Short answer: AI candidate matching is software that compares what a candidate brings with what a role needs, then shows the result to a recruiter. Tools match on keywords, on similar meaning, or requirement by requirement. For frontline roles, trust the version that shows the evidence behind each requirement and leaves every decision to a person.
Key takeaways
- Most matching tools compare text: keywords, or the meaning of a CV against a job description. Similar text is weaker evidence than a met requirement.
- Frontline roles turn on facts you can check, such as shifts, certificates, start dates and right to work. A CV often leaves them out.
- A single match percentage hides which requirement was met, and how. Ask to see the evidence for each requirement.
- In the EU, AI that filters or evaluates candidates is listed as high risk under the AI Act, and those rules apply from 2 December 2027. The GDPR and equal treatment law apply today.
On this page: What is AI candidate matching? · How does AI candidate matching work? · Why does AI candidate matching fail for frontline roles? · What does a match percentage hide? · What does EU law say about AI candidate matching? · What should you ask before you trust a matching tool? · How should agencies and frontline employers use AI matching day to day? · What can AI candidate matching not tell you? · Conclusion · How Radius helps · FAQ
What is AI candidate matching?
AI candidate matching compares candidates with a role's requirements and reports the fit. On one side sits the role: a job description, a list of must haves and sometimes a set of screening questions. On the other side sits the candidate: a CV, application answers, interview answers and sometimes documents such as a certificate. The software compares the two.
The result may be a percentage, an ordered list or, for each requirement, whether the candidate meets it and on what basis. That shape decides what the recruiter looks at.
How does AI candidate matching work?
Most tools use one of three methods, or a mix of them.
Keyword matching looks for the same words in the CV and the job description. If the role asks for "reach truck" and the CV says "reach truck", the candidate gains points. It is simple and easy to explain, but it misses anyone who describes the same experience in other words and rewards candidates who copy the advert.
Semantic matching compares meaning rather than exact words. A language model turns each CV and each job description into a long list of numbers, often called an embedding, so that texts about similar things end up close together. The tool then orders candidates by how close their text sits to the text of the role. This catches synonyms and varied phrasing. It also inherits patterns from the model's training text, and similar meaning is no proof that someone can do the job.
Requirement based matching breaks the role into separate requirements and checks each one against evidence. "Valid reach truck certificate", "available for night shifts" and "can start within two weeks" become separate lines, each answered by a document, an interview answer or nothing at all. Some tools first map skills to a shared vocabulary such as the European Commission's ESCO classification, which describes 3,039 occupations and 13,939 skills in 28 languages and which the Commission says online platforms can use for matching jobseekers to jobs on the basis of their skills (European Commission, ESCO, checked 22 October 2026).
| Method | What it compares | Handles well | Where it fails in frontline hiring | What you can check |
|---|---|---|---|---|
| Keyword | Words in the CV against words in the advert | Clear, exact terms | Misses people who write differently or have no CV | Which words count, and whether synonyms are listed |
| Semantic | The overall meaning of a CV against the job description | Varied wording and related experience | Treats similar text as fit; learned bias carries over | Which text goes in, and whether you can see why two texts were judged close |
| Requirement based | Each requirement against a piece of evidence | Musts such as certificates, shifts and start dates | Only as good as the requirements you write and the evidence you collect | Which evidence backs each line, and what is still open |
Many products combine the three: a semantic search to find people, then a requirement check on the shorter list. Ask which step does what.
See it on a real requirement list: how Match shows the evidence behind each requirement.
Why does AI candidate matching fail for frontline roles?
It fails where the CV says little and the job turns on facts. An order picker, a reach truck driver, a cleaner or a security officer is hired on a short list of concrete requirements. Can they work the shift? Do they hold the certificate the site needs? May they work in the Netherlands? Can they start on Monday? Frontline applicants often apply from a phone, with a short CV or none at all. A CV text match then compares two documents that barely describe the job.
Three failures follow.
Similar is mistaken for qualified. A CV that mentions six years "in a warehouse" sits close in meaning to a warehouse job description. It says nothing about whether the reach truck certificate is still valid or whether the person can work nights.
