EU AI Act  /  Annex III · Point 4
Annex III · Employment & HR

Is your HR AI high-risk? Annex III point 4, decoded

CV screening, candidate scoring, targeted job ads, performance monitoring — employment is one of the eight Annex III high-risk areas. Here is the use-case-by-use-case answer: what is high-risk, what escapes, and what is banned outright.

Short answer

If an AI system decides or materially influences who gets recruited, hired, promoted, assigned work, monitored, evaluated or let go, it is high-risk under Annex III point 4 — whether you built it or bought it. Narrow helper tasks (parsing CVs into fields, detecting duplicate applications) can escape via Article 6(3), but never if the system profiles people, and the escape must be documented. One HR use case is not high-risk at all — because it is prohibited: emotion recognition at work (Article 5(1)(f)).

Employment AI sits in Annex III because hiring and workforce decisions shape people’s livelihoods — recital 57 of the Act says these systems have an “appreciable impact on future career prospects, livelihoods of those persons and workers’ rights”, and can perpetuate historical patterns of discrimination. The classification does not care whether your tool is called an ATS, a copilot or an agent: it looks at the intended use.

What Annex III point 4 actually covers

Point 4 of Annex III (“Employment, workers management and access to self-employment”) has two limbs. The first covers getting people into jobs:

The second covers what happens once they are working — including gig and platform work, which recital 57 explicitly pulls in:

Note the drafting: “intended to be used”. It is the purpose the system serves in your process that classifies it — a general-purpose tool wired into recruitment decisions is caught the same way as a dedicated HR product. And under Article 6(2), being on this list is the default route to high-risk; the burden is on showing an exception applies, not on proving the system is in scope.

Use-case verdicts: your HR AI, case by case

Verdicts below follow the text of the Regulation (Annex III point 4, Article 5(1)(f), Article 6(3) with recital 53) and the Commission’s draft classification guidelines — see the note under the table. Where a case genuinely depends on configuration, it is marked borderline rather than forced into a bucket.

HR use caseVerdictWhy
CV / application screening and filteringHigh-riskPoint 4(a): “analyse and filter job applications” — named in the Annex itself
Candidate scoring or ranking (ATS ranking, assessment platforms)High-riskPoint 4(a): “evaluate candidates”; scoring people on their traits is also profiling — no Art 6(3) escape
Targeted job advertisementsHigh-riskPoint 4(a) names “targeted job advertisements” explicitly
AI video-interview analysisHigh-riskPoint 4(a) (evaluating candidates). If it infers emotions, check Article 5(1)(f) first — the ban outranks the high-risk tier
Promotion / termination decision supportHigh-riskPoint 4(b): decisions on “promotion or termination of work-related contractual relationships”
Task or shift allocation based on behaviour or personal traits (incl. platform-work dispatch)High-riskPoint 4(b); recital 57 explicitly includes persons providing services through platforms
Performance monitoring and evaluationHigh-riskPoint 4(b): “monitor and evaluate the performance and behaviour” of workers
Emotion recognition of employees (call-centre sentiment monitoring, “engagement” analytics on workers)ProhibitedArticle 5(1)(f) bans inferring emotions in the workplace, except for medical or safety reasons — in force since Feb 2025
CV parsing into structured fields (no scoring or filtering)Borderline — assess under Art 6(3)Transforming unstructured data into structured data is a recital 53 example of a “narrow procedural task” — provided it does not materially influence the decision and does not profile; document the assessment (Art 6(4))
Duplicate-application detectionBorderline — assess under Art 6(3)Detecting “duplicates among a large number of applications” is a verbatim recital 53 example of a narrow procedural task; still document under Art 6(4)
Polishing the language of job postings or HR documentsNot high-riskImproving the language of previously drafted documents is a recital 53 example under Art 6(3)(b) — the human decision already happened
Interview scheduling bots, HR-policy chatbotsNot high-riskNo point 4 use case — no decision about a person’s recruitment, terms or evaluation. Chatbot disclosure duties under Article 50 may still apply

One tool, several use cases. Classification attaches to the use, not the product name. A single “talent platform” can contain a prohibited emotion-analytics feature, a high-risk ranking engine and a harmless scheduling bot at once. The Commission’s draft classification guidelines take an anti-circumvention line: a combined tool is assessed as a whole, and splitting an evaluative system into “procedural” slices does not launder it out of high-risk. Classify per use case — and record each verdict.

The Article 6(3) escape route — and the profiling trap

Annex III listing is the default, not the final word. Under Article 6(3), a point 4 system is not high-risk if it does not pose a significant risk of harm to health, safety or fundamental rights — “including by not materially influencing the outcome of decision making” — and meets at least one of four conditions:

  1. it performs a narrow procedural task (recital 53: structuring unstructured data, classifying incoming documents, detecting duplicate applications);
  2. it improves the result of a previously completed human activity (e.g. polishing language of a drafted document);
  3. it detects decision-making patterns or deviations from them, without replacing or influencing a completed human assessment absent proper review; or
  4. it performs a preparatory task to an assessment (indexing, search, translation, file handling).

