ideas, three developers, two weeks
- 16:22 The four arms of the product role
- 18:52 Closing advice: be the bottleneck

### Highlight moments

- 5:05 to 6:23: **Two folders of transcripts, one living research pipeline**. The concrete setup behind the whole talk: existing call transcripts, sorted by new and existing clients, scanned by AI with almost no human interaction.
- 6:23 to 7:40: **A five kilometer backlog cut in half**. A vivid before and after with the two year old ticket that came back and the week old ticket that died.
- 7:40 to 8:59: **By Friday nothing was left**. The turning point: agentic engineering made product the bottleneck, told through her Teams chat.
- 9:54 to 11:21: **Why story points had to go**. An honest account of losing her stakeholder argument, with the coffee pot Scrum Poker story as the punchline.
- 13:58 to 15:16: **Two years of roadmap in two weeks**. The strongest number of the talk and the method behind it: one big idea per developer, a sketch and two weeks with agents.
- 18:52 to 19:49: **Don't be afraid to be the bottleneck**. The closing advice that frames the whole story as a chance to redefine the role.

### Quotes

- "I don't know about your backlog, but ours was about five kilometers long, with tickets from years ago nobody remembered." (6:31)
- "It turned out product kind of is the bottleneck again. AI development outran product." (8:43)
- "What's the point of planning a two week block for development when I'm mostly busy putting in tasks that are getting done within minutes?" (9:27)
- "Our estimations that before looked like pretty little numbers for every task now looked like a café in the middle of a city, because coffee pots kept popping up all the time." (10:54)
- "What we planned on would take us about two years at the beginning of this year took two weeks." (14:58)
- "Don't be afraid to be the bottleneck. Really challenge it to go to this point and then use it as a chance." (18:52)

### Q&A (not in the video, summarised from the stream)

**Q: If you eliminated sprints, what replaced them, and do you still have checkpoints to review and improve?**

The team went back to Kanban, and groomings now also cover when something is needed and how much time a release buys. Releases currently outpace what clients can absorb, so the team builds and waits for the release. Dailies remain, and a ticket that shows up in dailies for more than three weeks is the marker that something is off.

**Q: How did the development team handle the responsibilities you handed over, eagerly or defensively?**

Eagerly. Developers now join ideation and design at the earliest stage, something she admits she had preached but not practiced before. A shared vision of the product helps; the open question is the testing stage, where some of their design choices will be reworked or deleted.

**Q: As product manager, do you still have control over the objective so the team does not spin off into something else?**

Not easy when you can build everything, and focusing on what should actually be built is daily work. Right now the direction is clear, and ongoing contact with sales, support and customers keeps plans honest, with enough fluidity to change course quickly.

**Q: In the design and exploration arm, how did you speed up validation with customers?**

By doing less of it. The team tests against the existing product, uses power users as feedback givers and collects feedback during customer onboarding, when clients are curious and asking questions anyway. That replaces scheduling separate validation appointments.

### Mentioned

- medflex, Medtech company using AI to support communication and automate processes for healthcare providers, where Verena is Product Manager
- Product Masterclass cohort on product management in the AI era, spring 2026, where the experiments started with three blocked hours a week
- Claude Design and Figma, used for design exploration with faster prototypes and more variants
- Scrum Poker, the estimation tool whose coffee pot card took over once agents did the work
- Kanban, what the team returned to after dropping sprints
- Microsoft Teams, the chat that filled with developers asking for new tickets
- Stefan, medflex product designer, present in the audience
- medflex CEO, who sketched the vision prototype in about two hours that seeded the three two week experiments
- Productipus, the internal name for the octopus metaphor at medflex

### About the speaker

Verena Graber is a Product Manager at medflex, a company that automates patient communication for medical practices. Over the past year she rebuilt how her team works with AI: every sales, customer success and support conversation now flows into a research pipeline she can query, and product news is drafted automatically when a ticket closes. She is medflex's internal AI champion.

- medflex: https://www.medflex.de

---

## The Six Traps of AI Adoption (and How to Get Out of Them)

**Faith Forster**, Founder and Chief Maker, Discoveree. AI Mindset Shift, 11:00, Just Product 2026.

> AI made building fast, so teams skip the thinking; name the six traps and use the Makers Manifesto principles to rethink instead.

### Summary

Faith Forster opens with a familiar Monday morning. Engineers or their agents have built three features over the weekend, nobody knows who asked for them, sales want to demo, and the CEO has seen a competitor's release video. Her diagnosis: AI has made engineering much faster, which puts pressure on product to keep up, so people shortcut or skip discovery and nobody asks whether the feature should exist. It feels like progress and often is not. In spring she spoke to 96 product leaders about their AI use and found wildly different levels of understanding. The old rule book was gone, so she gathered 45 cross-disciplinary operators and wrote the Makers Manifesto: four values, 16 principles, an Agile Manifesto for the AI age.

The talk was announced with four traps and grew to six, each a kind of thinking teams skip when they rush, each paired with a principle. Tokenmaxxing is measuring activity instead of value, with Uber as the cautionary tale: a leaderboard ranked teams by AI usage, the year's budget was gone in four months, and when asked what customers got, COO Andrew Macdonald said that link is not there yet. Principle 7, measure what drives value. The Abandoned Junior is handing agents work with no context, then concluding AI does not work, as she saw on the board of an HR tech scale-up. The way we like it had never been written down. Principle 6, make context explicit. Onboard your agent the way you would onboard a graduate.

It's Live is calling a feature done when it ships. Faith's line is that distribution, adoption, monetization and retention are not what comes after the making, they are the making now. Omio put a subscription live within three months and then spent six months refining packaging and revenue impact until average order value and tickets per customer rose. Principle 9, make for adoption. The Fire Hydrant is shipping faster than customers can absorb. Fin, formerly Intercom, tripled throughput and shipped nine tier one releases in two months until customers complained that the product no longer made sense; they now run one major release per quarter. Principle 8, match pace to readiness. Humans are the bottleneck, and often those humans are your customers.

AI Ate My Homework is letting the model decide and hoping for the best. Deloitte Australia repaid half a million dollars after an unsourced AI statistic landed in client work, and Air Canada tried to blame its own chatbot in a tribunal, which called that a remarkable submission. Principle 16, humans sign the work: you can delegate decisions but not accountability. Lipstick on Legacy is adding a sparkle icon to last year's roadmap instead of asking what problem is worth solving. The winners rebuild: Fin sold to Salesforce for about 3.6 billion dollars, Pendo rethought analytics as an agent with Novus, and Faith rebuilt Discoveree in six weeks. Her closing ask: download the manifesto, give it to your agents, run a retro on one value, and name your trap.

