EU AI Act Recruitment Compliance: What's High-Risk, Who Owns What, and What Applies Now

Under the EU AI Act, recruitment AI that filters applications, ranks candidates or scores interviews is likely high-risk. The high-risk rules for employment don't apply until 2 December 2027. The rules that apply today are narrower: banned practices, AI literacy for staff and national enforcement powers. This guide maps EU AI Act recruitment duties by owner (employer or vendor) and by how much each tool actually shapes a hiring decision.

What Applies Now and What Starts in December 2027

Most early summaries got this date wrong, and it's the date recruiters care about most. The AI Act entered into force on 1 August 2024 and became generally applicable on 2 August 2026. Much of the early guidance told employers their hiring tools had to be ready by that August date. Then the AI Omnibus amendment, in force since 27 July 2026, moved the start date for Annex III high-risk systems, including employment, to 2 December 2027.

Here's the current timeline, based on the Commission's AI Act overview and timeline:

• Since 2 February 2025, the prohibited practices and the AI-literacy duty for providers and deployers have applied.
• Since 2 August 2026, the Act has been generally applicable and national authorities have had enforcement powers.
• From 2 December 2027, the high-risk rules for Annex III uses apply. Recruitment and selection are Annex III uses.
• From 2 August 2028, the high-risk rules for AI embedded in regulated products apply. That date rarely matters for recruiting software.

Two parts of the Act already apply to recruiters. The first is AI literacy. Anyone on your team who uses AI in hiring needs enough understanding of the tool to use it sensibly. The Commission doesn't require a fixed course or test. Training should match what your staff already know, how risky the system is and who it affects.

The second is the list of banned practices. If a tool claims to read candidates' emotions from their voice or face, get legal review now. Emotion recognition in workplace settings is a prohibited practice, with narrow medical and safety exceptions.

The Act also reaches beyond companies based in the EU. If you hire in Europe, or a tool's output is used on candidates in the EU, assume it can apply to you, even if your company is in the US. GDPR has applied throughout, so automated candidate evaluation already raises data-protection questions whatever the AI Act timeline says. This article is practical guidance, not legal advice.

EU AI Act Recruitment High-Risk Classification: Judge the Use, Not the Product

Annex III lists AI intended for recruitment or selection as high-risk. It names three examples: placing targeted job ads, analysing and filtering applications, and evaluating candidates. It also covers AI used for some decisions about current workers, such as promotion, termination, task allocation and performance monitoring. Using AI somewhere in the tool doesn't make it high-risk. What counts is what the system is intended to do and how much its output affects who moves forward.

Article 6(3) offers a narrow way out. An Annex III system may not be high-risk if it doesn't pose a significant risk or materially influence the decision, and it only does one of the following:

• a narrow procedural task
• improving the result of a human activity that's already finished
• spotting patterns without replacing or influencing a finished human review
• a preparatory task

There are two limits. A system that profiles people stays high-risk regardless, and it's the provider who documents the not-high-risk assessment. You can read the exact wording in the consolidated AI Act text.

On mobile, read the table row by row: what the tool does, where it probably lands, and what to do this quarter. "Depends" means your deployment decides the answer, so write down how you use the tool.

Edge cases that trip teams up

• One platform can include both a scheduling bot and a ranking model. Classify each function separately, not the vendor as a whole.
• A summarizer looks preparatory. If recruiters build shortlists from summaries without reading the actual answers, the summary materially influences the decision, and the preparatory argument gets weak.
• "The recruiter makes the final decision" is a vendor talking point, not a classification. When a tool ranks 300 applicants, the ranking decides who gets looked at, no matter who clicks the last button.
• A hard-coded knockout question ("Do you hold a forklift license?") may not be an AI system at all. A model that infers license status from free text is.
• Asking a general-purpose chatbot to rank a stack of CVs gives it a high-risk purpose. Under the Act, a deployer who does that can take on provider obligations.

Provider vs Employer: Who Owns What

The Act splits the work for high-risk AI systems in recruitment between the provider and the deployer. The provider is usually your vendor. The deployer is you, the employer. Vendor marketing often blurs this line, so keep it clear in contracts and in your own planning.

What the provider owns

• risk management across the system's lifecycle
• data governance for training, validation and test data
• technical documentation
• logging capability built into the system
• instructions for use that explain how to deploy it properly
• human-oversight design, so a reviewer can understand, override or stop the system
• accuracy, robustness and cybersecurity
• conformity assessment before the system goes on the market
• post-market monitoring once it's in use

You can't audit most of this yourself, but you can ask for evidence. A badge, a trust page or a contract clause promising compliance "with applicable law" doesn't settle any of it.

