How do you decide when AI genuinely enriches the user experience versus when it's just feature theater?
When AI outputs are probabilistic and sometimes wrong, how do you design for graceful failure without eroding user trust over time?
How do you think about the feedback loop between user behavior and model improvement?
At what point in the product development process should PMs get deeply involved in training data decisions?
How do you balance personalization with the "uncanny valley" effect, where AI knowing too much about a user starts to feel invasive?
When AI reduces friction or automates decisions, how do you preserve meaningful user agency without just adding friction back artificially?
How do you structure success metrics for AI features when the value is often indirect or long-horizon like better decisions the user makes downstream?
What's your approach to communicating uncertainty in AI outputs to users?
How do you manage the ethics conversation internally when an AI capability is technically feasible and commercially attractive but feels like it crosses a line?
As foundation models commoditize, where do you see the defensible PM craft in AI product work shifting to?
