The EU AI Act for US companies

The AI Act attaches to AI systems and models and to where they are used. Article 2(1)(a) covers providers that place AI systems or general-purpose AI models on the EU market, or put them into service there, "irrespective of whether those providers are established or located within the Union or in a third country".¹

Article 2(1)(c) reaches further. It also covers providers and deployers established outside the Union "where the output produced by the AI system is used in the Union".² A US company running a model on American servers can therefore fall within scope when its predictions, recommendations, or generated content reach users in Italy.

Most obligations depend on the role a company plays for each system. A provider develops an AI system or model, or has one developed, and places it on the market under its own name or trademark. A deployer uses an AI system under its authority in a professional capacity.³ A software company is often a provider for its own product and a deployer for the tools its staff use every day.

The Act entered into force on August 1, 2024, and applies in stages.⁴ On July 24, 2026, the Official Journal published Regulation (EU) 2026/1744, the Digital Omnibus on AI. It entered into force on July 27, 2026, and moved several of those stages.⁵ Guidance written before that summer is out of date on high-risk systems.

The general date of application stayed August 2, 2026. The high-risk rules moved to December 2, 2027, for the stand-alone systems listed in Annex III, and to August 2, 2028, for AI built into products covered by the EU harmonization laws listed in Annex I.⁶

Article 4 has applied since February 2, 2025, to every provider and deployer of an AI system, whatever the risk level. For many US companies it is the first obligation they meet in Europe, because it concerns the staff who use ordinary AI tools.

The Omnibus rewrote the article. Providers and deployers must now "take measures to support the development of AI literacy" of their staff and of other people who operate or use AI systems on their behalf. The new text adds that this obligation "does not require providers or deployers to guarantee any specific level of AI literacy of any individual".⁷ The Commission and the Member States support these efforts, in particular for small and medium-sized enterprises, and the Commission publishes practical examples.

In practice, a company with staff in Italy should be able to show what it did: who was trained, on what, and when. That record is what AIPIA's training produces. AIPIA is the first Italian AI association to issue European Digital Credentials, the European Commission's format for verifiable credentials, sealed under eIDAS and checkable on Europass, so a client or an authority can check each person's credential directly.

Article 5 lists practices banned since February 2, 2025. They include manipulative or deceptive techniques that distort behavior and cause significant harm; exploiting vulnerabilities linked to age, disability, or a social or economic situation; social scoring; predicting criminal offenses solely from profiling; building facial recognition databases by untargeted scraping; inferring emotions in workplaces and schools; and biometric categorization to infer sensitive traits such as political opinions or sexual orientation.⁸

The Omnibus added two prohibitions that apply from December 2, 2026. The first covers AI systems that generate realistic intimate or sexually explicit images of an identifiable person without explicit consent. The second covers systems that generate child sexual abuse material.⁹

Chapter V has applied since August 2, 2025, to providers of general-purpose AI models: models that display significant generality, perform a wide range of distinct tasks, and can be integrated into many downstream systems.¹⁰ Most large language models offered through an API fall here.

Article 53 requires technical documentation for the AI Office and national authorities, information for the companies that build on the model, a policy to comply with EU copyright law, and a public summary of the content used for training.¹¹ Models with systemic risk carry further duties under Article 55: evaluation with adversarial testing, assessment and mitigation of systemic risks, incident reporting, and cybersecurity.

A provider established outside the Union must also appoint an authorized representative. Article 54(1) reads: "Prior to placing a general-purpose AI model on the Union market, providers established in third countries shall, by written mandate, appoint an authorised representative which is established in the Union."¹² Models released under a free and open-source license are exempt unless they present systemic risk.

Since August 2, 2026, the Commission can fine providers of general-purpose AI models up to 3% of worldwide annual turnover or EUR 15 million, whichever is higher.¹³

An AI system is high-risk in two situations. It is a safety component of a product, or a product itself, covered by the EU harmonization legislation in Annex I and subject to third-party conformity assessment. Or it falls within a use case listed in Annex III, which includes employment, education, access to essential services such as credit, and certain law-enforcement and migration uses.¹⁴

Providers of these systems carry the heaviest duties in the Act: risk management, data governance, technical documentation, logging, instructions for deployers, human oversight, accuracy and cybersecurity, a quality management system, and a conformity assessment before the system reaches the market. Deployers have their own list under Article 26, starting with using the system according to its instructions and assigning human oversight to people with the competence and authority to exercise it.¹⁵

For a US provider, one duty is structural. Article 22(1): "Prior to making their high-risk AI systems available on the Union market, providers established in third countries shall, by written mandate, appoint an authorised representative which is established in the Union." The mandate must let the representative verify the declaration of conformity and the technical documentation, keep them for ten years, answer authorities, and cooperate on corrective action.¹⁶ Like the rest of that section, Article 22 applies from December 2, 2027, or August 2, 2028, depending on the annex.

From August 2, 2026, Article 50 applies to several kinds of systems regardless of risk class. Providers must design systems that interact with people so that those people know they are dealing with AI. Providers of generative systems must mark synthetic audio, images, video, and text in a machine-readable format. Deployers of emotion recognition and biometric categorization systems must inform the people exposed to them.¹⁷ Generative systems already on the market before August 2, 2026, have until December 2, 2026, to add the marking.

Article 99 sets three ceilings, each "whichever is higher" of a fixed amount and a share of worldwide annual turnover:

• EUR 35 million or 7% for the prohibited practices in Article 5
• EUR 15 million or 3% for most operator obligations, including those of providers, authorized representatives, importers, distributors, and deployers, and the transparency duties
• EUR 7.5 million or 1% for incorrect, incomplete, or misleading information given to authorities

For small and medium-sized enterprises, including start-ups, the lower of the two figures applies. The Omnibus extended that rule to small mid-cap companies for the second and third ceilings.¹⁸ Enforcement is national: in Italy the designated authorities are AgID and ACN, as described on our page on Italy's AI law.