AI resume screening: how it works, where it fails, and how to run it safely

A practical guide to automated resume screening for recruiting teams that get hundreds of applicants per role.

The short answer

AI resume screening reads every application, compares it with the job's requirements, and ranks or scores candidates so recruiters start with the strongest matches. Done well, it saves the hours spent on a first pass through 300 resumes and applies the same criteria to applicant 1 and applicant 300. Done badly, it repeats the shortcuts that made manual screening unfair, at scale. The safest setup scores against written, job-related criteria, keeps a human reviewing both the shortlist and a sample of rejections, and checks pass-through rates by group. Several laws, including NYC Local Law 144 and the EU AI Act, treat automated resume screening as a regulated use.

How AI resume screening works

Most tools follow the same four steps, even if they describe them differently:

Parse. The tool pulls structured data from each resume: titles, employers, dates, skills, education, certifications. Parsing quality matters more than most buyers think. Two-column layouts, tables and image-based PDFs can lose information here.
Understand the job. It reads the job description, or a list of criteria you give it, and works out what "good" looks like: must-have skills, years of experience, level, location rules.
Match and score. Older systems count keywords. Newer ones use language models to judge whether experience is relevant even when the words differ ("built payment APIs" matching "backend fintech experience"). The output is a score, a rank or a tier.
Surface. The top group goes to a recruiter or straight to the next step, often an assessment or a first-round interview.

The difference between tools usually comes down to step 3: whether the score is explained (which criteria matched, which didn't) or just a number.

Where automated resume screening goes wrong

Most failures trace back to the criteria, not the model.

Proxy criteria. "Top-tier university," "no employment gaps" or a home zip code can stand in for traits you aren't allowed to select on. Illinois made this explicit: since January 1, 2026, its Human Rights Act makes it a civil rights violation for an employer to use AI that has the effect of discriminating, or to use zip codes as a proxy for protected classes.
Copying past hires. A model tuned to resemble the people you hired before will rank people like them higher. If your past hiring had a skew, the screen inherits it.
Keyword gaming and keyword misses. Candidates paste the job description into white text, while strong candidates who describe the same work in other words fall below the line.
Over-literal must-haves. "5+ years of React" rejects the engineer with 4 years of React and 6 years of Vue. Ask whether each must-have really is one.
Nobody looks below the line. If rejected resumes are never reviewed, nobody learns that the screen is wrong.

Automated resume screening and the law

This is not legal advice, and the summary below covers only the laws we checked at their primary sources.

New York City. The city's DCWP says "employment decision" includes screening at any point in hiring. A resume screen that scores or ranks applicants to a specific job, and that you rely on to decide who advances, can be an automated employment decision tool. That triggers the annual independent bias audit, a public summary and candidate notice 10 business days before use. The FAQ also says searching a resume database or doing outreach, before someone applies for a specific job, is not covered. See our Local Law 144 guide.
European Union. The AI Act's Annex III, point 4(a) lists AI used "to analyse and filter job applications" as high-risk. After the 2026 amendment, the high-risk duties for these systems apply from December 2, 2027. Our EU AI Act hiring guide covers what employers will owe.
Illinois. Beyond the zip code rule above, employers must notify employees when they use AI for recruitment, hiring and other employment decisions covered by the Human Rights Act amendment.
Colorado. SB26-189, signed May 14, 2026, repealed and reenacted the state's 2024 AI Act. It treats employment as a consequential decision and requires deployers of covered automated decision-making technology to give notice and, after an adverse outcome, a plain-language description of the technology's role within 30 days, plus a path to human review.

A safe setup in seven steps

Write criteria before you write prompts. List 4 to 6 job-related requirements, marked must-have or nice-to-have. Strike anything you can't defend as needed for the work.
Remove proxies. Drop school names, gap penalties, addresses and photos from what the tool scores on.
Calibrate on real resumes. Have two recruiters rank 30 past applicants blind, then compare with the tool. Where they disagree, read why.
Set a review band, not a hard cut. Send the top band forward, but have a human read the band just below the cutoff.
Sample rejections weekly. Read 20 rejected resumes per role. If you would have advanced two or more of them, revisit the criteria.
Check pass-through rates by group. Where you collect voluntary demographic data, compare advance rates across groups the way a bias audit would. Our AI bias audit guide shows the math.
Tell candidates. Say that AI helps screen applications, what it looks at, and how to request an accommodation or a human review.

What to expect at real volumes

Here's a simple example. A mid-level backend role gets 400 applicants. A recruiter skimming at 90 seconds each spends about 10 hours on the first pass. With AI resume screening, the recruiter reviews the top 40 ranked resumes, the next 20 in the review band, and a weekly sample of 20 rejections: 80 resumes, or about 2 hours. The other 8 hours go into calls with the shortlist.

But a resume only shows what someone wrote about themselves. The bigger gain usually comes from pairing the screen with a structured first-round interview, so the ranking decides who gets interviewed, not who gets hired. That is how we built NTRVSTA: AI resume scoring ranks every resume against your job description and surfaces the top candidates, then an adaptive AI interview checks the skills the resume claims.

Which tools include resume screening

Resume screening shows up in three kinds of products, based on what each company publishes:

Type of product	Examples that list resume screening	Good fit when
ATS with built-in AI review	Manatal; Spark Hire (AI resume review on its Recruit ATS plans)	You want screening inside the system of record
AI interviewer plus resume screening	Alex, Eightfold, HeyMilo, Interviewer.AI, Metaview, Ribbon, Tenzo	You want ranked applicants to flow straight into an AI first round
Sourcing tools	Juicebox (search and outreach rather than applicant screening)	You need to find people who haven't applied

For a ranked comparison, see the best AI resume screening tools and our broader list of candidate screening software.

Questions

Is AI resume screening accurate?

It is consistent, which is not the same thing. It applies the same criteria to every applicant, so its accuracy depends on how good and how job-related those criteria are. Calibrate it against experienced recruiters and keep reviewing a sample of rejections.

Can candidates beat an AI resume screener?

Some try, for example by pasting the job description into hidden text. Tools that judge described experience rather than counting keywords are harder to game, and a structured interview after the screen catches claims that don't hold up.

Do we need a bias audit for resume screening in New York City?

If the tool scores or ranks applicants to a specific job and you rely on that output to decide who advances, it can be an automated employment decision tool under Local Law 144, which requires an annual independent bias audit and candidate notice. Check your setup with counsel.

Should AI make the final rejection?

We don't recommend it. Use the screen to order the pile and decide who gets a closer look, keep a human accountable for each decision, and review the band just below the cutoff by hand.

Want to see a ranked shortlist flow into a live, adaptive interview? Book a 20-minute demo, or try the candidate side with a sample interview.

Sources

NTRVSTA