EU AI Act: High-Risk AI Systems Explained (Annex III)
Not all AI systems are high-risk under the EU AI Act- but if yours is, the obligations are substantial. This guide maps all eight Annex III domains, explains the classification rules, and clarifies what providers and deployers must do.
Article written with the help of artificial intelligence — learn more about Article 50
Dernière mise à jour : mai 2026 · Temps de lecture : 11 minutes
Most AI Is Not High-Risk. But Some Is.
The EU AI Act's risk-based approach means that the vast majority of AI systems- recommendation engines, chatbots, internal productivity tools- face no mandatory compliance obligations beyond basic transparency rules. The heavy compliance burden sits with a defined category: high-risk AI systems.
If your system is high-risk, you must complete a full conformity framework before placing it on the EU market. The obligations are substantial. The fines for non-compliance reach 3% of global annual turnover or €15 million. The enforcement deadline is 2 December 2027 (postponed from 2 August 2026 by the Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026).
This guide explains exactly which systems qualify, why, and what to do about it.
Two Routes to High-Risk Classification (Article 6)
A system is high-risk under either of two paths:
Path 1- Annex I products (Article 6(1)): The AI system is a safety component of, or is itself, a product covered by EU harmonisation legislation listed in Annex I. This covers medical devices, machinery, aircraft, motor vehicles, toys, lifts, and similar regulated products. If AI is embedded in such a product, it inherits the product's regulatory treatment. This path applies from 2 August 2028 (postponed from 2 August 2027 by the Digital Omnibus).
Path 2- Annex III use cases (Article 6(2)): The AI system falls into one of eight high-risk use-case domains listed in Annex III, regardless of the product category it sits in. This is the path that applies to most software-based AI systems and is the focus of this article. It applies from 2 December 2027 (postponed from 2 August 2026 by the Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026).
Exception within Annex III (Article 6(3)): An AI system that falls into an Annex III domain is not high-risk if it poses only a limited risk to fundamental rights, health, or safety- provided the provider documents this assessment before market placement. The Commission published guidelines on this exception in February 2026. In practice, this exception is narrow.
The Eight Annex III Domains
DILAIG automates this step- free AI Act audit in 20 min, FRIA generation, Annex IV Technical Documentation and EU Declaration of Conformity included. Start the audit →
1. Biometrics
What it covers:
- Remote biometric identification systems- systems that identify individuals at a distance using biometric data such as gait, voice, face, or fingerprint. Real-time use in public spaces is largely prohibited (Article 5); post-remote identification is high-risk.
- Biometric categorisation systems- AI that infers sensitive or protected attributes from biometric data: race, political opinion, trade union membership, religious beliefs, sexual orientation, or health status.
- Emotion recognition systems- AI that infers emotional states from biometric indicators. Note: emotion recognition in workplaces and schools is entirely prohibited under Article 5.
Practical impact: Facial recognition software, biometric access control systems used to classify employees, and HR tools that infer affect from video interviews all land here.
2. Critical Infrastructure
What it covers:
AI used as a safety component in the management or operation of:
- Critical digital infrastructure (telecom networks, cloud infrastructure, data centres)
- Road traffic management systems
- Supply networks for water, gas, heating, or electricity
Practical impact: AI that controls traffic signals, manages grid load balancing, monitors water treatment systems, or detects cyberattacks on critical network infrastructure qualifies. General IT monitoring tools that do not make safety-relevant decisions do not.
3. Education and Vocational Training
What it covers:
- Systems that determine access to educational institutions or training programmes
- Systems that evaluate learning outcomes affecting progression decisions
- Systems that assess the appropriate level of education for a person
- Systems that monitor and detect prohibited behaviour during tests (cheating detection)
Practical impact: Automated university admissions tools, AI-driven proctoring systems, and adaptive learning platforms that make or influence final academic decisions qualify. Basic personalisation tools that recommend content without gating access generally do not.
4. Employment, Workers Management, and Self-Employment
What it covers:
Recruitment and selection:
- Targeted job advertisement systems
- CV filtering and ranking tools
- AI-driven interview assessment systems
Work-related decisions:
- Systems for promotions, task allocation, and work monitoring
- Systems that inform or make termination decisions
- Performance evaluation tools that feed into employment decisions
Practical impact: ATS (Applicant Tracking Systems) with AI scoring, video interview analysis tools, and workforce analytics platforms used to make or influence hiring or management decisions are high-risk. A scheduling tool that simply assigns shifts based on availability, without making performance-related judgements, is not.
This is one of the most commercially significant Annex III domains. Many HR tech vendors who assumed their products were outside scope are high-risk providers.
