How to interview someone who will make AI useful at work

Choose three job-related capabilities: diagnosing a workflow, checking an AI result, and helping people adopt a safer process. Give interviewers a one-to-four scale with concrete examples. Ask every candidate the same core questions; allow the same planned follow-ups. The U.S. Office of Personnel Management describes this as a structured interview and reports stronger agreement among interviewers than free-form interviews.

Tell candidates what the session contains and how to request an accommodation. Do not feed resumes to an opaque ranking tool and assume a human at the end cures every screening problem. Keep a record of the criteria and why each criterion belongs to the job.

Spend five minutes on the role and assignment; 15 on a past rollout; 20 on a work sample; 10 on a colleague’s pushback; and 10 on candidate questions and next steps. Ask for a failed or incomplete example, not only a victory. “What did users do instead?” often reveals whether the candidate noticed adoption at all.

The work sample should be short, job-related, safe, and paid if it would create useful business work. Give a fictional or fully de-identified queue, one flawed AI output, and one operational constraint. Ask the candidate to mark what is usable, what needs a source check, and what a person must decide.

Ask: “Which real job did you improve, and what was the old baseline?” “What did you do when the tool was wrong?” “Who could reject a result?” “What did a reluctant user teach you?” “Which measure made you change course?” Strong answers use ordinary nouns: source record, owner, exception, correction, accepted work, customer, and reversal. They include a moment when the candidate changed their own plan.

Listen for whether the candidate can explain the work without hiding behind acronyms. Technical language is welcome when it helps; it is not a substitute for explaining the business consequence. Do not demand one rehearsed phrase. A thoughtful candidate may describe the same capability in different words.

Say: “Three weeks after launch, two experienced employees copy the AI output into their own spreadsheet and continue working there. What do you do?” A good candidate investigates why: missing fields, slow approvals, poor trust, an audit requirement, or a manager’s unofficial request. They do not begin by punishing the users or declaring adoption complete.

The strongest response proposes a small observation period, a fix to the approved system, and a measurement plan. If the spreadsheet protects a legitimate exception, the right answer may be to keep it temporarily with controls while the workflow is redesigned.

Have interviewers record scores before group discussion. Use four dimensions: problem framing, evidence and safety, delivery, and adoption. A one means no concrete example; two means a plausible partial approach; three means a complete approach with ownership and checks; four means the candidate anticipates failure and explains how they learned from it.

Do not turn this into a personality score. Record the evidence sentence supporting each rating. If the panel disagrees, return to the work sample. OPM’s research is general selection evidence; this rubric is our suggested adaptation for an AI-adoption role, not a validated employment test.