TL;DR: CV screening software does four jobs that vendors blur: parse, score, match, and a human decision. The right tool passes three tests: you can run it on your own CVs, its score explains itself, and rejection stays with a person you choose. This page includes a 50-CV parse test, a buyer's scorecard, and the UK, EU and India rules that bear on automated screening.
General information, not legal advice. Have counsel confirm how this applies to you.
Updated 28 October 2026. Research as of 15 October 2026. Refreshed quarterly and on legal triggers: the ICO's final automated-decision guidance, EU AI Act changes, India DPDP milestones (13 Nov 2026) and the Mobley class-certification hearing on 9 Mar 2027.
Changelog: 28 October 2026 — first publication.
Key takeaways
- CV screening software is four separate stages (parse, score, match, decide), and a vendor strong at one is not necessarily strong at the others.
- The 2021 Harvard Business School and Accenture "Hidden Workers" study found that ATS filters can screen out high-skilled candidates who don't match the search criteria (Harvard Gazette). A rigid filter can reject people a recruiter would have shortlisted.
- A 2024 University of Washington study found three LLMs favoured white-associated names 85% of the time when ranking resumes. It tested raw LLM ranking, not any product, and it is a reason to require explainable criteria and bias monitoring.
- In the UK, the ICO's 2024 audit of AI recruitment tools made almost 300 recommendations (ICO). In the EU, CV filtering is high-risk under the AI Act (Annex III, point 4), and Regulation (EU) 2026/1744 moved the application date for those high-risk obligations to 2 December 2027. EU AI Act obligations for recruiters sets out what that means for a hiring team.
- A 50-CV parse test across five formats and three languages tells you more than a vendor accuracy figure. The sheet is on this page.
Vendors use "CV screening" for four different jobs, and each stage fails in its own way: the Hidden Workers finding above is about filters that exclude people who could do the work. This page defines the four stages, gives you a test to run on your own files, and sets out what the law in the UK, the EU, the US and India says about automated screening.
This page covers the CV stage; top AI interview tools compared covers the interview stage.
How we evaluated. Criteria: the ten-point scorecard in the fourth section. Sources: vendor and ranking pages, regulator pages and court records, read on 15 October 2026 and linked where used. Labels: Published means the vendor's own documentation states it; Claimed means a marketing page or third-party ranking states it without detail; Not stated means we found no statement. We tested ZeaHire directly in a QA audit on 15 October 2026. We have not tested any other tool's parser or scoring, so this page makes no accuracy comparison between products. We excluded Jobscan, which one ranking page lists but which is a candidate-side tool.
What is CV screening software?
Definition. CV screening software extracts information from CVs, scores or ranks candidates against job criteria, and presents a shortlist for review. It works in four stages (parse, score, match, decide), and the decision stays with a person.
- Parse: extract structured fields (titles, dates, skills, education) from the CV file.
- Score: rate one candidate against the job's criteria.
- Match: rank an existing pool of candidates against a job.
- Decide: a person accepts or rejects.
| Stage | Takes in | Puts out | Who answers for it, in the model this page uses |
|---|---|---|---|
| Parse | A CV file | Structured fields | The parser's vendor for extraction; the buyer for testing it on their own formats |
| Score | Parsed CV and job criteria | A score, ideally with reasons | The employer that sets the criteria; the vendor for how they are applied |
| Match | A candidate pool and a job | A ranked list | The employer that chooses the job's criteria; the vendor for the ranking method |
| Decide | The score or rank, and the CV | Approve or reject | The employer's reviewer |
Parsing is the first stage. Vendors that specialise in it include Textkernel, Daxtra and RChilli, as Sapia's ranking page names them. We name them as cues only and quote no accuracy figures. Parsing is also distinct from an applicant tracking system; the FAQ on Skima's page separates the two.
This page says "CV" in its headings. US readers searching for resume screening software will find the same four stages here.
Resume parsing: how it works and how to test it
A parser reads a CV file and extracts job titles, dates, skills and education into structured fields. Errors at this stage cascade: a misread date changes the years of experience, which changes the score, which changes the rank. Everything downstream inherits the parse.
How accurate is CV parsing?
On your formats, not theirs. Accuracy percentages quoted on the vendor and ranking pages we read on 15 October 2026 came without a method or a source, so this page repeats none of them. The only figure that predicts your results is one you measure on your own CVs.
The 50-CV parse test
Take 50 anonymised CVs that represent your real intake, across five formats and three languages. Cover these formats:
- multi-column layouts;
- scanned documents;
- photo CVs;
- Indian-style CVs with date-of-birth and marital-status fields;
- non-English CVs, including names in Arabic or Hindi.