5. Essential Private and Public Services
What it covers:
- Creditworthiness assessment- AI systems that evaluate whether individuals or businesses qualify for loans, credit, or similar financial products. Exception: AI used purely for fraud detection is not high-risk under this heading.
- Public benefits eligibility- AI determining access to social assistance, housing, healthcare benefits
- Life and health insurance risk assessment and pricing- AI that evaluates risk and sets premiums
- Emergency call evaluation and prioritisation- AI that assesses the urgency of emergency service calls and dispatches resources
Practical impact: Automated credit scoring (not just fraud screening), AI-driven benefits assessment in public administration, and underwriting tools in insurance are all high-risk. A bank's transaction fraud detection model, by contrast, is explicitly excluded.
6. Law Enforcement
What it covers:
- Victimisation risk assessment- predicting the likelihood of a person becoming a crime victim
- Polygraph-like tools- AI that assesses the reliability or truthfulness of statements
- Evidence reliability evaluation- AI that assesses the credibility of evidence
- Re-offending risk assessment- predicting the likelihood of criminal recidivism
- Criminal profiling- AI used to profile, predict, or detect criminal activity or behaviour patterns
Practical impact: Predictive policing tools, AI-assisted interrogation support, and recidivism prediction systems used by courts or probation services are high-risk. Surveillance cameras without AI processing are not.
7. Migration, Asylum, and Border Control
What it covers:
- Security, health, and migration risk assessment at borders
- Asylum and visa application examination- AI that assesses or assists in the processing of claims
- Polygraph-like tools used in migration contexts
- Biometric identification of persons in migration and border control contexts (excluding travel document verification)
Practical impact: AI systems used by immigration authorities to assess visa applications, screen asylum claims, or conduct risk profiling at borders. This domain applies primarily to public sector entities and their technology suppliers.
8. Administration of Justice and Democratic Processes
What it covers:
- AI assisting judicial authorities in researching and interpreting facts, the law, or applying law to facts
- AI designed to influence elections or referendums- including systems that influence voting behaviour of individuals or the outcome of elections
Practical impact: Legal research AI used to recommend case outcomes to judges, and political micro-targeting tools designed to shift voting behaviour, are high-risk. Standard legal research assistants that surface relevant case law without recommending outcomes operate in a grey zone- the Article 6(3) exception analysis is critical here.
Am I a Provider or a Deployer?
Your obligations depend on your role in the AI value chain.
| Question | If Yes, You Are Likely... |
|---|---|
| Did you build the AI system and place it on the EU market? | Provider |
| Did you integrate a third-party AI model into a product you sell? | Provider (of the integrated system) |
| Do you use a third-party AI system in your operations? | Deployer |
| Did you rebrand or substantially modify an existing AI system? | Provider (Article 25) |
Provider obligations (Articles 9–49): the full compliance framework- risk management, data governance, technical documentation, logging, transparency, human oversight, conformity assessment, EU Declaration of Conformity, registration in the EU database.
Deployer obligations (Article 26): implement human oversight, conduct a Fundamental Rights Impact Assessment (FRIA) before deploying certain high-risk systems, monitor performance, report incidents, inform workers when AI is used to monitor them.
The Conformity Assessment Process
Before placing a high-risk AI system on the EU market, providers must complete a conformity assessment demonstrating the system meets all AI Act requirements.
For most Annex III systems: Self-assessment is permitted. The provider conducts and documents the conformity assessment internally, then draws up an EU Declaration of Conformity and registers the system in the EU database.
For biometric identification systems and certain Annex I products: Third-party assessment by a notified body is required.
The conformity assessment must be documented and kept for 10 years after the system is placed on the market or put into service.
What the Compliance Framework Requires in Practice
For providers of high-risk AI systems, the obligations map onto ten requirements:
- Risk management system- a documented, ongoing process to identify, estimate, evaluate, and mitigate risks throughout the system's lifecycle
- Data governance- training, validation, and test data must be relevant, representative, error-free, and complete for the system's purpose
- Technical documentation (Annex IV)- full documentation of system design, development methodology, architecture, and evaluation results
- Automatic logging- the system must generate logs that enable post-market monitoring and incident reconstruction
- Transparency and information to deployers- a clear instructions manual enabling deployers to implement appropriate human oversight
- Human oversight- the system must be designed so that humans can monitor, understand, override, and intervene
- Accuracy, robustness, cybersecurity- documented performance metrics and security measures against adversarial manipulation
- Conformity assessment- internal (or third-party for biometrics) verification of compliance with all the above
- EU Declaration of Conformity- formal document declaring compliance, signed by the provider
- Registration- entry in the EU-wide database for high-risk AI systems (publicly accessible for most Annex III systems)
Key Dates for High-Risk Compliance
| Date | Milestone |
|---|---|
| 2 December 2027 (postponed from 2 August 2026 by the Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026) | Full Annex III framework applies; AI Office enforcement powers active |
| 2 August 2028 (postponed from 2 August 2027 by the Digital Omnibus) | Annex I product AI systems must comply |
| 2 August 2030 | Public sector high-risk systems compliance deadline |
How DILAIG Handles High-Risk Classification
Determining whether your system is high-risk requires mapping your specific use case against eight Annex III domains, then assessing whether the Article 6(3) exception applies. That analysis is non-trivial- and getting it wrong in either direction is costly.
