EU AI Act Rules for AI in Hiring and HR

Hiring software is not automatically high-risk just because it uses automation. Annex III covers systems intended for recruitment or selection, including targeted job advertisements, analysis and filtering of applications, and evaluation of candidates. It also covers systems used to make decisions affecting work-related terms, promotion, termination, task allocation based on individual behavior or personal traits, and monitoring or evaluation of workers. The purpose and deployment context matter: a general scheduling tool may be different from a system that scores workers for consequential decisions. Where a system falls in the high-risk category, the Act divides responsibilities. Providers must meet the applicable system requirements, including risk management, data governance, technical documentation, logging, transparency, human oversight, accuracy, robustness, and cybersecurity. Deployers must use it in accordance with instructions, assign competent oversight, monitor operation, and keep relevant logs under their control. Their use of output must also comply with employment, equality, privacy, and other applicable law. The AI Act does not itself authorize an employer to make a particular decision. The AI Omnibus, which entered into force on 27 July 2026, sets 2 December 2027 for Annex III high-risk obligations. The prohibition on emotion recognition in workplaces already applies from 2 February 2025, except where the Act’s medical or safety exception applies. This prohibition is separate from the high-risk classification: a prohibited practice is not made acceptable by adding a human reviewer or documenting it as high-risk. Before procurement, an employer should identify the exact task, affected workers, decision authority, vendor role, and any downstream action. Ask the provider for intended purpose, known limits, input requirements, and instructions. Test for disparate errors using lawful, representative evidence, define when a person can override or stop use, and provide a channel for correction or contest. Do not treat a model score as a neutral fact about a person.