The job description says the candidate needs AI skills. The recruiter sends people who list ChatGPT, Claude, Copilot, a prompt-engineering course, and a few certificates. In the interview, each says they use AI every day.
That tells you almost nothing about whether they will change the work.
The hire you want has already rearranged a piece of their job around the tool. They can describe what used to happen, what now happens, where the result still needs judgment, and what became easier for somebody else. That is the evidence. A model name is a line on a résumé.
The distinction matters because the person you bring in for this role will set the ceiling for the team around them. A capable user makes their own day shorter. The stronger hire leaves behind a workflow, a check, or a habit that other people can use without asking them to repeat the demonstration.
Tool familiarity expires before the role does
Hiring teams often begin with the wrong questions. Which tools have you used? How often? Have you taken a course? The answers are easy to rehearse and go stale quickly. A candidate who knew every useful feature in January may be describing a product that has changed twice since then.
The changed-workflow question does better work. A good answer names the work, the input, the output, and the friction that remains. A marketing candidate might explain that they turned weekly competitor monitoring into a recurring brief, then kept the final message with the person who understands the account. A finance candidate might have made a first pass through invoices faster, while keeping an exception review for someone who knows the contracts. The details vary. The shape does not.
That answer gives you something to inspect. The useful follow-ups cover what they tried first, where the model failed, and what they changed after it failed. Someone who treats a bad first result as the end of the experiment will not get much further after you buy them a better plan. Someone who can explain the revision usually has a working relationship with the tool already.
The Selecting Talent guide has three interview questions that expose this pattern. They work because they ask for a story the candidate has lived through, rather than an opinion they can borrow from a podcast.
A useful hire can name the handoff
AI work still has a boundary. The model can gather, draft, sort, compare, and make a first pass. The person has to decide what needs context, taste, authority, or accountability. A strong candidate knows where that boundary is in their own work.
The useful follow-up is: “Where do you still have to be careful?” The answer tells you more than a polished demo. If they say the tool handles everything, they have not spent enough time with it. If they can name the point where a manager, reviewer, or subject-matter expert has to step in, they have a better chance of building work your team can trust.
This is the same question that determines whether a workflow is worth automating at all. Verification is the bottleneck: work compresses where a team can cheaply tell a good result from a bad one. The candidate who recognizes that constraint will choose smaller, useful work first. The candidate who ignores it will sell you a large, fragile automation with a cheerful demo.
The interview should include a small piece of real work
The cleanest screen is a paid exercise using an ordinary task from the role. It has enough context to make the work real, the tools the candidate would use on the job, and a chance to observe the work. The final document matters. The route there matters more.
You are looking for three things. Did they choose a sensible place for AI to help? Did they improve the result after the first pass disappointed them? Could they explain how a colleague would check the output before relying on it?
One question belongs after the exercise: what part of this would you put on a schedule or hand off to the team, and what would stop you from doing that? This separates a fast user from someone who can redesign a function. It also keeps the exercise honest. Some work should remain a supervised session because the judgment is too expensive to encode.
The same screen works for people already on your team, with less ceremony. Their evidence is sitting in the work they have changed. A one-to-one can cover the last workflow they rebuilt and the next one they would take on if a constraint disappeared. The Talent Screen skill turns that conversation into a role-specific set of questions and a rubric, which is useful when several managers need to judge candidates the same way.
Hiring for AI is also an internal promotion decision
The best candidate may already be in the building. They may be a mid-level operator whose manager has noticed that the recurring report arrives early now, or an analyst who quietly made an intake process easier for everyone around them. They rarely announce themselves with the right vocabulary.
That is why the search should begin inside the team as well as with a recruiter. The people who produce disproportionate output are often visible in usage and work artifacts before they are visible on an org chart. Recognizing leverage explains how to find that concentration without mistaking message volume for value.
Once you find them, the hiring question changes. They need more than a title and a bigger tool budget. They need room to make their methods shareable, a manager who will protect the time, and a path that does not require them to leave for the work to count. The retention problem begins when the person carrying a new way of working has no visible future where they are.
The hiring brief should describe changed work
The useful hiring brief asks for a candidate who can show a workflow they changed, the judgment they kept in the loop, and the work their team can now do differently because they were there.
That gives the hiring manager a better brief than “AI experience required.” It gives them a way to identify the person whose work will remain useful after the novelty wears off.