Resume Screening in 2026: How Manual, ATS and AI Screening Work – and How to Do It Fairly

TL;DR

• Resume screening is the step where applications are checked against a role’s requirements to decide who moves forward – by a recruiter, by rules in an applicant tracking system (ATS), or by AI that reads and scores CVs.
• Most “the ATS rejected me” stories are rules people set up, not a hidden robot. The “75% of resumes are rejected by ATS” figure traces back to a 2012 vendor sales claim with no published method.
• A CV predicts less than most teams assume. Years of experience correlates with job performance at about .07; structured interviews reach .42 (Sackett et al., 2022).
• Bias shows up in both human and AI screening, and anonymising CVs has mixed results.
• The fix: use the CV as a gate for genuine minimums, and base the ranking decision on evidence every candidate produces the same way. Ask any AI vendor what its score is built from and whether it has a role-level bias audit.

Resume screening used to be a volume problem. It’s now a volume, trust and legal problem at once.

Applications per role have roughly doubled in a few years, recruiting teams are smaller, and many candidates use AI to write their applications. Your ATS vendor is probably selling AI screening features, and regulators are watching automated hiring decisions closely.

So the real question isn’t how to read CVs faster. It’s whether to screen them differently, automate the job, or stop relying on them for decisions they can’t support. This guide covers how each method works, what a CV can predict, where bias gets in, the rules that now apply, and a five-step process you can run this quarter.

A disclosure: Sapia.ai builds screening technology. We’ve assessed 10M+ candidates (vendor-reported) and publish an independent bias audit of our own AI, so we’ve tried to hold this guide to the standard it asks you to apply.

What is resume screening?

Resume screening (called CV screening in the UK, Australia and Europe) is the process of reviewing job applications to decide which candidates meet a role’s requirements and should move to the next stage. It can be done manually by a recruiter, automatically by rules in an applicant tracking system, or by AI tools that read and score CVs against the job.

It decides who gets seen by a person, which makes it the highest-leverage step in most hiring processes – and the least examined. It also helps to separate screening (checking minimum requirements – a gate) from ranking (ordering candidates by likely fit). A CV is reasonable evidence that someone holds a licence; it’s much weaker evidence of who will perform best. Most of the risk lives in ranking.

How resume screening works in 2026 – manual, ATS and AI

There are three common ways to screen resumes, plus a fourth approach that replaces the CV as the main input.

Manual screening

A recruiter skims each CV for titles, employers, dates and keywords. The best-known figure is about 7.4 seconds on the first pass, from an eye-tracking study by the job board Ladders (HR Dive, 2018) – a vendor study with no published sample size. People read context well but inconsistently, and, as the bias section shows, the evidence of discrimination is strongest for human screeners.

What ATS resume screening actually does (and the 75% myth)

By default, an ATS parses each CV into fields, stores it and lets recruiters search it. It doesn’t quietly reject people on its own.

Automatic rejections come from rules your team sets up. In Greenhouse, auto-reject fires only on answers to yes/no or multiple-choice application questions, configured for each job by an admin (Greenhouse); Workday offers configurable steps that advance or decline candidates. Because auto-rejected applicants don’t trigger notifications, rejections feel invisible – hence the “ATS black hole”.

That feeling keeps the most repeated statistic here alive: that 75% of resumes are rejected by ATS before a human sees them. Later analysts trace it to a 2012 sales claim by Preptel, a resume-optimisation company, which published no method.

The real filtering problem is better documented, and it’s human. In Harvard Business School and Accenture’s Hidden Workers study, 88% of employers agreed that qualified high-skill candidates are screened out because they don’t exactly match the job description, and 48% screen out CVs with gaps of more than six months (HBS/Accenture, 2021). The damage comes from rules people choose – so you can fix it.

AI resume screening – three generations

AI resume screening uses machine learning or large language models (LLMs) to read CVs and score or sort them against a job. It has arrived in three waves, and most tools now mix them.

1. Semantic matching. CVs and job descriptions are converted into numerical representations of meaning, so “customer success” can match “account management”. Some vendors also train on past hiring outcomes – Eightfold, for example, describes models “trained on tens of millions of historical candidate-position pairs with known outcomes”, scoring matches from 0 to 5 (Eightfold, vendor-reported). Training on past outcomes is powerful, and it’s the design that went wrong at Amazon (see below).

2. Requirement grading. The tool grades how closely a CV matches the skills, education and experience in the job description. Workday’s HiredScore, acquired in 2024, assigns grades from A to D (HiredScore, vendor-reported).

