Recruitment: why an AI score must always explain its choices

A score looks objective. "82/100", "strong match", "not recommended": three words, one number, a decision that already feels made. Recruiters are busy, so a score saves time and feels like a neutral sort. But why 82? What did the AI see, what did it miss, and which criteria actually mattered? If the recruiter cannot answer that, the score is a black box. And recruitment — a career, an income, fair treatment — is far too sensitive to decide behind one.

A number is not enough

A CV is not a whole person, and a job title is not a skill set. A candidate may lack the exact title you searched for while having done, in a previous project, precisely the work the role requires. A well-written profile, by contrast, can look perfect on paper and prove nothing. A number on its own flattens that nuance.

A transparent agent does not return only a figure: it shows the reasoning behind the score. Which parts of the CV it relied on, which criteria are covered, which are missing, and what it could not verify. The score then becomes a starting point for a human, not a verdict.

Scores create false certainty

Numbers feel precise. A 91 looks clearly better than a 76, even when the gap comes mostly from how the CV was written — a candidate who sells themselves well, a layout that happens to contain the right words. Apparent precision is not accuracy.

In high-volume hiring the score quickly becomes a shortcut: sort, filter, move on. That shortcut is only useful if the reasoning stays visible. A number you follow without being able to explain it does not save time — it simply pushes the mistake further down the process.

Good candidates hidden by weak wording

Strong people sometimes write imperfect CVs. They use different terms, come from another industry, describe their work without the expected vocabulary. Keyword screening drops them without ever seeing them. This is exactly where a well-designed AI adds value — not by filtering faster, but by surfacing what a filter misses.

It should highlight transferable skills: "no customer success keyword, but onboarding, retention and renewals are described". A candidate from a different field may hold the very capabilities you need under a different name. The agent's job is to make that visible, not to reject it over a wording mismatch.

Bias can hide inside the score

An unexplained low score can penalise an employment gap, a non-traditional path, a qualification earned abroad, or a criterion that never needed to be there. No one decides this on purpose: bias settles in quietly when the number does not say what it is measuring.

Explaining the factors that make up the score makes those patterns visible — and correctable. If you can see that an unjustified criterion is sinking whole groups of profiles, you can remove it. Without explanation, that bias stays invisible and repeats with every campaign.

Misuse and the right to an explanation

A busy manager will happily say "send me everyone above 80". Yet a lower-scored candidate may deserve a human look, and a candidate above the line may just be a well-written CV. A threshold is not a decision.

The firm also has to be able to explain itself. To a candidate, a manager or a compliance requirement, it should be able to say which criteria were met, what was missing, and that a human reviewed the file — never "the system gave a low score". A candidate has a right to an explanation that holds up, not to a number no one in the company can justify.

What an explainable score should include

A useful score is more than a total. It shows the overall match, then separates must-have criteria from nice-to-have ones — which are met, which are missing. It surfaces transferable skills, flags information absent from the CV, and ties each conclusion to concrete evidence drawn from the document.

It also suggests targeted screening questions to close the gaps, and ends with a recommendation for human review rather than an automatic decision. In this form the score imposes nothing: it prepares the recruiter's work instead of replacing it.

Where BeLogic fits

At BeLogic, our recruitment agent Nova reads every CV, summarises the profiles, compares them against the role criteria, flags missing information, highlights transferable skills and suggests screening questions — and, above all, explains why a candidate is highlighted. The reasoning is visible, not hidden behind a figure. Nova is designed, hosted and run in the EU, GDPR-focused, like all of our agents. The final decision stays human: the agent sorts and clarifies, it does not rule. A person's opportunity should never depend on a number nobody is able to explain.