Silence is mistaken for a gap. A candidate who never wrote down that they hold a VCA safety certificate looks like a weaker match than one who did. A text match cannot tell "does not have it" from "did not mention it".
Old patterns are repeated. Models learn from text that people wrote. In a 2024 University of Washington study, researchers ran more than three million comparisons between real CVs and job descriptions with three language models, varying the names on the CVs. The models preferred white associated names 85% of the time and female associated names 11% of the time, and never preferred Black male associated names over white male associated names (US names and data; University of Washington News, 31 October 2024; Wilson and Caliskan, AIES 2024). A matching tool built on such a model can repeat those patterns without anyone choosing them.
Three language models comparing real CVs with job descriptions
More than three million comparisons, with the names on the CVs varied; US names and data
Research on job performance points the same way. A 2022 reanalysis of decades of selection studies found that structured interviews had the highest average validity of the methods it compared, an estimated correlation of .42 with job performance (a review of published studies, not specific to the Netherlands; Sackett, Zhang, Berry and Lievens, Journal of Applied Psychology, 2022). A structured interview asks every candidate the same questions about the job, so it produces evidence you can match against requirements. A CV mostly produces text to match against more text.
Illustrative example. A staffing agency fills reach truck roles on night shifts at a distribution centre. Two people apply.
| Requirement | Candidate A (detailed CV) | Candidate B (no CV, short structured interview) |
|---|---|---|
| Valid reach truck certificate (must) | Not mentioned | Said in the interview; certificate document uploaded |
| Available for night shifts (must) | Not mentioned | Said in the interview |
| Can start within two weeks (must) | Not mentioned | Said in the interview |
| Two years of warehouse experience (nice to have) | On the CV: six years | Said in the interview: one year |
| Text similarity to the job description | High | Low |
A text match puts Candidate A first. A requirement view shows that A has not yet answered any of the musts, while B has answered all three and backed one with a document. Neither view decides the hire. The second tells the recruiter what to ask next, and of whom.
What does a match percentage hide?
It hides which requirement was met, and on what evidence. An 80% match can mean the candidate meets everything except a nice to have. It can also mean they meet every nice to have and miss the one certificate the site cannot do without. The number is the same. The next step is not.
What an 80% match can hide
The candidate meets everything except a nice to have.
They meet every nice to have and miss the one certificate the site cannot do without.
The number is the same. The next step is not.
A percentage also hides the choices behind it: which requirements carry more weight, how text was turned into numbers and where the cut off sits. Whoever set up the tool made those choices, and the recruiter reading the result rarely sees them.
There is a human side too. A neat number invites people to trust it. From 2 December 2027, the EU AI Act will require high risk hiring tools to be built so that the people overseeing them stay aware of the tendency to rely too much on the output, and can disregard or override it (Article 14; Regulation (EU) 2024/1689). That is hard with a single figure and easier with a list of requirements a recruiter can read and check. For the wider debate on whether software should order candidates at all, see our article on counting evidence instead of ranking candidates.
What does EU law say about AI candidate matching?
Matching that evaluates candidates is classed as high risk under the AI Act. Annex III, point 4(a), covers AI systems intended for recruitment or selection, in particular to "analyse and filter job applications, and to evaluate candidates" (AI Act Service Desk, Annex III, checked 22 October 2026). A narrow exception exists for tools that do not materially influence hiring decisions, but it never applies where the system profiles candidates (Article 6(3); AI Act Service Desk, Article 6). The European Commission's draft guidelines on high risk classification, published for consultation on 19 May 2026, are not final and not binding. They give a system that an employment agency uses to rank candidates for vacancies as a high risk example that cannot use that exception (European Commission, draft guidelines, 19 May 2026).
The Digital Omnibus on AI (Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026) moved the high risk rules for employment from 2 August 2026 to 2 December 2027 (Regulation (EU) 2026/1744; European Commission, AI Act page, checked 22 October 2026).
Two sets of rules already apply.
