What AI screening can and cannot judge from a CV

Nazmus Sakib · 21 September 2026 · 7 min read

A CV with some lines highlighted and others greyed out, showing what a parser reads

A CV score is a measurement of one document against another document. That is a narrower thing than it sounds, and knowing exactly how narrow is the difference between using a score well and being misled by it. Here is what ours reads, what it separates people on, and the three things it cannot see at all.

What is actually being compared

Not a person against a job. A parsed CV against a parsed job description. Both documents get turned into structured fields first — roles, dates, skills, education, employers — and the scoring happens on those fields.

That has an immediate consequence worth sitting with: anything true about a candidate that never made it into the document does not exist as far as the score is concerned. Not "is weighted lightly". Does not exist.

The six things it reads

Every score is the sum of six dimensions, each of which has to quote the line of the CV that earned it. The weights are set per role, because "analyst" at one company is not "analyst" at another.

Dimension	What it is actually reading
Skills Match	Requirements from the posting that appear anywhere in the CV
Experience Relevance	Whether previous work resembles this work
Education & Certifications	Qualifications, where the role asks for them
Achievement & Impact	Evidence that something changed because of them
Role Alignment	Whether the CV reads like someone aiming at this job
Stability & Tenure	Pattern of time in roles

Which of them actually separate people

This is the part nobody publishes, so here is ours. Across all 78 real CVs we have scored — one company, six jobs, September 2026, and yes that is a small sample — the six dimensions behave very differently.

Experience Relevance, Role Alignment and Stability spread widest. They run the full range from zero to the maximum, and they are doing most of the work of telling candidates apart.

Achievement & Impact is the flattest thing we measure. Its average is roughly half of what Education averages, and almost everybody clusters at the bottom. Not because the applicants achieved nothing — because CVs almost never say what changed. "Handled the monthly reporting" earns very little. "Cut the monthly reporting cycle from five days to one" earns a great deal, and describes the same person.

A low Achievement score is usually a fact about the CV, not about the candidate. Read it as "this person has not been taught to write about impact", which is a coachable problem, not a disqualifying one.

The same effect shows up in skills coverage, which is simply how much of what the posting asked for appears anywhere in the document. In our set the median was 38%, and 62% of candidates matched under half of the requirements listed.

Three things it cannot see

1. Anything that is true but unwritten

Someone who has run the monthly close for four years but wrote "assisted with finance operations" scores as an assistant. The model is not being fooled — it is reading what is there. Which is why the score belongs next to the CV, not instead of it.

2. Why a gap exists

An eighteen-month gap looks identical whether the person was ill, caring for a relative, running a business that failed, or studying. A recruiter asks in the first minute of a call and gets an answer in a sentence. A scoring model has nothing to ask.

3. Requirements nobody wrote down

Job descriptions leave out what everyone on the team already knows: that this role sits between two departments that disagree, that the last person burned out, that the tool in the posting is being replaced next quarter. The score is measured against what was written, which is always less than what is true.

What it is genuinely better at than a person

None of the above makes the score useless. It is very good at the parts people do badly:

Consistency. CV number 400 gets the same attention as CV number 4. Honest recruiters know they cannot do this.
Order independence. In a manual pile, arriving early is an advantage — we have measured what that costs. In a scored list it is not.
Showing its working. Every dimension points at the sentence that produced it, so your team can disagree with a specific line instead of vaguely.

How to read a score without being misled

Read the dimensions, not the total. Two candidates on 55 can be completely different people — one strong on experience and weak on credentials, one the reverse.
Treat a low Achievement score as a writing signal. Ask in the interview instead of discounting the person.
Check what the posting asked for before blaming the candidates. If everyone scores badly on skills coverage, the requirements list is often the problem — nine tools when the work needs three.
Never let it be the last step. The score is evidence for a conversation, and a person should press every button that affects someone's application.

The honest summary

It reads a document carefully, consistently, and in the open. It cannot read the person behind it, and it is worst precisely where CVs are worst written — which is most of them. Used as a ranking and an argument-starter, that is genuinely useful. Used as a verdict, it is a liability, and the fact that it shows its working does not change that.

Questions we get asked

What can AI actually tell from a CV?

How well one document matches another: skills named, experience that resembles the work, qualifications, evidence of impact, role alignment and tenure. Every point has to quote the line of the CV that earned it.

What can AI not tell from a CV?

Anything true but unwritten, the reason behind a gap, and any requirement the job description left out. All three are things a person gets in one phone call.

Why do candidates score low on achievement?

Because CVs rarely say what changed. It is the flattest dimension we measure, and a low score there is usually a fact about the writing rather than the person.

If you want to see this on your own roles, request a demo — we set up a live job with your job description and show you the shortlist. You can also read the six dimensions.