Tell me about a time you disagreed with a stakeholder or PM about a data or model decision.
This question tests whether you advocate for data-driven thinking without becoming combative, and whether you know when to escalate versus adapt. Strong answers show you led with evidence, listened to the stakeholder's actual concern, and reached a resolution that was better for the product than either original position.
How to think about it
What the interviewer is actually testing
Interviewers want to see that you are neither a pushover (who silently ships what a PM requests even when the analysis is wrong) nor a bulldozer (who lectures stakeholders until they surrender). The ideal candidate is diplomatically persistent: grounded in evidence, genuinely curious about the other person’s concern, and willing to adapt when the stakeholder surfaces a constraint the data doesn’t capture.
How to structure a strong answer
Set up the disagreement clearly. What did you believe the data showed? What did the stakeholder want to do instead? Make the tension legible — vague conflicts produce vague answers.
Describe how you raised it. Did you bring data to a meeting? Request a one-on-one? Draft a short memo with two scenarios? The medium matters; going directly to your manager’s manager without trying to resolve it first reads as poor judgment.
Show that you listened. Strong candidates find out what the stakeholder actually cares about underneath the surface request. Often the PM is optimizing for a launch timeline you didn’t know about. Often the business constraint changes your recommendation.
Describe the resolution. Ideally, your evidence changed the decision. If the stakeholder overrode you, describe whether you flagged it as a risk, documented your recommendation, and what happened after. Intellectual honesty about outcomes you didn’t control demonstrates maturity.
Skeleton example: “A PM wanted to ship a model with 70% precision because the recall was excellent. I believed that in our fraud context, false positives destroyed customer trust. I prepared a one-pager quantifying the customer support cost per false positive and proposed a threshold experiment. We ran a two-week pilot at both thresholds; the PM agreed to the higher-precision cutoff when the complaint-rate data came back.”