AI Act: penalties and enforcement

The sanctions of the AI Act make news for maximum amounts, up to 35 million euros or 7% of the world's turnover for prohibited practices (Article 99 of Regulation (EU) 2024/1689). For a normal Italian company though the concrete risk is another, and reasoning only on figures leads to wrong decisions: one buys conformity where it is not needed and neglects where it would be needed.

The sanctioning system is graded like the rest of the regulation: The more severe the violation, the higher the band. At the top are prohibited practices; at the base, inaccurate or misleading information provided to authorities. And for small businesses the cap is calculated backwards from large ones.

The amounts of Article 99, band by band

The sanctioning logic follows that of the risk classification, explained in the four levels of risk. These are the maximums set by the Article 99 of the AI Act, with the modifications of the Digital Omnibus on AI.

For businesses the mechanism is reminiscent of the GDPR: the highest value applies between fixed figure and percentage, so for very large companies it bites the percentage. For SMEs and start-ups the rule overturns and is worth the lowest value (Article 99(6)). In one example: a company with €2 million in turnover that violates the transparency of Article 50 risks at most 3%, i.e. €60,000, not €15 million.

What changed the Omnibus about sanctions

The Regulation (EU) 2026/1744, in force since 27 July 2026, changed Article 99 in three areas. When setting penalties, Member States must also take account of the interests and economic viability of small mid-cap enterprises (paragraph 1). For these enterprises, the lower of the fixed amount and the percentage applies to the bands in paragraphs 4 and 5, but not to prohibited practices (new paragraph 6 bis). The 15 million band also covers cooperation obligations along the value chain under Article 25, paragraphs 2 and 4 (new point d bis).

The postponement of the high risk also shifts the time when those obligations become punishable: for the systems of Annex III from 2 December 2027, for those of Annex I from 2 August 2028. Prohibitions and transparency remain where they were: from 2 February 2025 and from 2 August 2026 (Article 113 of the AI Act). All dates are in AI Act schedule.

Who watches

The control system is two-tiered. At European level, the Commission's AI office, which supervises model providers for general purposes and, after the Omnibus, also has exclusive competence over systems built on those models by the same supplier (Article 75(1), replaced by Reg. (EU) 2026/1744). For companies that use those systems only, national authorities remain competent.

In Italy the Law 132/2025 has designated the National Cybersecurity Agency (NCA) as a market surveillance authority, with inspection and sanctioning activities, and the Agency for Digital Italy (AgID) as a notifying authority, dealing with bodies responsible for verifying compliance. For banks, intermediaries and insurances remain Bank of Italy, CONSOB and IVASS; the Privacy Guarantor retains all its powers.

The power to apply the sanctions of Article 99 comes with the legislative decrees that the same law has delegated to the Government, to be adopted within twelve months of its entry into force (Art. 24). If you need to know to whom you would answer in concrete, check on Normattiva if those decrees have been published.

What is the real risk for an SME

Here a distinction must be made which is usually lacking.

The right column is the one that really materializes, and it is also the one that no one tells. The typical cost of non-compliance for an SME is not a fine: it is a skipping contract. More and more B2B chapters and more and more structured customers are asking to state which AI systems are being used and with what guarantees. Not having a ready answer is a business problem before sanctioning.

Three situations and the band that risk

Management software provider that integrates a scoring. The company sells to other businesses a form that assigns scores to people, for example for credit or selection. It is a supplier of an Annex III system: the obligations apply from 2 December 2027 (Reg. (EU) 2026/1744) and violating them falls into the band up to 15 million or 3%. The risk, however, comes first: to the first structured customer who requests the documentation of compliance during the tender phase, the absence results in lost sale.

Company using a CV screening tool. A discarded candidate asks on what basis the decision was made. The company is user: from 2 December 2027 will have the obligations of Article 26, including human surveillance and information to people, punishable in the same band (Article 99(4)). If it cannot prove that a competent person has been able to evaluate and flip, the position is weak, regardless of who wrote the software.

Small e-commerce with assistant on site. The obligation is transparency, from August 2, 2026. It is not the lowest range: the violation of Article 50 lies in the same range of obligations for high risk, up to 15 million or 3%, and for an SME it is worth the lowest of the two (Article 99(4) and (6)). Fulfillment costs a quarter of an hour: it is the classic case in which the relationship between compliance cost and risk makes sense to postpone.

How it proves to be in order

The moment someone asks for an account, whether it's a client in the race, an interested party contesting or an authority verifying, what matters is not what you've done, but what you can show.

The minimum documentation that makes an SME defensible is lighter than feared. Time estimates are indicative, for a company with few tools:

An overall working day, updated once a year. It is disproportionately little compared to the value it has at the time it is needed: in most cases it never serves an authority, but it does so when a structured customer sends their supplier questionnaire.

The front that strikes first: the supplier questionnaires

Many Italian SMEs will encounter the AI Act not through an authority, but through a module. More and more structured companies include questions in the supplier qualification processes such as: "Do you use artificial intelligence systems in the delivery of the service?", "What guarantees do you offer on data processing?", "Do you have an internal policy on the use of AI?".

These are questions that must be answered in the commercial phase, with short times and no margin to improvise. Those with the inventory and policy respond in ten minutes. Who doesn't have them, either loses time or loses the race.

• Prepare a standard half-page answer to typical questions
• Keep policy and inventory ready in shareable format
• Be precise about what you do and what you don't do: generic answers generate more questions
• If you use third-party providers, know which and with what guarantees

It is the reason why it is better to do this work now and not at the deadline: the commercial deadline comes before that regulation. The complete framework is in AI Act: The Business Guide and the contractual part in contract with an AI provider.

Five Risk Assessment Mistakes Penalties

• Deciding based on the maximum amount. It leads to overinvesting where it is not needed and ignoring the trivial obligations that really apply.
• Reading the best as if it applies to everyone. For SMEs and start-ups the cap is the lowest between fixed figure and percentage of turnover.
• Assuming that only the provider is liable. The obligations of the user are autonomous.
• Ignoring the contract front. For many SMEs compliance becomes a business requirement before a controlled obligation.
• Not writing anything in the contract with the supplier. See contract with an AI provider: that's where the responsibilities are distributed.

Five moves to be defensible

The most effective precaution measure is not a compliance budget: it is Being able to prove what you do.

1. Keep inventory of AI systems in use, up-to-date and dated.
2. For each note risk range and role (supplier or user).
3. Where AI talks to someone, it retains evidence of having declared the automatic nature.
4. Where AI supports decisions about people, it documents who exercises human surveillance and with what power.
5. In contracts with suppliers, ask in writing for guarantees of compliance and who answers for what.

These five points cost a few hours and cover both the sanctioning and, more useful, the commercial front. The method of assessment is in How to Assess the Risks of an AI System.

The real risk is not knowing how to answer

The penalties of the AI Act are serious but proportionate: the maximums they make relate to behaviour that a normal SME does not keep, and for small businesses the cap is still the lowest value between fixed figure and percentage (Article 99(6)). Building your own strategy on the maximum amount means spending badly.

The correct way to look at risk is to ask yourself: if tomorrow someone asked me what AI systems I use, for what, and with what guarantees, could I answer? If the answer is yes, you are covered on the front that matters. If it is no, the problem will present itself first as a lost contract than as a sanction.

If you want to get to the next supplier questionnaire with inventory, classification and policies ready, we can prepare them together starting from a AI consulting for SMEs. The general picture is in AI Act: The Business Guide.