AI CV screening: what to check before you rely on it
Echipa HR 365 · reviewed 2026-09-10 · 5 min read
The difference that matters is not between “using AI” and “not using it”, but between ranking and rejecting. A model that puts the top twenty CVs in front of you is a tool; the same model, in a process where nobody reads the rest, has become a selection filter — and the transition happens without any explicit decision.
Where the line gets crossed, in practice
Not through configuration, but through habit. In week one, the recruiter reads the tail of the list too. In week three, they have five open roles and no longer get there. By week six, “I reviewed them” means “I read the first fifteen”. Nobody decided anything, but the process has become one of automatic rejection.
| Ranking | Rejecting | |
|---|---|---|
| Who decides the rejection | A person, for each candidate | The score, by virtue of nobody reading |
| What happens to the bottom of the list | They are read, even if faster | They are never opened |
| Can a rejection be explained | Yes, with a reason | No, only with a score |
| Is a systematic error noticed | Yes, because somebody sees every CV | No, because nobody sees what was filtered out |
What you test beforehand
A one-hour test, worth doing once
- 1 — Take twenty CVs from a completed hire, where you know who was hired and how they got on
- 2 — Run them through the model and compare the ordering with what you now know
- 3 — Check where it placed the person you hired — if they are at the bottom, the model does not measure what you think
- 4 — Check what it penalised: career breaks, unusual phrasing, experience from another field
The fourth step usually produces the most useful discoveries. Models systematically penalise atypical profiles — people who changed field, who took a break, who describe their experience differently from the standard. Sometimes those are precisely the candidates who did best.
What you never delegate
- Automatic rejection below a score threshold. It is the one rule that turns ranking into rejecting, by definition.
- Assessing cultural fit or personality from text. Nobody can do that from a CV.
- Inferring characteristics that should not matter — age, gender, origin — from indirect clues.
- Communicating the decision to the candidate. An automatically generated rejection is felt immediately.
The third point is the most technically delicate: a model can infer such characteristics from a graduation year, a name or a turn of phrase, without anyone asking it to. You cannot directly verify that it does not; you can only make sure nobody is rejected automatically.
Where it genuinely helps
| The task | Why it is safe |
|---|---|
| Structured extraction from a CV: experience, education, tools | It is visually checkable, and an error is immediately obvious |
| Finding candidates in a database from a description | It rejects nobody, it only brings people forward |
| Summarising a long CV for the interviewing manager | The manager has the full CV alongside |
| Flagging the absence of a stated hard requirement | The criterion is objective and a person verifies it in five seconds |
The last row is the most useful in practice and the one most often confused with automatic rejection. The difference: the model flags “no driving licence, which was a requirement”, and the recruiter confirms in five seconds. It does not reject — it draws attention.
How do I know if I have crossed into rejection?
Check, on your last hire, how many CVs from the bottom half of the list were actually opened. If the answer is almost none, you have crossed, whatever the procedure says.
Should I tell candidates I use AI?
Transparency is generally a good choice and increasingly expected. Practically: it is easier to explain a process where AI ranks and people decide than one where you cannot say who rejected them and why.
Where to start
Run the test on a completed hire. It is the only way to find out whether the score has anything to do with what reality proved, and it costs an hour.
Then set one simple, verifiable rule: nobody is rejected without a person having opened the CV. Every other detail matters far less than this one.
What you tell candidates
Transparency is easier than it looks, if the process is built correctly. “We use a tool that helps us rank applications; the decision to go forward is made by a person, and every CV is read” is a sentence you can say without hesitation — if it is true.
If it is not true, the difficulty of phrasing it is itself the information: a process you cannot describe to candidates without sounding evasive is usually one you did not design but allowed to form.
The case of volume hiring
At hundreds of applications per role, “a person reads everything” becomes impractical — and this is where automatic filtering is most often justified. The reasonable compromise: filter only on objective criteria stated in the advert (availability, schedule, a required authorisation), not on fit scores.
The difference is verifiable: a candidate rejected for not being available for shift work can be told exactly why; one rejected for a score of 61 cannot. And the first form survives any question, including your own, six months later.
CV analysis with an explained score, not just a number
Every score comes with its reasons, and all candidates stay in the list — so the ranking helps without anyone dropping out of the process unread.
You can create an account in a few minutes and use every module for 7 days, no card required.