3. LLM criteria checks and agents. The recruiter sets the criteria; an LLM marks each one as met or not and cites evidence from the CV, usually without rejecting anyone automatically. Examples include Ashby’s AI-Assisted Application Review (September 2024), with personal details redacted and a FairNow audit (Ashby), and Greenhouse’s Talent Matching (June 2025), which sorts applicants into tiers, “does not auto-reject or auto-advance” and is audited by Warden AI (Greenhouse) – both vendor-reported. LinkedIn’s Hiring Assistant, Indeed’s Talent Scout and SmartRecruiters’ Winston (SAP-owned since September 2025) push further towards agents that source and screen.

Two things are worth noticing. First, look at what the score is built from. Every approach above still scores the CV – text the candidate wrote, or increasingly an AI wrote for them. The tools differ on whether the criteria come from you or from past hires, and on whether a person makes the call.

Second, “assistive” is a legal position as well as a feature. Describing a tool as helping a human decide keeps it outside some automated-decision rules – a position two lawsuits covered below are testing.

Comparing specific tools? See our guide to resume screening software.

What a resume can – and can’t – tell you

Industrial-organisational (I/O) psychologists have spent decades studying which selection methods predict job performance. The most current large meta-analysis, by Sackett, Zhang, Berry and Lievens, revised many of the older estimates most articles still quote (Journal of Applied Psychology, 2022).

Operational validity – the corrected correlation with job performance. Independent research.

Experience fares badly in other studies too. Across 81 samples, Van Iddekinge, Arnold, Frieder and Roth found that pre-hire experience correlates with job performance at about .06 and with staff turnover at .00. Even experience in closely related jobs reached only .07, and the authors caution against selecting on such measures (Personnel Psychology, 2019). Grades predict early-career performance modestly, and the link fades within a few years of graduation.

The takeaway: most of what a CV shows – years, titles, degrees, employer names – predicts little about who will do the job well. The strongest predictors, such as structured interviews, job knowledge tests and work samples, aren’t on a CV at all.

That doesn’t make the CV useless. It makes it a gate for genuine minimum requirements, not a ranking of who will perform.

Bias in resume screening – human and AI

Neither side of this debate starts from neutral. Human screeners show well-documented bias, and AI screeners can reproduce it, scale it or introduce new kinds.

The evidence on human screening

• Names change callbacks. In a classic US field experiment, identical CVs with white-sounding names received about 50% more callbacks than those with Black-sounding names (Bertrand & Mullainathan, 2004).
• It’s concentrated, and structure helps. A study of 83,000+ fictitious applications to 108 large US employers found distinctively Black names cut contact rates by about 2.1 percentage points, with a minority of firms accounting for much of the gap. Gaps were smaller at firms with more centralised recruiting (Kline, Rose & Walters, 2022).
• It’s not just a US problem. In the UK, ethnic-minority applicants needed about 60% more applications to get a positive response than white British applicants, with no improvement since the late 1960s (Nuffield College CSI, 2019).

The evidence on AI resume screening

• Amazon’s scrapped tool. A model trained on ten years of mostly male CVs learned to penalise the word “women’s” and was abandoned (Reuters via CNBC, 2018).
• Name bias. A language-model retrieval system favoured white-associated names 85.1% of the time and Black-associated names 8.6% of the time across 500+ CVs (Wilson & Caliskan, 2024). Bloomberg reported similar effects in GPT models (Bloomberg, 2024), though a reanalysis of its data found no significant race effect (interviewing.io).
• Disability bias. GPT-4 ranked CVs lower when they mentioned disability-related awards (Glazko et al., 2024).
• Bias can run either way. A test of 22 LLMs found they favoured female-named candidates and whoever was listed first (Rozado, 2025).
• AI favours AI. Candidates who used the same model as an LLM screener were 23–60% more likely to be shortlisted (Xu, Li & Jiang, 2025).

The lesson: AI screening bias depends on the model, the task and the context. Test it in your roles – don’t assume it in either direction.

Proxies – how bias gets in without protected data

Removing race, gender and age fields doesn’t remove bias, because postcode, school, employment gaps and names can stand in for them. The UK Information Commissioner’s Office found AI recruitment tools estimating candidates’ gender and ethnicity from their names (ICO, 2024). In the US, Harper v. Sirius XM (2025) alleges an AI screen used schools, employment history and ZIP codes as proxies for race – allegations not yet ruled on (Epstein Becker Green).

Does blind CV screening fix it?

The evidence is mixed. In an Australian Public Service trial, de-identifying CVs removed an existing tilt towards women and minority candidates (BETA, 2017). In France, anonymous CVs widened the gap for minority applicants among firms that opted in (J-PAL). In Sweden, they raised interview rates for women (Åslund & Skans, 2012). Redaction helps most where the starting point is biased, and does nothing about proxies or later stages. More in our guide to blind screening.

