AI CV screening
Also called AI resume screening, automated shortlisting, candidate ranking, AI applicant scoring
What independent studies show about bias in AI CV screening, what POPIA, the EEA, UK GDPR and NYC law require, and where it fails small firms.
Independent audits keep finding that AI CV screeners are biased in ways you cannot see. One study found text-embedding models favoured White-associated names in 85.1% of tests. Another, scoring about 361,000 CVs, found most of the five leading language models it tested marked Black men down. A third, testing 22 models, found the first-listed CV was picked 63.5% of the time. Users report saving time, but we found no independent evidence that off-the-shelf screening improves who small firms hire. Use it to sort, never to reject.
At a glance
| Status | In pilots |
|---|---|
| What it does | Reads incoming CVs, scores or ranks them against the job description and suggests a shortlist. Ranges from pasting CVs into a chatbot to a scoring feature inside an applicant tracking system. |
| Typical cost | From US$15 per user per month, billed annually, for an applicant tracking system with AI candidate scoring (vendor list prices, checked Oct 2026). General chatbots cost less but carry the same bias risks. |
| Data it needs | CVs, a written job specification with job-related criteria, and a record of who was shortlisted, by whom and why. |
| Setup effort | Low to switch on; Medium to do properly, because you need a privacy notice, a human review step and a decision log. |
| Evidence | Moderate. Several independent audit studies, all finding bias; one enterprise study of a custom-built screener; no independent study of small-firm results. |
What is AI CV screening?
AI CV screening uses software to read job applications and decide, or suggest, who moves forward. It comes in three forms:
- An owner pastes CVs into a general chatbot and asks for a ranking.
- A scoring feature inside an applicant tracking system gives each CV a match score against the job description.
- Dedicated tools add knockout questions or short assessments.
Adoption is real but uneven. In a 2024 survey of 2,366 US HR professionals, 26% said their organisation used AI in HR, and 64% of those used it in recruiting. Only 34% of recruiting users had it screen CVs. Writing job descriptions was far more common.
How does it work?
Most tools use large language models or text-embedding models. The software turns the job description and each CV into numbers that represent meaning, then ranks CVs by how closely they match. Chatbot-style tools also write a score or a short justification.
Neither approach knows your business. They match patterns learned from public text, which is why names and even document length can move a score.
A few large employers build screeners from years of their own hiring data. At one Fortune 500 firm, a custom algorithm designed to explore less familiar applicants picked better interviewees and a more diverse group than its recruiters, while a conventional model picked far fewer Black and Hispanic applicants. Small firms do not have that data.
What it needs to work
AI screening only works as a sorting aid, and only if you can show afterwards what it did. That depends on a clear job specification, one place where applications arrive, and a record linking each score to a human decision.
- A written job specification with job-related must-haves (qualifications, licences, right to work, location) that you could defend if challenged.
- All applications arriving in one place, an applicant tracking system or a single inbox, not split across email, WhatsApp and paper.
- A log per vacancy: each applicant, the AI score, who reviewed it, the decision and the reason, kept for as long as a claim could be brought.
- Voluntary equality data asked directly of applicants, never inferred from names, stored apart from CVs, if you want to check outcomes by group.
- A privacy notice saying AI is used and how to ask for human review, plus a data protection impact assessment for UK hiring.
- To test whether it works: at least a year of hiring outcomes (probation passed, early leavers) joined back to the original scores.
Most use cases in this guide succeed or fail here, on the data underneath them, before any AI is involved.
What does the evidence show?
The independent evidence is mostly about bias, and it is consistent. A University of Washington audit of more than 500 CVs across nine occupations found text-embedding models favoured White-associated names in 85.1% of tests and female-associated names in only 11.1%. Black male names were disadvantaged in up to 100% of comparisons. A 2025 study scored about 361,000 fictitious CVs with five leading language models: women scored slightly higher and, in most models, Black men lower. For GPT-3.5 Turbo, the example the authors work through, that moved their chance of passing an 80-point cut-off by 1.4 to 1.7 percentage points.
Some bias is not demographic. Across 22 models, the first of two CVs was picked 63.5% of the time, and leading models showed a self-preference bias of 67% to 82% for CVs they had written themselves over human-written ones.
On benefits, 88% of US HR professionals using AI in recruiting say it saves time or increases efficiency, but that is self-reported. We found no independent study showing off-the-shelf screening improves who small firms hire.
AI-ranked shortlist vs human-only shortlist
| Compared on | AI-ranked shortlist | Human-only shortlist |
|---|---|---|
| Speed | Minutes for a large pile of CVs | Hours per vacancy |
| Bias profile | Hidden and repeated at scale | Varies by reviewer; easier to question |
| Explaining a rejection | Hard: scores rarely come with defensible reasons | Possible if reasons are noted at the time |
| Cost model | Per-user subscription | Owner or manager time |
| Best for | Sorting high-volume, entry-level applications before a person reads them | Small pools, senior or specialist roles |
Status and evidence
Rated pilots because independent audits consistently find demographic and order bias in the models these tools rely on, and the only independent evidence of better hiring comes from a custom-built screener at a Fortune 500 firm. Evidence is moderate: several solid studies exist, but they measure risk rather than small-firm results, while New York and the UK now regulate or audit these tools, and EU high-risk rules apply from December 2027.
Where does it go wrong?
- You paste two CVs into a chatbot and it picks the first. Swap the order and ask again; if the answer flips, the ranking is noise.
