AI Act risk classification: how to classify your AI systems step by step
The AI Act does not treat every AI system the same: what it asks of you depends on the risk level. Classifying each system correctly is the first step of the plan, because everything else depends on it. Here is a simple four-question method.
The four levels
- Prohibited (Art. 5): practices the Regulation does not allow, such as harmful manipulation or social scoring. They have applied since February 2025.
- High-risk (Art. 6 and Annex III): systems that affect people’s rights or safety. They carry the most obligations.
- Limited risk (Art. 50): chatbots, generated content and deepfakes. The main duty is to inform and label.
- Minimal risk: everything else, such as spam filters or simple recommendations. No specific obligations, other than general ones such as AI literacy.
The four questions, in order
1. Is it among the prohibited practices?
Check the Art. 5 list: harmful subliminal manipulation, exploiting vulnerabilities, social scoring, sensitive biometric categorization, emotion recognition at work or school, among others. If it fits, it cannot be used in the EU.
2. Is it in Annex III, or a safety component of a regulated product?
Annex III covers uses such as employment and worker management (candidate selection, performance evaluation), access to essential services (credit scoring, life and health insurance), education, critical infrastructure, biometrics, migration and justice. Systems that are safety components of regulated products (toys, machinery, medical devices) are also high-risk, on their own timeline.
There is an exit: under Art. 6(3), an Annex III system may not be considered high-risk if it poses no significant risk (for example, if it only performs a narrow or preparatory task). It never applies if the system profiles individuals, and the conclusion must be documented.
3. Does it interact with people or generate content?
If it is a chatbot, or generates text, images, audio or video, or creates deepfakes, the Art. 50 transparency duties apply: tell people they are dealing with AI and label generated content.
4. What is your role?
Obligations differ depending on whether you are a provider (you develop the system or place it on the market under your name) or a deployer (you use it in your business). A company using Copilot or ChatGPT Enterprise is normally a deployer. If you adapt a system and offer it under your brand, you may become a provider.
Common mistakes
- Classifying by technology instead of use. The same model can be minimal risk in an internal chat and high-risk if it screens résumés.
- Forgetting the AI you already use. Third-party systems and AI that staff use without telling anyone count too; start with the inventory.
- Leaving no record. If you conclude something is not high-risk, keep the reasoning and the date: it is what you will be asked for.
- Treating the result as final. Review the classification when the system’s use or the regulatory timeline changes.
What to do with the result
- Record the level, the role and the justification in your AI systems inventory.
- For each high-risk system, open its file: documentation, human oversight and logs.
- For limited-risk systems, prepare the notices and labelling.
- Set a review date, for example every six months.
The application dates for high-risk and transparency duties have moved with the Digital Omnibus; check the current timeline in our timeline article and in the official text.
Frequently asked questions
How do I know whether I am a provider or a deployer?
If you develop the system or place it on the market under your name, you are a provider. If you use it under your authority in a professional activity, you are a deployer. You can be both for different systems.
Is a customer-service chatbot high-risk?
Usually not: it is normally limited risk and requires telling the user they are talking to AI. It changes if the chatbot makes decisions that fall under Annex III, such as granting or denying credit.
What about general-purpose models such as ChatGPT?
General-purpose AI models carry their own obligations for their providers. As a user company, you classify the specific use you make of them.
Informational content, not legal advice. AI regulation changes quickly: always check the current official text before deciding. Information reviewed as of October 15, 2026.
This gives you the method. If you want the finished document: AI Act Risk Classification Checklist.
What's your Data Governance maturity?
Free assessment with your priority gaps, plus the self-assessment quiz and savings calculator on the Data Governance path.