Run the batch through each tool. For every CV, check four fields: job title, dates, skills and education. A field-level miss is any of the four extracted wrongly or not at all. Record misses per field and per format, out of 50.
| CV ID | Format | Language | Job title | Dates | Skills | Education | Notes |
|---|---|---|---|---|---|---|---|
| 001 | Multi-column | English | hit / miss | hit / miss | hit / miss | hit / miss | |
| 002 | Scanned | Hindi | hit / miss | hit / miss | hit / miss | hit / miss |
No ranking page we read gives a reproducible parse test.
Parser, screener or ATS: which do you need?
- If you need clean structured fields from CVs for another system, you need a parser.
- If you need a scored or ranked shortlist for a job, you need a screener, and it contains a parser.
- If you need to track candidates through stages, you need an ATS, and some tools combine two or all three.
What we know about ZeaHire's parsing
In our own testing we saw resumes with a Failed parse status, and we have not established which languages ZeaHire's parser supports. We do not claim a parse accuracy figure. The test sheet applies to ZeaHire exactly as it applies to every other tool on this page.
How does AI CV screening work, and where does it go wrong?
AI screening scores a parsed CV against criteria and ranks candidates by the result. It goes wrong in two documented ways: filters that exclude people who could do the job, and models that carry bias into the ranking.
Exclusion by filter. The Harvard Business School and Accenture "Hidden Workers" study (2021) surveyed more than 2,250 executives and more than 8,000 workers across the US, Germany and the UK. It found that ATS filters can vet out high-skilled candidates because they don't match the search criteria, and it estimates around 27 million hidden workers in the US (Harvard Gazette, 2021). The finding is about filter design, not about how many candidates were rejected.
Bias in the model. University of Washington researchers, presenting at AAAI/ACM AIES on 22 October 2024, tested three LLMs ranking resumes for more than 500 real job listings: over 3 million comparisons using 120 names. The models favoured white-associated names 85% of the time and female-associated names 11% of the time, and never preferred Black male names over white male names. The study tested raw LLM ranking. It did not test any commercial product, and it says nothing about any vendor named on this page.
What to do with this evidence:
- Use structured criteria. Set the criteria before anyone reads a CV, and score against them.
- Ignore fields that do not bear on the job: name, photo, date of birth, gender, address and graduation year.
- Set a bias-monitoring cadence. The ICO's 2026 engagement with employers asked for monthly monitoring (see the legal section).
Screening is only the first filter. What happens after the CV stage decides whether the shortlist holds up, which is the subject of evidence-based AI interviews.
What should you evaluate? A buyer's scorecard
Ten criteria, in the order we would test them.
| # | Criterion | What to check | Why it is on the list |
|---|---|---|---|
| 1 | Parse accuracy on your formats | Run the 50-CV test | Parse errors feed every later stage |
| 2 | Explainable score | Matched and missing skills, with evidence from the CV | A score a reviewer cannot read cannot be checked |
| 3 | Fields ignored | Name, photo, DOB, gender, address, graduation year | The UW study above; the ICO audit found tools inferring gender and ethnicity from names |
| 4 | Configurable weights | Can you set weights per job | Criteria come before reading |
| 5 | Human gate before any rejection | Who can reject, and whether automation can be switched on | The ICO expects |
| 6 | Bias-monitoring cadence | How often, and by whom | The ICO asks employers for monthly bias monitoring |
| 7 | Data retention and deletion | Retention periods; deletion on request | The ICO audit found over-collection and indefinite retention |
| 8 | Candidate notice | What candidates are told | The ICO asks for transparency to jobseekers; NYC requires notice (legal section) |
| 9 | ATS integration | Which systems, and which direction | Determines where the shortlist lands |
| 10 | Price | Per resume or per month, published or not | Prices are mostly absent from the ranking pages we read (Sapia, Leelu, Gitnux, 15 Oct 2026) |
Where a human decides
The map below records, for each tool, at which step (parse, score, rank, reject) a human is required. For ZeaHire we used the product itself, so its row says what we observed.
| Tool | Parse | Score | Rank | Reject | Source |
|---|---|---|---|---|---|
| ZeaHire | No human step observed; Failed files are listed in a Resume Processing queue | Automatic; the recommendation is labelled "based on resume" and advisory | Recommended tab ranks the pool deterministically; AI Fit is shown separately | In ZeaHire, candidates wait in Pending Review until a person approves or rejects them. | Observed in our audit, 15 Oct 2026 |
| Every other tool on this page | Ask the vendor at which step a human is required | Ask | Ask | Ask | Vendor documentation |
A note on vendor-reported results
Some vendors publish customer outcomes. Sapia, for example, publishes customer figures on hiring time and turnover, with no sample sizes or methods disclosed on the page we read. We do not repeat them here. Any such figure should be read as vendor-reported.
