What is AI risk classification under the EU AI Act?
AI risk classification is the step in EU AI Act compliance where a company decides which of the Act’s tiers an AI system falls into. The four tiers are unacceptable risk, which is prohibited; high risk, which carries the full set of provider and deployer obligations; limited risk, which carries transparency obligations; and minimal risk, which carries none. The tier depends on what the system is used for and the product it sits in, and it does not depend on how sophisticated the model is. A simple rules-based tool that ranks job applicants is high-risk. A large language model that drafts marketing copy is minimal risk.
How AI risk classification works
The assessment runs in a fixed order. First, check the system against the prohibited practices in Article 5, such as social scoring, manipulation that causes harm, and emotion recognition in the workplace. Second, check Annex I: if the system is a safety component of a product covered by EU product legislation, such as medical devices, lifts, toys, or vehicles, and that product needs a third-party conformity assessment, the system is high-risk. AI safety functions in machinery are handled through the EU Machinery Regulation. Third, check Annex III, which lists use areas: biometrics, critical infrastructure, education, employment, access to essential services including credit scoring and life and health insurance pricing, law enforcement, migration, and justice.
A system in an Annex III area can escape the high-risk tier if it performs only a narrow procedural task, improves the result of a completed human activity, detects decision patterns without replacing human judgment, or performs a preparatory task. Any profiling of people removes that exemption, and a company relying on it must document its reasoning. Finally, systems that interact with people or generate synthetic content pick up the transparency obligations whatever their tier.
AI risk classification in practice
A regional bank classifies three systems. A model that scores loan applications is high-risk under Annex III because it determines access to credit. A tool that summarizes call transcripts for compliance review performs a preparatory task with no profiling, so the bank documents the exemption and treats it as minimal risk. The customer chatbot on the website is limited risk and needs a disclosure. The classification memo for each system goes into the AI inventory alongside the owner’s name.
What to watch for
- Classify by use, because the same model can be minimal risk in one deployment and high-risk in another.
- Revisit the tier whenever the purpose changes, since a reporting tool that starts driving decisions moves up.
- Record the exemption reasoning in writing, because an undocumented exemption is treated as a missed classification.
- Remember that general-purpose AI models have their own regime, separate from the four tiers that apply to systems.