HR and Recruitment AI Under the EU AI Act: A Use-Case Map
HR and recruitment AI mapped to the EU AI Act: prohibited (Art 5), high-risk Annex III point 4, and minimal-risk — deployer duties and 2 Dec 2027 deadline.
The EU AI Act does not treat HR as a single category. It draws three distinct lines across the sector: some tools are outright prohibited, a large cluster of tools that influence hiring and employment decisions are high-risk, and routine administrative software sits at minimal risk. Understanding which line your specific tool crosses is the first compliance task — and for HR, the stakes for getting it wrong are among the highest in the Act.
This page maps HR and recruitment AI systems to their risk classification under Regulation (EU) 2024/1689. For the deployer obligations that apply once you have classified a system — human oversight, worker notification, logging — see the EU AI Act for HR teams governance guide.
The Three Zones in HR and Recruitment AI
Zone 1: Prohibited — Emotion Recognition in the Workplace
Article 5(1)(f) prohibits AI systems that infer the emotional states of natural persons in workplace settings. This is not a high-risk classification with additional paperwork requirements. It is a ban, in force since 2 February 2025.
The prohibition covers any system that reads facial expressions, analyses vocal patterns, or interprets physiological signals to make inferences about how an employee or candidate is feeling. It extends to educational institutions by the same provision.
In HR practice, this catches more tools than compliance teams often expect:
- Affect-analysis features in video interview software. Several video interview tools include scoring modules that claim to assess a candidate's "communication style," "energy," or "engagement" from facial movement and vocal tone during the recorded interview. Where that analysis produces an inference about emotional state — including proxies for emotional state — it falls under Article 5(1)(f).
- Employee sentiment monitoring via audio or video. Tools that analyse meeting recordings, team-call video feeds, or in-office cameras to assess workforce mood, stress levels, or "engagement signals" are prohibited when the analysis reaches the level of emotional inference. Marketing language that avoids the word "emotion" does not change the legal classification.
- Productivity monitoring with affect signals. Some productivity-monitoring tools include features that flag anomalous patterns of behaviour as potential indicators of disengagement or distress. Where these involve actual emotion inference from physiological or audiovisual signals, Article 5(1)(f) applies.
The penalty for deploying a prohibited system — or continuing to operate one already live — is up to €35,000,000 or 7% of total worldwide annual turnover, whichever is higher (Article 99(3)). For SMEs and start-ups, Article 99(6) caps the fine at the lower of the percentage or the fixed sum, but that is a ceiling reduction, not a safe harbour.
If you are currently running any tool marketed as detecting employee sentiment, engagement, or emotional state from audio or video, audit it against this prohibition now. The obligation is not future-dated.
Zone 2: High-Risk — Annex III, Point 4
Article 6 sets the classification rules. A system that falls within one of the eight areas listed in Annex III is presumptively high-risk. For HR and recruitment, the relevant area is Annex III, point 4: employment, workers management, and access to self-employment.
Point 4 has two sub-categories.
Point 4(a) — Recruitment and Selection AI
Point 4(a) covers AI systems used in the process of recruiting or selecting natural persons. The scope is deliberately broad and covers the whole pipeline from reaching candidates to deciding who advances.
Targeted job advertising based on profiling. An AI system that determines which individuals receive a job advertisement — where that targeting relies on profiling from personal data — falls within point 4(a). A tool that simply places an ad on a general job board does not; a tool that selects which individuals within a job board's user base should see the ad based on inferred characteristics does.
CV parsing and application filtering. Systems that extract structured data from CVs and filter applications against job-requirement criteria are high-risk when they produce ranked or filtered outputs that a human then acts on. The system does not need to make the final hiring decision — it is enough that it influences which candidates advance. A pure keyword-search function without scoring or ranking sits closer to a database query than an AI system; a machine-learning model that scores candidate-fit and generates a shortlist is squarely within scope.
Candidate ranking and scoring. Any model that assigns a score, ranking, or probability estimate to candidates — whether labelled "fit score," "match score," "predictive hire quality," or similar — is high-risk under point 4(a). The label on the output does not change the classification.
Interview analysis and scoring. AI systems that evaluate candidate responses in recorded or live interviews and produce assessments — of communication quality, answer relevance, competency indicators, or similar — are high-risk. This includes text-analysis of interview transcripts and structured-interview scoring engines that do not involve emotional inference (those that do are prohibited under Article 5(1)(f) as described above).
Point 4(b) — Employment Management AI
Point 4(b) covers AI systems used to make or support decisions about the in-employment lives of workers. The Act specifically names promotion, termination, task allocation, and performance monitoring.
Promotion and career-progression AI. Systems that identify employees for promotion, flag candidates for accelerated development programmes, or rank internal applicants for open roles are high-risk where the AI output materially influences the decision. A system that surfaces a list of "high-potential" employees from performance and behavioural data is a point 4(b) system.
Termination and performance-management support. AI tools that flag underperformers, generate recommendations for performance-improvement plans, or score employees against termination criteria fall within point 4(b). The link between the AI output and the consequential employment decision — even where the human manager makes the final call — is enough to bring the system within scope.
