EU AI Act compliance for recruitment & HR AI

AI that screens CVs, ranks candidates, or manages workers is explicitly listed as high-risk under the EU AI Act. If you build or deploy it, the high-risk obligations apply to you.

If your company screens CVs, ranks applicants, or uses software to allocate tasks and evaluate performance, you are almost certainly inside Annex III point 4 — the employment category. That holds whether you built the tool or bought it. What changes is which set of duties is yours: the Act splits obligations between the provider who puts the system on the market and the deployer who uses it, and most SMEs are deployers of a vendor's tool. Being a deployer is lighter than being a provider. It is not nothing.

Why it is in scope

Annex III(4) covers AI used for recruitment, candidate evaluation, and decisions affecting work relationships — a high-risk category with the full obligation set.

What this does not trigger

Not every HR tool is caught. Payroll, benefits administration and scheduling sit outside Annex III point 4 only where they neither make decisions affecting the terms of the work-related relationship, nor allocate tasks on the basis of behaviour or personal traits, nor monitor and evaluate performance — only the task-allocation limb carries a behaviour-or-traits condition, so a rostering tool that sets working time can be caught even though it evaluates nobody. A CV tool that extracts fields into a database without ranking or scoring is the textbook narrow procedural task, but that is the Article 6(3)(a) derogation from a listed use, not a use that was never listed: it escapes high-risk only on a documented assessment. And that assessment is the provider's — Article 6(4) requires it to be documented before the system is placed on the market or put into service, with registration under Article 49(2) on top. A system that profiles natural persons is never exempt. As a deployer, ask your vendor to show you that file rather than assuming the exemption.

Provider or deployer?

You are a provider if you develop the system, or if you put your own name on it, modify it substantially, or repurpose a general-purpose tool into a hiring tool. You are a deployer if you use a vendor's system under your own authority. A recruitment agency running screening for its clients is a deployer; the vendor selling the screening engine is the provider. Fine-tuning a bought model on your own hiring data can tip you into being a provider of a new high-risk system where it amounts to a substantial modification or changes the system's intended purpose — those, not fine-tuning as such, are the triggers Article 25(1) lists.

Key obligations

• Risk management system (Article 9)
• Data governance for training data quality and bias (Article 10)
• Annex IV technical documentation (Article 11)
• Human oversight of automated decisions (Article 14)
• Employers: inform workers' representatives and affected workers before putting the system into use (Article 26(7))
• Conformity assessment and EU database registration (Articles 43, 49)

What this looks like in practice

• A CV-parsing tool that ranks applicants against a job description is in scope even when a recruiter makes the final call. Annex III point 4(a) covers systems used to filter applications and evaluate candidates, not only systems that decide on their own.
• Software that assigns shifts or tasks on the basis of measured behaviour or personal traits, or that monitors and evaluates performance, falls under point 4(b). This catches workforce-management tools that were never marketed as recruitment AI.
• Targeted job advertising is named explicitly: placing job ads in front of specific groups is part of point 4(a). A recruitment-marketing tool that optimises audience selection is not outside the regime merely because it never touches an application.
• Interview-analysis tools that score candidates on tone, expression or word choice are evaluating candidates — and some of them carry more than high-risk weight. Where the tool infers emotions from biometric data such as face or voice, Article 5(1)(f) prohibits it outright in a hiring process, outside narrow medical or safety uses. Emotion inferred from written text alone stays outside that prohibition and inside the Annex III high-risk regime.
• Internal mobility and promotion tools sit in point 4(b) alongside recruitment. The obligations do not begin at the hiring gate and stop there: decisions on promotion and termination are named in the same category.

Where SMEs get this wrong

• Assuming a human in the loop takes the system out of scope. It does not. Human oversight is an obligation under Article 14, not an exemption from the regime.
• Assuming the vendor's compliance covers you. A provider's conformity assessment does not discharge your deployer duties under Article 26 — using the system according to its instructions, assigning competent oversight, keeping the logs, and making sure the input data you feed it is relevant.
• Missing Article 26(7). Before putting into service or using a high-risk system at the workplace, a deployer who is the employer must inform the workers' representatives and the affected workers that they will be subject to it. It is purely procedural, and it is routinely skipped.

What actually proves compliance

For a deployer the evidence is unglamorous: the instructions for use you received, a record of who exercises oversight and what competence they have, the automatically generated logs you retained, and the notice you gave to workers. If you are the provider, add the Annex IV technical file, the Article 9 risk-management records, and the Article 10 evidence on training, validation and testing data.

What getting it wrong costs

Employment systems are Annex III high-risk, so breaches of the provider duties in Article 16 or the deployer duties in Article 26 fall in the Article 99(4) band: up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. Two things soften that for smaller companies. Article 99(6) caps fines for SMEs and start-ups at whichever of the two is lower, and the Digital Omnibus added Article 99(6a), extending the same lower-of rule to small mid-caps for the Article 99(4) and (5) bands. Emotion inference in the workplace is a different matter: prohibited under Article 5, and in the 7% band.

When it applies

Annex III high-risk obligations apply from December 2, 2027 following the Digital Omnibus reschedule. That is not permission to wait: the Annex IV file, the risk-management system and the data-governance evidence all describe work done during development, so they are assembled while the system is being built — not written up afterwards.

Questions we get asked

We only use an off-the-shelf applicant tracking system. Are we really in scope?

If it ranks, scores or filters candidates, yes — as a deployer. Your duties come from Article 26: use it as instructed, assign human oversight to someone competent, keep the logs it generates, ensure the input data is relevant, monitor the system in use and suspend and report it if it turns out to present a risk, inform candidates that they are subject to it (Article 26(11)), and — as an employer — inform your workers' representatives before you start. You do not owe the Annex IV technical file; your vendor does.

Do we need a fundamental rights impact assessment?

Probably not. Article 27 requires one from deployers that are bodies governed by public law or private entities providing public services, and from deployers of credit-scoring and life or health insurance pricing systems (Annex III points 5(b) and 5(c)). A private company hiring for itself is outside that list. A public employer, or a private body running a public employment service, is inside it.

What if we fine-tune the model on our own hiring history?

Then read Article 25 closely. Putting your name on the system, modifying it substantially, or changing its intended purpose can make you the provider of a high-risk system, with the full provider obligation set — conformity assessment, technical documentation and registration included.

Can we rely on the Article 6(3) exemption?

Sometimes, but never by assertion — and never by you, if you bought the tool. Article 6(3) takes a listed system out of high-risk where it performs a narrow procedural task, improves a previously completed human activity, detects decision patterns without replacing or influencing human assessment, or is purely preparatory. A system that profiles natural persons is never exempt. And Article 6(4) makes it the provider's call to document, before the system is placed on the market or put into service, with registration under Article 49(2) on top — an undocumented exemption is not an exemption, and a deployer's own opinion is not one either.

Check the law yourself

• Regulation (EU) 2024/1689 — consolidated text
• Regulation (EU) 2026/1744 — Digital Omnibus on AI