How AI Resume Screening Works, and How to Use It Fairly

Post a mid-level HR generalist role on a couple of job boards and you can easily get a few hundred applications in a week. A recruiter who reads every one properly isn't doing much else that week. AI resume screening is meant to fix that. It also creates a new problem: a badly set-up screen filters out good people quietly, consistently and at scale.

What the software actually does

There are three steps, and it's worth knowing them because each one fails in its own way.

First, parsing. Resumes arrive as PDFs, Word files, two-column designer templates and scanned images. Parsing turns each one into structured fields: job titles, employers, dates, skills, education, certifications, languages. Language models handle odd layouts much better than the old keyword parsers, but tables, columns and scans still cause errors. If someone's eight years at one employer get read as eight months, everything after that is wrong.

Second, matching. The parsed profile is compared with the job description. Modern matching is semantic, so "ran payroll for 200 staff across two entities" counts as payroll experience even though the exact phrase in the job description never appears. This is also where the job description starts to matter more than the software. If it lists ten "required" skills and only three are truly required, the screen will faithfully reward the other seven.

Third, scoring. Each candidate gets a score, ideally broken down by criterion, and the pool is ranked. Treat the ranking as a reading order, not a measurement. A 78 and a 74 are not meaningfully different.

Many teams add a few explicit rules on top. "Send to shortlist review if the score is above 70 and the candidate holds the required license." "Flag if the candidate needs visa sponsorship and this role can't offer it." Rules make the screen predictable and easy to audit. Keep them few, make sure each one is about the job, and record who agreed to it.

Every score should come with a short reason in plain language: what was met, what was missing and where in the resume the evidence was found. Without that, a recruiter can't spot a parsing error, and you can't give a candidate a straight answer if they ask why.

Where bias gets in

The model doesn't invent bias. It picks up patterns from what it's given and repeats them, quickly.

If the screen is tuned on your past hiring decisions, it learns whatever preferences those decisions had. Details like employment gaps, graduation year, home address or a particular university can stand in for age, gender, nationality or caring responsibilities. Inflated requirements, such as a degree for a role that doesn't need one or "10+ years" where four would do, shrink the pool for no good reason. Uneven parsing hurts quietly too. If non-English resumes or certain templates parse badly, those candidates are behind before matching even starts.

What helps is mostly unglamorous. Strip photos, age, marital status and similar fields out of what the scoring step can see. Read the job description with the hiring manager before screening and cut anything that isn't really needed. Take 50 resumes your recruiters have already assessed, run them through the screen and look hard at the cases where the two disagree. Where it's lawful and practical, compare pass-through rates across groups from time to time. Log scores, rules and overrides so that any decision can be reconstructed later.

Who does what

The goal is for recruiters to spend less time on applications that clearly don't fit and more on the ones that deserve a careful read. Automatic rejection should only apply to clear, objective requirements, like a license the job legally requires.

Telling candidates

A short note on the job page or application form is enough. Say that automated tools help review applications, what they look at, that a person makes the decisions, and how to ask a question or request a different process. Be specific. "We use AI responsibly" tells a candidate nothing.

The regulatory side

The rules differ a lot between countries and are still changing. Two come up often.

In the EU, the Artificial Intelligence Act (Regulation (EU) 2024/1689) lists AI systems used for recruitment or selection, including filtering applications and evaluating candidates, as high-risk in Annex III. High-risk systems carry obligations around risk management, data quality, documentation, human oversight and transparency.

In New York City, Local Law 144 restricts the use of automated employment decision tools unless the tool has had a bias audit within the previous year, the audit results are published and candidates receive the required notices.

Resumes are full of personal data, so data protection law applies as well, such as the GDPR in Europe and the Personal Data Protection Law in the UAE. If you hire in the UAE or the wider GCC, our UAE adoption guide sets out the local rules.

None of this is legal advice. Check what applies where you hire with a qualified lawyer.

Before you switch it on

1. Tighten the job description and agree the must-haves with the hiring manager.
2. Decide which fields the screen may use and which it must ignore.
3. Write down any auto-shortlist or auto-reject rule and who approved it.
4. Run it alongside human review on a past or current pool and compare the results.
5. Publish the candidate notice.
6. Every month or two, look at overrides and complaints and revisit the rules.

Screening is one stage in a longer process. Our end-to-end recruitment automation guide shows what comes before and after it, and our recruitment automation page explains how we set these stages up for HR teams. In DataPoint HR, every resume score shows its reasoning, screening rules are configured by the recruiting team and recruiters make the final call. You can see how that works on the HRMS page.