Why a pooled audit isn’t enough

Stanford, Chapman and Northeastern researchers analysed more than 4 million applications screened by pymetrics, a game-based assessment, between 2018 and 2022. Pooled figures looked acceptable; at the level of individual positions, they found racial disparities the pooled numbers had hidden (Stanford Report, 2026). The lesson applies to every screening tool: ask vendors for adverse-impact results by role, not an average across customers. That includes us.

The 2026 problem – AI-written applications and screening integrity

• Volume. LinkedIn reported about 11,000 applications per minute in mid-2025, up 45% in a year (eWeek). Greenhouse data puts the average at about 244 applications per job in 2025, against about 115 in 2022 (Greenhouse, 2026). Both vendor-reported.
• AI in applications. 39% of candidates told Gartner they’d used AI in their applications (Gartner, 2025).
• Hidden text. 41% of US job seekers said they’d hidden instructions or keywords in a CV to game AI screeners (Greenhouse survey, 2025). When researchers scanned about 197,000 real CVs, they found it in about 1%, mostly invisible skills or invented experience (Zhang et al., 2026, preprint).
• Fraud. Gartner forecasts that one in four candidate profiles will be fake by 2028, and the US Department of Justice has reported North Korean IT-worker schemes at more than 136 US companies (DOJ, 2025).

When a CV may be written, tailored and mass-submitted by AI, scoring CV text faster mostly rewards whoever optimised the CV best. The more durable signal is what candidates produce in response to the same questions – backed by integrity controls whose limits are published.

The rules that now apply to automated resume screening

As of September 2026. This is general information, not legal advice – check with counsel for your jurisdictions.

United States

• New York City Local Law 144 covers automated tools that substantially assist or replace hiring decisions, including resume screeners: an annual independent bias audit, a published summary and candidate notice. A December 2025 state audit found enforcement weak (NY State Comptroller), but the obligations stand.
• Illinois HB 3773, in force since January 2026, bans AI use with a discriminatory effect, names ZIP codes as a prohibited proxy and requires notice (Seyfarth).
• California’s automated-decision regulations have applied since 1 October 2025; vendors can be liable as the employer’s agent, and anti-bias testing counts towards a defence (Paul Hastings). Separate privacy-law rules apply from January 2027.
• Colorado’s SB 26-189 applies from January 2027: notice before use, an explanation after an adverse decision, and a route to human review (Crowell).
• Federal level. Disparate-impact enforcement has been deprioritised since Executive Order 14281 (April 2025). In June 2026, a Department of Justice opinion called the EEOC’s disparate-impact guidelines unconstitutional (DOJ), and the EEOC plans to rescind the Uniform Guidelines, home of the “four-fifths rule” (Reginfo.gov). Title VII hasn’t changed; the live exposure is state law and private lawsuits.
• Litigation. In Mobley v. Workday, a court conditionally certified a nationwide age-discrimination collective in 2025 that includes HiredScore features; the case is in discovery (Clearinghouse). Kistler v. Eightfold AI (January 2026) argues that AI match scores are “consumer reports” under credit-reporting law (Fisher Phillips).

European Union and United Kingdom

• Under the EU AI Act, tools that filter applications and evaluate candidates are high-risk. Those obligations were delayed, not dropped – they apply from 2 December 2027 (Hunton).
• The UK’s Data (Use and Access) Act 2025 allows solely automated decisions on more grounds since February 2026, but candidates must be told and be able to challenge the decision and get human intervention (legislation.gov.uk). The ICO found inconsistent human involvement and weak bias monitoring among employers using these tools (ICO, 2026). Under the Equality Act 2010, proxy and gap filters can amount to indirect discrimination whoever – or whatever – applies them.

Five questions to ask of any screening method – human, ATS or AI

Whether you’re reviewing your own process or a vendor’s product, these five questions cut through most marketing:

1. What does the candidate actually produce? A CV (possibly AI-written), answers to knockout questions, structured interview answers, a work sample, code – or nothing beyond a profile?
2. What is the score built from? CV text and keywords, similarity to past hires, criteria you set, or the candidate’s answers only, with the CV kept out of the score?
3. What validity evidence is published, and of which type? Content, construct or criterion – that is, does it predict performance or retention? “None published” is an answer.
4. Is there an independent, published bias audit of this specific product, analysed by role rather than pooled? Look for the auditor’s name, the date and the scope. An audit of a different product, a pooled audit or a self-assessment isn’t the same thing.
5. What are the integrity controls, and are their limits published? Defences against hidden text, AI-written content and fraud – with detection and false-positive figures, not adjectives.

Then ask the practical follow-ups, drawn from the UK government’s Responsible AI in Recruitment guide and ICO guidance: Does the tool auto-reject anyone, and where exactly does a person decide? How does it handle proxies such as postcode and career gaps? Can candidates get an explanation or a human review? Does it add third-party data to profiles?