- The GDPR. It does not ban automated screening. It gives candidates the right not to be subject to a decision based solely on automated processing that significantly affects them, such as an automatic rejection, unless an exception applies (necessity for a contract, authorisation by law, or explicit consent). Where an employer relies on the contract or consent exception, candidates must at least be able to get human intervention, give their view and contest the decision (Article 22; Regulation (EU) 2016/679).
- Equal treatment law. The Netherlands Institute for Human Rights (College voor de Rechten van de Mens) says that if an algorithm used in recruitment discriminates, the employer is responsible and cannot shift that responsibility onto the algorithm or the software vendor (College voor de Rechten van de Mens, 8 December 2021). Its guidance for employers suggests asking vendors how their algorithms work and whether they exclude or discriminate against groups of people (College voor de Rechten van de Mens, Q&A for employers).
For the full duties and dates, read our guide to the EU AI Act and recruitment.
What should you ask before you trust a matching tool?
Ask questions you can check in a demo, not in a brochure. Bring a real vacancy and ask the vendor to show the answer to each question on screen.
| Question | Why it matters | A good answer shows you |
|---|---|---|
| 1. What goes in: CV text, application answers, interview answers, documents? | A match is only as good as its inputs | The exact inputs used for this role |
| 2. Which method does the matching: keywords, meaning, requirements, or a mix? | Each method fails in a different way | Which step uses which method |
| 3. Can I see, for each requirement, what the candidate met and on what evidence? | A total hides the missing must | One line per requirement, with its source |
| 4. How does the tool tell "does not have it" from "did not mention it"? | Silence is not a gap | Unanswered requirements shown as open, not as failed |
| 5. Who writes the requirements, and can we separate musts from nice to haves? | The tool inherits your job criteria | Requirements your team sets and edits per role |
| 6. Does the tool ever reject or hide a candidate before a person sees them? | Solely automated rejections raise GDPR questions | Every applicant stays visible, and a person decides |
| 7. What has been tested for unfair outcomes, on what data, and when? | You stay responsible if the outcome discriminates | A dated test with its method, not a slogan |
| 8. Does it use names, photos, age or other personal details it does not need? | These details can carry bias into the match | A list of the inputs that are left out |
| 9. Are candidates told that AI is involved, and what is recorded? | Since 2 August 2026, the AI Act has required AI that talks to people to be built to tell them so, unless that is obvious | The exact words candidates see or hear |
| 10. Can a decision be explained later from the record? | Candidates, clients and auditors ask | The evidence, its source and its date, kept per candidate |
How should agencies and frontline employers use AI matching day to day?
Use it to sort evidence, then check it against your own roles. Five habits help.
- Write requirements as facts you can check. "Valid reach truck certificate" can be answered. "Strong work ethic" cannot. Keep the musts few and separate from the nice to haves.
- Collect evidence before you match. A short structured interview with the same questions for every applicant gives the tool something real to match against. It also gives applicants without a CV a fair way in.
- Read the near misses. A candidate who meets four of five requirements, with the fifth simply unanswered, is often one phone call away from a placement. Review them on purpose.
- Test the tool on a role you know. Before you rely on it, run it next to your recruiters on one live role and look at where they disagree. Read a sample of the candidates the tool rated lowest and ask whether a recruiter would have called them.
- Keep the reason with the decision. Clients ask agencies why a candidate was put forward. A record of which requirement was met, with which evidence and when, answers that question and helps if a candidate asks for a review.
Agencies need one more check. The same job title can carry different requirements at two clients, or at two sites of one client. A tool that cannot hold separate requirements per client, site and role matches people against the wrong standard.
What can AI candidate matching not tell you?
It cannot tell you whether someone will turn up for the first shift. It cannot judge motivation, and it cannot turn a claim into a fact. An interview answer that says "I have my certificate" is still a claim until someone sees the certificate. A requirement view only shows what was asked and answered, so an open requirement may simply be a question nobody asked. Matching narrows where a recruiter looks. The recruiter still has to look.
Conclusion
AI candidate matching is worth using when it makes a recruiter's next step clearer. Whatever you choose, ask to see the evidence behind each requirement, keep a person on every decision and test the tool on your own roles first.