Which CV screening tools are worth shortlisting?
Ranking pages name dozens of tools. We group them by what they do. Prices are mostly absent from the ranking pages we read.
| Group | Tools named on ranking pages |
|---|---|
| Parser specialists | Textkernel, Daxtra, RChilli |
| AI screening and matching | Eightfold, Sapia.ai, HireVue, Paradox, CVViZ / TurboHire, hireEZ / SeekOut |
| ATS-embedded | Workable, Greenhouse, Recruitee, Lever, Breezy, Ashby, Teamtailor, Manatal, Zoho Recruit |
We have not verified any of these vendors' prices, explainability or bias evidence; read each vendor's own pricing and documentation pages. Ask every AI screening vendor, Sapia.ai included, whether it has published an independent bias audit and whether it holds an ISO 42001 certificate, and ask for the link to each.
For ZeaHire, here is how the same columns read:
| Tool | Price | Score explainability | Human-gate step |
|---|---|---|---|
| ZeaHire | Custom pilot pricing: talk to us | Skills, Experience and Education sub-scores; matched and missing skills; strengths; AI summary | Pending Review: a person approves or rejects |
On HireVue's screening, the page to read is alternatives to HireVue's screening. On ATS-embedded review caps, tools that compete with Spark Hire covers the field. According to Spark Hire's pricing page, its tiers list 6,000, 24,000 and 36,000 AI resume reviews a year.
What ZeaHire does at the CV stage
From our audit of 15 October 2026:
- Intake: up to 25 resumes per upload (PDF, DOC, DOCX or TXT, 5 MB each), a Resume Processing queue, and imports of candidates from Workday, SAP SuccessFactors and Kombo, and of jobs from JazzHR, JobDiva, SuccessFactors and Kombo.
- Scoring: an overall match with Skills, Experience and Education sub-scores; matched and missing skills; key strengths and areas for improvement; and an AI summary. Weights and thresholds are configurable per organisation and per job.
- Human gate: new applicants land in Pending Review, where a person approves or rejects. The AI recommendation is labelled "based on resume" and is advisory.
- Ranking: the Recommended tab ranks the existing pool against a job. The ranking is deterministic, AI Fit is displayed separately, and each row carries a written rationale and explicit gaps.
Limits: many parse failures on the QA build; parsing languages unknown; no bias-audit status to cite; no fairness reporting.
What do the law and regulators say about automated CV screening?
This section summarises the rules that bear on automated CV screening in the UK, EU, US and India.
| Jurisdiction | What applies | Key date or status | Source |
|---|---|---|---|
| UK | ICO audit of AI recruitment tools; "Recruitment Rewired"; Data (Use and Access) Act 2025 | ICO audit published 6 Nov 2024; "Recruitment Rewired" published March 2026 | ICO audit; ICO 2026 |
| EU | CJEU SCHUFA judgment; EU AI Act | 7 Dec 2023; high-risk obligations apply from 2 Dec 2027 | Regulation (EU) 2026/1744; EU AI Act obligations for recruiters |
| US | NYC Local Law 144; Mobley v. Workday; Kistler v. Eightfold AI | Enforcement from 5 Jul 2023; Mobley class-certification hearing 9 Mar 2027 | nyc.gov DCWP; Mobley docket; Kistler docket |
| India | DPDP Act 2023 and Rules | Rules notified Nov 2025; consent managers from 13 Nov 2026 | HLC |
UK. The ICO's audit of AI recruitment tools, published 6 November 2024, made almost 300 recommendations. The issues included filtering by protected characteristics, inferring gender and ethnicity from names, over-collection of data, indefinite retention and weak transparency. In "Recruitment Rewired", published in March 2026, the ICO reported engaging more than 30 employers and writing to 16; it asks for monthly bias monitoring, transparency to jobseekers and explained recourse, including human review. A consultation on draft automated-decision-making guidance was open until 29 May 2026 (ICO). In our reading, the Data (Use and Access) Act 2025, which received Royal Assent on 19 June 2025, replaces UK GDPR Article 22 with Articles 22A to 22D and keeps safeguards for solely automated decisions; check with counsel which provisions are in force for your use.
