EU AI Act Risk Classification: 4-Tier System, Obligations & Compliance Guide
The EU AI Act sorts every AI system into one of four risk tiers. Each tier carries specific obligations, timelines, and penalties. This guide breaks down exactly where your systems fall and what you need to do about it.
The EU AI Act (Regulation 2024/1689) is the first binding horizontal regulation on artificial intelligence. Its core mechanism is a risk-based classification system that assigns obligations proportionate to the potential harm an AI system can cause.
This is not optional. If your AI system touches EU citizens, operates in the EU market, or produces outputs used within the EU, you fall within scope. The regulation entered into force on August 1, 2024, with phased enforcement beginning February 2, 2025 (prohibited practices) and August 2, 2025 (GPAI obligations), through to August 2, 2027 (full high-risk compliance).
Understanding where each of your AI systems sits within this classification is the first step toward compliance. Get it wrong, and you face fines up to 35 million EUR or 7% of global annual turnover for prohibited practice violations.
The 4 Risk Tiers
Every AI system falls into one of these categories. Click each tier to see the full details, examples, and what the regulation demands.
Unacceptable Risk
BANNEDAI practices that pose a clear threat to fundamental rights. Prohibited outright under Article 5, with enforcement starting February 2, 2025.
High Risk
REGULATEDAI systems with significant potential impact on health, safety, or fundamental rights. Full compliance requirements under Articles 6-15, Annex I and III.
Limited Risk
TRANSPARENCYAI systems that interact with people or generate content. Transparency obligations only: users must know they are interacting with AI or viewing AI-generated content.
Minimal Risk
NO OBLIGATIONSThe vast majority of AI systems fall here. No specific regulatory requirements, though voluntary codes of conduct are encouraged.
How to Determine Your Risk Tier
Answer these questions to narrow down which tier applies to your AI system. This is a simplified guide. Final classification should involve your legal and compliance teams.
Unacceptable Risk - PROHIBITED
This AI practice is banned under Article 5. It cannot be placed on the EU market or used within the EU. Violations carry fines up to 35M EUR or 7% of global annual turnover. Consult legal counsel immediately.
High Risk - Full Compliance Required
Your system must meet all requirements under Articles 9-15 and 17: risk management, data governance, technical documentation, logging, transparency, human oversight, accuracy, robustness, and security. Conformity assessment required before market placement. Full compliance deadline: August 2, 2027.
Limited Risk - Transparency Obligations
Your system must clearly disclose its AI nature to users. For generated or manipulated content, label it as AI-generated. The specific disclosure format depends on the interaction type. Lighter than high-risk, but still mandatory.
Minimal Risk - No Specific Obligations
No mandatory regulatory requirements apply under the EU AI Act. The Commission encourages voluntary codes of conduct. Consider adopting good practices from the NIST AI RMF or ISO 42001 anyway. Governance is smart business even when not legally required.
High-Risk Obligations Breakdown
If your AI system is classified as high-risk, these are the specific requirements you must meet. Each maps to a numbered Article in the regulation.
Establish, implement, document, and maintain a continuous risk management system throughout the AI system's entire lifecycle.
- Identify and analyze known and reasonably foreseeable risks
- Estimate and evaluate risks that may emerge when the system is used as intended and under conditions of reasonably foreseeable misuse
- Adopt suitable risk management measures, including design choices and testing
- Ensure residual risk is acceptable, with appropriate mitigation and communication
- Test the system to identify the most appropriate risk management measures
Training, validation, and testing data sets must meet quality criteria appropriate to the intended purpose of the system.
- Implement data governance and management practices covering design choices, data collection, preparation, and labeling
- Ensure training data is relevant, sufficiently representative, and free of errors to the extent possible
- Account for the specific geographical, contextual, behavioral, or functional setting of the system
- Take measures to detect, prevent, and mitigate possible biases in data sets
- Where special categories of personal data are processed, ensure appropriate safeguards
Maintain technical documentation that demonstrates compliance and provides national authorities with the information needed for assessment.
- General description of the AI system (intended purpose, developer, version)
- Detailed description of elements, development process, and system architecture
- Monitoring, functioning, and control of the system, including human oversight measures
- Description of the risk management system applied
- Documentation must be kept up-to-date throughout the system's lifecycle
High-risk AI systems must include automatic logging capabilities to ensure traceability throughout the system's lifecycle.
- Record events relevant to identifying situations that may result in risk, including events relating to the functioning period
- Logging must enable monitoring of the system's operation and post-market monitoring
- Logs must be accessible to deployers for compliance with their own obligations
- Retention period must be appropriate to the intended purpose and applicable law
Design and develop the system to ensure its operation is sufficiently transparent for deployers to interpret and use the output appropriately.
- Instructions for use in an appropriate digital or non-digital format
- Identity and contact details of the provider
- Characteristics, capabilities, and limitations of performance (accuracy, robustness, cybersecurity)
- Intended purpose and any known or foreseeable circumstances of misuse
- Human oversight measures, including technical measures to facilitate interpretation
High-risk AI systems must be designed to allow effective human oversight during the period the system is in use.
- Humans must be able to fully understand the capacities and limitations of the system
- Humans must be able to correctly interpret the system's output
- Humans must be able to decide not to use, override, or reverse the output
- Humans must be able to intervene or interrupt the system through a "stop" button or similar
- For biometric identification: at least two qualified humans must verify results before action
High-risk AI systems must achieve appropriate levels of accuracy, robustness, and cybersecurity, and perform consistently throughout their lifecycle.
