What AI Resume Fit Scores Actually Learn From

These days the applicant list for a busy role often comes with a number next to each name. One candidate is a 73% match, another is 41%. There are 200 of them and you have a morning to get through the list. What do you do with the numbers?

You use them, of course. The number is there to save you time, and nobody told you not to trust it. Most teams never asked for fit scores. They came bundled into ATS upgrades and got switched on by default, and I suspect very few of the people relying on them have looked closely at what they measure.

So what does 73% mean? It's a prediction against a pattern the model learned from its training data, and that data is usually a record of past hiring decisions: who got an offer. It's much more rarely a record of who turned out to be good at the job. Those sound close, but they're very different things to learn from.

A model trained on who got hired learns what made past hiring managers say yes. Someone who worked at a recognizable company under a familiar job title will score well, because people like them got offers before. The model hasn't measured anything about their ability. If your past hires came mostly from certain schools and companies, the score will favor people who look like them, whether or not that background had anything to do with how those hires did. What the score ends up rewarding is familiarity.

There's a second problem in the data, and it's subtler. The resumes that went far enough to become hires had already passed earlier screens. So the model never sees the excellent candidate who was rejected in the first round. It learns from a sample that was filtered once already, then applies what it learned to the whole unfiltered pool. Whatever the earlier screens got wrong gets baked in.

None of this would matter much if people treated the score as a rough hint. But a number doesn't invite doubt the way a hunch does. If a colleague said they had a good feeling about someone, you'd ask why. When the ATS says 73%, you move on to the next name. The recruiter working through those 200 applicants is acting on the model's learned preferences without knowing that's what they're doing. That's why I think the confidence is a bigger problem than the inaccuracy. An uncertain tool makes you hesitate, and a confident one makes you act.

The people who lose out are strong candidates from unusual backgrounds. They get filtered before anyone reads their resume, because they don't resemble the people who were hired before them.

What would you want to know about an applicant early on, if you could know anything? I'd want to hear how they talk about a problem they've solved. Do they get specific, or stay at a comfortable level of abstraction? What happens when you ask a follow-up they didn't see coming? None of that is on a resume, so none of it can be in a score built from resumes. You only find it out by talking to the person.

This is where a voice conversation, even an automated one, has an edge over resume parsing. The candidate with an odd CV who explains their work clearly and in detail shows you something a fit score would never find. So does the candidate with a polished background who can't explain their own decisions in plain words. In a conversation both of them become readable in a way their documents didn't allow.

We built Asendia around reversing the usual order. Fit scores exist partly because nobody has time to talk to 200 people. A recruiter can only hold so many screening calls in a day, so something has to decide who gets one. Software doesn't have that limit, so Asendia talks to everyone first. When someone applies, it calls them within hours, day or night, and has a spoken conversation that adapts to what they say and follows up on the specifics. What comes out is a qualification summary based on how the person communicated, what they said about their background and how they handled the follow-ups, rather than on how their resume was formatted. The summaries and a ranked shortlist go into the ATS. For an agency running a high-volume campaign, that means the shortlist was filtered by what candidates actually said, and nobody sat in a queue waiting for a recruiter to get to them.

There's a broader point here about AI that moves candidates through the next step of a pipeline versus AI that only sorts them, which the post on agentic recruiting goes into.

I don't think the answer is to stop using AI in recruiting. Fit scores will keep spreading, and most will keep being presented as objective. What they really are is a mirror of your past decisions, shown with a precision it hasn't earned. The better use of AI is to point it at the candidate directly. If your ATS has fit scores turned on, here's a small test: look up the scores your best recent hires got when they applied. If some of them scored low, you've learned what the number is missing.

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