Note Notes 2026
AI Act penalties: how they are calculated and who pays
Translated from the Italian original.
In May 2026 the EU AI Office published the first operational guidelines on AI Act penalties, supplemented by a joint interpretative note with the European Data Protection Board. The picture that emerges is more nuanced than press summaries suggested. The penalties of up to 35 million euros or 7% of global annual turnover, quoted everywhere, are the maximum ceiling for the most serious tier. For other violations the parameters are lower but remain substantial. And the split of liability between provider, deployer and distributor follows specific logic that Italian entrepreneurs need to know before signing contracts.
The point is that most Italian SMEs do not know which risk tier they are taking on with their AI contracts.
The three penalty tiers of Article 99
Comparative data on enforcement across EU countries is in the enterprise summary of the Stanford AI Index 2026, useful for anticipating the orientation of the Italian authorities. Article 99 of Regulation (EU) 2024/1689 distinguishes three levels of severity with proportionate penalties. The first tier is the most serious: violation of the prohibitions of Article 5 (prohibited AI practices such as social scoring, subliminal manipulation, emotion recognition in the workplace or in education, profiled predictive policing). Maximum penalty: 35 million euros or 7% of total worldwide annual turnover for the preceding financial year, whichever is higher.
The second tier covers violations of obligations on high-risk systems (Articles 8-15, 16, 22, 24, 25, 26, 49, 50) and on GPAI (Articles 53-55). Maximum penalty: 15 million euros or 3% of turnover. The third tier covers incorrect or misleading information supplied to national competent authorities or to the AI Office. Maximum penalty: 7.5 million or 1%.
To frame the specific obligations in marketing activities, the AI Act guide for digital marketing is useful. For GPAI providers there are specific penalties under Article 101: up to 15 million or 3% for violations of the Code of Practice or of specific transparency and governance obligations for foundation models.
How the reference turnover is calculated in practice
Global annual turnover is that of the entire group, not just the single entity involved in the violation. For an American multinational with global turnover of 30 billion, 7% is 2.1 billion euros of maximum theoretical exposure. For an Italian SME with consolidated turnover of 8 million, 7% is 560 thousand euros.
Article 99(1) establishes that the higher of the absolute ceiling (35 million) and the percentage of turnover applies. For large multinationals, turnover almost always applies. For Italian SMEs the absolute value applies, which is enormously higher than their turnover and therefore becomes the theoretical maximum penalty.
Paragraph 6 of the same article introduces a principle of proportionality for SMEs and startups: authorities must take size and economic viability into account when determining the actual penalty. It is interpretative guidance, not an exemption. An Italian SME may receive penalties well below the ceilings, but it can still receive penalties that are substantial for its economic structure.
Who pays in practice: provider, deployer, distributor
The AI Act distinguishes roles with different responsibilities. The provider is whoever develops, or has developed, an AI system and places it on the market or puts it into service. The deployer is whoever uses the AI system under its own authority (excluding purely personal use). The distributor is whoever makes the system available on the market without developing it or substantially modifying it. The importer is whoever introduces into the EU market systems developed outside the EU.
Primary responsibility for technical requirements (quality system, documentation, risk management, accuracy, robustness) lies with the provider. Responsibility for use in accordance with the provider’s instructions lies with the deployer. Distributor and importer have obligations of documentary verification and chain of responsibility.
Translated into concrete Italian cases. A software house selling an Italian HR AI system to a manufacturing company: the software house is the provider, the manufacturing company is the deployer. An Italian company that integrates OpenAI’s GPT-5 into its own product and sells it to customers: the company becomes provider of the integrated product, OpenAI is provider of the base model. An Italian reseller of an AI SaaS platform developed in the US: the reseller is probably a distributor or importer, with the related obligations.
When the deployer becomes a provider and doubles the risk
Article 25 of the AI Act contains a critical and often overlooked provision. A deployer that substantially modifies a high-risk AI system already placed on the market, or uses it for purposes other than those intended, becomes a provider in every respect for that system. It inherits the provider’s obligations.
