EU AI Act Penalties and Fines: Tiers, SME Caps and Enforcement

EU AI Act fines are set in Article 99 and reach up to €35 million or 7% of worldwide annual turnover, whichever is higher, for prohibited AI practices. Breaches of most other obligations, including high-risk and Article 50 duties, carry up to €15 million or 3%. Misleading information to authorities costs up to €7.5 million or 1%.

Key takeaways

• There are three main fine tiers for operators, plus a separate regime for general-purpose AI providers enforced by the European Commission.
• For large companies the higher of the fixed amount and the turnover percentage applies; for SMEs and start-ups it is the lower.
• National market surveillance authorities enforce most of the Act; the AI Office enforces GPAI rules.
• Authorities must weigh ten factors in Article 99(7), including cooperation, self-reporting and mitigation, before setting a fine.
• The penalties chapter has applied since 2 August 2025, so banned practices are already sanctionable as of October 2026.

How the EU AI Act penalty system works

The AI Act is an EU regulation, but most of its enforcement is national. Article 99 sets maximum fines and the criteria for using them; each Member State lays down its own rules on penalties and other enforcement measures, notifies them to the Commission and designates the authorities that apply them. The amended Article 99(1) states that these measures may include administrative fines, warnings and non-monetary measures, and must be effective, proportionate and dissuasive while taking into account the interests and economic viability of SMEs, start-ups and small mid-caps. The text is on EUR-Lex, and a readable version of Article 99 is a useful reference.

That split matters in practice. The ceilings are harmonised, but the procedure, the authority you deal with and its appetite for enforcement can differ from country to country. For a broader overview of the regulation itself, see our EU AI Act guide. To find out which obligations, and therefore which fine tiers, apply to your systems, the free EU AI Act readiness assessment works through your AI use in about 40 minutes.

EU AI Act fine tiers at a glance

"Worldwide annual turnover" means the total for the preceding financial year, at undertaking level. For a group, that can be much larger than the revenue of the subsidiary that runs the AI system.

What each tier covers in practice

Top tier: prohibited practices

The highest fines are reserved for practices the EU considers unacceptable: social scoring, manipulative or exploitative AI that causes significant harm, untargeted scraping of facial images to build recognition databases, emotion recognition at work or in education (with narrow medical and safety exceptions), certain biometric categorisation and predictive-policing uses, and real-time remote biometric identification in public spaces for law enforcement outside strict exceptions. From 2 December 2026, AI used to create non-consensual intimate imagery or child sexual abuse material joins the list. Our article on prohibited AI practices under Article 5 explains each one with examples.

An example: an employer that rolls out a video-interview tool which scores candidates' emotional states from facial expressions is not running a high-risk system with paperwork gaps. It is using a banned practice, and the exposure sits in the top tier.

Middle tier: high-risk, transparency and operator duties

Most compliance failures will land here. A provider placing a high-risk credit-scoring model on the market without a conformity assessment, a deployer using an AI recruitment tool without human oversight or log retention, or a company running a customer-service chatbot that never tells users they are talking to AI would all fall under Article 99(4). The high-risk duties apply from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I products; Article 50 has applied since 2 August 2026. See high-risk AI systems and their obligations for the detail.

Lower tier: misleading information

Providing incorrect, incomplete or misleading information in reply to a request from a notified body or a national authority is a separate infringement. It is a reminder that documentation is not a formality: an inaccurate model card, an incomplete technical file or an overstated claim in a supervisory response can be fined in its own right.

General-purpose AI providers

Providers of GPAI models answer to the Commission, acting through the AI Office, not to national authorities. Article 101 allows fines of up to €15 million or 3% of worldwide turnover for breaches such as failing to provide documentation, ignoring a request for information or not giving access for an evaluation. The GPAI obligations have applied since 2 August 2025, while the Commission's fining power under Article 101 applies from 2 August 2026.

The SME "lower of" rule

For large companies, each ceiling is "whichever is higher". For SMEs, including start-ups, Article 99(6) reverses this: the fine is capped at whichever of the percentage or the fixed amount is lower. The Digital Omnibus extended tailored penalty caps to small mid-caps (fewer than 750 staff, with turnover up to €150 million or a balance sheet up to €129 million).

A worked example: a company with €40 million in annual turnover breaches a high-risk obligation. Under the middle tier, 3% of turnover is €1.2 million and the fixed amount is €15 million. If the company is large, the ceiling is €15 million. If it qualifies for the SME rule, the ceiling is €1.2 million. Those are maximums, not likely outcomes; the actual fine depends on the factors below. Check the consolidated Article 99(6) for the exact wording as it applies to small mid-caps.