### Key takeaways

1. Replace AI usage leaderboards and percent of code written by AI with a metric that ties AI spend to customer outcomes or revenue.
2. Write down what good looks like for your codebase and documents, then onboard agents with that context before judging their output.
3. Share the rationale behind each feature and the outcome you expect with your agents, not just the ticket.
4. Move the definition of done from shipped to adopted, and plan the months of refinement after launch into the roadmap.
5. Set release pace by what customers can absorb, for example one major release per quarter with small refinements between.
6. Put a named human on every consequential AI decision so the post-mortem never says the root cause was the model.
7. Re-ask what problem is worth solving before adding AI to every roadmap line, because build speed now allows a fundamentally different solution.
8. Download the Makers Manifesto, pick one value for your next retro, and ask the team which thinking you are skipping.

### Chapters (timestamps in the talk video)

- 0:01 Intro and the Monday morning story
- 3:11 96 product leaders, the Makers Manifesto
- 6:03 Trap 1: Tokenmaxxing, the Uber story
- 8:38 Trap 2: The Abandoned Junior
- 11:01 Trap 3: It's Live, Omio
- 13:39 Trap 4: The Fire Hydrant, Fin
- 15:53 Trap 5: AI Ate My Homework
- 18:11 Trap 6: Lipstick on Legacy
- 21:14 Recap and the smallest next step

### Highlight moments

- 2:23 to 3:11: **Faster builds, skipped thinking**. The core diagnosis of the talk: engineering speed pushes product to skip discovery, and it only feels like progress.
- 7:39 to 8:38: **Uber burned a year's AI budget in four months**. The token maxing story with the COO quote that made executives start asking what the spend buys customers.
- 9:54 to 11:01: **You would never abandon a graduate**. The onboarding analogy that reframes agent setup as teaching, and the PM's job as sharing rationale and outcomes.
- 14:54 to 15:53: **Fin shipped faster than customers could absorb**. A concrete 3x productivity story that turns into a lesson about customers as the real bottleneck.
- 16:19 to 16:56: **Humans sign the work**. Principle 16 in one breath, backed by the Deloitte case and the post-mortem that blamed the model.
- 21:06 to 21:55: **Six traps, one instruction: rethink**. The recap of all six trap names and the closing argument that the build will only get faster.

### Quotes

- "If there's one thing you take away from today, it's this. Don't rush to keep up. Rethink instead." (5:51)
- "You've hired a brilliant graduate, you give them no onboarding, no examples, no feedback, and then by Friday you decide they're not good enough. That's exactly what we're doing to our agents." (10:05)
- "Distribution, adoption, monetization, retention, they don't come after the making. That is what making is now." (11:17)
- "Yes, humans are the bottleneck, but often not the humans within your own business. It's your customers." (15:35)
- "You can delegate decisions, but you cannot delegate accountability." (16:31)
- "If the definition of done is shifted from shipping to adoption, that is a whole different way of working." (25:06)

### Q&A (not in the video, summarised from the stream)

**Q: Any tips for the fire hydrant problem when even internal teams cannot keep up with the shipping pace?**

Shift the definition of done from shipping to adoption and organise roadmap, planning and teams around it; she calls this the biggest mindset shift in the manifesto. Less discovery is fine if you build to learn and verify with customers that the problem and the impact are right. It has to come from the executive level, because a PM told to push features out cannot fix it alone.

**Q: When the backlog is empty and developers ask what to do next, do you have a slow-down list for them too?**

Start with no-regrets moves, the work you know needs doing with low risk of being wrong. Then stop treating product as PMs and designers doing all the thinking up front and handing off; involve developers in building to learn, since you can create ten prototypes now where one used to be expensive, and test them with internal users and close customers.

### Mentioned

- Makers Manifesto (makersmanifesto.org): four values, 16 principles, written by 45 cross-disciplinary operators; downloadable in a version for humans and a version for agents
- Discoveree.com: Faith's company, a shared context layer or product brain for teams, rebuilt from scratch in six weeks
- Uber: internal AI usage leaderboard, 2026 AI coding budget spent by April; COO Andrew Macdonald: that link is not there yet
- Fin (formerly Intercom): 3x throughput, nine tier one releases in two months, now one major release per quarter; CTO Darragh Curran quote on humans as the bottleneck; Salesforce deal of about 3.6 billion dollars in June 2026
- Omio: first subscription product in its category live in three months, six months of refinement after
- Dext: Faith's former company, nine months to redo pricing and packaging
- Air Canada (Moffatt v. Air Canada): tribunal ordered C$650 refund after the airline blamed its chatbot
- Deloitte Australia: repaid half a million dollars over an unsourced AI generated statistic
- Pendo Novus: analytics rethought as an agent, covered by Dave Killeen later in the day
- Agile Manifesto: the rule book the Makers Manifesto replaces for the AI age
- Simonetta: introduced Faith to Sebastian and Thomas and co-authored the manifesto
- Martin: earlier speaker who mentioned the Salesforce and Fin deal
- An HR tech scale-up where Faith sits on the board: the 'so we just do it ourselves' story

### About the speaker

Faith Forster is a product leader in B2B SaaS and fintech, including Wise, Dext and Legl, with two exits. She is the founder of Discoveree, a shared context layer for product teams, co-hosts a weekly podcast on building AI native products, and organises the Makers Manifesto.

- Discoveree: https://discoveree.com
- Makers Manifesto: https://makersmanifesto.org

---

## Run Your Product From Chat

**Thomas Hartmann**, Co-Founder, Product Masterclass. Author of From Project to Product Mode. AI End-2-End PM Workflows, 13:00, Just Product 2026.

> Connect AI to your tools and context, encode your workflows as skills, and the operational grind disappears, freeing you for strategy.

### Summary

Thomas opened by interrupting his own talk. He had two things to do first: answer an email from Julian, who wanted a team ticket for Just Product and needed an offer to start his purchasing process, and act on Sebastian's complaint that the website button was boring. He typed both requests into chat, asked for three playable button prototypes, told the AI not to ask questions and let it run. Then he explained the year behind that casual move. In October 2025, at a client hackathon, an engineer and product manager interviewed a colleague, pulled the photo book selection algorithm from GitHub, brainstormed changes with AI and had a prototype in two hours. That used to take Thomas four weeks. He went home convinced everyone was going to be replaced.

