Structured Interview Guide for AI-Assisted Work

Setup

Do these four before the first candidate sits down. Two of them cost an email.

1. Name the task in the invitation. Ask the candidate to bring one task they finished in the last month with an assistant, and say the round stays on that task. A rule nobody states gets guessed at, and the guessing tracks how much interview coaching someone has had.
2. Say that using an assistant is expected, and that saying so costs nothing.
3. Write what a strong answer sounds like for each block, in the vocabulary of this role. A marketer opening the study behind a quoted figure and an analyst opening the filing are different acts, so the level text is different text. That belongs on the rating instrument rather than in this guide.
4. Decide who asks and who writes. One person can do both.

Then fix the order below and leave it fixed. Every candidate gets these five questions in this order, with these follow-ups. Ask the written follow-ups every time, and probe freely only after. Federal guidance for hiring panels asks for the same three things: the same questions in the same order, one common rating scale, and agreement in advance on what an acceptable answer contains ¹.

Running order

The round is forty-five minutes on one task.

Write each answer down as it is given, in the candidate's own words. A note written after the round is a memory of an answer rather than the answer.

Opening script

Read this out the same way to every candidate.

"This round is about how you work with an assistant on real work. Using one is expected here, and saying you used one costs you nothing. Keep every answer on the one task you brought. There are five questions, each has two set follow-ups, and every candidate gets the same ones in the same order. Your answers get written down as you give them, in your words. Take the time you need."

The five blocks

Each block asks about one behaviour. A question that asks about two gets one answer and loses both.

1. Framing

Ask. What was the first thing you typed to the assistant on that task, the actual first message rather than the tidy version?

Then. What would have made the answer wrong, and did you tell it that?

Then. What did you leave out of that first message on purpose?

Evidence of. Whether the constraints of a task get stated before any of it is handed over.

2. Sourcing

Ask. Which specific claim in what it gave you did you check, and where did you go to check it?

Then. What did that source turn out to say?

Then. How did you pick that claim as the one worth checking?

Evidence of. Whether checking is an act on a named claim, or an impression of the output as a whole.

3. Delegation

Ask. What did you keep and do yourself, when the assistant could have done it?

Then. Why that piece and not another one?

Then. What did you hand over that you would keep next time?

Evidence of. Whether the candidate can say where the assistant was the wrong instrument.

4. Rejection

Ask. What did you throw out, and on what grounds?

Then. What did you put in its place?

Then. What made you look at it twice?

Evidence of. Whether a wrong answer gets caught by a habit or by luck.

5. Effect

Ask. What changed because of the check: a number, a recommendation, a caveat you added?

Then. Who had already seen the version before it changed?

Then. What would have gone out if you had not checked?

Evidence of. Whether the checking reached the work, or stopped at the reading.

Those five sit where the thinking sits. In a survey of 319 knowledge workers describing 936 first-hand examples of AI-assisted work, the effort that remained shifted toward verifying an answer, fitting it into the task, and stewarding the work, rather than toward execution ². Nothing above asks which tools someone uses, how often they use them, or how they phrase a request. A tool list is evidence of nothing, and phrasing is the part an assistant supplies on the candidate's behalf.

Limits

• It rates nothing. The questions live here. The rating instrument is a separate blank document the interviewer fills in.
• It cannot tell you whether the candidate would do any of this on your work. Every answer is a report about work rather than the work, and reports about working with an assistant are unreliable in both directions: in a randomised trial, sixteen experienced open-source developers took 19% longer to finish issues with early-2025 AI tools available, and believed afterwards that the tools had made them about 20% faster ³.
• The cheapest closure is fifteen minutes of live work at the end of the same round: unfamiliar material, assistant open, one act watched instead of five accounts.
• It applies no law to your facts. A hiring decision made off these answers may have to be explained later. Once a complaining party shows a particular practice causes disparate impact, demonstrate that it is job related for the position in question and consistent with business necessity. 42 U.S.C. 2000e-2(k), effective November 21, 1991. Questions written from the work the role does, and answers recorded in the candidate's own words, are what leaves something to explain it with.