What the employer owns

• using the system according to the provider's instructions
• assigning human oversight to people with the competence, authority and support to do it
• making sure input data under your control is relevant and representative enough for the intended purpose
• monitoring the system in use and acting on risks you notice
• keeping the logs under your control for at least six months, unless other law sets a different period
• notifying the provider, and in some cases the authorities, when you see a serious incident or a risk
• informing workers and worker representatives before using a high-risk system in the workplace
• telling candidates and employees when a high-risk system makes or assists decisions about them

People affected by some high-risk decisions may also have a right to a clear and meaningful explanation of the role the AI played. Separate transparency rules mean people should know when they're interacting with an AI system, unless that's obvious from context. Our AI hiring disclosure templates cover candidate notice and accommodation wording.

Fundamental-rights impact assessments are an extra AI Act duty for public bodies and some deployers that provide public services. Not every private employer has that duty under the AI Act. A GDPR data protection impact assessment is a separate question and often still applies.

Human oversight that actually counts

A recruiter who clicks "approve" on 200 AI rankings isn't providing oversight. The Act expects the people doing oversight to understand the system's limits and watch for automation bias. They also need to be able to disregard, reverse or stop the output. In practice:

• Reviewers can see why a candidate got a score or rank, not only the number.
• They have enough time. Ten seconds per candidate is a rubber stamp.
• They can overrule the tool without asking a manager for permission.
• They review a sample of rejected candidates, not only the shortlist.
• Overrides are logged, and someone looks at the override rate every month.

A structured interview scorecard gives reviewers their own basis for checking AI output. Without one, reviewing tends to turn into agreeing. If you use scoring models, read how automated interview evaluation produces its numbers before you decide how much weight to give them.

Vendor Checklist for Recruitment Tools That May Be High-Risk

Send these questions to every vendor whose tool filters, ranks, scores or summarizes candidates. A vague answer still tells you something. Use the same list when comparing AI candidate screening tools of any kind, whether that's a CV ranker, an interview-scoring model or an async voice-interview platform like Kira.

1. What is the system's intended purpose, and do you classify it as high-risk under Annex III? If not, can we see your Article 6(3) assessment? 2. Can we see the instructions for use now, including known limitations and conditions where accuracy drops? 3. What does a reviewer see behind each score, rank, or summary? How do they override, disregard, or stop it? 4. What is logged automatically, who controls the logs, how long are they kept, and can we export them? 5. How and how fast do you notify customers of serious incidents or malfunctions? Who do we report problems to? 6. What data was the system trained and tested on? What do you need from the inputs we control, such as job criteria and question sets? 7. What performance and bias metrics do you monitor after deployment, and which can we see for our own roles? 8. If a candidate asks why they were rejected, what can the system tell us about the role it played? 9. What candidate notice do you provide or recommend, and does the candidate-facing experience disclose that it's AI? 10. Which third-party models or subprocessors do you use? Will you tell us before changing the model or scoring logic? 11. What is your plan and timeline for conformity assessment before 2 December 2027?

Ask for written answers. A model update six months from now can change how a tool classifies candidates, so question 10 matters more than it looks.

A 30-Day EU AI Act Recruitment Compliance Plan

You don't need a legal team to make progress this month. You need a spreadsheet, named owners and vendor answers in writing.

1. Days 1 to 5: list every AI feature in your hiring process. Include ATS features that were switched on by default and any chatbot a team member set up on their own. Note the vendor, the function, the hiring stage and who uses it.
2. Days 6 to 10: map how much each function influences decisions. Mark each one as procedural, preparatory, shapes the shortlist, or decides. Anything in the last two categories goes on the high-risk track.
3. Days 11 to 15: assign an owner for each tool and design the oversight. Name the reviewers, what they check, how much time they get per candidate and how they override the tool.
4. Days 11 to 20: train the people who use these tools. AI literacy already applies. Cover what each tool does, where it tends to be wrong, automation bias and the override process. Keep a record of who was trained.
5. Days 16 to 25: collect evidence from vendors. Send the checklist above and flag any vendor that can't answer questions 1, 3 and 10.
6. Days 21 to 30: prepare candidate notices, worker communications and an incident path. Decide who spots a problem, who can pause the tool and who tells the vendor.

If you also hire in the US, the AI Act applies alongside state and city rules. Our guide to AI interview laws for employers covers those.

Key Takeaways

• The high-risk rules for employment start on 2 December 2027, not in August 2026. AI literacy and the prohibited practices have applied since February 2025.
• Classification depends on intended purpose and how much the tool influences decisions. Scheduling and transcription are likely not high-risk. CV ranking and interview scoring likely are.
• "The recruiter makes the final decision" doesn't remove high-risk status when the AI output shapes the shortlist.
• Vendors own design, documentation and conformity assessment. Employers own use according to instructions, oversight, input data, logs, monitoring and notices.
• Oversight only counts when reviewers have the time, the authority and enough information to disagree with the tool.
• Start now with a tool inventory, a map of how much each tool influences decisions and written vendor answers to the checklist.