EU AI Act risk tiers, explained.

The EU AI Act does not regulate "AI" as one thing. It sorts AI systems into four tiers by the risk their use poses to people, and the tier decides what you have to do and by when. Most products land in the bottom two. The work is knowing which one is yours, and being able to show why.

The tier follows the use case, not the technology. The same model can be minimal risk in a spam filter and high-risk when it ranks job applicants. Run one system through the free 2-minute scan to see where it likely lands.

The four tiers at a glance

Unacceptable risk: banned practices

Article 5 lists practices the Act treats as a clear threat to safety or fundamental rights. They include social scoring that leads to unfair treatment, emotion recognition in the workplace and in education (except for medical or safety reasons), untargeted scraping of facial images to build recognition databases, manipulation that exploits people's vulnerabilities, and predicting crime based solely on profiling. Real-time remote biometric identification in public spaces for law enforcement is banned too, with narrow exceptions.

There is no compliance route for a prohibited practice: no documentation makes it lawful. The bans have applied since 2 February 2025, and regulators have been able to fine for them since 2 August 2026. If a feature comes close to one of these, it is worth a qualified review before launch.

High risk: allowed, with the heaviest obligations

A system is high-risk in one of two ways. Either it is used in one of the eight areas listed in Annex III, such as recruitment and worker management, creditworthiness, education, access to essential services, biometrics or critical infrastructure. Or it is a safety component of a product already covered by EU product law, such as machinery or medical devices (Annex I).

Providers of high-risk systems carry most of the work: a risk management system, data governance, technical documentation, logging, human oversight, accuracy and cybersecurity, a quality management system, a conformity assessment, CE marking and registration in the EU database. Deployers have their own, lighter duties, such as using the system according to its instructions and assigning human oversight.

After the June 2026 Omnibus, these obligations apply from 2 December 2027 for Annex III systems and from August 2028 for Annex I products. Deferred is not distant: the conformity work realistically takes 12 to 18 months.

Limited risk: tell people when AI is involved

Article 50 covers AI that interacts with people or generates content. The duties are about transparency, not about how the system is built:

• Chatbots and voice agents must tell people they are dealing with an AI, unless that is obvious. This has applied since 2 August 2026.
• Providers of generative systems must mark synthetic text, images, audio and video in a machine-readable way, from 2 December 2026.
• Deployers must disclose deepfakes, and AI-generated text published to inform the public on matters of public interest, unless a person reviewed it and holds editorial responsibility.

For most chatbots this is a line of copy in the greeting. The Article 50 guide covers wording and placement, and this one covers when a chatbot is more than limited risk.

Minimal risk: AI literacy, and little else

Spam filters, product recommendations, search and most internal tools sit here. The Act adds no obligations specific to this tier. One duty applies whatever the tier: under Article 4, providers and deployers must take measures to ensure a sufficient level of AI literacy among the people who build or use the system. That has applied since 2 February 2025. Other law, such as the GDPR, applies as usual.

How your tier is decided

Four questions, in this order, settle most cases.

1. What is your role? Providers, who develop a system or put it on the market under their own name, carry most of the obligations. Deployers, who use a system under their own authority, carry fewer. Provider or deployer?
2. What is the system used for? The intended purpose decides the tier, not the model or the vendor. Check it against the Article 5 bans first: if a practice is prohibited, nothing else matters.
3. Is that use listed in Annex III, or is the AI a safety component of a regulated product? If neither, the system is not high-risk.
4. Does the Article 6(3) exemption apply? An Annex III system is not high-risk if it only performs a narrow procedural or preparatory task, improves the result of work a person has already done, or detects patterns in decisions without replacing human assessment, and it does not profile people. A provider that relies on the exemption must document that assessment. How the exemption works.

Then check Article 50 separately. Its transparency duties apply on top of any tier, so a recruitment chatbot can be high-risk and still owe the chatbot disclosure.

General-purpose AI models: a separate track

General-purpose AI models, the large models behind most AI products, are regulated separately from the four tiers. Their providers have duties under Article 53, including technical documentation, information for the companies that build on the model, a copyright policy and a summary of the training content. These have applied since 2 August 2025.

If you call a model through an API, or build on it with prompts and retrieval, you are very likely not its provider. Your own product is classified in the four tiers like any other AI system. A substantial modification of the model is where that can change.

What happens if you get it wrong

Fines depend on which duty was breached. Under Article 99, prohibited practices carry fines of up to €35 million or 7% of worldwide annual turnover; most other operator duties, including Article 50 and the high-risk obligations, up to €15 million or 3%; supplying incorrect or misleading information to authorities, up to €7.5 million or 1%. For large companies the higher of the two amounts is the ceiling; for SMEs and start-ups, the lower (Article 99(6)).

These are ceilings, not tariffs: authorities weigh the gravity of the breach, whether it was intentional or negligent, and what you did to mitigate it. A dated, documented self-assessment is likely a better position than having never looked. The penalties guide works through real numbers.

Frequently asked questions

How many risk tiers does the EU AI Act have?

Four for AI systems: unacceptable (prohibited), high, limited and minimal. General-purpose AI models are regulated on a separate track under Article 53, alongside the tiers rather than inside them.

Who decides which tier my AI system is in?

In most cases you do. There is no pre-approval: providers classify their own systems, and many high-risk systems are also self-assessed, with only some needing a notified body. Market surveillance authorities can check the classification later, so it is worth writing your reasoning down.

Can one AI system fall into more than one tier?

Yes. The Article 50 transparency duties sit on top of the tiers, so a high-risk recruitment tool with a chatbot interface likely owes both the high-risk obligations and the chatbot disclosure. Different features of one product can also land in different tiers.

Is a customer service chatbot high-risk?

Usually not. A chatbot answering customer questions is typically limited risk: it must tell people they are talking to an AI. It can become high-risk if it is used for an Annex III purpose, such as screening job applicants or deciding access to essential services.

Does minimal risk mean no obligations at all?

Not quite. Article 4 AI literacy applies whatever the tier, and has since 2 February 2025. Other laws, such as the GDPR and consumer protection rules, apply as usual.

When do the high-risk obligations apply?

From 2 December 2027 for Annex III systems and from August 2028 for AI in regulated products (Annex I), after the June 2026 Omnibus deferred them. The Article 5 bans and the Article 50 transparency duties were not deferred.