EU AI Act Risk Categories Explained: The 4 Risk Classes with Examples
The EU AI Act (Regulation 2024/1689) regulates artificial intelligence by risk, not by technology — every obligation in the law follows from which of four classes an AI system falls into. At the top sits 'unacceptable risk': practices banned outright, such as social scoring by authorities, manipulative techniques that materially distort behavior and cause harm, emotion recognition in workplaces and schools, and untargeted scraping of facial images. These prohibitions already apply, and using a banned system carries the Act's highest fines — up to €35 million or 7% of global turnover.
The second class, 'high-risk', is where most corporate compliance work happens. It covers AI used as a safety component of regulated products and AI in sensitive use cases listed in the Act's Annex III: recruitment and worker management (CV screening, promotion decisions), credit scoring, education admissions and exam scoring, essential services, critical infrastructure, and several law-enforcement and migration contexts. High-risk systems demand risk management, data governance, technical documentation, logging, human oversight and conformity assessment — for providers, and lighter but real duties for deployers.
Classes three and four are lighter: 'limited risk' triggers transparency duties — chatbots must reveal they are AI, deepfakes and AI-generated content must be labeled — while 'minimal risk' (spam filters, game AI, most recommendation and productivity features) carries no specific obligations beyond general law. The catch for international companies: the classification follows the use case in the EU, so the same model can be minimal-risk in one deployment and high-risk in your German HR department. Classify your systems with our AI Act audit tool, and see the AI Act fines overview for what each class risks.
Your checklist
- 1Inventory your AI systems first: you cannot classify what you have not catalogued — include AI features embedded in vendor SaaS your EU entities deploy.
- 2Screen against the prohibited practices list: social scoring, harmful manipulation, emotion recognition at work, and similar banned uses must be stopped immediately — these prohibitions are already in force.
- 3Check Annex III use cases for high-risk status: anything touching recruitment, worker management, credit, education scoring, essential services or critical infrastructure is a high-risk candidate.
- 4Check the product route to high-risk too: AI acting as a safety component of products under EU harmonization law (machinery, medical devices, vehicles) is high-risk via that path.
- 5Apply transparency duties to limited-risk systems: disclose chatbots as AI, and label AI-generated or manipulated audio, image and video content where the Act requires it.
- 6Document the classification and reasoning per system: regulators and enterprise customers will ask; a recorded, defensible classification is the first artifact they expect.
- 7Assign roles per system: provider, deployer, importer or distributor — obligations differ sharply, and rebranding or substantially modifying a system can move you into the provider role.
- 8Re-classify on every change of use: a minimal-risk tool redeployed into hiring becomes high-risk; classification is a living process, not a one-off exercise.
Frequently asked questions
What are the four risk categories of the EU AI Act?
Unacceptable risk: practices banned outright, such as social scoring and manipulative techniques causing harm. High risk: AI in sensitive use cases like hiring, credit scoring or safety components, subject to extensive obligations. Limited risk: systems with transparency duties, such as chatbots and AI-generated content that must be disclosed. Minimal risk: everything else — spam filters, game AI, most productivity features — with no specific AI Act obligations.
Which AI systems count as high-risk?
Two routes lead there: AI used as a safety component of products covered by EU harmonization law (machinery, medical devices, vehicles), and AI in the use cases of Annex III — including recruitment and worker management, credit scoring, education admissions and scoring, essential public and private services, critical infrastructure, and certain law-enforcement, migration and justice contexts. A CV-screening tool used by a German subsidiary is a textbook high-risk example.
Are chatbots high-risk under the AI Act?
Ordinarily no. A customer-service chatbot is typically a limited-risk system: the main obligation is transparency, meaning users must be informed they are interacting with an AI. It can escalate to high-risk if it is used for a sensitive Annex III purpose — for instance effectively deciding on access to essential services or screening job applicants — which is why classification follows the use, not the product label.
What happens if we misclassify an AI system?
Classification errors carry the consequences of the class you should have been in: operating an unrecognized high-risk system means missing conformity, documentation and oversight duties, each sanctionable, and using a prohibited practice triggers the top fine tier of up to €35 million or 7% of worldwide turnover. A documented, reasoned classification per system is the strongest evidence of good faith if a supervisory authority disagrees.