How Radius helps
Radius Hire is the AI hiring platform for frontline roles in the Netherlands. Its AI Interviewer, Match and Radius Passport sit on one candidate record. Radius is new, so this section explains how the product works rather than claiming results.
- The AI Interviewer runs a short, structured voice interview that candidates can take on their phone, whenever it suits them, in the language set for the role. Every applicant answers the same questions before anyone compares them.
- Match reads each interview against the role's requirements. For each requirement, Radius shows whether it is backed by a document or a check, rests only on what the candidate said, or is still missing, with the moment in the interview that answered it.
- Where the employer uses Radius Passport, it checks a candidate's identity and right to work with a trusted verification partner and records certificates with their expiry dates, so a recruiter can review the evidence before an offer. It does not replace the employer's own identity check.
- Radius never rejects a candidate on its own. A recruiter reviews the evidence, can overrule anything Radius presents and makes every hiring decision. Candidate data is hosted in the EU.
- Staffing agencies recruiting for several clients and sites can see how this works on Radius for agencies.
FAQ
Is a higher match percentage a better candidate?
Not necessarily. A percentage adds up requirements with weights that someone chose, so two candidates with the same figure can differ on the one requirement that matters. Look at which musts are met, and on what evidence, before you compare totals.
Can AI candidate matching work without a CV?
Yes, if the tool matches against evidence other than CV text. Answers from a structured interview, and documents such as certificates, can be matched against each requirement. For frontline roles, where many applicants have a short CV or none, this is often the better input.
Does AI candidate matching remove bias?
No. Applying the same requirements to everyone helps consistency, but models can carry bias from the text they learned from. Test outcomes on your own roles, keep requirements related to the job and keep a person on every decision. Our article on AI bias in recruitment goes further.
What is the difference between AI job matching and AI candidate matching?
They look at the same fit from opposite sides. AI job matching usually recommends jobs to a job seeker. AI candidate matching shows an employer or agency which applicants fit a role. The methods overlap. The Commission's draft guidelines (19 May 2026, not final or binding) treat a system that an employment agency uses to rank candidates for vacancies, or to suggest vacancies to jobseekers, as high risk.
General information, not legal advice. Position as of 22 October 2026.
Sources
- European Union. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal L, 12 July 2024; Annex III point 4, Articles 6, 14 and 50. Official Journal text. Checked 22 October 2026.
- European Commission, AI Act Service Desk. "Annex III" and "Article 6: Classification rules for high-risk AI systems" (consolidated text as at 27 July 2026). Annex III; Article 6. Checked 22 October 2026.
- European Union. Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal L, 24 July 2026. Official Journal text. Checked 22 October 2026.
- European Commission. "AI Act" (policy page on the regulatory framework for AI). Commission page. Checked 22 October 2026.
- European Commission. "Draft Commission guidelines on the classification of high-risk AI systems", published 19 May 2026, updated 23 July 2026 (draft for consultation, not binding); Annex III document, example of an AI system used by employment agencies to assign candidates to vacancies. Commission page; Annex III document. Page and Annex III document rechecked 23 October 2026: still a draft, no final version published.
- European Union. Regulation (EU) 2016/679 (General Data Protection Regulation), Article 22. Official Journal text. Checked 22 October 2026.
- College voor de Rechten van de Mens. "Werkgever zelf verantwoordelijk als algoritmes bij sollicitaties discrimineren", 8 December 2021. mensenrechten.nl. Checked 22 October 2026.
- College voor de Rechten van de Mens. "Vraag en antwoord over werving- en selectie-algoritmes voor werkgevers" (undated). mensenrechten.nl. Checked 22 October 2026.
- Wilson, K. and Caliskan, A. "Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval", Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 2024. arxiv.org/abs/2407.20371. Summary: University of Washington News, 31 October 2024, washington.edu. Checked 22 October 2026.
- Sackett, P. R., Zhang, C., Berry, C. M. and Lievens, F. "Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range", Journal of Applied Psychology, 107(11), 2040 to 2068, 2022. doi.org/10.1037/apl0000994. Checked 22 October 2026.
- European Commission. "What is ESCO?" esco.ec.europa.eu. Checked 22 October 2026.