Last updated: October 2026 · Legal state checked against the consolidated AI Act after the Digital Omnibus, Regulation (EU) 2026/1744
The EU AI Act sorts AI systems into four risk categories: unacceptable risk (banned), high risk (strict obligations), limited risk (transparency duties) and minimal risk (no specific obligations). A system is high-risk if it is a safety component of a product covered by Annex I, or if it is used in one of the eight Annex III areas, such as hiring, credit scoring, education or biometrics. After the Digital Omnibus, the Annex III obligations apply from 2 December 2027, and the Annex I obligations from 2 August 2028.
Most pages on this topic list the four tiers and stop there. That does not answer the question compliance teams actually ask: is our system high-risk? Before writing, we compared six top-ranking guides against 22 practical questions. None of them explained who must carry out a fundamental rights impact assessment. Only one said plainly whether a chatbot is high-risk, and only one explained where general-purpose AI models fit. This guide adds a seven-step classification test and a table of common use cases. For the full regulation, including timelines and the GPAI regime, see our EU AI Act explainer.
What are the four EU AI Act risk categories?
| Category | Legal basis | What it means | Applies from | Maximum fine |
|---|---|---|---|---|
| Unacceptable risk | Art. 5 | Banned outright | 2 Feb 2025; added bans from 2 Dec 2026 | €35m or 7% of turnover |
| High risk | Art. 6, Annex I and III | Risk management, data governance, logging, human oversight, conformity assessment, registration | Annex III: 2 Dec 2027 · Annex I: 2 Aug 2028 | €15m or 3% |
| Limited risk | Art. 50 | Transparency: tell people they are talking to AI, label deepfakes and synthetic content | 2 Aug 2026 | €15m or 3% |
| Minimal risk | — | No specific obligations; voluntary codes of conduct | — | — |
Two clarifications competitors often skip. First, the categories describe uses, not technologies. The same model can be minimal-risk in a spam filter and high-risk when it ranks job applicants. Second, general-purpose AI models such as large language models sit outside the four tiers. They have their own regime (Articles 51–55, with extra duties above 1025 FLOPs of training compute). A product built on one is classified by what it is used for. Fines are capped at the lower of the two amounts for SMEs and start-ups. Our EU AI Act fines guide covers the details.
Which AI practices are banned under Article 5?
- subliminal, manipulative or deceptive techniques that materially distort behaviour and cause significant harm;
- exploiting vulnerabilities linked to age, disability or social or economic situation;
- social scoring that leads to unjustified or disproportionate detrimental treatment;
- predicting a person’s risk of committing a crime based solely on profiling or personality traits;
- untargeted scraping of facial images to build facial recognition databases;
- emotion recognition in workplaces and schools, except for medical or safety reasons;
- biometric categorisation to infer race, political opinions, trade union membership, religion, sex life or sexual orientation;
- real-time remote biometric identification in public spaces for law enforcement, apart from narrow exceptions;
- from 2 December 2026, added by the Omnibus: AI systems that generate non-consensual intimate imagery or child sexual abuse material.
What makes an AI system high-risk?
Article 6 gives two routes:
- Annex I route (products). The AI is a safety component of a product, or is itself a product, covered by EU harmonisation law listed in Annex I (machinery, toys, medical devices, lifts, radio equipment and others), and that product needs a third-party conformity assessment. The Omnibus clarified that features which only assist users or optimise performance, without a safety function, are not safety components.
- Annex III route (use cases). The AI is intended for one of eight sensitive areas:
| Annex III area | Typical high-risk uses |
|---|---|
| 1. Biometrics | Remote biometric identification (not simple identity verification), biometric categorisation by sensitive traits, emotion recognition |
| 2. Critical infrastructure | Safety components in digital infrastructure, road traffic, and water, gas, heating or electricity supply |
| 3. Education | Admissions, grading that steers learning, assessing education level, detecting cheating in tests |
| 4. Employment | Targeted job ads, CV filtering, candidate evaluation, promotion, termination, task allocation, performance monitoring |
| 5. Essential services | Eligibility for public benefits, credit scoring (fraud detection excluded), life and health insurance pricing, emergency call triage |
| 6. Law enforcement | Victim-risk assessment, polygraph-like tools, evidence reliability, re-offending risk, profiling in investigations |
| 7. Migration and borders | Polygraph-like tools, security or health risk assessment, asylum and visa applications, detecting persons |
| 8. Justice and democracy | Assisting courts with facts and law, systems intended to influence elections or voting behaviour |
When is an Annex III system not high-risk? The Article 6(3) exception
An Annex III system escapes the high-risk category if it poses no significant risk to health, safety or fundamental rights and meets at least one of four conditions. It must:
- perform a narrow procedural task, for example sorting incoming documents into categories;
- improve the result of a human activity that is already finished, for example polishing the language of a drafted letter;
- detect decision-making patterns or deviations without replacing or influencing the human assessment;
- perform a preparatory task for an assessment, for example translating application files.
Two traps. If the system profiles natural persons, it is always high-risk and the exception does not apply. And using the exception still leaves work to do: the provider must document the assessment and register the system in the EU database (Art. 49(2)). The Commission’s Omnibus proposal would have removed that registration duty. Parliament and Council kept it, and only trimmed the required data fields.
How do you classify your AI system? A 7-step test
Run the test on the intended purpose stated in your documentation and marketing, not on the model’s technical capabilities. If a deployer repurposes a general tool for an Annex III use, for example using a general chatbot to pre-screen job applicants, the deployer can become the provider of a high-risk system under Article 25.