Task allocation driven by monitoring of personal traits or behaviour. Workforce-management systems that assign work tasks, shift schedules, or workload levels based on AI analysis of individual performance data, past behaviour, or predicted capacity are high-risk. Warehouse management and logistics systems that set individual pick-rate targets using algorithmic productivity data are textbook point 4(b) systems.
Employee performance monitoring and evaluation. Broader performance-management AI — systems that aggregate productivity signals, attendance patterns, system-usage data, and behavioural indicators to produce a performance score or evaluation input — fall within point 4(b) where they influence consequential decisions. The monitoring system need not itself make the employment decision; if it feeds into one, it is in scope.
The Article 6(3) Filter
A system that appears to fall within Annex III point 4 is nonetheless not high-risk if it does not pose a significant risk of harm to health, safety, or fundamental rights. Article 6(3) identifies four situations where the exemption can apply: the system performs a narrow procedural task; it improves the result of a previously completed human activity; it detects decision patterns without replacing or influencing human assessment; or it performs preparatory work that has no direct significant effect on persons.
In practice, this exemption is a narrow escape in HR. Any system that profiles natural persons is always high-risk, regardless of the Article 6(3) conditions. Most CV-ranking, candidate-scoring, and performance-monitoring tools profile individuals by definition. Providers claiming the exemption must document their assessment and register the system (Article 49).
The genuine exemption cases in HR are narrow: a formatting tool that restructures CV text for readability without scoring or ranking candidates, or an analytics dashboard that reports aggregate departmental statistics without generating individual assessments, may plausibly pass the filter. These cases should still be documented.
Zone 3: Minimal Risk — Routine HR Administration
Most AI functionality in HR administration falls outside the high-risk perimeter: payroll calculation, shift scheduling based on declared availability, policy Q&A chatbots, and benefits-election tools do not profile individuals for consequential employment decisions.
Two qualifications apply. First, a scheduling tool that uses AI analysis of individual behavioural data to make allocations crosses into point 4(b); the distinguishing test is whether the system analyses personal traits, not just stated preferences. Second, a chatbot that attempts emotional inference from employee responses re-engages the Article 5(1)(f) prohibition — the minimal-risk label does not protect that function.
Minimal-risk classification does not mean no obligations: Article 4 AI literacy obligations have applied since 2 February 2025, and Article 50 transparency duties apply from 2 August 2026 for any customer-facing generative AI interaction.
The GDPR Article 22 Overlap
GDPR Article 22 gives individuals the right not to be subject to decisions based solely on automated processing that produce significant legal or similarly significant effects — including employment decisions. It runs alongside the EU AI Act, not instead of it.
For recruitment and employment-management AI, a genuine human review step that can override the AI output satisfies both Article 14 of the EU AI Act (human oversight) and the GDPR Article 22 safeguard simultaneously. Document the oversight process under Article 14 and the Article 22 basis in the same record. Employers must be able to show the AI output was not the sole basis for any consequential decision.
Provider and Deployer Obligations at a Glance
| Role | Who | Core obligation set |
|---|---|---|
| Provider | The HR-tech vendor building the CV parser, interview scorer, or performance-monitoring tool | Articles 9–16 stack: risk management (Art 9), data governance (Art 10), technical documentation (Art 11/Annex IV), deployer instructions (Art 13), human oversight by design (Art 14), subgroup accuracy (Art 15), conformity assessment (Art 43), DoC (Art 47), CE marking (Art 48), registration (Art 49) |
| Deployer | The employer using the vendor's tool | Article 26: follow provider instructions; implement human oversight (Art 14); keep logs ≥ 6 months; notify workers' representatives before deployment; monitor and flag serious incidents to the provider |
Two deployer obligations are frequently overlooked.
Worker-representative notification (Article 26). Before deploying a high-risk AI system in the workplace, you must inform workers' representatives — your works council, staff committee, or equivalent co-determination body. This is a pre-condition for deployment. In Germany, Austria, the Netherlands, and other co-determination jurisdictions, national works-council rights may additionally require consultation or consent beyond the Article 26 baseline.
Informing individuals subject to AI evaluation (Article 26). Where a high-risk system evaluates individuals, the deployer must inform those individuals. For recruitment AI, candidates must know AI is being used to assess them.
The Conformity Assessment Route
For Annex III point 4 systems, the conformity assessment follows the Annex VI internal self-assessment route — no notified body is required. This differs from Annex III point 1 (biometrics), where the Annex VII notified-body route generally applies.
The provider conducts the assessment against Articles 9–15, documents the outcome in accordance with Annex IV (the nine technical-documentation sections), draws up the Article 47 EU Declaration of Conformity, and registers the system in the EU database under Article 49 before placing it on the EU market.
Does a Private Employer Need a FRIA?