How to do resume screening well – a 5-step process

1. Define criteria from the job before anyone opens a CV. Evaluators who hadn’t committed to criteria shifted them to favour the candidate they preferred; committing in advance removed the effect (Uhlmann & Cohen, 2005). Start from a job analysis – tools like Sapia’s Job Analysis Studio turn a job description into a competency framework.
2. Knock out only on genuine, defensible minimums – licence, right to work, shift availability with a route to request an adjustment. Drop years-of-experience cut-offs and career-gap filters.
3. Score against a rubric, not a gut read. Look for evidence someone has done the thing, not pedigree. If you use AI, use criteria you set, require cited evidence and keep a person in the decision.
4. Base the ranking decision on evidence candidates produce – the same structured interview questions or work sample for everyone, especially where a CV says little. See predictive assessment and chat-based interviews.
5. Monitor adverse impact by role, and pilot before you scale. Track selection-rate ratios by role family with significance tests – the four-fifths comparison remains the practitioner benchmark, and NYC requires impact ratios. Track completion and quality of hire too, and re-audit after model updates.

When resume screening belongs in your process – and when it doesn’t

It belongs where a credential or specific experience is a genuine requirement: licensed and regulated roles, senior specialist hires, and jobs with hard technical prerequisites. Use it as a gate, then follow it with a structured evaluation.

It adds little for frontline, high-volume, graduate and early-career roles. Most applicants have thin or similar CVs, and the qualities that matter – communication, reliability, customer focus, willingness to learn – aren’t on the page. Here, screening CVs, by hand or by AI, mostly sorts people by how well they write a CV.

The middle ground is experienced and professional hiring. Experience matters, but counting years and titles doesn’t capture it. Ask about experience directly and score the answers consistently instead.

Where Sapia fits (and where it doesn’t)

Sapia takes a different route from CV screening. Every applicant gets the same untimed, mobile, text-first structured chat interview at the point of application, scored against a competency framework, with results written back to your ATS (an async video interview is available as a second stage). Here’s how it answers the five questions:

• Candidate produces: written answers to structured, job-related interview questions, on any phone, with no time limit.
• Score built from: the candidate’s answers only. The score doesn’t use their CV, school or employers.
• Published validity: Sapia publishes research on its science page. SHL publishes deeper validity documentation for its assessments, and we don’t claim parity with the meta-analytic figures earlier in this guide.
• Independent audit: BLDS audited Sapia’s scoring in late 2022, running 72 tests (23 standardised mean difference and 49 adverse impact ratio tests), and found no evidence of practically significant disparate impact. The audit covered US and Canada models only (FAIR framework). Ask us – and every vendor – for current, role-level results.
• Integrity controls: built-in detection of AI-generated answers. Sapia reports 98% detection with a 1% false-positive rate (vendor-reported). Published research shows AI-text detectors can over-flag non-native English writers (Liang et al., 2023), so ask us for the false-positive rate for that group.

For experienced roles, Experience Scoring [confirm product name with Sapia] in Job Analysis Studio adds up to two structured questions about the work someone has owned, scored on depth, outcomes and relevance – without using the CV as the assessment. It’s being validated in customer pilots now; it isn’t a hard-skills test or a lie detector, and people still make the decisions. The Talent Intelligence Assistant (TIA) then lets recruiters ask plain-language questions across everyone interviewed, with answers traceable to what candidates said.

In practice (customer-reported, via Sapia case studies): Starbucks saved 1,900 hours a month in screening time, and LNER maintained 30% ethnic-minority representation through its hiring stages.

Where Sapia doesn’t fit: it isn’t an ATS, so you’ll still need one for parsing, knockout questions and right-to-work checks. If you want to keep CV screening but make it more consistent, the LLM criteria checks inside ATSs such as Greenhouse and Ashby are the closer fit. For hard technical skills, use a coding assessment such as HackerRank or Codility; for the deepest published psychometric validity, SHL; for sourcing people who haven’t applied, SeekOut or hireEZ.

Either way, the principle holds: screen CVs for the minimums a role truly requires, and decide on evidence every candidate produces the same way. To see what that looks like for your roles, book a demo.

Conclusion

Resume screening is a gate, not a crystal ball. Most “invisible” rejections come from ATS rules people configured, and those can be fixed. AI can make screening faster and more consistent, but it still scores the CV and has to be tested for bias role by role. And when many applications are written by AI, the most durable signal is what candidates say and do in response to the same questions.

Use the five questions to review your process, run the five steps on one role family, and keep the CV for what it does well: confirming the minimums. If a CV-blind structured interview with a published audit is the gap in your process, talk to us.

FAQs

About Author

Laura Belfield

Head of Marketing