The second moment came at the end of January. Thomas, who has never written a line of code, built a working app for a side business over one weekend, with filters, a map and listing pages. Sebastian, a software engineer, asked how long it took and estimated two to three months for himself. From then on the two co-founders agreed they had a problem, including the uncomfortable question of whether anyone would still come to the Product Masterclass if AI can explain everything. At the same time, daily work looked nothing like superintelligence. A voucher request meant hunting through Eventbrite, checking Fastbill, downloading a PDF and drafting a mail. Twenty to thirty minutes, or two days of putting it off.

So they decided to automate every workflow. Publish a blog post: does Webflow have an API? Build a skill. Write an invoice: does Fastbill have an API? Connect it to Claude, build a skill. Tools without full API access, such as the newsletter tool and the old website builder, got replaced with ones that had it, and some were replaced with self-built applications. Step by step that became the Product Masterclass operating system: connections to tools and GitHub repos, the business context, and skills that encode the processes, all handed to an agent. Back on stage, the first task was done. The AI had created the Eventbrite voucher, looked Julian up in Pipedrive, created a five ticket offer in Fastbill and drafted the Gmail reply.

Thomas counted the gain: five to ten times faster for that one chore, and they have around 46 such skills. The bigger effect is new capacity. After each participant's half hour review call, the AI turns the recording into an alumni story page, picks the angle and the video segments, cuts the clips, writes subtitles and the blog post. Two hours on the side instead of three days of focused work. Questions like who visited the website or whether someone is already registered get answered on the spot. Then the second live result: three button prototypes, the audience voted for the third, and Thomas told the AI to write a ticket, implement it and push to main, with 250 people as QA. Normally two sprints.

For product managers the recipe is short. Connect your AI to Jira, Confluence and GitHub and you have a first product operating system. AI will not replace PMs, but PMs who use AI will replace those who do not. Sort your workload into protect, augment and automate, master the protected part (customer understanding, strategy, product sense) and automate the rest. Fundamentals matter more than ever because AI is a lever: a fool with a tool is still a fool, a great PM with AI feels superhuman. The new job looks like a three star chef at the pass, inspecting every plate the agents cook and sending back anything with a hair in the soup. Master AI in your field, and remember a tiny team can have huge impact.

### Key takeaways

1. Pick one recurring chore, check whether every tool it touches has an API, and encode the full process as a reusable skill.
2. Replace tools that lack full API access; manual copy and paste steps block automation for everyone on the team.
3. Start your product operating system by connecting your AI to Jira, Confluence and GitHub, then add context and skills.
4. Sort your workload into protect, augment and automate; automate anything the AI can finish without you, review what leaves your name on it.
5. Invest in the protected part yourself: customer understanding, strategic decisions and product sense are where you must beat the AI.
6. Use freed capacity for work that was impossible before, such as turning every customer review call into a published story.
7. Treat reviewing agent output as the core job: inspect every result like a chef at the pass and send back anything flawed.
8. Let designers own the design system and engineers own code maintainability so AI prototypes stay on brand and shippable.

### Chapters (timestamps in the talk video)

- 0:01 Intro and two live AI tasks
- 3:07 Two moments that changed everything
- 7:20 Superhuman AI, thirty minute voucher chores
- 8:30 Automating every workflow into an OS
- 11:11 Live result: offer, voucher, email draft
- 13:25 What freed capacity makes possible
- 15:04 Your first product operating system
- 18:35 AI is a lever for PMs
- 22:25 Three takeaways

### Highlight moments

- 3:18 to 4:36: **Two hours instead of four weeks**. The hackathon story where an engineer PM went from interviews to prototype in two hours and sent Thomas home with sleepless nights.
- 4:57 to 6:23: **A weekend app worth three developer months**. Thomas builds a working app over a weekend and Sebastian estimates two to three months for the same thing.
- 12:47 to 13:41: **Task one done: 10x and 46 skills**. The live demo pays off and Thomas does the math on speed and freed capacity.
- 13:41 to 14:46: **Three days of work in two hours**. The alumni story pipeline shows AI enabling work that was impossible before, plus data questions answered on the spot.
- 20:07 to 21:08: **A fool with a tool is still a fool**. The core argument that fundamentals matter more with AI, delivered as a memorable line.
- 21:28 to 22:25: **The chef at the pass**. The vivid image of the PM's new job: reviewing everything the agents cook and finding the hair in the soup.

### Quotes

- "Holy cow, that used to take me four weeks, and he did it in two hours." (4:14)
- "I've never written a line of code, and I created an app for which a professional software developer would say: I need two to three months." (6:07)
- "Connections, context, you give that to an agent, you build workflows on top of that, and you will get amazing output. That's an operating system." (10:57)
- "AI won't replace PMs. PMs who use AI will replace those who don't." (18:49)
- "A fool with a tool is still a fool. A bad product manager with AI is an even worse product manager. A great product manager with AI is a freaking machine." (20:25)
- "Your AI agents cook something, everything passes by, and you look for the hair in the soup. If you find it, you send it back to the kitchen. That's your job now." (21:41)

### Q&A (not in the video, summarised from the stream)

**Q: Where is the line between augmenting and automating your work?**

Automate everything you are confident the AI can finish without you. Keep a review step on things like outgoing email, where you want to see the result before it goes out.

**Q: How do you work with a designer when the PM prototypes with AI?**

Product Masterclass has a design system and the AI pulls from it. In a bigger company the designer owns and maintains the design system, the engineer keeps the codebase maintainable, and the AI builds within both. Asked what the PM does then, Thomas joked: nothing.

### Mentioned

- Claude, the AI model the whole Product Masterclass operating system is connected to
- Eventbrite, where the demo created the voucher codes for Julian's team ticket
- Fastbill, invoicing tool that created the five ticket offer via API
- Pipedrive, the CRM the agent checked to identify the customer
- Gmail, where the offer email was drafted for review before sending
- GitHub, repos the agent pulled to prototype and ship the new enroll button
- Webflow, the website builder that got replaced because its API did not cover page structure changes
- n8n, named among the glue tools that still needed manual pasting
- Jira and Confluence, with GitHub the suggested first connections for a PM's operating system
- Protect, Augment, Automate, the framework for sorting a PM's workload
- Verena Graber, whose morning talk showed the same push toward automation at medflex
- Sebastian Borggrewe, co-founder and software engineer, estimated two to three months for Thomas's weekend app
- Julian, the customer whose team ticket request became the live demo
- Engineers are the babysitters of the AI, a line Thomas picked up at another conference
- AI Product Masterclass, the program with two upcoming start dates, pitched as the starting point
- Alumni stories page on product-masterclass.com, generated by AI from review call recordings

### About the speaker

Thomas Hartmann co-founded Product Masterclass and hosts Just Product together with Sebastian. He is the co-author of From Project to Product Mode. In 2026 he rebuilt the whole company to run on an AI operating system.