Article 27 requires certain deployers to complete a Fundamental Rights Impact Assessment before deploying a high-risk AI system. The obligation applies to:
- Public bodies
- Private deployers of creditworthiness-scoring systems (Annex III, point 5(b))
- Private deployers of health/life-insurance risk-assessment systems (Annex III, point 5(c))
A private employer deploying Annex III point 4 AI — recruitment tools, performance monitoring, promotion support — does not automatically owe a FRIA under Article 27. This is a common misstatement. Article 27 is precise: employment AI (point 4) is not on the list for private-sector deployers.
A public-body HR department — a city council, a public hospital, a national regulator — does trigger Article 27 because of its public-body status, not because of the AI system's category.
The Deadline
High-risk obligations under Annex III, point 4 apply from 2 December 2027 for stand-alone systems. The Digital Omnibus, agreed between Parliament and Council in May 2026, pushed back the original 2 August 2026 deadline. That is useful preparation time — assembling Annex IV technical documentation, running bias testing across demographic subgroups, establishing six-month log-retention processes, and working through works-council notification procedures takes longer than most companies expect.
The Article 5(1)(f) emotion-recognition prohibition is not deferred. It applies now, and has done since 2 February 2025.
How Confir Helps
Most employers deploying third-party HR tools are deployers with a narrower obligation set than the vendor. The classification step determines what applies.
Confir's rule-based classification engine works through the Annex III point 4 questions in plain English: does the system rank or filter candidates? does it produce individual performance scores? does it influence promotion or termination decisions? The output is a determined risk tier — prohibited, high-risk, or minimal — with the obligation set scoped to your role.
For a deployer, the resulting scope is typically two or three controls: human-oversight verification (Article 14), log-retention setup (Article 26), and the worker-representative notification workflow (Article 26). Confir generates the documentation for each. For providers building HR AI, the full Article 9–15 stack runs through the same workflow, ending in the Annex IV technical documentation pack and the Article 47 Declaration of Conformity.
The classification is deterministic — same intake, same finding, the rule that fired is human-readable — which matters when you need to show a market surveillance authority how you reached your compliance conclusion.
Frequently Asked Questions
Which HR AI tools are always prohibited, regardless of safeguards?
AI systems that infer the emotional states of workers or job candidates from audiovisual signals — facial expressions, vocal tone, physiological data — in a workplace or educational context are prohibited under Article 5(1)(f). No technical safeguard or consent mechanism makes them lawful. The ban applies to "emotion AI" features in video interview software, mood-monitoring tools, and any employee-engagement product that reaches the level of emotional inference from audio or video. The prohibition has been in force since 2 February 2025. Fines reach €35,000,000 or 7% of worldwide turnover (Article 99(3)).
Is a CV-screening tool always high-risk?
If it produces a ranked or filtered list of candidates that influences who advances in a hiring process, yes — it falls within Annex III point 4(a) and is high-risk. The Article 6(3) exemption is theoretically available but practically narrow: any tool that profiles individual candidates (which scoring and ranking tools do by definition) is always high-risk regardless of that filter. A pure keyword-search function with no scoring or ranking may not qualify as an AI system at all under the Act's definition, but that classification should be documented and not assumed.
We use an AI tool to monitor warehouse workers' productivity. Is that high-risk?
Almost certainly yes. Workforce-management systems that use AI to set individual productivity targets, monitor workers against those targets, or allocate tasks based on analysis of individual performance data fall within Annex III point 4(b). The system does not need to make a termination decision itself — the influence over task allocation and performance assessment is enough. Your obligations as the employer deploying the tool are those of a deployer under Article 26, including the worker-representative notification requirement before the system goes live.
Do we need to tell job candidates that AI is evaluating them?
Yes. Under Article 26, deployers of high-risk AI systems must inform the individuals who are subject to those systems. For recruitment AI, this means candidates must know that automated evaluation is being used in the hiring process. The specific information to be provided is set out in the provider's instructions for use (Article 13). This obligation sits alongside GDPR transparency requirements, which have applied independently since 2018.
Does our HR-tech vendor's CE marking mean we have no compliance work to do?
No. CE marking (Article 48) and the EU Declaration of Conformity (Article 47) are the provider's obligations. As the employer deploying the tool, you are the deployer under Article 26 with independent duties: follow the provider's instructions, implement genuine human oversight, retain logs, notify your works council before deployment, and monitor for problems. Request the provider's instructions for use and Declaration of Conformity as evidence of their compliance — and document your own deployer obligations separately.
When exactly do the high-risk obligations apply to our recruitment tools?
Under the Digital Omnibus agreed in May 2026, stand-alone high-risk Annex III systems must comply by 2 December 2027. If your recruitment AI is embedded in a product regulated under EU product safety law (Annex I), the date is 2 August 2028. The prohibited-practices obligation (Article 5, including the emotion-recognition ban) already applies — since 2 February 2025. Use the time to 2027 for real preparation: audit your AI inventory, complete classification analysis for each tool, begin the works-council notification process, and verify your vendors' compliance documentation.
Related guides
- EU AI Act for HR teams: governance obligations
- Article 9 risk management system
- CV screening high-risk classification
- promotion and termination decisions
- Article 6 high-risk designation
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 7-day free trial.
Start free trial →