The profiling trap. None of the four conditions helps if the system profiles natural persons: “an AI system referred to in Annex III shall always be considered to be high-risk where the AI system performs profiling of natural persons” (Art 6(3), final subparagraph). Most candidate scoring, fit prediction and behaviour-based task allocation is profiling — automated evaluation of personal aspects like work performance, reliability or behaviour. This is where HR tools most often get misclassified as exempt.

Claiming the exception is a documented act, not an opinion. Under Article 6(4), a provider that considers its Annex III system not high-risk must document that assessment before the system is placed on the market or put into service, register the system in the EU database under Article 49(2), and hand the documentation to national authorities on request.

On guidance: Article 6(5) required the Commission to publish classification guidelines with practical examples by 2 February 2026. That deadline was missed — draft guidelines were published on 19 May 2026, with a consultation running to 23 July 2026, and no final version had been adopted as of early July 2026. Directionally, the draft confirms the reading above: HR systems that analyse, filter, score or rank candidates are typically high-risk; purely procedural steps like CV data structuring can qualify for the derogation. Treat the draft as a signal, not settled law.

If you are high-risk: what you must do

Most companies do not build their HR AI — they buy it. That makes you a deployer, and your duties live in Article 26:

And one duty specific to the workplace — before the system ever runs:

Fundamental rights impact assessment (FRIA): here private employers get relief most summaries miss. Article 27(1) requires a FRIA only from deployers that are bodies governed by public law, private entities providing public services (education, healthcare, social services, housing, administration of justice — recital 96), and deployers of the credit-scoring and life/health-insurance systems in Annex III points 5(b) and 5(c). Deploying a point 4 HR system does not by itself trigger a FRIA for a private company — even a bank’s FRIA duty attaches to its scoring and pricing models, not its recruiting tool. Public-sector employers, however, are in scope for their HR AI.

If you build or resell HR AI (or substantially modify a bought system, or put your name on it — Article 25), you carry the provider stack instead: risk management system (Art 9), technical documentation, transparency to your customers (Art 13), human-oversight design (Art 14), conformity assessment and registration.

When this applies

The Regulation as enacted set 2 August 2026 for the Annex III obligations. The Digital Omnibus on AI — adopted by the European Parliament on 16 June 2026 and by the Council on 29 June 2026, but as of early July 2026 still awaiting publication in the Official Journal — defers the Annex III high-risk obligations to 2 December 2027. Two things do not move: the Article 5 prohibitions (in force since February 2025, workplace emotion recognition included) and the practical reality that classification, vendor evidence and worker-information processes take longer than a compliance quarter.

Classifying your HR AI portfolio

The unit of compliance here is the use case, and most organisations have more of them than they think: the ATS ranking module, the sourcing tool’s ad targeting, the survey platform’s sentiment add-on, the workforce-management system’s scheduling optimiser. Each needs a recorded verdict — prohibited screen, Annex III area, Article 6(3) claim and its rationale — that survives the next feature release and the next auditor.

Track this in your Jira

Give every HR AI use case a classification record

Model Inventory for Jira turns each AI system into a work item in the Jira your team already uses, with a built-in EU AI Act category field and dynamic risk tiering. Record the Annex III point 4 verdict, the Article 6(3) rationale, and who signed it off — with an immutable change history, so when the vendor ships an “AI-powered insights” update, the record shows the classification needs a re-check. The legal judgement stays with your compliance team; the inventory makes sure no tool skips the question.

See how it works

FAQ

Is CV screening high-risk under the EU AI Act?

Yes, in almost all configurations. Annex III point 4(a) names analysing and filtering job applications and evaluating candidates. The Article 6(3) exception realistically covers only procedural steps (parsing, duplicate detection) — and never systems that profile people.

Is AI monitoring of employee performance high-risk?

Yes — point 4(b) covers monitoring and evaluating the performance and behaviour of workers, and decisions on promotion, termination, terms and task allocation. If the monitoring infers emotions, it is not high-risk but prohibited (Art 5(1)(f)), unless it serves medical or safety purposes.

Is emotion recognition in the workplace banned?

Yes, since 2 February 2025, with the Act’s top fine tier (€35M / 7% of turnover). The Commission’s guidelines on prohibited practices (C(2025) 884, February 2025) clarify the edges: inferring physical states like fatigue for accident prevention is not emotion recognition, and the “safety” exception protects life and health — not property or productivity.

Do private employers need a FRIA for HR AI?

Usually no. Article 27 covers public bodies, private entities providing public services, and deployers of Annex III 5(b)/(c) credit and insurance systems. A private employer’s point 4 HR system does not itself trigger a FRIA — the Article 26 deployer duties, including informing workers under 26(7), still apply in full.

When do these rules start to apply?

Prohibitions: since February 2025. Annex III high-risk obligations: 2 August 2026 as enacted, deferred to 2 December 2027 by the Digital Omnibus adopted in June 2026 — which was still awaiting Official Journal publication as of early July 2026.

This page is a practical explanation, not legal advice. Always confirm classification against the official text of Regulation (EU) 2024/1689 and, where the stakes warrant it, qualified counsel.