How AI CV screening reduces time-to-hire without hiding bias

A recruiter screening 300 CVs for one backend role spends around 23 hours before a single interview is booked. Automating that step is easy. Automating it without quietly baking in yesterday's hiring patterns is the hard part.

The real cost of slow screening

Across twelve EvalCV beta customers, the median time from application to first interview was 11 working days. Most of that was queueing: CVs waiting for someone with enough context to read them properly. Strong candidates rarely wait that long. By day eight, a third of the shortlisted applicants had already accepted another process.

Where bias hides in automated screening

Most screening tools learn from past hiring decisions. If those decisions favoured certain universities, postcodes or career shapes, the model learns to favour them too, and presents the result as an objective number.

Proxies are the problem, not the model

Names and photos are obvious. Career gaps, graduation years and home addresses are not, but they correlate just as strongly with protected characteristics. EvalCV strips them before scoring and logs that it did.

“A score nobody can explain is just a faster way to make the same mistakes.”

Scores you can explain

Instead of training on past outcomes, EvalCV scores each CV only against the requirements written in the job description. Every point in the score maps to a requirement and a quoted line of evidence from the CV.

In the recruiter portal, the score breakdown lists each requirement with the sentence in the CV that satisfied it. Recruiters can click through from a number to the words behind it, and disagree with it if they want to.

What the API returns

The same evidence is available programmatically, so teams plugging EvalCV into their ATS can show it to hiring managers too.

POST /v1/score
{ "job_id": "job_8f2c41", "cv_id": "cv_31d0a7" }

200 OK
{
  "score": 78,
  "matched": ["Spring Boot", "PostgreSQL", "5+ years backend"],
  "missing": ["Kubernetes in production"],
  "evidence": [{
    "requirement": "Spring Boot",
    "quote": "Led migration of 14 services to Spring Boot 3"
  }],
  "excluded_signals": ["name", "photo", "date_of_birth", "address"]
}

The excluded signals list

The excluded_signals field is there for auditors. It records which fields were removed before the model saw the CV, on every single request.

Measuring time-to-hire honestly

After eight weeks, median time to first interview dropped from 11 to 4 working days. More useful: the share of interviewed candidates from non-traditional backgrounds rose slightly rather than fell, which is the opposite of what we feared.

Try it on your own roles. EvalCV's free tier lets you score CVs against a live job description. Plans and limits are on evalcv.com.

What we'd do differently

We shipped scoring before we shipped the evidence view. Recruiters trusted the numbers less in those first weeks, and they were right to. Next time, the explanation ships first.