- Product Masterclass: https://www.product-masterclass.com

---

## It's Time to Lead Your Team of AI Agents

**Gopika Mohanan**, Director of Product at BestSecret Group. AI End-2-End PM Workflows, 13:30, Just Product 2026.

> Working with AI agents forces every product manager, even the most junior, to practice classic product leadership skills from day one.

### Summary

Gopika Mohanan opens with a thought experiment. On Monday your boss tells you five new people joined your team. They research a market in minutes, pull customer insights, write code, take no bathroom breaks and have no working hours. The catch: sometimes they hand you something with full confidence and it is wrong. Would you check every output, write a 15 page instruction and hope? Or would you think about roles, context and decision rights? Gopika leads about eight product managers at BestSecret, where multi-agent systems have moved into the daily work of product teams. Her observation: this shift teaches less about technical skills and more about old product leadership skills, and it pushes them onto junior PMs who used to wait years for a team.

The mental model has moved from question and answer to goal and result, since around May or June. An agent observes, reads expectations, picks tools, proposes, inspects and loops until the task is done. You are directing work, so output quality depends on what you put in. Skill one is delegation and context. Her example: leadership asks for a strategy to reduce returns caused by size and fit. The prompt "look at our size returns and tell me what you find" fails because the agent does not know which markets, which customers, whether return rate or revenue matters, or the time horizon. Like a new employee, it does not know BestSecret is off-price or that an experiment failed six months ago. The fix is an environment of structured context.

Skill two is decision boundaries. Gopika calls "how autonomous should my agents be" the wrong question, since nobody asks that about a human team. Instead you define the points where the agent must involve you. An agent analysing return data may trigger a second analysis on its own, but it must not launch a live test. With agents you can hard code this: the research agent gets web access, the analysis agent gets read access. Skill three is judgment and the quality bar. Agents do not tire after the fifth idea, so fire several at one problem with different lenses such as time to market, risk and user experience, and welcome disagreement. The agent expands the decision space, the PM narrows it. BestSecret now hires PMs mainly for critical thinking.

Skill four is organisational capability. When an agent keeps returning superficial customer analysis, you do what you would do with a team member: teach it how to think, which principles to follow and what good looks like, then encode that into its instructions or memory and test whether outputs improve. Every piece of feedback is a permanent upgrade, and once it works for one agent you can build it into a whole workflow. In production, BestSecret runs product classification with one agent classifying and another challenging the output. Her closing: define the outcome, provide context, set boundaries, own the judgment while we still can, and build a system that improves as a whole. In Q&A she stresses that not everything is agentic, and cost is weighed every time.

### Key takeaways

1. Treat a new agent like a new hire: give it the company context it cannot know, including past failed experiments and what the business is.
2. Replace vague asks with a brief that names markets, customer types, the metric that matters and the time horizon.
3. Build a reusable context environment with strategy, constraints, history and evidence instead of rewriting a long prompt for every request.
4. Stop asking how autonomous agents should be. Write the checklist of decisions where they must escalate to you, such as launching a test.
5. Give agents hard-coded permissions per role: web access for research, read access for analysis, no write access to live systems.
6. Run several agents on the same problem with different lenses such as time to market, risk and user experience, and value their disagreement.
7. When an agent repeats a mistake, teach it what good looks like and encode that in its instructions or memory, then test the next output.
8. Scale agentic setups only where the benefit beats the cost. Build the system around one concrete problem, not around everything.

### Chapters (timestamps in the talk video)

- 0:01 Host intro: four applications later
- 1:48 Monday exercise: five new team members
- 5:16 From question and answer to goal and result
- 7:32 Skill 1: delegation and context
- 13:26 Skill 2: defining decision boundaries
- 16:54 Skill 3: judgment and the quality bar
- 20:35 Skill 4: building organisational capability
- 23:40 What actually changes for a PM

### Highlight moments

- 2:07 to 3:31: **Five new hires who never take breaks**. The opening thought experiment sets up the whole talk: fast, confident, sometimes wrong, so how do you manage them?
- 5:30 to 6:52: **You are no longer using a tool**. The core shift from asking questions to delegating outcomes, ending on the line that you are now directing work.
- 8:48 to 9:42: **Why the vague prompt fails**. A concrete size-and-fit example shows exactly what context the agent lacks and the PM takes for granted.
- 13:55 to 14:57: **How autonomous should my agents be?**. Her sharpest reframe: nobody asks that about a human team, so define escalation points instead.
- 18:45 to 20:03: **Agents expand, the PM narrows**. The drill-me agents and the division of labour between agent and PM, the most quotable practical idea of the talk.
- 24:08 to 25:25: **Own the judgment while we still can**. The five-step closing that ties the four skills together and sends the audience off to lead their new team.

### Quotes

- "You're no longer just using a tool. You're directing work. And so the quality of output very much depends on what you're putting inside." (6:37)
- "You know all of these things as a product manager because you live inside of the problem, but the agent does not." (9:24)
- "If you draw these parallels again with your human team, you will never ask yourself how autonomous should my team be." (14:06)
- "The agent gives you the possibility to expand the decision space and the product manager narrows it." (19:47)
- "The benefit with an agent is every piece of feedback that you give it, it is a constant permanent upgrade." (21:48)
- "Truly work with them to flesh out the problem and solution, but at the end, you own the judgment while we still can." (24:42)

### Q&A (not in the video, summarised from the stream)

**Q: How do you handle enablement, since you cannot just install Claude locally and hand it all your customer data, plus the legal and technical questions that come with it?**

Not everything a product person does is agentic. BestSecret builds systems around specific problems in isolation: for size and fit the setup connects meeting transcripts, Qualtrics data, app review outputs and supply analytics, while customer-facing product classification uses entirely different tooling. Internal and external use cases get their own environment, and because feeding agents more data costs money, every setup is weighed against its true benefit.

