Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: Choose a story where the disagreement was substantive and you handled it through evidence: you made the criteria explicit, looked at the data together (evaluation results by segment, failure modes, cost), found a decision both sides could support (often a narrower launch or a staged rollout), and kept the relationship strong. Show that you could have been wrong and how you checked.
How to approach it
STAR-L, with the action section focused on how you moved the disagreement from opinion to evidence. Be fair to the other side; a story in which the PM was simply right and everyone else wrong tells the interviewer more about the narrator than about the decision.
The norm most employers describe is to argue before the decision and commit after it. Amazon's leadership principles put it directly: leaders "are obligated to respectfully challenge decisions when they disagree" and, once a decision is made, "commit wholly" (Have Backbone; Disagree and Commit, checked November 2026). Your story should show both halves.
A strong answer
What panels listen for.
| They want | Red flags |
|---|---|
| Clear understanding of the other side's view | Caricaturing engineers or researchers |
| Explicit criteria and evidence | "I convinced them" with no mechanism |
| A creative resolution | Winning by escalation |
| Willingness to be wrong | No acknowledgement of uncertainty |
| Relationship intact afterwards | Lingering conflict |
An illustration of the structure (adapt to your own experience):
Situation. An ML team believed a new ranking model for a marketplace search was ready: offline metrics improved several points. The PM was worried because user complaints about irrelevant results in regional languages had been rising.
Task. The PM owned the launch decision with the engineering lead.
Action. Rather than arguing, the PM proposed agreeing launch criteria first: offline improvement overall, no regression on any major language segment, and an online test with guardrails. The PM pulled 100 queries each in Hindi and Tamil and had two native-speaker reviewers label relevance with the ML team in the room. (In your own story, say how many cases, who labelled them and how much the labellers agreed; a regression on 100 queries is a signal to investigate, not proof.) The segment breakdown showed the overall gain came from English queries, while Hindi and Tamil relevance dropped. The ML team found a tokenisation issue in their training pipeline for those languages. Together they agreed to launch for English queries behind an A/B test while fixing the issue.
Result. The English launch improved conversion in the test; the fix for regional languages shipped later with a measured improvement. The PM noted that their own initial instinct (to block the whole launch) would have delayed real gains.
Learning. "I now push for segment-level launch criteria to be agreed before results arrive, and I bring data rather than opinions to disagreements. It turns arguments into a shared investigation." Anthropic makes the same point about agent evals, calling them potentially "the highest-bandwidth communication channel between product and research teams" (January 2026).
Tell it in the first person. The illustration is written in the third person so you do not borrow it. In the room, say "I proposed", "I pulled", "I was wrong about".
Follow the flow, one stage at a time.
Illustrated sequence; animation speed does not represent processing time.Read diagram description
The story arc: a substantive disagreement about launch readiness, agreeing the criteria before looking at results, examining evidence together by segment and failure mode, finding an option both sides can support such as a narrower or staged launch, the result with the relationship intact, and the learning that criteria come before results.
Variations. Sometimes the PM is the one pushing to launch and research urges caution. The same principles apply: criteria, evidence, staged options, and genuine openness to the safety or quality concern.
What interviewers probe next
"What if you couldn't agree even after looking at the data?" Escalate jointly with a shared write-up of positions and evidence, so leadership decides on facts, and commit to the decision.
"How much evidence did you need before deciding?" Enough to make the decision reversible-safe, not certain. A staged launch behind an experiment let us act on incomplete evidence because we could undo it.
"Were you ever wrong in a situation like this?" Have an honest example ready; it strengthens credibility.
How it's scored
| Signal | Weak | Hire | Strong Hire |
|---|---|---|---|
| Fairness | Other side caricatured | Understands their view | Represents their view fairly, acknowledges own uncertainty |
| Resolution | Escalation or persistence | Data-informed discussion | Criteria agreed first, shared investigation, creative staged option |
| Learning | None | Personal | Process-level change applied later |
Common mistakes
Stories where the PM was right and the team was wrong.
Resolving through seniority rather than evidence.
No mention of how the working relationship fared.