Who Is Taking Your Skills Assessment?

Take two applicants for an analyst job who get the same take-home exercise. One works through it alone over an evening and hands in something competent but a little rough. The other pastes the prompt into a chatbot, tidies up the output, and hands in something that reads like it came from a consultant. Which one gets the interview?

If you rank by score, you already know. And you've learned almost nothing about which of them can do the job.

A lot of hiring teams are in this position now, whether or not they've noticed. Skills assessments were supposed to be the fix for résumés that no longer meant much. The reason they looked like a good idea is worth going over, because it's also the reason they stopped working.

Résumés had drifted away from reality. Titles got inflated, summaries got polished, and credentials were only loosely connected to how well someone would perform. A work sample seemed to cut through that. Instead of reading claims about what a person could do, you gave them a small version of the job and looked at what came back. For a few years that was a real improvement.

But think about what a take-home measures. It measures a document that arrives in your inbox. It never measured the person directly; it was just safe to assume the person had made the document. Once AI tools could produce good documents, that assumption broke, the same way it had broken for résumés a little earlier.

One piece of research from 2025 put numbers on it: 67% of candidates applying for technical and professional roles said they had used AI tools to help with pre-hire assessments, and 41% had let AI do most of the task. I wouldn't lean hard on the exact figures, but the direction is what you'd expect. A coding exercise designed to take an afternoon goes much faster with a model's help. A scenario prompt meant to test strategic thinking comes back fluent and well organized, because the model has read a great many examples of what strategic thinking sounds like on paper. The work is good. It just doesn't tell you much about the candidate.

So what your top-scoring submissions increasingly reward is skill at using AI. For a few jobs that's a useful skill. For most, it isn't the one you set out to test.

The part your metrics won't show

What makes this worse than the résumé problem is that nothing in the pipeline looks wrong. Completion rates are fine. Scores may even have gone up.

Consider who ends up at the bottom of the ranked list. Some candidates do the exercise without help, because they think that's the point, or they don't have the tools, or the task is specialized enough that a model can't do it cleanly. Their work tends to be rougher. When you sort by score they fall below the cutoff, and the people above them are disproportionately the ones who used AI well. The filter can end up doing close to the opposite of what you built it for, and no report will tell you.

The assessment vendors know this, and some have added detection: behavioral biometrics, clipboard monitoring, timing analysis. I'm skeptical it can work for long. The thing being detected is a kind of tool that gets better every few months, and each detection method gives people one specific thing to work around. My guess is that detection mostly buys a feeling of integrity, while recruiters go on making shortlist decisions from scores they treat as trustworthy.

What a model can't do for the candidate

If you step back, no particular test is at fault. Anything a candidate completes on their own time, with nobody on the other end, can be handed to a model. That covers the take-home and the written scenario, and mostly the timed task in a monitored browser window as well. Better detection won't change it, because the weakness is in the format.

What a model can't easily do on someone's behalf is hold up their end of a live conversation. When the next question depends on what you just said, and the conversation turns somewhere you didn't expect, there's no quiet moment to go and ask a chatbot. Either you understand the thing or you don't, and it shows fast.

That's the idea we built Asendia AI on. Instead of assigning a task and grading what comes back, it phones the candidate, usually soon after they apply and often the same evening, and holds a structured screening conversation with them. From the candidate's side it's a real back-and-forth. They answer a question, and the next one depends on what they just said, which is how you find out whether their reasoning holds up under a question they didn't prepare for. The summary goes into your ATS with the rest of the pipeline, so it's there the next morning rather than in some separate review queue.

People prepare for conversations too, of course, and some now rehearse interviews with AI. If you've seen that, this piece on how AI interview practice has changed what you're really measuring is worth reading alongside this one. Still, rehearsal makes someone better at talking about what they know. A follow-up question they didn't see coming tests whether they know it.

I don't think anyone is going to fix the take-home. The more useful move is to stop relying on formats where the candidate doesn't have to be present. If your process currently ranks people by an asynchronous test, here's something to try: pull the assessment scores for your last several hires and ask whether they told you anything the first real conversation didn't.

Ready to transform your hiring strategy? Schedule a Demo with our founders today!