### Mentioned

- BestSecret Group, the off-price fashion e-commerce company where Gopika leads about eight product managers
- Size and fit returns, the running example for a product strategy brief delegated to agents
- Drill-me agents, agents some BestSecret colleagues set up whose only job is to challenge every input
- Product classification agents in production, one classifying products, another challenging the output
- Qualtrics, customer feedback data connected to the size-and-fit agent system
- Claude Code, shown on the slides as the multi-layer agent development kit behind the team's setup
- Claude, which host Sebastian used to check that Gopika had applied to speak at every Just Product edition
- TIER, Zalando and Wayfair, where Gopika led product before BestSecret

### About the speaker

Gopika Mohanan is Director of Product at BestSecret Group, where she leads the search, browse, recommendations and infrastructure product organisation. She has led products at scale at TIER, Zalando and Wayfair, and speaks at Product at Heart, Product Management Festival, Turing Fest and Product Led Summit.

- LinkedIn: https://www.linkedin.com/in/gopikaem/

---

## When Everyone Can Build, What Makes Great Product?

**Dave Killeen**, Field CPO EMEA at Pendo. AI Transformation and Leadership, 14:45, Just Product 2026.

> When AI makes building cheap, the advantage shifts to how fast you spot signal, redesign work around moving bottlenecks, and prove impact.

### Summary

Dave Killeen opened by taking away the industry's favourite comfort blanket. He used to tell product people that AI was the sous chef and human taste was the moat. He now thinks he was wrong. In December 2025 Andrej Karpathy said AI had passed a threshold of coherence: the barrier to building was no longer code but upfront clarity. Around the same time Claude Code added session start hooks, and developers began feeding lessons files from one session into the next, a practice called compound engineering. Dave asked why that should stop at software. He built Dex, a local, markdown based AI chief of staff that knows his goals, projects and people, plans his day and tells him off when he runs off on three tangents on a Monday morning.

He open sourced Dex and three weeks later the head of AI at a top five tech company asked whether their executives could use it instead of an internal tool that 50 engineers had spent a year building. Building fast taught him that when AI is in the kitchen you get hundreds of ideas, and taste becomes the bottleneck. The head chef judges what is on the plate; the truffle hunter decides what is worth cooking. So he sends out what he calls his dog robots: agents that scan X and GitHub every night against his ICP, vision and backlog, and brief him each morning. His conclusion: when intelligence is effectively free, the advantage is the speed of the loop from signal to software to proof.

From there he laid out three moves. Out hunt: treat GitHub Trending as early radar on your market, because categories start as scrappy repositories, and open source alternatives to products built in stealth for two years now appear within days. Bend the system: when engineering gets faster the bottleneck moves to product, legal or evidence, so keep asking where work is stuck and redesign around it. He borrowed the shipping container from Sangeet Paul Choudary's book Reshuffle: the value was the reconfiguration of trade around the box. He then showed his software factory from Mind the Product London in June: a raw ideas column, a Steve Jobs 10x column and an agent ready column, each triggering a Claude Code skill, plus agents sharing his written truths about the coming year.

The embarrassing part: that whole setup is now in the bin. He had become the traffic controller moving cards between columns, and as models got better at coordinated work he was the bottleneck. Today he sits in one calm chat while an orchestration layer (herder.dev and the First Mate skill) runs agents in parallel and brings back only the decisions he needs to make. At Pendo, 500 people, engineering throughput jumped 50 percent and the bottleneck moved to product and design, with 30 squads each inventing workflows. The answer is Rex, a shared skills catalogue behind an internal MCP server, plus a conviction document that captures the evidence for a bet once and feeds spec, OKRs and GTM. Nobody at Pendo may log in to Pendo during Headless October.

The third move is prove impact, and Dave admitted he has no idea who uses Dex, because terminal products resist analytics and love is not data. Pendo's bet here is Novus: connect the code base as product context and combine it with behavioural analytics as user context, so humans and agents share a living picture of what exists and how people use it. He connected Dex's GitHub repo in five minutes and Novus worked out the ICP, tagged pages and events and found the start my day journey unprompted. He closed with the lemming problem: remove one constraint and another pops up. Last week his bottleneck moved to CI, roughly 200 pull requests a day, and he cut the cost per PR from 50 to 17 cents by switching runners.

### Key takeaways

1. Treat GitHub Trending as market radar: run a nightly agent that scans repos against your ICP and backlog and briefs you each morning.
2. Write down your truths, the hypotheses about customers, market and technology 12 months out, and share them as context with every team and agent.
3. When one function gets faster, find where work piles up next and redesign the system around that constraint, then repeat.
4. Replace the PRD with a conviction document: evidence, problem, audience, what would change your mind, then generate spec, brief, OKRs and GTM from it.
5. Standardise how the company works with AI through a shared, evolving skills catalogue, as Pendo does with Rex, and let people pick their model.
6. Pick a forcing function for the change you know you need, as Pendo did by banning logins to its own product for Headless October.
7. Define what finished means before handing work to an agent; verification is the most important part of agentic workflows.
8. Instrument your product so you can prove value; qualitative praise from people you admire tells you nothing about retention.

### Chapters (timestamps in the talk video)

- 0:01 Host intro by Sebastian
- 1:36 Smart glasses and the taste comfort blanket
- 6:20 Dex: compound engineering for knowledge work
- 8:00 Truffle hunter and the three moves
- 10:25 Out hunt: GitHub as market radar
- 13:54 What a moat still means
- 15:12 Bend the system of work
- 17:21 The software factory, then the bin
- 22:13 Pendo: Rex and Headless October
- 25:00 Prove impact: Novus and moving bottlenecks

### Highlight moments

- 4:09 to 5:09: **Taste is not the moat**. Dave retracts his own sous chef argument and names the December 2025 turning point where clarity, not code, became the barrier.
- 8:05 to 9:15: **From head chef to truffle hunter**. The central metaphor of the talk: taste becomes a bottleneck when ideas explode, so he sends out agents to hunt for him.
- 9:15 to 10:00: **The advantage is the speed of the loop**. The core argument in one block: when intelligence is free, winning means spotting signal, shipping and proving customer value fastest.
- 11:39 to 12:28: **200 brilliant ideas is a coordination problem**. A shared backlog with an AI makes him more creative and shows why the bottleneck moves to the system of work.
- 15:26 to 16:27: **Bend the system, not the task**. The shipping container analogy from Reshuffle explains why AI is a reconfiguration of work rather than a faster horse.
- 24:23 to 25:00: **Headless October at Pendo**. A vivid forcing function: nobody at Pendo may log in to Pendo for a month, and nobody knows how it will go.

