Show Notes
Hot take: most companies building AI agents right now are setting themselves up to fail — and it's not the technology's fault. Deloitte says 25% of enterprises are piloting AI agents in 2025. Salesforce data shows agents succeed on multi-step tasks only 35% of the time. Sendbird's analysis puts the overall project failure rate at 80–90%. We're watching billions flow into systems that fail nearly half the time. The culprit? Applying deterministic product thinking to non-deterministic systems. In Episode 2 of Product Mastery in the Age of AI and ChatGPT, I pull back the curtain on: → Why the CrewAI vs. LangGraph debate is the THIRD most important decision you'll make → The 5-pillar AGENT Framework I built from real fraud detection systems at Walmart → The whitespace opportunity almost no PM is talking about → 3 things you can do THIS week to build agents that actually work This one's spicy. I want to hear where you disagree.
More from Product Mastery in the Age of AI and ChatGPT
Most teams don't decide how much autonomy to give their AI agent. It just happens — by default, by accident, by whoever wrote the prompt. Then three weeks later you're in an incident review asking "wait, why did it do that?" In this week's episode of Product Mastery in the Age of AI and ChatGPT, I break down what I call the Autonomy Dial — the spectrum between "agent suggests, human approves" and "agent acts with zero human touchpoint," and the three variables that should actually decide where any given decision sits on it: → Reversibility — can you undo it, and at what cost? → Blast radius — who and what does it touch when it's wrong? → Verifiability — can someone cheaply check it, before or after?
In the newest episode of Product Mastery in the Age of AI and ChatGPT, I break down why "Who Decides When the Machine Says No" is a product problem before it is ever an ethics problem, including: The real job to be done in a fraud decision, and why most systems only build for a third of it A segmentation model based on motivation and behavior, not risk score, that surfaces the exact users your roadmap is currently underserving A five-step exercise you can run this week with any ChatGPT or Bard-type tool to find the gap on your own roadmap If you build, manage, or design any system that makes decisions on someone's behalf, this one is worth listening to.
The best product managers are not the ones who build the most. They are the ones who can resist building for the longest. That single idea anchors the newest episode of Product Mastery in the Age of AI and ChatGPT, and it lands harder than ever in a world where a working prototype takes an afternoon. When building has never been cheaper, building the wrong thing has never been faster. The episode, The Technical PM Playbook, breaks down five techniques for using AI to understand the problem space so deeply that the right product becomes obvious: Problem laddering to find the need hiding underneath the feature The steelman and the gravestone for pressure-testing any idea Segmenting users by motivation and jobs to be done, never by demographics Mapping the journey that starts long before a user finds the product A negative space map for finding the segment every competitor ignores The most useful takeaway may be the simplest one. AI is a brilliant partner for generating hypotheses, and a terrible substitute for talking to a real person. The whole playbook is the loop between the two.
A practice I had maintained for over 13 years — gone overnight. I couldn't walk. Surgery. Months of physical therapy. Six months later, I'm still in rehabilitation, still working from home, still rebuilding. And yet — my work as a Staff PM at Walmart Marketplace continued. Full ownership. Full accountability. High-stakes deliverables in seller risk and data infrastructure. Even when showing up was the hardest thing I did that day. This season has been one of the most difficult of my life. It has also been one of the most clarifying. In this episode I talk about what I've learned — as a leader, and as a human.
The product managers who are quietly outperforming their peers are not working harder. They have changed the question they ask AI. Most people use ChatGPT the same way they used Google five years ago. Ask a question. Get an answer. Move on. The PMs building better roadmaps, finding underserved user segments their competitors have missed, and walking into executive reviews with arguments that actually hold up under scrutiny — they are using it differently. They are using AI not to produce content, but to pressure-test thinking. There is a meaningful difference between asking AI to write a PRD and asking AI: "What is the strongest argument that my core product hypothesis is wrong?" One saves thirty minutes. The other changes the product. The latest episode of Product Mastery in the Age of AI and ChatGPT makes the case for that shift with something rare in AI content: specific, tested, immediately usable techniques. Not theory. Actual prompts. With an explanation of why they work and what they surface that solo thinking tends to miss.