AI product management: Familiar enough to start, different enough to fail

Benedikt Kantus Sep 01, 2026 Please like ❤️ and restack 🔁 this article so others can find it, too. Thank you! You've been asked to own an AI feature. Maybe you raised your hand. Maybe it landed on your plate because nobody else wanted it. Now you're responsible for something that feels like product management, but occasionally behaves like something else entirely. Most PMs approach AI tools the same way they’d approach any other software. That works, until it doesn’t. And when it doesn’t, the failure mode is usually the same: you built something technically functional that nobody trusts, produces wrong answers at the wrong moments, or falls apart the second it faces a situation nobody anticipated. Petra Wille wrote an article that stuck with me. Four ideas in particular. Here they are. How would a human do this task? Before you think about AI, reverse-engineer the human workflow. What information would a person look at? What judgment would they apply? In what order? When would they do the task, and when wouldn’t they? This matters more than it sounds. If you skip it, you end up automating a process you don’t fully understand. That’s a recipe for an AI tool that works in demos and breaks in practice. What context does the AI need, and how does it get there? AI tools don’t just need instructions. They need a context delivery system. A system prompt is part of it, but only part. Think about retrieval: when does the tool search for additional information, and how? What other tools can it access? What should it remember between interactions, and what should it intentionally forget? These are product decisions. Not engineering decisions. Most PMs underestimate how much design work lives here, and then wonder why the tool behaves inconsistently. Who’s in the loop, and when? Full automation isn’t always the right answer. There’s a spectrum: human in the loop, human on the loop, fully automated. Where you land depends on how much users trust the machine, how costly mistakes are, and how easy it is to correct the tool when it gets something wrong. This is a deliberate product choice. Don’t let it become a technical default. How do you measure if it’s actually working? This is where AI diverges most clearly from traditional software. Iteration for AI includes evaluations. You need to measure accuracy, latency, and cost. Hallucinations are an accuracy problem. Slow responses are a latency problem. Both affect whether users trust the tool enough to keep using it. Search for “evals” and you’ll find a growing body of practice. Get familiar with it before you need it. AI product management is familiar enough to feel approachable. Different enough to catch you off guard. These four ideas are a useful place to start. Did I write the exact opposite recently ? No! Read again, there is no contradiction. AI won't kill the PM role. Same job, better tools. Benedikt Kantus · Aug 4 Please like ❤️ and restack 🔁 this article so others can find it, too. Thank you! Read full story What I read As usual, I will list some of the best articles I read on the Internet. I will keep a list of the best articles (currently >900) at https://www.digital-product-management.com. These are today’s picks: New Manager: Now the team is your product: Transition from individual product delivery to coaching and developing high-performing teams as a core leadership responsibility. The Multiplier: What Leaders Are Actually Paid For: Leaders must generate returns above baseline expectations by designing systems and coaching teams rather than maintaining technical superiority. Laws of Software Engineering: Collection of principles and patterns shaping software systems, team structures, and technical decision-making processes. Benefits Save 20 USD when subscribing to Proton: Secure, end-to-end encrypted email, calendar, drive and more. Save 10% on the first month when subscribing to Krater.ai: Every AI You Need. One Subscription.