### Quotes

- "The head chef judges what's on the plate, but the truffle hunter decides what's worth cooking in the first place." (8:29)
- "When intelligence is effectively free, the advantage no longer becomes who has the most data, the most engineers, the biggest roadmap. The advantage becomes the speed of the loop." (9:22)
- "200 brilliant ideas is not a roadmap, it's a coordination problem, it's chaos." (12:20)
- "AI fluency is a sharing problem, not a training problem." (23:04)
- "Love is not data. People coming back for seconds? I have no idea. Are they just being nice to me? Maybe." (25:56)
- "When everyone can build, the advantage isn't who produces the most software, it's who learns fastest about what's worth building and who can keep bending the system around that learning." (30:11)

### Mentioned

- Dex (heydex.ai): Dave's open source AI chief of staff, local markdown files for goals, projects and people; desktop and mobile beta at heydex.ai/beta
- Pendo: Dave's employer, 500 people, 30 squads; mascot is a dinosaur
- Rex: Pendo's shared AI chassis, a catalogue of reusable skills served from an internal MCP server
- Headless October: Pendo's forcing function, no logins to Pendo for a month, all work via Claude and MCP connectors
- Novus: Pendo's bet that connects the code base as product context with behavioural analytics as user context
- Conviction document: Pendo skill that assembles evidence around a bet and feeds spec, engineering brief, OKRs, board slide and GTM
- Andrej Karpathy: said in December 2025 that AI had passed the threshold of coherence
- Claude Code session start hooks and compound engineering: loading a lessons file into each new session
- Reshuffle by Sangeet Paul Choudary: AI as shipping container, the value is the reconfiguration around the box
- Stevenson at Index Ventures: questioned whether a moat dug once and held forever is still intellectually honest
- Martin Eriksson: earlier Just Product speaker whose assumptions up the stack framing Dave tied his truths to; Dave recommended his book
- GitHub Trending: real time view of what engineers are building, used as early market radar
- Twenty: fully functioning open source CRM surfaced by Dave's agents
- Open source alternative to Claude Design built by one developer in a week
- Linear, Jira, OpenAI Symphony: orchestration tools Dave tried in May and June before building his own board
- herder.dev and First Mate: open source agent orchestration Dave uses today, First Mate by a former Meta engineer
- Mind the Product London, June: where Dave first showed his software factory to a thousand product managers
- Pandemonium: Pendo's conference, where Sebastian first saw Dave speak
- SAP and MasterCard: examples of executives Dave works with as Field CPO
- GitHub Actions, WarpBuild and a cheaper CI runner (transcript: 'jog'): CI cost per PR dropped from 50 to 17 cents

### About the speaker

Dave Killeen is Field CPO at Pendo with 25 years in product across media, dating, health and fintech, previously at Productboard, BBC and Badoo/Bumble. He built and open sourced Dex, an AI chief of staff for product people.

- Dex: https://heydex.ai
- LinkedIn: https://www.linkedin.com/in/davekilleen/

---

## Building an AI Operating System

**Felix Geelhaar**, Director of Product and R&D Operations at PandaDoc. AI Transformation and Leadership, 15:15, Just Product 2026.

> AI only speeds up the wiring you already have, so rewire the front with problem, proof and prototype first.

### Summary

Felix Geelhaar runs product and R&D operations at PandaDoc, after years in engineering and product and a stint at Snyk. That seat lets him watch how product, engineering, design and go-to-market actually hand work to each other, and his talk opens with a scene everyone in the room recognised. Monday, right before sprint planning: quick one before the planning, I have an idea. A ticket appears. Build the smart filter, dropdown, tags, CSV export, design attached, this is final. The one pager was written by Claude from a couple of Gong calls and some metrics. It is six pages long, so nobody reads it. The ticket goes straight to an agent, and engineering never gets to ask why.

His thesis: AI speeds up the wiring you already have. However product, engineering, design and go-to-market connect today, AI makes that connection faster without fixing it. He names three joints. The handoff, where work changes hands. The check, what stands between the work and the customer. The signal, how the customer's reaction travels back. Agile optimised for time to market. With AI the goal becomes time to signal. And the front was thin before AI arrived: Atlassian's 2026 State of Product survey found that 82 percent of teams have no engineers involved at ideation, and only 4 percent can trace a roadmap decision back to customer evidence.

Then he speeds up every station. Engineering ships 42 issues faster, 420 pile up at review, and the natural reflex is an agent named Steve that stamps looks good to me on all of them. Go-to-market hears about the release when a customer asks why the button moved. He quotes Geoff Charles of Ramp: AI simply removes the bottleneck, but moves it. Backlog refinement, the slow ritual where engineering got to ask why, was what kept teams out of the build trap. Three or four hands in the room still do it. Claire Vo's observation completes the picture: your roadmap empties faster, which means every bad idea still on it ships too.

The fix starts at the front. Product brings the problem, the proof and a prototype, the practice Ramp uses and Marty Cagan describes in Transformed: a clickable truth that falsifies an assumption before engineering touches it. Then wire the check on purpose. Linear runs Quality Wednesday, one fix per person per week to train the eye for quality, and a feature roast before each beta where fresh eyes look and the beta owner decides whether to ship. Gene Kim's Wiring the Winning Organization calls this slowifying: slow down at chosen moments to go faster overall. Add deliberate launch moments so go-to-market stays wired in, then keep tracking the signal after launch instead of forgetting the feature.

The Q&A ran long because the keynote was moved back 15 minutes. On people and pushback, Felix described a cross-functional pilot at PandaDoc that first drew the company's interaction model, who talks to whom and why, then fixed one link at a time with retrospectives in between. On regulated environments, drawing on FedRAMP work at Snyk, he said evidence collection is the real gap and no tool collects it in real time yet. On where context should live, he argued against markdown piling up in repos and pointed to a queryable company layer that answers whether there is customer evidence for an assumption, with sources attached.

### Key takeaways

1. Map your handoffs, checks and signals before adding AI; it will speed up whatever wiring you already have, gaps included.
2. Measure time to signal, not only time to market: how fast does a customer reaction reach the team that built the feature?
3. Bring the problem, the proof and a clickable prototype to every handoff so engineering can still ask why.
4. Falsify assumptions with a prototype before a ticket exists; many ideas should never reach engineering at all.
5. Put a deliberate human check in the pipeline, like Linear's Quality Wednesday or a feature roast before each beta.
6. Decide what not to build before you speed up delivery, or the bad ideas on your roadmap ship first.
7. Create explicit launch moments so go-to-market learns about releases from you, not from a confused customer.
8. Pilot the rewiring with one cross-functional team, run retrospectives at each stage and watch where the bottleneck moves.

