A score is a sort order, not a verdict

AI matching has become a standard ATS feature: drop a candidate against a position and get back a number — 78% match, 91% match — often with sub-scores and a sentence of reasoning. Used well, it's one of the best time-savers in recruiting, collapsing a manual triage queue into a ranked list. Used badly, it becomes an auto-reject machine that quietly launders bias, frustrates good candidates, and hands a plaintiff's lawyer a clean exhibit.

The difference is entirely in how the recruiter reads and acts on the score. This is a practical guide to doing that well. It pairs with AI resume screening best practices (which covers setting screening up) and AI, bias, and the EEOC (which covers the legal exposure) — here the focus is the daily judgment call: you're looking at a 64%, what do you do?

Read the breakdown, not the headline number

A single composite score hides more than it shows. A 70% can mean "great on everything, mediocre overall" or "perfect on skills, zero on clearance" — and those demand opposite actions. Any matching worth using gives you a per-dimension breakdown: skills, experience, location, certifications, and — in cleared work — clearance, scored separately.

Read the dimensions first:

  • A low score driven by one hard requirement (no clearance for a clearance-mandatory role; no license for a licensed role) is the model being correctly decisive. That's a knockout, and it's fine to treat it like one — the same logic as knockout screening questions.
  • A low score spread evenly across soft dimensions is the model being appropriately uncertain. That's a "human should look," not a "reject."
  • A high score with one glaring gap is a prompt to verify the gap, not to rubber-stamp the number.

The headline percentage is a sort order. The breakdown is where the actual information lives.

When a low score means "the model is missing context"

Models score what's in front of them, and resumes are lossy. A genuinely strong candidate scores low when:

  • The resume undersells. Career changers, military veterans whose experience doesn't map to civilian keywords (see reading a military resume), and people who describe outcomes instead of tools all score below their true fit.
  • The requirement is written badly. If the position lists ten "required" skills that are really nice-to-haves, every candidate scores low and the model is faithfully reporting your bad job description, not bad candidates. Fix the job description, not the candidate.
  • The signal is off-resume. A referral, a portfolio, a clearance the candidate didn't list — the model can't see what isn't in the data.

This is why the reasoning sentence matters as much as the number. A score that says "low: no AWS experience listed" is checkable in ten seconds — and often wrong because the candidate has it under a different name. A score with no explanation is just a number to distrust.

The bias and EEOC traps in acting on scores

Acting on AI scores is regulated behavior, and the exposure is real:

  • Disparate impact survives good intentions. If acting on the score produces a selection rate for a protected group below the four-fifths threshold, you have a problem regardless of whether anyone intended it — and "the AI did it" is not a defense. This is the four-fifths analysis your adverse-impact monitoring should be running continuously, and the legal frame is in AI, bias, and the EEOC.
  • Auto-reject on score is the riskiest configuration. A hard cutoff that rejects everyone below X, with no human in the loop, is exactly the pattern regulators have warned about. A score should rank and surface, and a human should decide — especially on rejections.
  • Local law may require disclosure or a bias audit. Some jurisdictions (NYC's Local Law 144 is the bellwether) require notice to candidates and independent bias audits of automated employment decision tools. Know your jurisdictions before you let a score gate anyone.
  • Records retention applies to scores too. The score and its reasoning are part of the hiring record; keep them per recruiting records retention so you can reconstruct why a decision was made.

The human review that keeps it defensible

A defensible, useful matching workflow looks like this:

  1. Use the score to rank and triage, never to silently auto-reject. Let it tell you who to look at first.
  2. Read the per-dimension breakdown and the reasoning before acting, not just the headline number.
  3. Verify the one thing the score hinges on — the missing skill, the location flag, the clearance gap — before you act on it. Half the time the data was incomplete, not the candidate.
  4. Keep a human decision-maker on every rejection. The model surfaces; the recruiter decides and can articulate why.
  5. Watch the aggregate. Even a fair-looking per-candidate process can produce a biased pattern; monitor selection rates across groups continuously.
  6. Tune the weighting deliberately. If clearance is genuinely mandatory, weight it so; if it's a nice-to-have, don't let it dominate. Garbage weighting produces confident garbage scores.

How the product is built for this

Hosting HR's AI talent matching is designed around exactly this discipline: every match comes with a per-dimension breakdown (clearance, skills, experience, location, certifications) and a one-to-two-sentence reasoning tied to evidence in the candidate's profile, so a recruiter can audit why before acting. The weighting is tunable in admin settings, scores rank rather than auto-reject, and the reports layer runs four-fifths adverse-impact analysis across your applicant flow so a biased pattern surfaces as a flag, not as a lawsuit. The model does the sorting. You — accountably, on the record — still make the call.