EU AI Act: the risk classification of AI systems
AI risk classification is the first step towards complying with the EU Artificial Intelligence Act, because every other obligation depends on it. This guide covers the risk pyramid, a decision table and typical business examples.
In 30 seconds: AI risk classification under the AI Act has four levels: prohibited (Article 5), high risk (Article 6), transparency (Article 50) and minimal risk. In addition, general-purpose AI models follow their own rules. After Regulation (EU) 2026/1744, Annex III high-risk rules apply from 2 December 2027; Annex I rules, by contrast, apply from 2 August 2028.
The AI risk classification pyramid
Regulation (EU) 2024/1689 does not regulate technology, but the intended purpose of each system. As a result, the same tool can land on different levels depending on what you use it for. AI risk classification therefore follows this scale:
- Unacceptable risk: prohibited practices.
- High risk: requirements on risk management, data, documentation and human oversight.
- Transparency: disclosure and labelling duties.
- Minimal risk: no specific obligations.
The levels are not mutually exclusive: a high-risk system that talks to people must also meet Article 50. Moreover, none of this replaces the GDPR if the system processes personal data.
Unacceptable risk: the prohibited practices in Article 5
Since 2 February 2025, the AI Act has banned these uses, among others:
- Manipulation or exploitation of vulnerabilities causing significant harm.
- Social scoring.
- Predicting crime based solely on profiling.
- Facial recognition databases built by untargeted scraping of images.
- Emotion recognition at work and in education, except for medical or safety reasons.
- Biometric categorisation to infer sensitive characteristics.
- Real-time remote biometric identification for law enforcement, save for narrow exceptions.
Regulation (EU) 2026/1744 (the Digital Omnibus on AI) adds a further ban from 2 December 2026. It targets systems that generate realistic intimate images, video or audio of an identifiable person without their consent, or child sexual abuse material. In fact, a prohibited practice can cost up to €35 million or 7% of worldwide turnover.
High risk: the two routes in Article 6
Route 1: regulated products in Annex I
A system is high-risk if it is a safety component of an Annex I product, or the product itself, and that product requires a third-party conformity assessment. Examples include medical devices, toys and lifts. Moreover, the Omnibus clarified that building AI into a regulated product is not enough: it must perform a safety function. For connected product manufacturers, this route applies from 2 August 2028.
Route 2: use cases in Annex III
Annex III lists eight areas: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and border control, and justice and democratic processes. However, only the specific use cases it lists count. This route applies from 2 December 2027.
The Article 6(3) exception in AI risk classification
An Annex III system is not high-risk if it poses no significant risk and meets one of these conditions:
- It performs a narrow procedural task.
- Its output improves the result of a completed human activity.
- The system detects decision-making patterns without replacing human judgement.
- It only prepares an assessment that a person then carries out.
However, the exception never applies where there is profiling. Moreover, the provider must document its AI risk classification before placing the system on the market and register it in the EU database (Article 49(2)). The Omnibus kept that registration, although it reduced the data required.
Transparency and minimal risk
Article 50 has applied since 2 August 2026, whatever the level of the system. So you must tell people when they are talking to AI and mark generated content; anyone publishing deepfakes must also disclose them. For generative systems placed on the market before August, by contrast, marking becomes mandatory on 2 December 2026. Furthermore, on 20 July 2026 the Commission published guidelines on Article 50.
Minimal risk covers most uses, such as spam filters or spell checkers. Even so, keep them in your inventory, because the AI risk classification changes if the use changes.
General-purpose models: outside the tiered AI risk classification
General-purpose AI (GPAI) models, such as large language models, have had their own rules since 2 August 2025. That said, if your organisation builds one into its own tool, the tool gets its own AI risk classification based on its use. An internal assistant may be minimal risk; if it screens candidates, on the other hand, it is high-risk.
A step-by-step AI risk classification decision table
First, apply these questions in order to every system. That way, your AI risk classification stays reasoned and documented:
| Question | Outcome |
|---|---|
| Is it an AI system under Article 3(1)? | If not, it is outside the Regulation. |
| Does it match a practice in Article 5? | If so, it is prohibited: withdraw it. |
| Is it a safety component of an Annex I product with third-party assessment? | If so, high risk from 2/08/2028. |
| Does its intended use appear in Annex III? | If not, skip to the last question. |
| Does it meet an Article 6(3) condition without profiling? | If so, document and register it; if not, high risk from 2/12/2027. |
| Does it interact with people, generate content or recognise emotions? | If so, Article 50 applies; if not, minimal risk. |
Then record each answer with a date and an owner, as we explain in how to automate compliance evidence.
AI risk classification examples in business
HR: the most sensitive AI risk classification
Screening CVs, scoring candidates, evaluating performance or deciding on promotions and dismissals is high-risk (Annex III, point 4). Moreover, because these tools usually profile people, the Article 6(3) exception rarely fits. Analysing staff emotions, on the other hand, is prohibited. Therefore, if you deploy one of these systems, assign human oversight, keep the logs and inform workers’ representatives beforehand.
Credit scoring and insurance
Assessing the creditworthiness of natural persons is high-risk, unless its purpose is to detect fraud. Pricing in life and health insurance is high-risk too. In both cases, the deployer must also carry out a fundamental rights impact assessment first. That is why, in banking and fintech, it pays to start the AI risk classification of decision models now.
Chatbots and assistants
A customer service chatbot usually falls under transparency: it discloses that it is AI and marks what it generates. Nevertheless, if it decides on credit or a public benefit, that use makes it high-risk.
Dates and Commission guidelines on AI risk classification
The dates for each level appear in the table above; in addition, the full timeline is in the key dates of the AI Act.
Article 6(5) gave the Commission until 2 February 2026 to publish guidelines. In the end, on 19 May 2026 it released draft guidelines on high-risk classification in three parts: general principles, Annex I and Annex III. The consultation was then extended to 23 July, and the Commission plans to adopt the final version by the end of 2026. Meanwhile, use the draft as a reference for your AI risk classification, bearing in mind that it is not binding.
Frequently asked questions
Who carries out the AI risk classification, the provider or the user?
The provider does, because it sets the intended purpose, but you should check that it matches your real use. Moreover, if you change the purpose and the system becomes high-risk, or you sell it under your own brand, you take on the provider’s obligations (Article 25).
Do you have to register an Annex III system that is not high-risk?
Yes. If the provider relies on the Article 6(3) exception, it must register the system in the EU database and document its analysis, although with less data than before the Omnibus.
Are there official guidelines on AI risk classification?
There is a Commission draft from 19 May 2026, with examples of systems that are and are not high-risk. However, the Commission does not plan to adopt the final version until the end of 2026.
What happens if you misclassify a system?
First, the market surveillance authority can review it and require you to bring it into line. If the error aimed to circumvent the Regulation, a fine follows as well: up to €15 million or 3% of turnover, with the lower amount for SMEs.
Reviewed on 14 October 2026. For information only, not legal advice.
How Atalaya GRC helps
Atalaya GRC helps you decide, service by service, which regulations apply and keep the AI risk classification of each system as evidence, with the Article 6(3) analysis. In addition, its regulatory clock warns you ahead of every deadline and, with “one control, many regulations”, you reuse what you already have for GDPR or ISO 27001. It also handles incidents, notifications and the third-party register. Finally, see the EU AI Act guide.