### Chapters (timestamps in the talk video)

- 1:20 Opening: an operations view on AI
- 3:10 Monday morning and the six pager
- 6:47 AI speeds up the wiring you have
- 9:33 The thin front: 82% and 4%
- 11:32 From 42 to 420: the bottleneck moves
- 15:00 Rewire the front: problem, proof, prototype
- 17:42 Wire the check: Quality Wednesday, feature roast
- 21:25 Slow down, launch moments, signal
- 23:33 What is product for now?
- 24:35 Q&A

### Highlight moments

- 6:47 to 8:03: **AI speeds up the wiring you already have**. The core thesis plus the three joints of the system: handoff, check and signal.
- 9:33 to 10:42: **82 percent and 4 percent**. The Atlassian numbers that show the front was thin before AI arrived.
- 11:32 to 12:52: **42 issues become 420 and Steve says ship it**. A vivid walk through what happens when every station speeds up but the wiring stays the same.
- 13:32 to 15:00: **Backlog refinement was the moment to ask why**. Three hands still do refinement, and an empty roadmap means the bad ideas ship too.
- 15:22 to 16:42: **The problem, the proof and a prototype**. The practical fix for the handoff, ending on the clickable truth line.
- 23:33 to 24:25: **What is product for when building is cheap?**. The closing question, the great question punchline and the call to rewire the front.

### Quotes

- "AI speeds up the wiring you already have in your organization." (6:47)
- "With AI, that changes, because you don't want that just something goes out the door. You want to get time to signal faster." (8:15)
- "This was the slow building that kept us out of the build trap. Because now what we have is the wiring. We are not asking the why anymore." (14:05)
- "All the stuff on your roadmap gets done faster. But it also means all the bad stuff is being shipped, because all the bad ideas are still on your roadmap and you haven't made the decision to not build it." (14:37)
- "With that prototype, you have essentially a clickable truth." (16:27)
- "I definitely didn't want to give that question to an AI, because the first thing I will get is: great question. Thinking. Looks good to me." (23:43)

### Q&A (not in the video, summarised from the stream)

**Q: What about the people side: fear, pushback, obstacles when you introduced this, and how did you solve them?**

Pilot in one area with a cross-functional team, set a goal and give people a target picture of what faster means for them, because installing AI as a tool only speeds up the wiring you have. At PandaDoc the pilot first drew an interaction model of who talks to whom, when and why, across internal enablement, support, customer enablement, AEs and SEs, then working groups fixed one link at a time with retrospectives at each stage. They watched the bottleneck move.

**Q: In a regulated environment development is fast but validation cannot keep up. Is there a solution?**

From FedRAMP work at Snyk: agents write GRC documents fast, but evidence collection is the gap and nobody collects it in real time while building yet, Vanta does it after the fact. Humans were still interviewed about how the system works, and the diagrams product security demanded helped teams understand what the agent had actually built. FedRAMP itself moved from around twenty thousand pages of documentation to one interview with evidence checked asynchronously by an agent.

**Q: Where should the context live? Right now I keep it in GitHub with the developers.**

Markdown files in repos blow up and go stale within hours without governance, and Boris Cherny says he regularly deletes his CLAUDE.md files for that reason. Felix expects a queryable company layer, which Y Combinator is already funding, that an MCP or CLI can ask for context including how fresh and trustworthy it is. His wish: a Slack bot that answers whether support signals or Gong calls hold customer evidence for an assumption, condensed and linked to the source.

### Mentioned

- PandaDoc, where Felix leads product and R&D operations
- Snyk, his previous employer, where he went through FedRAMP certification
- Dave Killeen and Dex, the preceding talk he says he will try
- Verena Graber, cited on story points no longer being needed
- Faith Forster and the Makers Manifesto, cited on go-to-market finding out from customers
- Thomas Hartmann's agent OS talk, cited on dogfooding
- Martin Eriksson, cited on assumptions and bets
- Atlassian State of Product 2026 survey, source of the 82% and 4% numbers
- Geoff Charles of Ramp at Lenny's and Friends Summit, source of the bottleneck quote
- Claire Vo of the How I AI podcast, on roadmaps emptying faster
- Ramp's practice of shipping prototypes with every handoff
- Marty Cagan's book Transformed, on prototypes that falsify assumptions
- Linear's Quality Wednesday and feature roast, reported by Karri Saarinen
- Gene Kim's book Wiring the Winning Organization, on slowifying
- Melissa Perri's Escaping the Build Trap
- Gong calls as raw customer context
- Claude as the one pager writer in the opening story
- Vanta, for after the fact compliance evidence
- Y Combinator, funding queryable company startups
- Boris Cherny, on deleting bloated CLAUDE.md files

### About the speaker

Felix Geelhaar is a technology and operations leader focused on operating leverage through systems, organisational design and AI enabled platforms. He leads product and R&D operations at PandaDoc and previously held leadership roles at Snyk.

- LinkedIn: https://www.linkedin.com/in/felix-geelhaar/

---

## Move Early, Commit Carefully

**Simina Pasat**, Former VP of Product, GitHub. Closing Keynote, 16:30, Just Product 2026.

> AI made building cheap but promising did not get cheaper, so build to learn and commit only to what deserves a promise.

### Summary

Simina Pasat closed the day with a homecoming. Fifteen years ago she came to Munich for a master's at TUM, worked at hybris, then moved to Seattle for Xbox and GitHub, where she was most recently VP of Product for GitHub and Copilot. She opened with Olympic skier Eileen Gu's line that hard work is overpraised and direction underrated: working harder means punching the wall harder, direction means noticing you can walk around it. Translated to product, hard work is building and direction is deciding what to build, for whom and why. Her warning: it has never been easier to build something nobody wants. The old cost of engineering forced product teams to ask the boring questions early. That forcing function is gone.

Her advice is to build to learn, not to commit. Write down the hypothesis and what success looks like first. Customers loving a demo is not success, because they can love ten other products like it. Two stories showed the gap between a demo and a product. At hybris labs her team built a digital wine shelf that lit up recommended bottles in the store. Retailers loved it, but moving it from the lab into the business was the hard part. At Xbox she worked on Game DVR while Twitch took off. The team built prototypes, but a bet that size needed someone at the very top to commit, and nobody did. AI does not fix that: you can build Twitch overnight, but not the distribution and network effects.

At GitHub she joined Special Projects, formed after the Microsoft acquisition to win back developer trust. In one year they shipped over a hundred changelog improvements, and betas drew tens of thousands of developers. One fast team stress tested the whole organization. Release processes, reviews and a comms team needing days for copy on something built in an afternoon started to break. AI now gives every team that speed, and when judgment cannot keep up, you build the wrong thing faster. Atlassian surveyed about 1,000 product people: 80 percent ship faster with AI, customers see no more value. Customers also stopped waiting. Banks, German carmakers and healthcare companies want to try a demo the same day. IBM found two thirds of CEOs invest before understanding the value.