DILAIG's audit covers this directly. Step 4 of the 50-question audit cross-references your system's actual function against all eight Annex III domains and the prohibited practices of Article 5. The result is a documented classification with the applicable obligations listed article by article.
For systems confirmed as high-risk, DILAIG then generates:
- Fundamental Rights Impact Assessment (FRIA)- required under Article 27
- EU Declaration of Conformity- required under Article 47
- Technical Documentation (Annex IV)- the core compliance dossier
- Transparency Notice- required under Articles 13 and 50
Each document is a professional draft structured to the exact requirements of the regulation, ready for legal review.
If you are a compliance consultant or AI lead managing multiple systems across different Annex III domains- biometrics, HR, credit, education- DILAIG also offers an AI agent connection via MCP. An agent (Claude, ChatGPT, Le Chat) can pre-fill an audit per system from your internal documents; the agent never sees the score and cannot submit the audit, and every draft requires human review before submission. Learn more about the MCP connection →
→ Start your free AI Act audit- 20 minutes, no credit card required.
FAQ: High-Risk AI Systems
How do I know if my AI system is in Annex III?
Map your system's actual function- what decision does it make or influence, about whom, in what context- against the eight domains above. The relevant question is not what sector you operate in but what the AI does. A credit scoring tool used by a retailer is still high-risk under Annex III §5.
Can I use the Article 6(3) exception to avoid high-risk status?
Only if your system poses a limited risk to fundamental rights and safety, and you can document that assessment. The exception is available but narrow. The Commission's February 2026 guidelines provide further criteria. It must be documented before market placement.
Does the AI Act apply to AI systems already deployed before 2 December 2027?
Systems deployed before the 2 December 2027 deadline (postponed from 2 August 2026 by the Digital Omnibus) and not substantially modified benefit from a transitional period. However, any substantial modification- changing intended purpose, risk level, or key technical components- triggers full compliance.
How does DILAIG's AI agent connection work for Annex III classification?
An agent connected via MCP can read your internal documents (product specs, use-case briefs) and pre-fill the audit questions relating to the system's function. The agent has no access to your score or the classification engine, and cannot submit the audit on your behalf. A human reviews and validates every answer before submission.
What is a Fundamental Rights Impact Assessment (FRIA)?
A FRIA is a documented assessment of the impact of a high-risk AI system on the fundamental rights of the people it affects. It is required from deployers before they deploy certain high-risk systems (Article 27) and from public sector deployers more broadly. DILAIG generates FRIA drafts as part of its document generation module.
Are open-source high-risk AI models exempt?
No. The open-source exemption in the AI Act applies only to GPAI models (Article 53). High-risk AI systems have no equivalent exemption based on licence type. If your open-source model is used in an Annex III context, full compliance obligations apply.
Key Takeaways
- High-risk AI systems are defined by what they do, not what sector you are in
- There are two routes to high-risk status: Annex I products (from 2 August 2028, postponed from August 2027) and Annex III use cases (from 2 December 2027, postponed from August 2026)- postponements confirmed by the Digital Omnibus
- Eight Annex III domains cover biometrics, infrastructure, education, employment, essential services, law enforcement, migration, and justice
- Providers must complete a full 10-step compliance framework before market placement
- Deployers must conduct FRIAs, implement human oversight, and report incidents
- The Article 6(3) exception exists but is narrow and requires documented justification
- DILAIG's audit classifies your system and generates all four mandatory documents
Further Reading
- Risk Classification: How to Determine Your Category
- EU AI Act Conformity Assessment: Step-by-Step
- Technical Documentation Guide (Annex IV)
- How to Fill Out a FRIA: Field-by-Field Guide
- Obligations for Deployers of High-Risk AI (Article 26)
Not sure if this applies to you? Check in 2 minutes, no account needed.
Am I a GPAI provider? → · Is my AI prohibited under Article 5? →
Take action
Is your AI system compliant?
Free audit in 20 minutes. Detailed report, no commitment.
Start the audit →Keep reading
Practical guides, regulatory analysis, DILAIG news.