If you prepare only one behavioural story for an AI PM loop, make it a launch you stopped. It is the story that carries the most information per minute about the judgment this job needs, and it is easy to reach the loop without one ready.
Why this story does so much work
Every AI product team lives with the same tension. A demo looks wonderful, leadership wants it in customers' hands, and the evaluation results are good but not quite good enough, or good on average and bad somewhere important. The PM is usually the person who has to say which of those facts wins.
A story about stopping, delaying or narrowing a launch shows an interviewer most of what they want to see in one place. You defined quality with numbers rather than impressions. You read the failures. You held a position against pressure, and you did it in a way that kept the people around you on side. You also found a path that still shipped something, because a PM who only ever says no is not much use either.
At frontier labs, where values conversations weigh candour and safety judgment heavily, the story also shows that you can disagree in the open and change course when the evidence says so.
The structure
Use STAR-L, and spend your time where interviewers will.
Keep the situation short. One or two sentences on the product, the users and why the launch mattered. "We were three weeks from launching an assistant that drafted replies to insurance claim emails, and sales had already promised it to two large customers."
The task is your responsibility in the decision, not the team's. "I owned the launch decision and the quality bar."
The action is most of the story, and it needs evidence. What did the evaluation show, and on how many cases? What did you see when you read the failures? Who disagreed, and what did they argue? What did you propose instead?
The result should include numbers and what happened to the relationships. Did the narrower launch hit its targets? Did the customers accept the delay? Did the person who pushed hardest for shipping end up agreeing, or not?
The lesson is what you do differently now. Not a slogan. A practice.
An illustration
The following is a composite example to show the shape, not a real case.
"The assistant passed 89% of our 300-case set, which was above the 85% bar we had agreed. When I read the 33 failures, eleven involved claims mentioning water damage, where the assistant stated coverage terms more confidently than the policy allowed. Water damage was 18 cases, 6% of our test set, but nearly a fifth of claim volume in the two regions those customers operated in. So the headline number was hiding the segment that mattered most to them."
"Sales wanted to ship and patch. I proposed launching for every claim type except water damage, which would route to adjusters with the draft hidden, and agreeing a separate bar for water damage on a new set of 150 cases. Our head of sales disagreed strongly because one of the customers had used water damage in their pitch to their board. We took it to the customer together. They accepted the narrower launch once they saw three of the failures."
"Six weeks later water damage passed its bar and went live. The lesson I took is that I now weight evaluation sets by production volume per segment before agreeing a launch bar, and I put the segment floors in the launch document before anyone sees results."
Arithmetic check: 89% of 300 is 267 passes and 33 failures; eleven of 33 is a third of the misses, and eleven failures in 18 water-damage cases is a pass rate under 40% for that segment. When you build your own story, check numbers like these before the interview, because a good interviewer will.
Read diagram description
A STAR-L story with most of the time on the action. In the composite example, 267 of 300 cases passed, 89 percent, above an agreed 85 percent bar, but water damage passed only 7 of 18 cases, under 40 percent, and held a third of the misses while making up 6 percent of the test set. The decision was to launch every claim type except water damage, which got its own bar on 150 new cases.
What makes it credible
Specific numbers that hang together. A date, or at least a quarter. Named roles rather than "stakeholders". A cost to you: an argument you had to have, a relationship you had to repair, a target you missed. And a lesson that is a habit, not a value statement.
Stories fail when they are hero narratives (everyone else wanted to ship recklessly and only you saw the truth), when they blame research or engineering for the quality gap, or when the candidate cannot answer a simple question about the data.
The follow-ups to expect
Interviewers probe this story to find out whether it happened the way you told it. Prepare for these.
How did you choose the bar, and when? If the answer is "after we saw the results", expect a harder follow-up.
What would have made you ship anyway? A strong answer names the evidence that would have changed the decision.
What did it cost? Revenue, a deadline, goodwill. If nothing, the stakes were probably lower than the story suggests.
What would you do differently? This is the lesson again, from another angle, and it should match.
Would you make the same call today? Sometimes the true answer is "probably not, because newer models handle that case", and saying so shows you are tracking how capability changes the decision.
If you have no such story
Look wider. A feature you argued should use a rule instead of a model. A prompt change you rolled back. A pilot you ended early. A sales promise you renegotiated into an outcome commitment. Any decision where you chose evidence over momentum, and paid something for it, will carry the same signal.
Questions about this topic
Use a launch you delayed, narrowed or rolled back, or a feature you argued should not use AI at all. The point of the story is a decision made on evidence against pressure to ship, and that comes in many sizes.