Her sharpest story came from launching GitHub Copilot Enterprise. While her team was still prototyping AI code review, a financial institution that never ran betas was already piloting pull request reviews with a ten-person startup. Two lessons: build ahead of what today's models can do, and startups may make mistakes a large vendor cannot afford. GA still means security reviews, compliance, SLAs, support, docs and pricing, plus a promise to keep it working for years. She then asked where product advantage still holds. Adoption starts bottom-up, as GitHub and later Cursor, Claude Code and Codex spread from personal accounts into companies, while commitment happens top-down. Value moved from access to usage to impact, and pricing followed from per seat to per token to Intercom's 99 cents per resolved ticket.

If outcomes are what matters, only features that drive the outcome deserve to exist. Anyone can add now; the skill is knowing what to take away. Trader Joe's stocks 4,000 products against a typical supermarket's 50,000 and sells about four times as much per square meter because customers trust the selection. Agents are becoming the customer, and they do not see your onboarding or landing page, only whether the tool call worked. The advantage goes to whoever learns fastest, compounded through context, trust and community. Simina left GitHub this summer and is building hands-on again, running a Friday retro with Claude on what pushed the limits and what she learned. Her closing line: move early enough to learn, and commit carefully enough to endure.

### Key takeaways

1. Before building anything, write down the hypothesis, what you want to learn and what success looks like; customers loving it is not success.
2. Ask the boring questions while the demo is still exciting: does it solve a real problem and will anyone pay for it?
3. Build ahead of what today's models can do, because the model you are using today is the worst one you will ever use.
4. When one team speeds up, check release, review, documentation and comms processes, because they were built for a slower pace and will break.
5. Launch with the outcome, not the feature: show measured impact from internal use and pilot customers, like review time dropping from days to hours.
6. Define one impact metric tied to the problem statement, because boards and CFOs now ask for ROI rather than activation or adoption rates.
7. Spend as much judgment on what to remove, combine or kill as on what to add; users and buyers cannot absorb more.
8. Pick one thing each week that pushes the limits with the goal of learning fast, and run a Friday retro on what you learned.

### Chapters (timestamps in the talk video)

- 0:00 Host intro: how Simina was found
- 1:47 Homecoming and Eileen Gu's direction
- 8:27 Build to learn, not to commit
- 9:19 hybris labs, Xbox and Twitch
- 12:55 GitHub Special Projects breaks the org
- 16:35 Customers stopped waiting
- 18:13 Copilot Enterprise and the startup beta
- 23:01 Adoption bottom-up, commitment top-down
- 26:01 Access, usage, impact and pricing
- 33:15 Agents as customers and closing

### Highlight moments

- 5:21 to 6:34: **Punching the wall harder is not progress**. The Eileen Gu analogy that frames the whole talk, landing on the line that it has never been easier to build something nobody wants.
- 11:06 to 12:29: **Xbox saw Twitch coming and never committed**. A vivid story of a company that had the prototypes but lacked a top-level commitment, and why AI still does not fix that.
- 19:03 to 20:14: **The bank that chose a ten-person startup**. Her biggest learning from Copilot Enterprise: a regulated customer took a risk because the pain was that big, so build ahead of the models.
- 21:31 to 23:01: **Building got cheap, promising did not**. The core argument plus her personal US and Europe synthesis: move fast, then decide what you commit to.
- 31:29 to 32:41: **Trader Joe's: 4,000 products, four times the sales**. A surprising number that makes the case for taking features away instead of adding them.
- 36:11 to 37:20: **Push it until it breaks**. Her closing advice, made credible by her own new chapter of building hands-on after leaving GitHub.

### Quotes

- "In product terms, it has never been easier to build something nobody wants." (6:27)
- "A demo is something people admire. A product is something they get to depend on, and in many cases pay for." (10:21)
- "When one team moves fast, every process built for a different speed starts to break." (14:46)
- "For a product team, AI made building cheap. But promising didn't get any cheaper." (21:31)
- "You can copy features, but you cannot copy where people are." (35:21)
- "Move early enough to learn, and then commit carefully enough to endure." (37:50)

### Mentioned

- Eileen Gu, Olympic freestyle skier and Stanford graduate, source of the hard work versus direction analogy
- Jay Shetty's On Purpose podcast, where Eileen Gu made the direction remark
- TUM and CDTM, where Simina did her master's in Munich 15 years ago
- hybris and hybris labs, Munich commerce company acquired by SAP for about 1.5 billion dollars, home of the digital wine shelf prototype
- Xbox Game DVR (Xbox Record That), the feature she worked on while Twitch took off
- Twitch, the live streaming platform Microsoft tried to buy before Amazon acquired it
- GitHub Special Projects, the small team created after the Microsoft acquisition to win back developer trust
- GitHub Copilot Enterprise, the AI offering she launched with AI-assisted code review in preview
- Atlassian product survey, about 1,000 product people, 80 percent ship faster but customers see no more value
- IBM CEO study, 2,000 CEOs, two thirds invest in AI before understanding the value
- Kevin Weil, former CPO at OpenAI, quoted that today's model is the worst you will ever use
- Facebook's move fast and break things slogan, retired in 2014
- Cursor, Claude Code and Codex, AI coding tools that spread bottom-up from personal accounts into companies
- BMW Welt, where she met GitHub's German customers last year
- Intercom, charging 99 cents per resolved ticket as the example of per-outcome pricing
- Nan Yu, at OpenAI's Codex and formerly Linear, on people's ability to absorb what you give them as the biggest constraint
- Trader Joe's, US grocery chain owned by Aldi Nord, 4,000 products versus 50,000 and about four times the sales per square meter
- Charles Lamanna, who runs Copilot and agents at Microsoft, on agents being very harsh customers
- Dex, Dave Killeen's open source AI chief of staff, which she plans to switch to for her Friday retro
- Claude, which she uses for a weekly retro on what pushed the limits and what she learned
- Earlier speakers Martin, Verena and Dave, whose points on strategy, backlogs and truffle hunting she built on

### About the speaker

Simina Pasat was VP of Product at GitHub, where she led product management for the developer platform and took GitHub Copilot Enterprise to general availability. Before GitHub she was at Microsoft, with earlier stops at Xbox Live, SAP hybris and 4PSA. She holds a Master's from TU Munich and now builds again hands on in Seattle.

- LinkedIn: https://www.linkedin.com/in/siminapasat/

---