EU. In our reading, the CJEU's SCHUFA judgment (C-634/21, 7 December 2023) means that producing an automated score can itself count as an automated individual decision under GDPR Article 22(1) where a third party's decision depends heavily on it, so a CV score that a recruiter almost always follows may be caught. Under the EU AI Act, CV filtering falls under Annex III, point 4, as high-risk, with deployer duties in Article 26. Regulation (EU) 2026/1744 of 8 July 2026 (the Digital Omnibus on AI) entered into force on 27 July 2026 and moves the application of those high-risk obligations from 2 August 2026 to 2 December 2027. EU AI Act obligations for recruiters sets out what that means for a hiring team.
US. NYC Local Law 144 covers automated employment decision tools: a bias audit within a year, a public summary, and candidate notice 10 business days before use, with enforcement from 5 July 2023 (nyc.gov DCWP). Check the rule text for penalties and for whether resume screening falls within scope.
In Mobley v. Workday (N.D. Cal. 3:23-cv-00770), the docket on CourtListener shows that on 12 July 2024 the court granted in part and denied in part Workday's motion to dismiss (the ruling allowing the claim that Workday acted as employers' "agent" to proceed), and on 16 May 2025 granted preliminary collective certification. A Fourth Amended Complaint was filed on 10 July 2026, and the plaintiffs filed a motion to certify a class on 14 September 2026; as of 28 October 2026 the docket showed a hearing set for 9 March 2027. These are rulings on who can be sued and how the case proceeds. The underlying claims are the plaintiff's allegations, not findings. Kistler v. Eightfold AI (N.D. Cal. 3:26-cv-01768) includes a claim under the Fair Credit Reporting Act. As of 13 October 2026, the docket showed Eightfold's motion to dismiss filed on 20 April 2026, the hearing vacated on 24 July 2026 and no ruling on the motion. We cite it as a named case and draw no conclusion about either party.
India. The Digital Personal Data Protection Act 2023 has Rules notified in November 2025, with phased commencement. The consent-manager framework applies from 13 November 2026 (HLC). We have not confirmed any rule specific to automated CV screening. What to check with counsel: how the Act's consent and notice provisions apply to the CVs you process, and which of its phased dates apply to your organisation.
How do you screen CVs at agency scale?
CV screening for recruitment agencies
An agency's version of the problem adds three requirements to the scorecard: volume, per-client criteria, and a human sign-off before anything is submitted to a client. Test each. Can you hold different criteria for different clients at once? Can you bulk-upload without a per-file step? Does a person approve before a candidate leaves your desk?
In ZeaHire, a Submit to Client action keeps a log, and a multi-organisation transfer feature exists. Both are described in our audit of 15 October 2026.
CV screening in the UK and in India
For UK desks, the checks come from the ICO: monthly bias monitoring, transparency to candidates, and a route to human review. For India, add Indian-style CVs to the parse test, because date-of-birth and marital-status fields are ones the model should ignore, and confirm the DPDP dates above with counsel.
To run the process rather than choose the tool, see how to screen 200 CVs with AI step by step. To try it, see an evidence pack in a demo.
FAQ
What is CV screening software?
It is software that extracts, scores or ranks CVs against job criteria, with the decision left to a person. It runs in four stages: parse, score, match and decide.
How does CV screening software work?
It parses a CV into structured fields, scores it against the job's criteria, matches or ranks the pool, and presents the result to a reviewer who decides. Each stage can be done by a different tool.
What is the difference between CV parsing and an ATS?
Parsing extracts fields from a CV file. An ATS tracks candidates through the hiring workflow. Many CV parsers integrate with an ATS, as Skima's FAQ frames it.
Is CV screening software biased?
It depends on the criteria and the monitoring. The University of Washington study found bias in raw LLM ranking, and it did not test any product. That is why the scorecard asks which fields the tool ignores and how often bias is monitored.
Can ATS filters reject qualified candidates?
They can. The Harvard Business School and Accenture "Hidden Workers" study found that ATS filters can vet out high-skilled candidates who don't match the search criteria. That describes how filters are built, not how many candidates were rejected.
Is automated CV screening legal in the UK or EU?
It can be, with a lawful basis and safeguards. In the EU, CV filtering is classed as high-risk under the AI Act; the legal section above and the EU AI Act page set out the detail.
How accurate is CV parsing?
Test it on your own formats. The accuracy percentages on the vendor pages we read came without a method or source, so we do not repeat them.
Does ZeaHire reject candidates automatically?
In ZeaHire, candidates wait in Pending Review until a person approves or rejects them.
Try it on your own CVs
See an evidence pack in a demo to view the match scores, the matched and missing skills, and the Pending Review queue. Custom pilot pricing is available; talk to us.
Comments
No comments yet.
Comments are closed on this post.