Build the answer around three things: the record, the bound and the reason. Say what the agent proposed and what the record showed, name the bound or source problem that made the proposal wrong, then give the reason you wrote down. Keep it under two minutes, finish with what happened next, and never claim the agent was simply "wrong".
What is the interviewer testing?
Not whether you distrust AI. Interviewers asking this want to know three things: that you look at evidence rather than tone, that you know the limits of what an agent may do, and that you leave a record a colleague could follow. A candidate who says "it just felt off" fails all three. A candidate who says "I always double-check AI" sounds cautious but says nothing.
They are also listening for proportion. Overruling an agent is sometimes right and sometimes an overreaction. The best answers make clear why this case needed a person.
How do you structure the answer?
STAR (situation, task, action, result) works, with three parts placed inside the action. For more on the format, see what a STAR answer is.
| Part | What to say | Example line |
|---|---|---|
| Situation | The queue and the case, briefly | "I was reviewing held supplier payments." |
| Task | What the agent proposed | "The agent drafted a payment of 4,000 to new bank details." |
| Action: the record | What the record showed | "The record showed the new details came from an emailed letter." |
| Action: the bound | Which bound or source rule applied | "That's an external source, and an external claim can't authorise a payment." |
| Action: the reason | What you decided and wrote | "I declined, with the reason recorded, and asked for confirmation via the supplier's verified contact." |
| Result | What happened next | "The supplier confirmed they'd never changed details. The request was fraudulent." |
If the result was less dramatic, say so. "The supplier confirmed the change, and the payment went out a day later with a clean record" is a good result too.
What does a full answer sound like?
"In my practicum I was working a queue of held refunds. The agent had drafted a refund of 600 for a customer who said a parcel never arrived. The record showed the customer's account history was sealed, from the system of record, but the non-delivery claim was declared by the customer and the carrier's tracking showed a signed delivery. The amount was also above the business's own limit for unsupported claims, which is why it was held. I declined the refund as drafted and escalated with both facts, recommending a carrier investigation first. My reason in the record was one sentence: 'Declared non-delivery contradicts signed carrier record; above limit for unsupported claims.' The approver agreed and opened an investigation."
Notice what it does: names the provenance of each fact, names the bound that held the action, and quotes the recorded reason. That is a grounded answer: every claim tied to a source.
What makes the answer weak?
- Blaming the agent. "The AI got it wrong" suggests you do not understand that agents act within bounds and on sources. Say which source or bound was the issue.
- No record. If you cannot say what you wrote down, the interviewer wonders whether anyone could check your decision.
- Too much screen detail. Interviewers do not need to know which button you pressed.
- No result. End with what happened, even if it was ordinary.
What if you have never overruled an agent?
Use the practicum. If your job does not use Pryme, you did the practicum on a simulated queue in a free workspace, and you made real decisions there to a standard of nineteen in twenty correct with no missed red line. Say that it was simulated, then tell the decision. Interviewers care about the reasoning more than the setting.
If you have only done labs so far, describe a lab where you traced a refusal to the bound that decided it. It is weaker evidence, but it is honest. A ready-to-fill version is in the STAR answer template for AI agent interview questions.
Where the Academy fits
Pryme Intelligence Academy gives you decisions to talk about: labs in a free Build workspace that trace verdicts to the bound that decided them, and a practicum where every decision carries a recorded reason. Certified Pryme Operator, issued through VerifyDoc.ai, lets the interviewer check the standard behind your answer at source.
Questions people ask
How long should the answer be?
Ninety seconds to two minutes. If the interviewer wants more, they will ask a follow-up, which is better than losing them halfway through.
Should I pick a dramatic case?
Pick a clear one. A fraud caught is memorable, but an ordinary case explained precisely shows more judgement than a dramatic one told vaguely.
Does this question come up for part-time roles too?
Yes, in much the same form. See how part-time workers explain AI agent skills, and the guide to AI skills for job seekers for the other questions employers ask.