Artificial intelligence in occupational safety and health: what the EU AI Act regulates, what Spain’s draft OHS reform proposes and uses with human oversight
Artificial intelligence has entered occupational risk prevention through two doors at once. As a tool: accident analysis, inspection prioritisation, reading contractor documentation, assistants for OHS professionals. And as a risk: algorithmic management of shifts and pace, monitoring with wearables and automated decisions about people. Regulation (EU) 2024/1689 regulates the first side and the draft reform of Spain’s Law 31/1995, not yet in force, proposes addressing the second. This article separates them, without promises.
What the EU AI Act says about work
Regulation (EU) 2024/1689, the AI Act, covers in Annex III certain AI systems used in employment: recruitment, decisions affecting the employment relationship, task allocation based on individual characteristics or behaviour, and monitoring or performance evaluation. They are high-risk unless the limited Article 6(3) exception applies; systems that perform profiling are always high-risk. For them it requires risk management, data quality, technical documentation, activity logging, transparency, human oversight, and accuracy and robustness. It also prohibits emotion recognition in the workplace, except for medical or safety reasons, since 2 February 2025. The regulation applies generally from 2 August 2026; Sections 1 to 3 of Chapter III apply to Annex III systems from 2 December 2027 and to systems associated with Annex I products from 2 August 2028, following the amendment by Regulation (EU) 2026/1744. Add to this Directive (EU) 2024/2831, which establishes algorithmic-management safeguards specifically for platform work, including human oversight, and must be transposed by 2 December 2026.
What Spain’s draft OHS law reform proposes
The draft approved on 28 April 2026 does not establish a separate, detailed duty to assess every effect of algorithmic management. It proposes that technological change should not increase risks, particularly those arising from digitalisation; that assessments should be reviewed where digitalisation or new forms of work organisation create risks; and that violence and harassment should include conduct through algorithms or AI. These proposals are not in force. See our guide to the OHS law reform.
Wearables and monitoring devices: the INSST study in progress
On 6 July 2026 the INSST, Spain’s national OHS institute, announced that it was developing a study on monitoring devices used in occupational risk prevention. The notice invites participation and does not publish conclusions on processing purposes, impact assessments or disciplinary use; those matters must be assessed under the GDPR and employment law. As a general criterion, devices that measure exposure (noise, heat, postures, fatigue) require a lawful basis, a preventive purpose, information to the worker and data minimisation. We have covered sensor integration in the Bosch-Sabentis Intelligence Center.
Real uses of AI in OHS with human control
- Accident and incident analysis: grouping causes, detecting patterns by site and job and prioritising measures. The decision remains with the professional.
- Prioritising safety inspections: ordering what to review first according to history, changes and accident rates.
- Reading and validating contractor documentation: checking expiry dates and consistency in contractor coordination, with human review of exceptions.
- Assistants for OHS professionals: drafting plans, summarising regulations, preparing reports from system data, always verified.
- Trend analysis: identify aggregated signs of fatigue or overload from relevant, minimised data, with a lawful basis, human-defined thresholds, professional validation and no solely automated decisions.
Limits and conditions
- Real human oversight, not nominal: someone reviews and can reverse.
- Traceability: what data went in, what came out, who decided.
- Bias: verify that the model does not penalise groups (age, gender, disability).
- Data protection: lawful basis, minimisation, preventive purpose and a data-protection impact assessment where processing is likely to result in a high risk to rights and freedoms.
- Participation: information and consultation of worker representatives before deployment.
How Sabentis approaches it
Sabentis treats AI as a support layer for the OHS professional, not a replacement: proposals that the professional accepts, modifies or rejects, with a record of every decision. Any future Sabentis AI agent will follow the principles of human oversight, traceability and professional validation. You can read the context in Prevention 5.0 and data as the nervous system of the OSH management system.
Frequently asked questions
Is using AI to manage prevention a high-risk system?
It depends on purpose and effect. It may be high-risk if it falls within Annex III and Article 6(3) does not apply; working with aggregated data or performing a preparatory task is not automatically sufficient to exclude it. Transparency and oversight remain required.
Does AI have to be included in the risk assessment?
Current law requires all present risks to be assessed, including those arising from digitalisation where they exist. The draft proposes expressly addressing them but does not automatically make every AI system a psychosocial risk.
Can AI sign off a risk assessment?
No. AI does not replace competent OHS personnel or displace the employer’s responsibility. It may assist in preparation, but the assessment must be validated and adopted by those holding the relevant legal competence and responsibility.
