One of the most important aspects of the EU AI Act is its risk-based regulatory model.
Instead of applying the same requirements to every AI system, the regulation categorizes AI systems based on their potential risk to individuals and society.
Understanding these categories is essential for determining which compliance obligations apply to your organization.
Why the EU AI Act Uses a Risk-Based Model
Artificial intelligence systems vary widely in how they affect people.
Some systems automate simple tasks with minimal impact. Others may influence hiring decisions, credit approvals, or access to essential services.
The EU AI Act focuses regulatory oversight on systems with the highest potential risk.
This approach allows regulators to protect individuals while still encouraging innovation.
The Four EU AI Act Risk Categories
The Act regulates prohibited practices, high-risk systems, and certain transparency-only systems, and leaves everything else largely unregulated. This is commonly summarized as four tiers. The Act itself does not use the labels "limited" or "minimal."
Prohibited Risk
Certain AI practices are banned entirely because they are considered incompatible with fundamental rights.
Examples include:
- Social scoring that evaluates people based on their social behavior or personal traits and leads to unjustified or disproportionate detrimental treatment
- Subliminal, manipulative, or deceptive techniques that distort behavior and cause, or are likely to cause, significant harm
- Exploiting vulnerabilities related to age, disability, or socio-economic circumstances
- Untargeted scraping of facial images from the internet or CCTV to build facial recognition databases
- Inferring emotions in workplaces and educational institutions, except for medical or safety reasons
- Biometric categorisation that infers sensitive attributes such as race, political opinions, or sexual orientation
- Predicting an individual's risk of committing a crime based solely on profiling or personality traits
- Real-time remote biometric identification in public spaces for law enforcement, except in narrow cases (such as searching for missing persons, preventing an imminent threat, or identifying suspects of serious crimes) and subject to authorization and safeguards
- Generating non-consensual intimate imagery and child sexual abuse material (applies from December 2, 2026)
These practices have been banned since February 2, 2025, with the last two bans applying from December 2, 2026. They carry the Act's highest fines: up to €35 million or 7% of worldwide annual turnover.
High Risk
High-risk AI systems face the strictest regulatory requirements.
AI systems are high-risk in two ways:
- They are a safety component of, or are themselves, a product covered by the EU product-safety laws listed in Annex I that requires third-party conformity assessment
- They fall within one of the use cases listed in Annex III
Annex III examples include AI used in:
- Employment: recruitment, CV filtering, candidate evaluation, promotion and termination decisions, and performance monitoring
- Essential services: creditworthiness assessment (excluding fraud detection), eligibility for public benefits including healthcare services, health and life insurance risk assessment and pricing, and emergency call triage and dispatch
- Education: admissions, evaluating learning outcomes, and detecting prohibited behavior during exams
- Law enforcement, migration, asylum, and border control
- Critical infrastructure: safety components for digital infrastructure, road traffic, and the supply of water, gas, heating, and electricity
- Biometrics that are not prohibited, including remote biometric identification and emotion recognition
- Administration of justice and democratic processes
An Annex III system can fall outside high-risk status if it only performs a narrow procedural task, improves the result of a previously completed human activity, detects decision patterns without replacing human review, or performs a preparatory task. This exception never applies if the system profiles individuals. Providers who conclude their Annex III system is not high-risk must document that assessment before placing it on the market.
High-risk requirements apply from December 2, 2027 for Annex III systems and August 2, 2028 for Annex I systems. Providers carry most of the obligations. Deployers have fewer.
Limited Risk
AI systems in this category must meet transparency requirements under Article 50, which apply from August 2, 2026. They cover four situations:
- AI that interacts directly with people, such as chatbots, virtual assistants, and AI agents: users must be told they are interacting with AI
- Generative AI that produces synthetic audio, image, video, or text: providers must mark outputs in a machine-readable format so they are detectable as AI-generated
- Emotion recognition and biometric categorisation: deployers must inform the people exposed to them
- Deepfakes, and AI-generated text published to inform the public on matters of public interest: deployers must disclose that the content is AI-generated (text is exempt where it has undergone human review and a person holds editorial responsibility)
Disclosure must be clear and distinguishable, and given no later than the first interaction or exposure. Generative AI systems already on the market before August 2, 2026 have until December 2, 2026 to meet the marking requirement.
Article 50 applies to any AI system used in these situations, regardless of risk tier.
Minimal Risk
Most AI applications fall into this category.
Examples include:
- Spam filters
- Video game AI
- AI-powered recommendations (unless used in a high-risk context such as employment or credit)
These systems have no specific obligations under the Act, beyond AI literacy measures for providers and deployers. Providers can voluntarily adopt codes of conduct.
General-Purpose AI
General-purpose AI (GPAI) models sit outside this tiered structure and have their own rules, in force since August 2, 2025. A GPAI model can also be built into a high-risk system, in which case both sets of rules matter.
Why Risk Classification Matters
Risk classification determines the level of compliance required.
Providers of high-risk systems must implement:
- A risk management system
- Data governance
- Technical documentation
- Automatic logging
- Instructions for use
- Human oversight by design
- Accuracy, robustness, and cybersecurity
- A quality management system
- Conformity assessment, an EU declaration of conformity, and the CE marking
- Registration in the EU database
- Post-market monitoring and serious-incident reporting
Providers must document their assessment when they conclude an Annex III system is not high-risk. For everyone else, documenting classification decisions and reassessing them when a system's purpose or design changes is strongly recommended.
How Organizations Should Approach Risk Classification
A structured approach typically includes:
- Determining your role (provider, deployer, importer, or distributor) for each system
- Identifying AI systems across the organization
- Mapping use cases and business impact
- Determining risk categories
- Documenting classification decisions
This process helps organizations determine where compliance efforts should focus.
Download our EU AI Act Guide and EU AI Act Compliance Checklist to move beyond risk classification and start building a clear compliance strategy. Use them to implement governance frameworks, monitoring processes, and documentation controls as you assess your AI systems.