Every week another headline tells me AI has broken hiring. The resume was written by a model. The candidate on the call has a second screen feeding them answers. The screening step we trusted for years has gone quiet. So the panic arrives right on cue, and it always takes the same shape: lock it down, add a detector, ban the tool, catch the cheats.
I understand the instinct.
Shraddha Sunil and Mudit Saraf made the case in Harvard Business Review in June 2026, and one line of theirs stuck with me.
the ability to perform well in interviews is becoming infinitely scalable and practically free.
A polished application can leave you with an unanswered question: how much does this person understand, and how would they handle the work themselves? We need evidence beyond the finished document.
Hiring’s signals were weak long before AI
I think the concern about AI points to an older question: what does the assessment actually tell us about the person’s ability to do this job? Restricting a tool may be appropriate for a particular task, but it does not answer that question on its own.
A resume gives us claims to check. An interview gives us a chance to ask about experience. A take-home gives us a piece of work to examine. Each can help, but a polished answer alone leaves gaps. We may still not know what the candidate understood, which help they used or how they would respond if the situation changed. That’s where watching relevant work and asking good follow-up questions can help.
And there’s a machine running underneath all of it that nobody sat down and designed. The tracking software reads for keywords, so people write for keywords. A recruiter bolts on an AI filter, so the candidate bolts on an assistant to beat it. Round and round it goes, each side automating against the other, and every loop drags the whole thing further from the only question that matters: can this person do the job. Clear rules still matter. We also need an assessment that gives us a useful view of the work, rather than relying on how well someone presents an answer.
Stop policing the tool. Watch the work.
We want to get closer to the work the person would actually do. That starts with agreeing what we need to understand before designing the exercise.
We call our version a Work Simulation. It’s a paid task built for the role, with technical and behavioural criteria agreed in advance. The candidate receives a task deck explaining the context, activity, time, payment and permitted tools. Some roles suit a live exercise. Others benefit from a prepared piece of work followed by a live discussion. The format follows what we need to assess.
For a growth strategist, that might mean examining a bounded part of an account with authorised read-only access, or a representative anonymised dataset. We want to hear how they interpret the business metrics, what they would investigate and why. Someone who understands that work asks follow-up questions and introduces relevant changes without feeding them the answer. If AI is approved for the exercise, we observe how they use it and check its output. Live work gives us more to examine, but it does not prove that every answer is unaided.
The finished answer (useful, with questions still open)
- The application they submitted
- The experience they described
- The finished take-home task
- The recommendation they presented
- The qualifications they listed
The Work Simulation (relevant work, observed and discussed)
- A task built for the role
- Approved tools and bounded data
- Technical understanding and behaviour
- Questions about choices and changes
- Evidence against agreed criteria
What a candidate’s AI use tells you
When AI is relevant to the role and approved for the exercise, its use becomes part of the work we can assess. Permission to use a tool does not mean permission to upload any company or client information. Make those limits clear before the task.
Watch the specifics. Does the candidate check what the model gives them? Can they explain the recommendation in their own words and spot a wrong assumption? When the situation changes, what do they reconsider? If the role needs a base level of technical understanding, ask questions that let them demonstrate it. Record what you observed without guessing at their motives from one answer.

For a role that uses AI, checking its output and knowing when to question it are part of doing the job. The assessment should make room to observe those skills alongside the understanding the person brings.
We map the real job before we test anyone
The task needs to be built for the actual role. Realistic context makes the evidence more relevant. It does not make the exercise immune to rehearsed answers or outside help.
Before writing the task, we use the Job Map to agree what the person will own and what good work looks like. Then we identify the technical understanding and behaviours we need to see. A useful way to describe a behaviour is: when this happens, they take this action to create this impact. For a growth strategist, that might mean checking costs when sales rise but profit doesn’t, then explaining what needs investigation before recommending more spend. We study how strong practitioners approach that work and agree the criteria with someone qualified to assess it.
The task is designed to bring those signals into view. We record observations against the criteria and use them alongside what we’ve learned about experience and Work Drivers. The simulation can support our earlier view or challenge it. Either is useful. We want the hiring team to understand the evidence and the gaps before deciding.
Where this breaks
I should be honest about the edges, because every method has them. This is not a screen for the top of a funnel. You can’t sit and watch five hundred applicants, and you shouldn’t try. We use it with a considered shortlist, with the timing and sequence agreed for the role. It also needs a real task to build from, so a role that is still vague on paper has to be defined before any of this works, and that definition is its own honest piece of effort.
Someone can still receive help during a live call. A changed condition and a useful follow-up question can show more of their understanding, but neither makes deception pointless or impossible. The assessor needs enough technical knowledge to examine the answer, recognise a reasonable alternative and record what remains uncertain.
Memory is another limit. After several candidates, the sessions can blur together. With appropriate consent and permitted data handling, use notes or a recording to support a prompt assessment against the agreed criteria. An approved AI tool can help organise observations from an authorised transcript. A person who understands the work checks the evidence and owns the assessment.
Expect to improve the exercise after the first runs. Check the timing, instructions and whether it brought out the evidence you needed. Record meaningful changes and account for them when comparing candidates.
Hiring is matchmaking, and the match needs context. For your next role, write down the technical understanding and behaviours you need to see. Build a realistic, paid exercise around them, explain it clearly to the candidate and have someone who knows the work assess it. Watch, ask why and document what happened. That gives both sides something more useful to discuss than a polished answer on its own, while leaving room for the uncertainty every hiring decision carries.


