AI Act for recruitment: the Annex III 4 path, step by step
Recruitment is the most common way a normal company becomes a high-risk AI operator. Not because the software is exotic, but because Annex III point 4 names hiring explicitly: targeted job ads, filtering applications, evaluating candidates.
If you use a tool that ranks CVs, scores interviews, drafts rejection reasons or decides who gets an interview slot, you are probably in scope. The test is not whether the software calls itself AI. The test is whether its output influences a decision about a person's access to employment.
What Annex III point 4 covers
| Use case | In scope? | Why |
|---|---|---|
| Targeted job advertisements | Yes | Explicitly named in Annex III 4(a) |
| CV parsing and filtering | Yes | Analysing and filtering applications |
| Interview scoring or ranking | Yes | Evaluating candidates |
| Chatbot answering FAQ on a careers page | No (limited risk) | Article 50 disclosure applies |
| Shift allocation and task assignment | Yes | Worker management, Annex III 4(b) |
| Performance scoring and promotion lists | Yes | Worker management |
| Automated termination recommendations | Yes | Explicitly named |
| Payroll or HR admin without decisions about people | Usually no | Not an Annex III function |
The tempting escape is Article 6(3): a system that only performs a narrow procedural task is not high-risk. It almost never works for recruitment, because the tool typically ranks or filters — which is an assessment, not a formatting step. And the exception is unavailable where the system performs profiling.
Who is responsible for what
- The vendor (provider) owes the technical file, the quality management system, the conformity assessment, the CE marking and the registration.
- You (deployer) owe oversight, relevant input data, monitoring, logs for at least six months, information to workers and candidates, and cooperation with authorities.
- If you configure the vendor's model heavily, or put your own brand on it, you may become a provider for that configuration. Read the contract before assuming you are only a deployer.
The obligations that actually bite
| Obligation | In recruitment terms |
|---|---|
| Human oversight (Art. 14) | A named recruiter reviews every rejection; the model cannot send a rejection alone |
| Record-keeping (Art. 12) | Who screened which candidate, with which model version, on which data, and who approved |
| Data governance (Art. 10) | Training data representative of the roles and languages you hire for; documented quality checks |
| Bias testing (Art. 10(2)(f), Art. 15) | Metrics by gender, age, nationality and disability proxies, before and after deployment |
| Transparency (Art. 13, Art. 26(11)) | Candidates informed that AI is used and what it does; instructions for use available |
| Worker information (Art. 26(7)) | Workers' representatives informed before a worker-management system goes live |
| Monitoring and incidents (Art. 72–73) | A named owner, a review cadence and a way to report serious incidents |
Automated rejection is the single highest-risk pattern in recruitment AI: it is a decision about a person, taken without oversight, often at scale, and it is trivially discoverable in litigation. Make rejection a human action, always.
GDPR runs in parallel, not instead
The AI Act does not replace the GDPR. Recruitment data is personal data, often special category data when health or disability comes up. A DPIA is effectively mandatory for this kind of processing, and Article 22 GDPR restricts solely automated decisions with legal or similarly significant effects. The AI Act's Article 14 oversight requirement and the GDPR's human-intervention right point the same way: a person must be able to look at the case.
A 30-day implementation path
- 1Week 1 — inventory. List every tool that touches hiring or worker management, what it decides and which vendor runs it. Include the ATS plugins nobody remembers buying.
- 2Week 2 — classify and contract. Confirm Annex III 4 for each, get the vendor's instructions for use and technical documentation, and check whether the contract makes you a deployer or a provider.
- 3Week 3 — instrument. Turn on event logging with model version and data categories, move approvals out of email, and set the retention rule (six months minimum, longer for employment claims).
- 4Week 4 — evidence. Document the oversight procedure with names, run a first bias check, issue the first evidence pack for the period, and fix what it exposes.
Twelve questions, no account. If the answer is Annex III 4, you get the exact control list and the three things to fix first.
Run the free risk checkTimeline: December 2027, but buyers move first
Regulation (EU) 2026/1744 moved the Chapter III deadline for Annex III systems to 2 December 2027. Enterprise customers, however, already put AI Act questions into procurement and vendor-risk questionnaires, and a vendor without a technical file loses points today. Article 50 transparency for any candidate-facing chatbot has applied since 2 August 2026.
There is also a market reason to move early: the same evidence file answers the buyer's security questionnaire, the candidate's transparency request and the auditor's questions. One artefact, three audiences.
Frequently asked
We only use ChatGPT to draft job descriptions. Is that high-risk?
Drafting a job description for a human to edit is not an Annex III function. The risk starts when the output sorts, scores or selects people, or when it targets the advertisement at specific groups.
Our ATS vendor says they are compliant. Are we covered?
No. As a deployer you still owe oversight, log retention, worker and candidate information, monitoring and incident handling. Ask for their instructions for use and technical documentation, and keep your own records.
Do we need to tell candidates that AI is involved?
Article 26(11) requires deployers of high-risk systems to inform natural persons that they are subject to the use of the system, where the decision produces legal effects or similarly significantly affects them. GDPR transparency duties apply on top.
What about bias testing if we did not build the model?
You still need evidence that the system performs acceptably on your population. Ask the vendor for validation results, then run your own outcome monitoring by group — selection rates, interview conversion, offer rates.
Sources
Keep reading
Article 14 in practice: who counts as an overseer, what they must be able to do, the automation bias trap, and the records that prove oversight happened.
Which AI systems Annex III makes high-risk, how the Article 6(3) filter works, the full obligation checklist under Articles 9–15, 17, 43 and 72, and the timeline after the AI Digital Omnibus.