- Accuracy levels must be declared in instructions for use and measurable
- Resilience against errors, faults, or inconsistencies within the system or its operating environment
- Technical redundancy including backup or fail-safe plans
- Protection against unauthorized third-party attempts to exploit vulnerabilities
- Resistance to adversarial attacks, data poisoning, model manipulation, and input perturbation
Providers of high-risk AI systems must establish a quality management system that ensures compliance with the regulation in a systematic and documented manner.
- Strategy for regulatory compliance, including conformity assessment procedures
- Techniques, procedures, and systematic actions for design, control, and verification
- Techniques for development, quality control, and quality assurance
- Examination, test, and validation procedures before, during, and after development
- Systems and procedures for data management, including collection, analysis, labeling, storage, filtering, and aggregation
- Risk management procedures, including post-market monitoring
Before placing a high-risk AI system on the market or putting it into service, the provider must ensure it undergoes the relevant conformity assessment procedure.
- For Annex III systems (standalone high-risk): internal control procedure (Annex VI) is sufficient in most cases
- For biometric identification systems: third-party conformity assessment by a Notified Body is required
- For Annex I systems (product safety): follows the conformity assessment procedure of the relevant product legislation
- CE marking must be affixed after successful conformity assessment
- EU declaration of conformity must be drawn up and kept for 10 years
Conformity Assessment (Art. 43)
Before any high-risk AI system can be placed on the EU market or put into service, it must pass a conformity assessment. The type of assessment depends on how your system is classified.
Internal Control (Annex VI)
Self-assessment by the provider. Applies to most Annex III high-risk systems. The provider verifies and documents compliance internally.
- ✓ No external audit required
- ✓ Provider draws up EU Declaration of Conformity
- ✓ CE marking affixed by provider
Notified Body (Annex VII)
Third-party assessment required for remote biometric identification systems. An independent Notified Body evaluates the system and QMS.
- • Assessment of QMS and technical documentation
- • Notified Body issues certificate of conformity
- • Periodic surveillance audits
Product Legislation (Annex I)
For AI embedded in products already regulated by EU safety legislation. Follows the existing product conformity assessment, with AI Act requirements added.
- • Integrated into existing CE marking process
- • AI Act requirements treated as additional criteria
- • Single conformity assessment covers both product and AI
Documentation retention: All conformity assessment documentation, EU Declarations of Conformity, and technical documentation must be kept for at least 10 years after the AI system has been placed on the market or put into service.
Fundamental Rights Impact Assessment (FRIA, Art. 27)
Beyond conformity assessment, certain deployers of high-risk AI systems must conduct a Fundamental Rights Impact Assessment before putting the system into use.
Who Must Conduct a FRIA?
- → Bodies governed by public law (government agencies, public authorities)
- → Private entities providing public services (utilities, healthcare, education)
- → Deployers using high-risk AI for creditworthiness assessment or risk pricing
- → Deployers using high-risk AI for life and health insurance risk assessment
What Must the FRIA Cover?
- ✓ Description of the deployer's processes using the AI system
- ✓ Time period and frequency of intended use
- ✓ Categories of natural persons and groups likely affected
- ✓ Specific risks of harm to identified groups
- ✓ Human oversight measures in place
- ✓ Measures to act on risks identified, including governance arrangements
The FRIA must be performed before first use and updated when circumstances change materially. Results must be notified to the relevant market surveillance authority. Where a Data Protection Impact Assessment (DPIA) is already required under GDPR Article 35, the FRIA can be conducted alongside it.
Incident Reporting (Art. 73)
Providers of high-risk AI systems must report serious incidents to market surveillance authorities. The timelines are strict and non-negotiable.
Critical Infrastructure Disruption
Report within 2 days of becoming aware. Covers widespread infringements or serious and irreversible disruption to the management and operation of critical infrastructure.
Death or Serious Harm to Health
Report within 10 days of becoming aware. Covers incidents causing death or serious damage to the health of a person.
Other Serious Incidents
Report within 15 days of becoming aware. Covers serious breach of fundamental rights obligations, serious damage to property, environment, or other serious incidents not resulting in death.
Who Must Report?
Providers of high-risk AI systems placed on the EU market. If the deployer identifies the incident first, they must notify the provider who then reports to authorities.
Required Documentation
- • Identification of the AI system and provider
- • Description of the incident, severity, and circumstances
- • Corrective actions taken or planned
- • Initial assessment of causal relationship
CSA alignment: The Cloud Security Alliance (CSA) 5-step AI incident response framework (Detect, Analyze, Contain, Remediate, Review) maps well to these EU AI Act obligations. Organizations that adopt the CSA model can use it as their operational IR process, with the Art. 73 reporting integrated as a mandatory step during the Contain/Remediate phases.
Proportionate Response per Tier
The EU AI Act is designed so governance effort matches risk. Here is how governance intensity scales across the four tiers.
| Governance Activity | Unacceptable | High Risk | Limited | Minimal |
|---|---|---|---|---|
| Risk Management System | N/A (Banned) | Mandatory | - | - |
| Data Governance | N/A | Mandatory | - | - |
| Technical Documentation | N/A | Mandatory | - | - |
| Logging | N/A | Mandatory | - | - |
| Transparency | N/A | Mandatory | Mandatory | Voluntary |
| Human Oversight | N/A | Mandatory | - | - |
| Accuracy/Robustness | N/A | Mandatory | - | - |
| Conformity Assessment | N/A | Mandatory | - | - |
| Incident Reporting | N/A | Mandatory | - | - |
| FRIA | N/A | Conditional | - | - |
Apply What You Have Learned
Put this classification knowledge into practice with tools built for the EU AI Act risk-based approach.