How high are the fines under the EU AI Act? Up to 35 million euros or 7 percent of total worldwide annual turnover, whichever is higher, for the most serious breaches. For small and medium companies, Article 99(6) softens this: each fine is capped at the lower of the two figures, not the higher.
Updated: 2026-09-15
The fine is the first thing a management team asks about once the AI Act lands on the table. The figure of 35 million circulates in every headline, but it describes the extreme case, not the daily reality of a company with fifty employees. Knowing the three tiers and when each one bites lets you size your own exposure realistically instead of fearing it.
The three fine tiers in Article 99
The AI Act grades its fines by the severity of the breach. The amounts sit in Article 99 of Regulation (EU) 2024/1689, the AI Act itself, and are explained in the European Commission guide to Article 99.
The top tier hits the prohibited practices in Article 5, such as inferring emotions in the workplace or social scoring. Here it is up to 35 million euros or 7 percent of total worldwide annual turnover for the preceding financial year, whichever is higher.
The middle tier applies to breaches of most other obligations: the provider and deployer duties for high-risk systems, the transparency duties, and the obligations of notified bodies. The ceiling is up to 15 million euros or 3 percent of worldwide turnover, again the higher of the two.
The bottom tier covers incorrect, incomplete, or misleading information supplied to the authorities. It carries up to 7.5 million euros or 1 percent of turnover. Answer a market surveillance request carelessly or evasively, and you risk a fine of your own, separate from the underlying breach. This tier is easy to overlook in the debate, yet it is the one a company is most likely to touch first.
The words up to matter. All three figures are ceilings, not standard amounts. Within the range the authority sets the level under Article 99(7): the nature, gravity, and duration of the breach, whether it was intentional or negligent, how cooperative the company was, and its size. A one-off lapse fixed quickly sits at the low end, a deliberate ongoing breach at the top. Treat the headline maximum as the default and you overestimate your exposure systematically.
Since when fines can apply
The power to fine has applied since 2 August 2025. Article 113 of the AI Act settles this: the chapter on penalties applies from that date, and member states had to lay down their national penalty rules and notify the Commission by then (Article 113, European Commission).
The distinction that matters is between the power and the trigger. A fine presupposes the breach of an obligation that is itself already in force. The prohibited practices in Article 5 have been banned since 2 February 2025, so they can be sanctioned since August 2025. The transparency duties in Article 50 apply since 2 August 2026. The heavier obligations for high-risk systems under Annex III, by contrast, were postponed to 2 December 2027 by the Digital Omnibus, Regulation (EU) 2026/1744, as we set out in our analysis of the high-risk deadlines. Where there is no obligation yet, there is no fine.
That eases the picture for smaller firms considerably. The great majority of obligations that carry a fine risk concern high-risk systems, and most companies run no such systems at all. What is already enforceable today is the prohibited practices and, since August 2026, the transparency duty for chatbots and AI-generated content. Everything else builds up only over the next two years.
What is different for SMEs
Small and medium companies and startups get their own rule. Under Article 99(6), the fine for them is capped at the lower of the two figures, not the higher. Large undertakings pay the higher of fixed amount and percentage, an SME the lower. The difference in wording is a single word; the difference in effect is enormous.
A worked example makes it concrete. Take a company with 4 million euros in annual turnover: 7 percent of that is 280,000 euros. Because the lower value applies to it, that is the ceiling, not the 35 million from the headline. The sanction stays tied to the actual size of the business, which, based on our reading, is exactly the point of the rule. For a group with turnover in the billions the relationship flips, 7 percent runs far above 35 million, and the percentage governs.
A cap is still not a free pass. Even 280,000 euros hits a mid-sized firm hard, and within the ceiling the authority sets the amount by severity, duration, and intent. The SME rule lowers the maximum; it does not make the breach disappear. Deploy a prohibited practice and you remain in the most expensive category, only with a limit scaled to your size.
Who imposes the fines
The fines are imposed not by Brussels but by national authorities. In Spain that is the Agencia Espanola de Supervision de la Inteligencia Artificial, AESIA, created back in 2023 by Real Decreto 729/2023 and therefore in place earlier than in most member states. In Germany the government draft of a market surveillance and innovation act, adopted by the federal cabinet in February 2026, names the Bundesnetzagentur as the central market surveillance authority. That draft is not yet in force, so the exact structure may still change. The pattern across the EU is the same shape: one or more designated national bodies enforce the rules on their territory.
Article 101 is the exception. Fines against providers of general-purpose AI models, the large language models, are imposed by the European Commission itself, with a ceiling of up to 15 million euros or 3 percent. That rule applies since 2 August 2026 (Article 101, European Commission). For a company that merely uses a finished or open model, it is not relevant: Article 101 targets the makers of the models, not their users.
How to reduce your exposure
Fine risk drops not through worry but through two sober steps. The first is classifying your role correctly. A company that only uses a system is a deployer, with a manageable set of duties; one that rebuilds it or passes it on under its own name becomes a provider. We covered that fork in detail in deployer or provider, and it decides which of the sanctionable obligations apply to you at all.
The second step is documentation. The bottom fine tier targets precisely incorrect or incomplete information given to the authority. A company that documents cleanly which systems run in house, what they are used for, and who is responsible answers a market surveillance request calmly and with evidence. Local AI on your own server often makes that proof easier, because data flows and model provenance stay in house rather than needing to be reconstructed from an external service.
The cheapest place to start is checking the prohibited practices in Article 5, since they sit in the most expensive tier and already apply. Which of your systems could even fall under them is usually settled quickly, and the answer, in our experience, is none. That leaves transparency for chatbots and published AI content as the main obligation you can be fined for right now.
This article sets out the legal position based on our reading; it does not replace individual legal advice. What the fine framework means for your specific business is best clarified against your actual systems in a pilot project.
Frequently asked questions
How high are the fines under the EU AI Act?
Up to 35 million euros or 7 percent of total worldwide annual turnover for the prohibited practices in Article 5, up to 15 million euros or 3 percent for most other obligations, and up to 7.5 million euros or 1 percent for supplying incorrect information to authorities.
Do smaller companies face a lower ceiling?
Yes. Under Article 99(6), for SMEs and startups each fine is capped at the lower of the fixed amount or the percentage, while large undertakings pay the higher of the two. At a small turnover the percentage usually applies.
Since when can the AI Act impose fines?
The power to fine has applied since 2 August 2025 under Article 113. A fine still requires a breach of an obligation that is itself already in force, such as the ban on certain practices since 2 February 2025.
Who imposes the fines?
National market surveillance authorities in each member state, for example AESIA in Spain. Fines against providers of general-purpose AI models are imposed by the European Commission under Article 101 instead.