AI in HR: which uses are high-risk and which obligations apply?
In brief
- Annex III of the AI Act classifies as high-risk AI systems used for recruitment (targeted publication of job adverts, screening and evaluation of applications) and for HR decisions (promotion, dismissal, task allocation, monitoring and evaluation of performance). The specific high-risk obligations will apply on 2 December 2027, the deadline set by Regulation (EU) 2026/1744, known as the "Digital Omnibus". But three frameworks already apply: the prohibited practices of article 5 (including emotion recognition at work), user training (article 4, an obligation of means to be documented) and the GDPR, in particular the regulation of wholly automated decisions (article 22) and the DPIA.
Which HR uses are classified as high-risk?
Annex III, point 4, covers two families:
- Recruitment and selection: systems intended to publish targeted job adverts, analyse and filter applications, evaluate candidates (automated tests, CV scoring, analysed video interviews).
- Decisions concerning the employment relationship: promotion or dismissal, task allocation based on individual behaviour or personality traits, monitoring and evaluation of performance and behaviour.
The criterion is not the tool but the use: the same generative assistant is minimal risk when it rephrases a job description, and prefigures high risk when it scores applications. This is the whole point of mapping actual uses (see the shadow AI method), because HR uses of AI are often invented on the ground, without validation.
What is already prohibited at work, since February 2025
Article 5 applies without waiting until 2027, with the highest penalty ceiling (€35 million or 7% of worldwide turnover):
- Emotion recognition at work (inferring the emotions of an employee or candidate from their biometric data: voice, face), except for medical or safety reasons.
- Social scoring leading to detrimental treatment that is disproportionate or out of context.
- Biometric categorisation based on sensitive data (origin, opinions, sexual orientation, etc.).
- Exploitation of vulnerabilities related to age, disability or social situation.
A video interview tool that claims to measure a candidate's "emotional engagement" via facial analysis falls within the scope of the prohibition: this is the first point to check with vendors.
What the GDPR already requires, without waiting for the AI Act
- Article 22: a candidate or employee may not be subject to a decision producing legal effects or significantly affecting them based solely on automated processing, except in strict exceptions. A rejection of an application decided by the algorithm alone, without real human intervention, is contestable; human intervention must be effective, not a rubber-stamping exercise.
- DPIA (article 35): the systematic evaluation of personal aspects of individuals in a position of dependence (employees, candidates) by automated processing is among the cases where an impact assessment is required; for an AI-based HR tool, it is in practice unavoidable.
- Transparency and minimisation: informing candidates and employees about the logic of the tool, limiting to data relevant to the position, retention periods (CNIL's reference framework on recruitment in France is authoritative).
- Works council consultation: the introduction of an AI tool affecting working conditions or enabling monitoring of activity falls under the information-consultation requirement (article L.2312-8 of the French Labour Code).
What will become mandatory on 2 December 2027
For each high-risk system deployed, the employer must in particular:
- Use the system in accordance with the provider's instructions and ensure the relevance of input data.
- Entrust human oversight to trained persons with the authority to override the machine.
- Inform employees and their representatives before putting into service, and inform candidates of the use of the system.
- Keep the logs generated by the system (at least six months).
- Suspend and report in the event of a risk or serious incident.
Public employers and certain bodies will also have to carry out a fundamental rights impact assessment (FRIA). Providers, for their part, will bear the bulk: risk management, data quality, documentation, CE marking, registration.
The six-point HR action plan for 2026–2027
- Inventory all AI uses affecting candidates and employees, including AI functions in HR SaaS already under contract (matching, scoring, analytics) often activated by default.
- Eliminate immediately any use falling under article 5 and require vendors to provide a written declaration on this point.
- Secure the present GDPR: DPIA, information, documented effective human intervention, retention periods, works council.
- Train HR teams (article 4) with a reinforced module: they are the company's most exposed users.
- Qualify contracts: who is provider, who is deployer, who bears which obligations at the 2027 deadline; avoid requalification as provider (substantial customisation, white label).
- Plan for 2027: logs, human oversight procedures, informing employee representatives: nothing insurmountable, everything simpler anticipated than improvised.
Take action
Manage your compliance with Data Comply One: all-in-one software and outsourced DPO/CISO experts.
Discover the AI Act software →Test my AI compliance →Sources
- Regulation (EU) 2024/1689 (AI Act), articles 4, 5, 26, 27 and Annex III (point 4), EUR-Lex; Regulation (EU) 2026/1744 (postponement of high-risk deadlines).
- Regulation (EU) 2016/679 (GDPR), articles 13, 22 and 35, EUR-Lex.
- CNIL (France's data protection authority), reference framework on personnel management processing and recommendations on recruitment; French Labour Code, article L.2312-8, Légifrance.