AI Act risk categories How the AI Act classifies AI: from prohibited to uncritical.
The AI Act regulates AI according to its risk to health, safety and fundamental rights. Most uses in companies fall into the lowest level. What matters is what you use AI for, not which program you use.
As of October 2026, checked against the text of the regulation. General information, not legal advice.
Four levels
-
Unacceptable risk
Prohibited practices under Art. 5. Since 2 February 2025.
-
High risk
AI as a safety component of regulated products (Annex I) or in sensitive areas (Annex III), Art. 6. Strict obligations for providers and deployers from 2 December 2027 and 2 August 2028 respectively.
-
Limited risk
Transparency obligations under Art. 50 for chatbots, AI‑generated content, deepfakes and emotion recognition. Since 2 August 2026.
-
Minimal risk
Everything else, such as draft texts, translations, spam filters or analyses. Here the Act mainly requires AI literacy (Art. 4).
The Act itself does not speak of four risk classes. The division is a common simplification of its structure. General-purpose AI models are subject to rules of their own (Chapter V), which mainly concern their providers.
Prohibited AI practices under Art. 5
- Subliminal, manipulative or deceptive techniques that materially influence decisions and can cause harm.
- Exploiting vulnerabilities due to age, disability or social or economic situation.
- Social scoring: evaluating people by behaviour or personality, with disadvantages in other contexts.
- Predicting criminal offences solely on the basis of profiling or personality traits.
- Untargeted scraping of facial images from the internet or from video surveillance for facial recognition databases.
- Emotion recognition in the workplace and in educational institutions, except for medical or safety reasons.
- Biometric categorisation to infer origin, political opinion, trade union membership, religion, sex life or sexual orientation.
- Real-time remote biometric identification in public spaces for law enforcement, with narrow exceptions.
- From 2 December 2026: AI that generates sexualised depictions of identifiable persons without consent or depictions of child sexual abuse.
Particularly important for companies: emotion recognition among employees is prohibited. Among customers it is not prohibited, but counts as high-risk AI (Annex III, point 1) and must be disclosed (Art. 50(3)).
Annex III Where AI counts as high-risk.
Eight areas in which AI systems are classed as high-risk. For companies, HR and essential services are the most relevant.
-
Biometrics
Remote biometric identification, categorisation and emotion recognition, where not already prohibited.
-
Critical infrastructure
Safety components in critical digital infrastructure, in road traffic and in the supply of water, gas, heating and electricity.
-
Education
Access and admission, assessment of learning outcomes, exam proctoring.
-
Employment and HR
Targeted job advertising, screening and evaluating applications, decisions on promotion, dismissal and task allocation, monitoring of performance and behaviour.
The most important area for many companies
-
Essential services
Assessing creditworthiness, risk and pricing in life and health insurance, access to public benefits, emergency calls.
Relevant for the finance and insurance sectors
-
Law enforcement
For example assessing evidence or the risk of becoming a victim of a crime.
-
Migration, asylum and border control
For example examining applications and assessing risks.
-
Justice and democratic processes
Support for judicial decisions and influencing elections.
Annex I additionally covers AI as a safety component of products that are already subject to EU product legislation, such as medical devices, toys or lifts.
Get in touchThe exemption in Art. 6(3)
A system from Annex III is not high-risk if it poses no significant risk. That is the case, for example, if it only performs a narrowly defined procedural task, improves the result of a completed human activity, detects decision patterns without replacing human assessment, or only carries out a preparatory task.
Profiling of natural persons always remains high-risk. Providers who rely on the exemption must document their assessment (Art. 6(4)).
Typical uses classified
- Preparing quotations, reports or emails with AI: as a rule minimal risk.
- Reading invoices, delivery notes or contracts automatically: as a rule minimal risk.
- Chatbot on the website: transparency obligation, people must be able to tell that they are writing with AI.
- Pre-sorting or evaluating job applications with AI: high risk under Annex III, point 4, obligations from 2 December 2027.
- Evaluating employee performance with AI: high risk; emotion recognition in the workplace is prohibited.
- Quality control in production without a safety function: since the Digital Omnibus expressly not a safety component (Art. 6(1a)).
The classification depends on the specific use. This list does not replace a review in the individual case.
How we support you
-
Classifying uses
We go through with you where and for what AI runs in your company and classify each use. More on AI consulting.
-
Preparing the team
We set out in your AI policy when an application becomes sensitive and whom your employees should ask then.
For high-risk uses we also recommend a legal review in each individual case.
Frequently asked questions about the risk categories
Which risk categories are there in the EU AI Act?
The usual division is into four levels: unacceptable risk (prohibited), high risk, limited risk with transparency obligations and minimal risk. In addition there are separate rules for general-purpose AI models.
What counts as a prohibited practice under the AI Act?
Among others manipulative techniques, exploiting vulnerabilities, social scoring, emotion recognition in the workplace and in educational institutions, untargeted collection of facial images and certain biometric methods (Art. 5). From 2 December 2026, prohibitions on sexualised deepfakes without consent are added.
What is high-risk AI?
AI that serves as a safety component of a regulated product (Annex I) or is used in one of the eight areas of Annex III, for example in recruitment or credit assessment (Art. 6).
Are general-purpose AI chat programs high-risk?
Not in themselves. It depends on what they are used for. Anyone who uses such a program, for example, to evaluate job applications can end up in the high-risk area.
From when do the obligations for high-risk AI apply?
After the Digital Omnibus, from 2 December 2027 for Annex III and from 2 August 2028 for Annex I.
Not sure how to classify your AI uses?
Tell us which uses you have or are planning. We will reply personally, free of charge and without obligation.