High-risk AI classification under the EU AI Act
Whether your AI system is classed as high-risk determines how heavy the obligations are that apply to you. The EU AI Act takes a risk-based approach: the greater the impact on people, the stricter the requirements. The European AI Office coordinates enforcement and published classification guidelines in 2026. On this page, you work through the categories, the concrete examples from Annex III and a practical self-check.
If you do not yet know which AI systems your organisation has, start with the AI inventory. If you already have an overview, this page helps you decide which systems deserve extra attention in your AI risk management.
Two routes to high-risk
The AI Act has two routes by which an AI system is classed as high-risk.
Route 1: Annex I (regulated products). If AI is part of a product that already falls under EU product safety rules (medical devices, vehicles, aviation systems, machinery), it is high-risk when it is itself a safety component. Think of an AI system that controls braking force in a car, or one that provides diagnostic support in a medical device.
Route 2: Annex III (specific areas of use). Regardless of the product, AI is high-risk if it falls within one of the eight categories below. The exception is a system that can demonstrably show it poses no significant risk to people.
The European Commission has published classification guidelines in 2026 that help you assess the “no significant risk” exception. Use them as additional guidance.
Annex III: the eight categories
1. Biometrics
AI that identifies people, categorises them or recognises emotions based on biometric characteristics.
| Example | Risk level |
|---|---|
| Real-time facial recognition in public spaces by law enforcement | Prohibited (with limited exceptions) |
| Facial recognition in an office for access control | High-risk |
| Emotion recognition in the workplace or in education | Prohibited |
| Biometric categorisation based on sensitive characteristics (religion, ethnicity) | Prohibited |
2. Critical infrastructure
AI used as a safety component in managing critical digital infrastructure, traffic networks, or energy and water supply.
Examples: AI that manages electricity grids, controls traffic lights or monitors water purification.
3. Education and vocational training
AI that determines access to education or has a significant influence on someone’s future.
Examples:
- Admissions software for universities of applied sciences and universities
- AI that assesses or ranks students on their performance
- Systems that detect prohibited behaviour during exams
4. Recruitment and workforce management
AI in HR is one of the most common high-risk categories for businesses.
| Application | High-risk? |
|---|---|
| CV screening and candidate ranking | Yes |
| Analysing job interviews (language, facial expressions) | Yes |
| Assessing employee performance | Yes |
| Recommendations for promotion or dismissal | Yes |
| Targeted job advertisements based on a profile | Yes |
| Generating work rosters automatically (not an HR decision) | No (limited risk) |
If you use an AI system that contributes to decisions on recruitment, promotion or dismissal, you most likely fall under high-risk. This also applies to purchased HR software with AI features.
5. Essential services and credit
AI that controls access to basic public services.
Examples:
- Credit scoring systems for loans and mortgages
- AI-driven assessment of benefit applications
- Risk classification for insurance
- Assessment of applications for social housing
6. Law enforcement
AI used by the police, the Openbaar Ministerie (OM, Public Prosecution Service) or other enforcement bodies.
Examples:
- Recidivism risk assessment for suspects
- Prediction of crime areas (predictive policing)
- Reliability assessment of witness statements
- Profiling for criminal investigations
7. Border control and migration
AI used to assess visa and asylum applications or for border security.
Examples:
- Automated screening of visa applications
- Polygraph systems in a migration context
- AI that verifies travel documents at border crossings
8. Administration of justice and democratic processes
AI used in legal decision-making or political processes.
Examples:
- AI that supports judges in interpreting facts
- Systems that try to influence voting behaviour or election outcomes
Self-check: is my AI high-risk?
Work through the steps below for each AI system in your AI register. The first question you answer with “yes” determines the category.
| Step | Question | If “yes” |
|---|---|---|
| 1 | Does the system do something that is explicitly prohibited (social scoring, real-time biometrics in public spaces, emotion recognition at work or school)? | Prohibited: stop using it |
| 2 | Is the system part of a regulated product (medical device, vehicle, machine) and does it control a safety function? | High-risk (Annex I) |
| 3 | Does the system fall within one of the eight Annex III categories? | High-risk (Annex III) |
| 4 | Does the system affect people directly but fall outside the categories above? | Limited risk (transparency obligation) |
| 5 | None of the above applies | Minimal risk |
Step 3 contains an exception: if an Annex III system can demonstrably show it poses no significant risk to people, it does not have to be treated as high-risk. This is a narrow exception. You must be able to justify and document it well.
Current deadlines for high-risk AI
The deadlines for high-risk AI have shifted through the Digital Omnibus package. On 7 May 2026, the European Parliament and the Council reached a political agreement. Formal publication is still to follow.
| Milestone | Date | Status |
|---|---|---|
| Prohibited AI (Article 5) | 2 February 2025 | In force |
| AI literacy requirement (Article 4) | 2 February 2025 | In force |
| Requirements for GPAI models | 2 August 2025 | In force |
| National enforcement powers | 2 August 2026 | In force |
| High-risk AI in Annex III | 2 December 2027 | Political agreement, not yet formally adopted |
| High-risk AI in regulated products | 2 August 2028 | Political agreement, not yet formally adopted |
The postponement to December 2027 is a political agreement, not an adopted law. Do not build your compliance timeline on the maximum postponement. Enforcement by national supervisory authorities begins in August 2026.
For current information, see the Dutch government (Rijksoverheid) page on the AI Act and the European AI Office .
What must you do if you have high-risk AI?
Providers (those who place the AI on the market) and, in some cases, users of high-risk AI must meet an extensive set of requirements. The core requirements are:
- Set up and maintain a risk management system
- Ensure data governance and data quality
- Keep technical documentation
- Build in logging and traceability
- Provide transparency to users
- Enable human oversight
- Guarantee robustness, accuracy and cybersecurity
- Carry out a conformity assessment
- Register in the EU AI database
ISO 42001 provides a solid basis for all of these requirements. It does not replace the formal conformity procedure, which is a separate process.
Further reading
- EU AI Act and ISO 42001 - The full relationship between law and standard
- AI inventory - Step 1: mapping your AI systems
- AI risk management - Assessing risks per system
- AI impact assessment - Mapping the societal impact
- Annex III at artificialintelligenceact.eu - The official list of high-risk AI
- EC classification guidelines - Guidelines for determining high-risk status