AI in hiring: the EU AI Act rules for recruitment and HR

Hiring is where the EU AI Act lands on ordinary companies. You may neverbuild a model, but if your applicant tracking system ranks candidates, your assessment platform scores them, or your HR analytics evaluates performance, you are deploying Annex III high-risk AI — with obligations applying from 2 December 2027 under the 2026 Omnibus, in the €15M / 3% penalty tier (workplace emotion inference is banned already). Annex III point 4 is the single most widely triggered category in the Act, and the one companies most often discover through their vendor rather than their compliance register.

Which HR systems are in scope

The derogation that usually does not apply

Article 6(3) lets a provider classify an Annex III system out of high-risk when it performs only a narrow procedural task, improves the result of a previously completed human activity, detects decision patterns without replacing human assessment, or performs a preparatory task. Vendors reach for it often. It carries a hard limit: a system is always high-risk where it performs profiling of natural persons. Scoring, ranking or predicting the performance of candidates is profiling, so the derogation is closed for the core of recruitment AI. Where a provider does claim it, they must document the assessment before market placement and still register the system under Article 49(2) — ask to see that document.

Provider or deployer — and how employers become both

Buy an ATS and use it as sold: you are the deployer, the vendor is the provider. Three common moves flip you into the provider seat with the full Articles 9–15 stack, per Article 25: putting your own name or brand on the system, substantially modifying it, or repurposing it into a high-risk use the provider did not intend. Fine-tuning a model on your own hiring history is the version of this that HR teams walk into without noticing — a screening model trained on your data is, for practical purposes, a system you provide to yourself.

The employer duty stack

What to ask your ATS vendor this week

Your Article 26 duties depend on information only the provider holds; their Article 13 duty is to give it to you. Send these eight questions as they are — the quality of the answers tells you as much as the answers.

1. Do you classify this product as a high-risk AI system under Annex III point 4 — and if not, on what documented basis?
2. Can we see the Article 13 instructions for use: intended purpose, known limitations, accuracy metrics and required oversight measures?
3. Is there an EU declaration of conformity and CE marking, and is the system registered in the EU database under Article 49?
4. What accuracy and bias metrics do you publish, measured on which population, and how often are they re-measured (Arts. 10, 15)?
5. What does the system log automatically, and how do we export and retain those logs for six months (Arts. 12, 26(6))?
6. Which controls let our recruiters override or stop a ranking, and how is that recorded (Art. 14)?
7. Which of our configuration or fine-tuning options would make us the provider under Article 25?
8. What is the serious-incident channel and your reporting timeline (Art. 73)?

The GDPR layer on top

Hiring AI is one of the clearest cases of two regimes on one system. Article 22 GDPR governs solely automated decisions with legal or similarly significant effects — a rejection qualifies — and a DPIA is very likely required for systematic candidate evaluation. National employment law adds its own layer: works-council co-determination over monitoring technology, and Article 88 GDPR opening the door to stricter national rules in the employment context. Passing one of these regimes tells you nothing about the others; the AI Act vs GDPR guide maps where they overlap and where they do not.

A defensible 30-day position

• Inventory every tool that touches hiring, promotion, task allocation or performance — including features inside HR suites you did not buy separately.
• Kill any workplace emotion-inference feature immediately and record the date you did it.
• Classify each remaining tool and write down the reasoning, including any derogation the vendor claims.
• Send the vendor questions above; file the answers with the classification.
• Name the oversight owners, train them, and record the training.
• Notify workers’ representatives and update candidate-facing notices; turn on log retention.

None of this requires a consultancy engagement — it requires an inventory, a defensible classification per system, and a dated record of both. That is the same sequence the compliance checklist applies company-wide, and the deployer guide covers for the non-HR systems you will find along the way.