AI Interview Questions That Actually Predict Performance

29 August 20265 min read

Twelve structured interview questions we run through AI interviews, why each one predicts on-the-job performance, and the follow-up probes that separate a rehearsed answer from a real one.

Most interview question lists are trivia. A question predicts performance when it forces the candidate to describe a decision they owned, under constraints, with a result someone else could verify. Everything else is conversation.

Here are the twelve we run, why they work, and the probe that does the real work.

The core twelve

#	Question	What it measures	The probe that matters
1	Walk me through the last thing you shipped end to end.	Ownership, scope	Who decided the scope, and what did you cut?
2	Describe a decision you made with incomplete information.	Judgement under ambiguity	What would have changed your mind?
3	Tell me about something you got wrong.	Calibration, honesty	What did you change afterwards, concretely?
4	What part of your last role were you slowest at?	Self-knowledge	How did you compensate?
5	Explain a technical or process trade-off you argued for and lost.	Reasoning independent of outcome	What was the other side's strongest point?
6	Which metric did your work move, and by how much?	Impact, numeracy	How was that measured, and by whom?
7	Describe the hardest handover you have done.	Communication	What broke after you left?
8	When did you push back on a request from a manager?	Integrity, spine	How did you frame it?
9	What did you learn in the last six months that changed how you work?	Growth rate	Where has it shown up since?
10	Describe a time you had to work with an unclear spec.	Initiative	What did you write down to make it clear?
11	What would you do in your first thirty days here?	Role comprehension	What information would you need first?
12	What kind of work do you not want to do?	Fit, honesty	What happens when you have to do it anyway?

Why the probe is the question

A rehearsed answer survives the first ask and collapses on the second. The probe is where evidence appears: names, numbers, constraints, and the parts that did not work. In an AI interview this is a structural advantage — the model never runs out of patience, never skips the follow-up because the meeting is over, and asks it the same way for every candidate.

Scoring without inventing a number

Score against a rubric written before the role opened, and require each rating to cite the transcript line that produced it. A score with no traceable reason is unusable in a hiring decision and indefensible if a candidate challenges it. Recruitment and selection systems are treated as high-risk under the EU AI Act, and NYC's Local Law 144 requires an annual bias audit and candidate notice for automated employment decision tools — a cited transcript is what makes either survivable.

What to skip

Brainteasers, culture-fit chat, and hypotheticals about ideal worlds. They correlate with confidence, not competence, and they widen the gap between candidates who interview well and candidates who work well.