EU AI Act Fines and Penalties: Article 99 Tiers, SME Caps and Worked Examples

EU AI Act fines are capped in three tiers by Article 99: up to €35 million or 7% of total worldwide annual turnover, whichever is higher, for prohibited AI practices; up to €15 million or 3% for breaches of operator and transparency duties; and up to €7.5 million or 1% for misleading information to authorities. For SMEs, the lower of the two figures applies.

Those numbers are ceilings, not tariffs. Since the Digital Omnibus on AI, Regulation (EU) 2026/1744, small mid-cap companies also get the lower cap in the second and third tiers, and providers of general-purpose AI models face a separate regime run by the European Commission.

Every figure below was checked on 14 October 2026 against the consolidated text of Regulation (EU) 2024/1689 on EUR-Lex, which incorporates the Omnibus. This is general information about the Regulation, not legal advice. For the dates on which each obligation starts, see our deployer readiness guide to the phased dates; to see which obligations reach a given system, run it through the EU AI Act checker.

EU AI Act penalties for non-compliance: the three tiers

Turnover here means total worldwide annual turnover for the preceding financial year, not revenue earned in the EU. The percentages apply only “if the offender is an undertaking”; for anyone else, the fixed amount is the ceiling.

The second tier is a closed list in Article 99(4): provider obligations under Article 16, which take in the high-risk requirements of Articles 8 to 15, documentation, logs, conformity assessment, CE marking and registration; authorised representatives (Article 22); importers (23); distributors (24); the Article 25(2) and (4) duties added as point (da); deployers (26); notified bodies; and the Article 50 transparency duties.

Two duties many organisations are working on are not on that list: Article 4 on AI literacy and Article 27 on fundamental rights impact assessments. Article 99 sets no specific ceiling for them, so their treatment depends on national rules under Article 99(1). The exception is the group of systems the AI Office supervises directly, where Article 75c(4) makes breaches of provisions missing from the Article 99(4) list subject to second-tier fines as well.

What the Digital Omnibus changed

Regulation (EU) 2026/1744, in force since 27 July 2026, amended Article 99 in three places:

• Paragraph 1 now names administrative fines alongside warnings and non-monetary measures, covers any infringement by operators, and adds small mid-caps to the interests Member States must weigh.
• Point (da) brings Article 25(2) and (4) into the second tier: an original provider’s duty to cooperate with a new provider of a high-risk system, and the written agreement between a high-risk provider and the suppliers of models, tools and components it builds in.
• Paragraph 6a gives small mid-caps the lower cap in the second and third tiers.

New Articles 75a to 75d also let the AI Office fine some providers directly, as described below.

Penalties for SMEs and small mid-caps

Article 99(6) reverses the rule for smaller companies. For SMEs, including start-ups, each fine is capped at the fixed amount or the percentage, whichever is lower, in all three tiers. Paragraph 6a does the same for small mid-cap enterprises in the second and third tiers only, so a small mid-cap that breaches an Article 5 prohibition still faces the full €35 million or 7% ceiling.

Since the Omnibus, the Act defines both categories by reference to Commission recommendations:

Both recommendations bring in the figures of partner and linked enterprises, so the EU subsidiary of a large group is unlikely to qualify however small it is on its own. For a genuine SME the percentage is always the lower figure, so its ceilings work out at 7%, 3% and 1% of turnover.

Worked examples at four company sizes

The companies are illustrative; an actual fine is set case by case inside these ceilings.

Two patterns stand out. The percentage overtakes the fixed amount at €500 million of turnover in the first two tiers and at €750 million in the third, so below those points the flat euro figure is the ceiling. And the Omnibus helps company C in two tiers, but not in the one that carries the largest fines.

Whose turnover counts

The AI Act does not define “undertaking”. The GDPR, whose fines are also capped as a share of worldwide turnover, says in recital 150 that the term follows EU competition law, and in Deutsche Wohnen (C-807/21, December 2023) the Court of Justice calculated a GDPR fine’s ceiling on the turnover of the undertaking in that sense, an economic unit that can span several companies. The AI Act has no matching recital, so the point is open, but a subsidiary should not assume only its own revenue will count.

How the amount is set in each case

Article 99(7) lists what an authority must take into account, as appropriate, when deciding whether to fine and how much:

1. the nature, gravity and duration of the infringement, including the system’s purpose, the people affected and the damage suffered;
2. fines other market surveillance authorities have already imposed for the same infringement;
3. fines under other law for the same activity or omission;
4. the operator’s size, annual turnover and market share;
5. other aggravating or mitigating factors, such as financial benefits gained or losses avoided;
6. cooperation with the authorities to remedy the infringement;
7. the operator’s degree of responsibility, given its technical and organisational measures;
8. how the infringement came to light, in particular whether the operator reported it;
9. intent or negligence;
10. action taken to mitigate harm to those affected.