For Italian companies doing fine-tuning of base models, custom RAG, advanced prompt engineering, integration with other systems, the consequence is serious. If the intervention is substantial, the company is a provider and has the provider’s obligations, not just the deployer’s. The threshold of substantiality is a matter of interpretation; the AI Office has published guidelines that provide criteria but not automatic answers.
Consequence: CE marking of the product, complete technical documentation, quality system, conformity certification. Compliance costs in the order of 60-90 thousand euros in the first year, with recurring annual costs estimated at 25-35 thousand.
Administrative penalties are not the only ones
The exposure of an Italian company that violates the AI Act is not limited to the administrative penalty. There are at least three other families of concrete consequences. Civil damages to injured parties (discriminated employees, manipulated consumers, harmed patients). Criminal penalties for specific conduct already provided for by the Criminal Code (e.g. unlawful processing of personal data, computer fraud, breach of secrecy). Exposure to national privacy authorities for concurrent GDPR violations.
Commercial and reputational fallout is often more serious than direct penalties. Exclusion from public tenders where AI Act compliance is becoming a requirement, loss of enterprise customers asking for guarantees, blocked certifications that matter for the sector of activity. For an Italian SME that has invested years building its market position, even a penalty of “only” 50 thousand euros can come with indirect consequences worth ten times as much.
The Italian penalty regime: still under construction
Italy, like all EU Member States, must adopt a national law adapting to the AI Act. The bill is under parliamentary review in 2026. The most substantial provisions concern the designation of national competent authorities (orientation toward AgID plus the Garante Privacy, the Italian Data Protection Authority, with coordination), inspection procedures, and cooperation mechanisms with European authorities.
Until the national legislation enters into force, some operational procedures remain unclear. How do you submit a serious incident notification to an authority that has not yet published operational protocols? How do you request a preliminary opinion on a system under development? The EU AI Office partly fills the gap with its own help desk, but national authorities are the first point of contact envisaged by the legislation.
In March 2026 AIPIA, the Italian association of AI professionals, which has a dedicated Legal Advisor, opened an observatory on the Italian implementation of the AI Act, with monthly monitoring of the practices emerging from the first decisions and from clarifications by the authorities.
An economically sustainable compliance strategy
Four principles that work for Italian SMEs. First, know your role in the chain (provider, deployer, distributor, importer) for every AI system in use or sold. It depends on the contract and on the modifications made, not on self-perception. Second, classify every system by AI Act risk level (prohibited, high-risk, limited-risk, minimal-risk). Only high-risk systems carry heavy documentation obligations; limited and minimal risk carry transparency obligations only.
Third, allocate compliance resources in proportion to risk. An SME with a single limited-risk system does not need a dedicated AI Compliance Officer; it can manage with an external consultant on an annual basis. An SME with three high-risk systems needs an internal structure. Fourth, build documentation as an asset, not as a formality. Well-made AI technical documentation is also the product’s operating manual, a source of internal training, a basis for technical marketing.
What the penalties will really signal
The first penalty cases under the AI Act, expected in 2027, will signal the orientation of European and national authorities. GDPR enforcement was relatively cautious in its first two years, with medium-sized penalties and more attention to education than to repression. Only from 2020-2021 did substantial penalties against large players arrive.
The AI Act will probably follow a similar trajectory. The first cases will be educational, aimed at clarifying the application framework. The most serious cases will hit structured providers with blatant violations. Italian SMEs will be hit almost exclusively for manifest violations of the Article 5 prohibitions, not for documentary imperfections on limited-risk systems.
This does not mean small businesses can ignore the legislation. To demonstrate staff qualifications in a formal and recognized way, the European credential for AI professionals is a tool that supervisory authorities recognize as qualifying evidence. It means that reasonable, proportionate, documented compliance is sufficient in most cases. Paranoid excess costs as much as negligence, just in different directions. The right measure, as is often the case, is what separates those who plan from those who react.