- Applicants who used the same chatbot to write their CV float to the top.
- A rejected applicant asks why and you cannot say. South African law protects applicants from unfair discrimination and puts the burden of proving fairness on the employer.
- A supplier guesses applicants' race or gender from their names, or lets you filter on them. The UK regulator found both in audits of recruitment AI providers.
- The score rejects people with no human looking. POPIA restricts decisions based solely on automated profiling; where an employer relies on the contract exception, the applicant must be able to make representations and learn enough of the logic to do so.
- A vague job specification lets the model fill the gaps with patterns from elsewhere.
Numbers it should move
Ask the demo business
Northgate Supply is a fictional wholesale distributor of catering and packaging supplies: 28 staff, R46.6m in sales over the last year. Its data is generated for this demo, to 30 September 2026.
Payroll took 12.6% of sales over the year. December's payroll was nearly double a normal month because of 13th cheques, but record sales held that month's ratio to 14.9%, about the same as February 2026, the quietest month for sales, at 14.7%. March's increases lifted monthly payroll by about 5.5%.
How this was worked out
Monthly payroll cost, including 13th cheques, divided by monthly sales.
Where the data usually lives
Where the records this use case needs are usually kept in a small business. It is not a list of recommended tools.
Frequently asked questions
How much does AI CV screening cost for a small business?
One applicant tracking system lists AI candidate scoring on plans from US$15 per user per month, billed annually (checked October 2026). A general chatbot subscription is cheaper, but it leaves no audit trail and shows the same order and name effects found in independent studies. The bigger cost is the review and record-keeping time you still need, because the software cannot make the decision for you.
Is it legal to use AI to screen CVs under POPIA and UK GDPR?
Yes, with conditions. POPIA section 71 bars decisions with legal or substantial effects based solely on automated profiling, with narrow exceptions. Under the contract exception, the person must be able to make representations and learn enough of the logic to do so. The Employment Equity Act covers applicants and puts the burden of proving fairness on the employer. In the UK, the ICO expects a data protection impact assessment before processing starts and says AI should not make recruitment decisions alone. This is not legal advice.
Do I need a developer or IT team to use AI screening?
No. Scoring is a setting in most applicant tracking systems, and a chatbot needs no setup. What you need is process: a clear job specification, one inbox for applications, a privacy notice and a record of who reviewed each shortlist and why. The UK regulator's audit lessons, such as asking suppliers for accuracy and bias test results, suit non-technical buyers.
Is AI less biased than a human recruiter?
Not reliably. The pitch is consistency, but audits show the models carry their own skews: favouring White-associated names in one study, marking down Black men in another and preferring whichever CV comes first. A machine applies the same skew to every applicant. One custom algorithm at a large firm improved diversity, but off-the-shelf tools are not built that way.
Do New York or EU AI hiring rules apply to a small business?
Only if you hire there. New York City's Local Law 144 requires an independent bias audit within a year before use, a published summary and at least 10 business days' notice to candidates. A 2025 state audit found the city flagged one compliance issue across 32 companies it reviewed, where state auditors found at least 17 potential ones. The EU AI Act lists AI used to filter applications and evaluate candidates as high-risk; after a 2026 amendment, those obligations apply from 2 December 2027.
More in people & payroll
Sources
- 1Gender, Race, and Intersectional Bias in Resume Screening via Language Model RetrievalAAAI/ACM Conference on AI, Ethics, and Society (University of Washington), 2024Study
- 2
- 3
- 4
- 5Hiring as ExplorationNational Bureau of Economic Research, 2020Study
- 62024 Talent Trends: Artificial Intelligence in HRSHRM, 2024Survey
- 7Enforcement of Local Law 144 – Automated Employment Decision Tools (Report 2024-N-6)Office of the New York State Comptroller, 2025Government
- 8AI Act, Annex III: High-risk AI systemsEuropean Commission, AI Act Service Desk, 2024Government
- 9AI tools in recruitment: Audit outcomes reportInformation Commissioner's Office (UK), 2024Government
- 10Protection of Personal Information Act 4 of 2013South African Government, 2013Government
- 11Employment Equity Act 55 of 1998South African Government, 1998Government
- 12Pricing and PlansManatal, 2026Vendor
- 13The Digital Omnibus on AI enters into force todayLewis Silkin, 2026Commentary
This entry is educational. It is not legal, financial or professional advice, and it does not recommend any vendor. Figures attributed to vendors are their own claims.
- 1Gender, Race, and Intersectional Bias in Resume Screening via Language Model RetrievalAAAI/ACM Conference on AI, Ethics, and Society (University of Washington), 2024Study
- 2
- 3
- 4
- 5Hiring as ExplorationNational Bureau of Economic Research, 2020Study
- 62024 Talent Trends: Artificial Intelligence in HRSHRM, 2024Survey
- 7Enforcement of Local Law 144 – Automated Employment Decision Tools (Report 2024-N-6)Office of the New York State Comptroller, 2025Government
- 8AI Act, Annex III: High-risk AI systemsEuropean Commission, AI Act Service Desk, 2024Government
- 9AI tools in recruitment: Audit outcomes reportInformation Commissioner's Office (UK), 2024Government
- 10Protection of Personal Information Act 4 of 2013South African Government, 2013Government
- 11Employment Equity Act 55 of 1998South African Government, 1998Government
- 12Pricing and PlansManatal, 2026Vendor
- 13The Digital Omnibus on AI enters into force todayLewis Silkin, 2026Commentary