What authorities consider when setting a fine (Article 99(7))

Authorities must take into account all relevant circumstances and give due regard, as appropriate, to these factors:

1. The nature, gravity and duration of the infringement and its consequences, taking into account the purpose of the AI system, the number of people affected and the level of damage they suffered.
2. Whether other market surveillance authorities have already fined the same operator for the same infringement.
3. Whether other authorities have already fined the same operator for infringements of other EU or national law arising from the same activity, for example a GDPR fine.
4. The size, annual turnover and market share of the operator.
5. Other aggravating or mitigating factors, such as financial benefits gained or losses avoided through the infringement.
6. The degree of cooperation with national authorities to remedy the infringement and mitigate its effects.
7. The degree of responsibility of the operator, taking into account its technical and organisational measures.
8. How the authorities found out, in particular whether the operator notified the infringement and to what extent.
9. Whether the infringement was intentional or negligent.
10. Any action the operator took to mitigate the harm suffered by affected persons.

Several of these factors reward preparation. An organisation that can show a risk assessment, documented human oversight, staff training and a record of acting quickly on incidents is in a far better position than one that cannot. That is the practical case for governance work now, well before high-risk deadlines.

Who enforces the EU AI Act?

• National market surveillance authorities supervise and enforce the rules for AI systems in each Member State. Each country also designates a notifying authority for conformity assessment bodies. In some countries an existing regulator, such as a data protection authority or a product-safety body, takes on the role; others have created or extended a digital regulator.
• The AI Office, within the European Commission, supervises GPAI model providers. Since the Digital Omnibus it also has exclusive supervision of AI systems built on a GPAI model by the same provider or group, and of AI integrated in very large online platforms and search engines.
• The European Data Protection Supervisor supervises EU institutions, bodies and agencies.
• Sector regulators keep their roles where AI is part of a regulated product (Annex I) or a regulated financial service.

Non-EU companies are not outside this system. Providers established outside the EU must appoint an authorised representative for high-risk systems, and authorities can direct enforcement at that representative. Our guide on how the EU AI Act applies to non-EU companies covers this in detail.

How to reduce your exposure

1. Rule out prohibited uses first. They carry the largest fines and are already enforceable. Review emotion recognition, biometric and behavioural-scoring features in tools you buy.
2. Build an AI inventory with roles and tiers. You cannot estimate exposure for systems you do not know about.
3. Close Article 50 gaps. Chatbot disclosures and deepfake labelling are cheap to fix and have applied since August 2026.
4. Document as you go. Risk assessments, oversight assignments, training records and vendor due diligence support the Article 99(7) mitigating factors.
5. Prepare an incident and authority-response process. Accurate, prompt answers avoid the separate misleading-information tier.

A structured EU AI Act gap analysis is the fastest way to work through these steps in order.

Frequently asked questions

What is the maximum penalty for violations under the AI Act?

The maximum is €35 million or 7% of total worldwide annual turnover for the preceding financial year, whichever is higher, for engaging in a prohibited AI practice under Article 5. For SMEs and start-ups the lower of the two figures applies. Fines on EU institutions are capped separately at €1.5 million.

Is the EU AI Act being enforced?

Yes. The penalties chapter has applied since 2 August 2025, so national authorities can sanction prohibited practices, and since 2 August 2026 they can also act on Article 50 transparency breaches. The Commission can fine GPAI providers from 2 August 2026. Enforcement activity and national procedures vary by Member State, so check your local authority's guidance.

Can deployers be fined, or only AI providers?

Deployers can be fined. Breaches of deployer obligations under Article 26, such as failing to ensure human oversight, keep logs or inform workers before using high-risk AI in the workplace, fall in the €15 million or 3% tier. Deployers using a prohibited practice face the top tier, regardless of who built the tool.

Can a company be fined under both the AI Act and GDPR?

Yes, the regimes are separate, and one AI system can breach both. Article 99(7) requires authorities to consider fines already imposed under other laws for the same activity, which helps avoid disproportionate double penalties but does not rule out two fines. Personal-data issues in AI training and use remain subject to GDPR's own fine ceilings.

Do the fines apply to companies outside the EU?

Yes. The AI Act applies to providers placing AI on the EU market and to non-EU providers and deployers whose AI output is used in the EU. Fines are calculated on worldwide turnover, and authorities can address the EU authorised representative that non-EU providers of high-risk systems must appoint.

Find your gaps before an authority does

The factors that reduce fines are the same ones a good governance programme produces. Answer 14 steps of questions and get your AI governance documents written: an AI Governance Framework, AI Acceptable Use Policy, AI Vendor Due Diligence Checklist and AI Risk Assessment Template, in editable Word, plus a maturity score across ten governance areas. It is free and takes about 40 minutes.

This article is general information, not legal advice. Check the consolidated text of Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744. Statements reflect the position as of October 2026.