AI in Recruitment: Is Our Applicant Tracking System High-Risk AI? Practitioner Q&A
Is your applicant tracking system high-risk AI? What the EU AI Act, UK GDPR and the Equality Act ask of employers using AI to screen candidates.
Many HR teams now run AI without ever having bought any. Vendors add matching, ranking and scoring features to recruitment software that organisations already use and nobody makes a purchasing decision. This page is for HR directors, compliance leads and DPOs in mid-sized UK organisations who need to find out what their recruitment system does and how much human involvement each step needs. It is general information and is not legal advice.
Is an applicant tracking system an AI system?
It depends on what the system does, and the answer may have changed since you bought it. Article 3(1) of the EU AI Act defines an AI system as a machine-based system that infers from its inputs how to generate outputs such as predictions, recommendations or decisions. An applicant tracking system that applies fixed keyword rules and knock-out questions may fall outside that definition. Ranking, matching and scoring features may fall inside it and vendors often add them through routine updates.
The definition matters less under UK data protection law. The UK GDPR rules on automated decision-making apply to automated decisions whether or not the tool counts as AI. Start by listing what each feature does to a candidate’s application.
Is AI used in recruitment high-risk under the EU AI Act?
Usually, yes, where the tool screens, ranks or scores candidates. Annex III point 4(a) of the Act lists AI systems intended for recruiting or selecting people, in particular placing job adverts for specific audiences, analysing and filtering applications and evaluating candidates. Under Article 6(2), an Annex III use is a high-risk AI system.
Article 6(3) lets a system fall outside high-risk if, for example, it only performs a narrow procedural or preparatory task, but an Annex III system that profiles people is always high-risk. CV ranking, filtering and candidate scoring will usually stay high-risk. Interview scheduling or drafting a job advert may fall outside. Classification turns on the provider’s intended purpose, and the Commission’s guidelines on Article 6 are still a draft.
Regulation (EU) 2026/1744 moved the start date for Annex III high-risk obligations to 2 December 2027 and left the obligations themselves in place.
Does the EU AI Act apply to UK employers?
It can, where the system’s output is used in the EU. Article 2(1)(c) covers providers and deployers located outside the Union where the output the AI system produces is used in the Union. In recruitment, that could mean screening candidates for roles in an EU member state or a UK agency sending shortlists to EU clients. The Act does not define when output is used in the Union and there is no settled Commission guidance on the boundary.
A UK employer hiring only for UK roles, with no output used in the EU, is unlikely to be caught. UK data protection and equality law apply either way. Our FAQ on the EU AI Act and UK businesses covers the wider question.
Can employers reject candidates automatically using AI?
UK law does not ban automatic rejection outright, but it attaches safeguards. Articles 22A to 22D of the UK GDPR have been in force since 5 February 2026. They allow a significant decision made with no meaningful human involvement where the Article 22C safeguards are in place: information about the decision, the chance to make representations, human intervention and the ability to contest it. Article 22B keeps tighter limits where the decision rests on special category data, such as health data relevant to disability.
The harder question is whether your human review counts as meaningful. In its March 2026 report Recruitment Rewired, the ICO said most employers it spoke to thought it unlikely they were using automated decision-making, yet it concluded several were. Token or rubber-stamp review does not amount to meaningful involvement, which is how automation bias weakens a human in the loop control. The ICO’s updated guidance on automated decision-making is still a draft.
Who is responsible if AI recruitment software is biased?
Under UK equality law, responsibility sits with the employer. Section 39 of the Equality Act 2010 is aimed at employers and covers the arrangements they make for deciding who is offered a job. Section 19 on indirect discrimination applies whether a person or software applies the criterion, so AI bias in a vendor’s tool is tested as part of the employer’s own process. The employer is the respondent in an employment tribunal, where compensation for discrimination is uncapped.
Where the EU AI Act applies, Article 26 places separate duties on the deployer and the provider’s own conformity does not discharge them. A contractual indemnity may move cost to the vendor while legal responsibility stays with the employer. The vendor’s own exposure is unsettled.
What should we ask a recruitment software vendor about AI?
Ask what the tool does to a candidate’s application and what evidence supports it. For recruitment, five questions matter most:
- Which features rank, match, score or filter candidates? Which are switched on by default?
- Can any step reject a candidate before a person sees the application?
- Does the tool infer characteristics such as gender or ethnicity, or let recruiters filter by protected characteristics?
- What bias testing have you run and on which populations? Can we see the results?
- What candidate data do you collect and how long do you keep it?
The ICO’s audits of AI recruitment tools, reported in November 2024, found tools that inferred characteristics, allowed filtering by protected characteristics and kept candidate data indefinitely. The ICO also published its own key questions for buyers. Our questions to ask suppliers about AI cover the wider list.
How do you test AI recruitment tools for bias?
Compare outcomes across groups at every stage where the tool filters, ranks or scores, using your own applicant data where you can. The government’s Responsible AI in Recruitment guide, which is non-binding, lists bias audits, performance testing, piloting and impact assessments, including an equality impact assessment. Where the buyer lacks in-house expertise, the guide says the supplier should do the performance testing.
Two limits apply. Testing for AI fairness needs protected-characteristic data that an employer may not hold, and the ICO has criticised tools that infer those characteristics instead of collecting them. Small applicant volumes weaken the statistics, and a vendor’s results from other populations may not reflect your applicant pool.
Check too how automated assessments, video analysis and timed tests handle candidates who need reasonable adjustments under the Equality Act.
Do we have to tell candidates we use AI in recruitment?
UK data protection law already asks for transparency, and the EU AI Act adds a specific duty for deployers in scope. The UK GDPR requires transparency about how organisations use personal data, and the Article 22C safeguards include information about significant automated decisions. The ICO found poor transparency was a common failing in recruitment.
Under Article 26(11) of the EU AI Act, deployers of Annex III systems that make or assist in decisions about people must inform those people that they are subject to the system. Article 26(7) asks employers to inform workers’ representatives and affected workers before putting a high-risk system into service at the workplace. Both apply from 2 December 2027. A worry that disclosure will deter applicants does not change what either law asks for.
What should we check before using AI to screen candidates?
Check five things before you switch AI screening on, or leave it on:
- What each AI feature does and whether any infers emotions. Where the EU AI Act applies, Article 5(1)(f) has banned emotion inference in the workplace since 2 February 2025, and that covers candidates assessed for work.
- Every point where the system can reject a candidate before a person sees the application.
- Whether the people who review the tool’s output have the competence, training and authority to overrule it.
- Whether candidates are told about automated decisions and can ask for human intervention and contest the outcome.
- Whether a data protection impact assessment, an equality impact assessment and bias test results are on file.
UK data protection and equality law apply now, so the December 2027 date for EU high-risk obligations is no reason to wait.
AI Security Projects
An AI tool risk assessment is one of our focused AI Security Projects, each with a defined deliverable.