Is my use case high-risk? Common examples
| Use case | Category | Why |
|---|---|---|
| Customer-service chatbot | Limited | Must disclose it is AI (Art. 50(1)). Not high-risk unless it decides an Annex III matter |
| CV screening or candidate ranking | High | Annex III 4(a); usually involves profiling, so Art. 6(3) cannot apply |
| Bank credit scoring of consumers | High + FRIA | Annex III 5(b); deployers must do a fundamental rights impact assessment |
| Payment fraud detection | Not high-risk under 5(b) | Explicitly carved out of the credit-scoring entry |
| Exam proctoring that flags cheating | High | Annex III 3(d) |
| Emotion recognition of employees | Prohibited | Art. 5(1)(f); outside work and education it is high-risk under Annex III 1(c) |
| Diagnostic AI in a medical device | High (Annex I) | Medical device law requires notified-body assessment; applies from 2 Aug 2028 |
| AI image generator | Limited | Outputs must be machine-readably marked; deepfakes must be disclosed |
| Spam filter, recommendation engine for a shop | Minimal | No specific obligations |
What do providers and deployers of high-risk AI have to do?
Providers, who build the system or sell it under their own name, carry most of the burden. They need a risk management system, data governance, technical documentation, automatic logging, instructions for use, human-oversight design, accuracy and cybersecurity, a quality management system, a conformity assessment, CE marking and registration in the EU database (Arts. 9–17, 43, 49). Deployers, who use the system under their own authority, must follow the instructions, assign competent human oversight, monitor operation, keep logs for at least six months and inform affected workers (Art. 26).
The fundamental rights impact assessment (FRIA) is narrower than many guides suggest. Under Article 27, it applies to deployers that are public bodies or private entities providing public services, and to any deployer of credit scoring (Annex III 5(b)) or life and health insurance pricing (5(c)). It must be done before first use, and the results go to the market surveillance authority. Critical-infrastructure systems (Annex III point 2) are exempt. Public-authority deployers also register their use in the EU database (Art. 49(3)).
What is still unsettled in 2026?
Article 6(5) required the Commission to publish practical guidelines on high-risk classification by 2 February 2026. It missed that date. Draft guidelines appeared on 19 May 2026, a targeted consultation closed on 23 June 2026, and the final version is expected by the end of 2026. Until then, borderline cases, especially the Article 6(3) conditions, depend on your own documented reasoning. Harmonised standards for high-risk requirements are also still being finalised. That is one reason the Omnibus moved the Annex III deadline to December 2027.
What does this mean in practice?
Start with an inventory, not a policy. List every AI system you build or use, write down its intended purpose in one sentence and run the seven-step test. Most organisations find that most of their systems are minimal or limited risk. The work concentrates on a few systems in hiring, credit, insurance, education or biometrics, and that is where the 14 months before December 2027 should go. To link the classification to an existing risk programme, see NIST AI RMF vs ISO 42001 vs the AI Act. If credit is your domain, our explainer on AI credit scoring shows what the models actually do. A Polish version of this guide is available: systemy wysokiego ryzyka w AI Act.
FAQ
What are the 4 risk levels of the EU AI Act?
Unacceptable risk (banned practices under Article 5), high risk (Annex I products and Annex III use cases with strict obligations), limited risk (transparency duties under Article 50) and minimal risk (no specific obligations).
Is ChatGPT or a chatbot high-risk under the EU AI Act?
Not by default. A chatbot is limited risk and must tell users they are interacting with AI. It becomes high-risk only if it is used for an Annex III purpose, such as screening job applicants or deciding eligibility for public benefits. The underlying model is regulated separately as a general-purpose AI model.
When do the high-risk obligations apply?
After the Digital Omnibus (Regulation (EU) 2026/1744), obligations for Annex III high-risk systems apply from 2 December 2027, and those for AI in Annex I regulated products from 2 August 2028.
Can an Annex III system avoid being high-risk?
Yes, under Article 6(3), if it poses no significant risk and only performs a narrow procedural, improving, pattern-detecting or preparatory task. It never applies to systems that profile natural persons, and the provider must still document the assessment and register the system.
Who needs a fundamental rights impact assessment?
Deployers of high-risk systems that are public bodies or private entities providing public services, plus any deployer of AI credit scoring or life and health insurance pricing. Critical-infrastructure systems are exempt.
What is the fine for using a prohibited AI system?
Up to €35 million or 7% of worldwide annual turnover, whichever is higher. For SMEs and start-ups, the lower of the two amounts applies.
Sources and methodology
- Regulation (EU) 2024/1689 (AI Act), consolidated version of 27 July 2026 — EUR-Lex.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) — Official Journal of the EU.
- AI Act, Article 6 — classification rules for high-risk AI systems.
- AI Act, Annex III — high-risk AI systems.
- AI Act, Article 27 — fundamental rights impact assessment.
- AI Act, Article 49 — registration.
- European Commission — targeted consultation on draft guidelines for high-risk classification (2026).
- Golpayegani, D., Pandit, H. J., & Lewis, D. (2024). AICat: An AI Cataloguing Approach to Support the EU AI Act. JURIX 2024, arXiv:2501.04014 (on registering high-risk AI systems in the EU database).
Method: on 1 October 2026 we compared six top-ranking guides on EU AI Act risk categories against 22 practical questions. A local reranking model screened them, and a language model verified each candidate answer. Legal statements were checked against the consolidated AI Act text. This article is not legal advice.