Several of these are within an organisation’s control long before anything goes wrong. Points 7, 8 and 10 reward documented controls, self-reporting and remediation, and point 3 lets a GDPR fine for the same conduct be weighed. Paragraph 10 adds procedural safeguards, effective judicial remedies and due process.

Who imposes the fines

National authorities, under national law. Article 99(1) leaves the penalty rules to each Member State, which notifies them to the Commission and reports its fines every year. Each country designates at least one market surveillance authority under Article 70, and depending on its legal system, fines may be imposed by those bodies or by national courts (99(9)). Member States also decide how far their own public authorities can be fined (99(8)), and under Article 85 anyone can complain to a market surveillance authority.

The European Data Protection Supervisor, for EU bodies. Under Article 100, the EDPS can fine Union institutions, bodies, offices and agencies up to €1.5 million for a prohibited practice and up to €750,000 for other non-compliance, with their annual budget among the factors.

The AI Office, for some systems. Since the Omnibus, Article 75(1) gives the AI Office exclusive competence over AI systems built on a general-purpose model where the same provider, or its group, develops both, and over AI systems that are or sit inside very large online platforms and search engines designated under the Digital Services Act. Systems linked to Annex I products, critical infrastructure, the administration of justice and some law enforcement, border and financial uses are carved out of the first group. For these systems, Article 75c lets the AI Office fine at the Article 99 levels and add periodic penalty payments of up to 5% of average daily income or worldwide annual turnover in the preceding financial year, per day, subject to Court of Justice review and a five-year limitation period.

The Commission, for model providers. Fines on providers of general-purpose AI models follow Article 101, covered next.

Fines on general-purpose AI model providers

Article 101 gives the Commission its own power over providers of general-purpose AI models. It can fine a provider up to 3% of annual total worldwide turnover in the preceding financial year or €15 million, whichever is higher, where the provider intentionally or negligently:

• infringed the relevant provisions of the Regulation;
• failed to comply with a request for a document or information under Article 91, or supplied incorrect, incomplete or misleading information;
• failed to comply with a measure requested under Article 93;
• failed to give the Commission access to the model, or to a model with systemic risk, for an evaluation under Article 92.

The Commission must share its preliminary findings and hear the provider first (101(2)), and it takes account of commitments made under Article 93(3) or in codes of practice under Article 56. The Court of Justice has unlimited jurisdiction to cancel, reduce or increase the fine (101(5)). Article 101 contains no SME rule of its own.

Using another company’s model through an API does not make you that model’s provider. It makes you the provider or deployer of your own system, and any fine for that system falls under Article 99.

From when each penalty applies

A fine needs two things: a penalty rule in force and an obligation that already applies. The penalties chapter came first, and several of the duties it backs arrive later.

In practice, the original eight bans have been backed by the penalties chapter since 2 August 2025, six months after they started to apply. Two transition rules soften the high-risk rows: high-risk systems already on the market before the relevant date are caught only if their design changes significantly afterwards, and systems intended for public authorities must comply by 2 August 2030 regardless (Article 111(2)).

Keeping your exposure low

Most of the exposure sits in duties you can already evidence. A practical order:

1. Screen for prohibited practices. The top tier concerns Article 5 alone, so check every system against the bans, including the two added from 2 December 2026. Our overview of the Article 5 bans includes a screening checklist.
2. Fix transparency now. Article 50 sits in the second tier and already applies; our note on Article 50 for chatbots and agents covers disclosure and marking.
3. Record your role for each system. Provider, deployer, importer and distributor duties are fined separately, and a deployer that repurposes a system into a high-risk use becomes its provider under Article 25(1).
4. Put supplier terms in writing. Point (da) makes the Article 25(4) agreement between a high-risk provider and its suppliers a second-tier duty.
5. Prepare answers you can stand behind. Misleading replies to an authority are a separate third-tier breach, so keep documentation, logs and decisions in a form you can produce accurately; an evidence pack is a practical format.
6. Plan impact assessments where they apply. Public bodies, private entities providing public services, and deployers of credit-scoring and life and health insurance systems must assess fundamental-rights impact before deploying a high-risk system; see our FRIA template.
7. Check your reach from outside the EU. The Act covers non-EU providers and deployers whose system output is used in the EU; our guide for US companies explains the scope.

How VDF AI fits

Several of the Article 99(7) factors turn on what you can show: the measures you had in place, how quickly you noticed and reported a problem, and what you did about it. VDF AI is a self-hosted platform, so in an on-premises or air-gapped deployment the audit logs and approval records it produces stay in infrastructure you control, ready to hand over when an authority asks.

VDF AI Networks provides Human Approval nodes for putting review steps into agent workflows, role-based access control is available on every plan, and VDF AI Router registers on-premises model deployments alongside the cloud models you approve, so the set of models in use is a known list. None of this classifies a system, decides your role or makes you compliant; that work stays with your legal and compliance teams.