Three Sanction Tiers, One Principle: Severity Determines the Amount
Art. 99 AI Act tiers sanctions according to the severity of the infringement. The highest tier concerns breaches of the prohibited AI practices under Art. 5: here, fines of up to EUR 35,000,000 or up to 7% of worldwide annual turnover for the preceding financial year apply – whichever is higher (Art. 99(3)).
The second tier (Art. 99(4)) covers a whole range of specifically named obligations: provider obligations under Art. 16, obligations of authorised representatives (Art. 22), importers (Art. 23) and distributors (Art. 24), deployer obligations under Art. 26, requirements for notified bodies (Art. 31, Art. 33(1), (3) and (4), Art. 34), as well as the transparency obligations for providers and deployers under Art. 50. Here, the framework is up to EUR 15,000,000 or up to 3% of worldwide annual turnover.
The third tier concerns false, incomplete or misleading information supplied to notified bodies or competent national authorities in response to their requests: up to EUR 7,500,000 or up to 1% of worldwide annual turnover (Art. 99(5)).
In practice, this means that an infringement of a prohibition under Art. 5 carries considerably more legal weight than incomplete technical documentation – the level of the fine follows this logic consistently.
SME Privilege: The Lower Value Applies
A special rule applies to small and medium-sized enterprises, including start-ups: under Art. 99(6), for each of the three fines mentioned, the lower amount – whether the percentage or the fixed sum – is applied. An SME therefore never pays more than the percentage of its turnover requires, even if the fixed sum would arithmetically be lower. This does not exempt start-ups from liability, but it does limit the economic risk compared with large enterprises.
What Authorities Take Into Account When Assessing Fines
Art. 99(7) lists the criteria that authorities must weigh in each individual case when deciding on the amount and the “whether” of a fine. These include, among others:
- the nature, gravity and duration of the infringement and the number of persons affected (lit. a)
- whether other authorities have already imposed fines for the same or a related matter (lit. b, c)
- the size, turnover and market share of the operator (lit. d)
- the degree of cooperation with the authorities to remedy the infringement (lit. f)
- intent or negligence (lit. i)
- measures already taken to mitigate the harm (lit. j)
In practice, this means that documented technical and organisational measures (lit. g) and a cooperative approach in the event of an infringement have a measurable effect on the level of the fine. Anyone who can demonstrate that they took risks seriously is in a better position than someone who only reacts under pressure.
The Practical Case: Transparency Obligations under Art. 50
Of particular relevance to many businesses is lit. g in Art. 99(4): breaches of the transparency obligations under Art. 50 fall within the second fine tier – up to EUR 15,000,000 or 3% of annual turnover. These obligations apply from 02.08.2026, with a grace period until 02.12.2026 for systems already placed on the market beforehand.
Anyone operating AI systems with labelling or information obligations – such as chatbots that must be identifiable as such, or synthetic media content that must be marked as AI-generated – should keep this deadline in mind. The grace period gives existing systems additional time, but does not change the fundamental relevance of fines from the standard deadline onwards.
Conclusion
The fine framework under Art. 99 is tiered, but sharp at the top end: EUR 35 million or 7% of worldwide turnover is no small matter. In practice, the more pressing question is often which obligations actually apply, and from when – particularly with regard to the transparency obligations under Art. 50 and their staggered start dates.
If you are not yet sure which obligations are relevant to your AI systems and what fine risk follows from them, our free risk check at /einstufung can help with an initial assessment.