What AI Can — and Can't — Tell You About a CV
We build AI CV screening for a living, so you'd expect this post to tell you AI reads CVs better than people do.
It doesn't. It reads them differently — and knowing exactly where the line sits is what separates recruiters who get real leverage from AI from the ones who get burned by it.
What AI is genuinely better at
Reading every CV with the same attention. The 4th CV of the day and the 84th get identical scrutiny. No human screener can say that — attention fades, and the strong candidate buried at the bottom of the pile pays the price. Consistency is the single biggest thing a machine brings to screening, and it's not close.
Holding the full criteria in mind, every time. A role has eight requirements. By CV number thirty, a human is pattern-matching on two or three of them. The AI checks all eight, every time, and tells you which ones matched and which didn't.
Reading past keywords. Traditional ATS filters miss the engineer who wrote "built payment infrastructure" because the job ad said "fintech experience." A language model understands those are the same thing. This is where modern AI screening genuinely differs from the keyword filters recruiters have (rightly) learned to distrust.
Writing down its reasoning. Every score comes with a written explanation you can audit in seconds. Your manual shortlist has no such paper trail — which matters the day a client asks why candidate X didn't make the cut.
What AI can't tell you
Whether the CV is true. AI evaluates what's written, not what happened. A well-written exaggeration outscores an honest but badly-written truth. Verification — references, portfolios, probing questions — remains entirely human work.
Whether the person will fit the team. Culture, temperament, how someone behaves under a deadline — none of it is in the document. Anyone selling you "AI culture-fit scoring" from a CV alone is selling astrology with a dashboard.
What the candidate could become. A career-changer's CV scores low against the role as written, and sometimes that's exactly wrong — the recruiter who knows the client would trade two years of experience for hunger will overrule the score. They should. That's the system working, not failing.
Anything about a bad job description. Give the AI vague criteria and it returns confident-looking nonsense. The quality ceiling of AI screening is set by how well the role is defined — which is, again, a human skill.
The honest division of labor
So the real workflow isn't "AI screens, human hires" or "human screens, AI assists." It's sharper than that:
AI does the reading. Humans do the believing.
The machine turns 60 CVs into a ranked list with written reasons in a few minutes — the part of screening that was never a good use of a recruiter's judgment, just a tax on their time. The recruiter then spends that reclaimed hour on the things the machine can't touch: verifying, sensing fit, spotting the diamond the criteria didn't anticipate, and talking to actual humans.
When we demo Nexu AI, we deliberately show a CV it scores low and walk through the written reasoning. Sometimes the recruiter agrees. Sometimes they say "actually, I'd interview this one anyway" — and that moment is the product working exactly as designed. The score started the conversation; it didn't end it.
The question to ask yourself
If your screening pile is small and your roles are unusual, manual screening may honestly be fine. But if your team spends hours a week reading CVs that mostly get rejected, you're paying senior-judgment prices for reading work.
That's the trade worth automating — and only that.
NodalNexus builds Nexu AI, a CV screening assistant that ranks candidates with a written reason for every score. If you want to see it on one of your real roles, start a conversation.