Asked in an AI engineering manager interview.
What they are really testing:
Whether you can turn a heated technical argument into a decision people will stand behind, without settling it by seniority or by whoever argues loudest.
A strong answer:
- Agree on criteria before options: quality on a shared eval set, latency and cost targets, data handling and compliance constraints, and the operational burden over the next year.
- Write the options into a short design doc with the trade-offs for each, so the debate is about evidence rather than preference.
- Where the disagreement is empirical, settle it with a time-boxed experiment, such as comparing a retrieval-based and a fine-tuned approach on the same golden set with the same grading rubric.
- Name one decision owner and a deadline. Escalate with the options and evidence, not with a request for someone to referee.
- Classify the choice as reversible or hard to reverse. Reversible decisions get made quickly and revisited on defined triggers; irreversible ones, like vendor lock-in or data residency, deserve more scrutiny.
- Record the decision, its rationale and the conditions that would reopen it, then communicate it to every affected team, including the ones who lost the argument.
Common mistakes:
- Mediating indefinitely, so the disagreement becomes the default state while the roadmap slips.
- Deciding by seniority or volume, which costs buy-in from the team that lost.
- Reopening a settled decision in every review because the rationale was never written down.
How would you handle a case where the evidence favours one team but the other holds a veto on compliance grounds? Curious how others would sequence that.
Originally shared on LinkedIn.
An AI interview question every day, leadership for tech people and hands-on AI learning:
Top comments (0)