EU AI Act extraterritoriality: what non-EU companies need to do now

The European Union’s Artificial Intelligence Act — Regulation (EU) 2024/1689 — entered into force on 1 August 2024. Most of its substantive obligations apply from 2 August 2026, with prohibited-practice rules already enforceable since February 2025 and general-purpose AI (GPAI) provisions in effect since August 2025. The regulation has been compared to the GDPR for its scope and ambition, but for non-EU organisations the practical impact is different in important ways.

The extraterritorial trigger

Article 2 of the AI Act sets out four scenarios where the regulation applies to organisations established outside the EU. Each is worth reading carefully.

First, the regulation applies to providers placing AI systems on the EU market or putting them into service in the EU. This includes selling, licensing, or making AI systems available to users in the EU through any channel — direct, through resellers, or via cloud-delivered services with EU users.

Second, the regulation applies to deployers established outside the EU when the output of the AI system they use is used in the EU. A US-based recruitment platform processing CVs of EU job applicants falls within scope, even if the company has no EU office and no European staff.

Third, importers and distributors handling AI systems on the EU market carry obligations to verify compliance and refuse non-compliant systems.

Fourth, providers and deployers of AI systems used to provide services to people in the EU are within scope, regardless of where they are established.

The cumulative effect is that few internationally operating AI organisations sit fully outside the regulation. The question is not “does the AI Act apply to us” but “which roles do we hold under the AI Act, and what obligations follow.”

The authorised representative requirement

Non-EU providers of high-risk AI systems must appoint an authorised representative established in the EU. The representative is the regulator’s point of contact, holds technical documentation on the provider’s behalf, and may be held liable jointly with the provider for certain obligations.

This is a meaningful operational requirement. The authorised representative cannot be a shell entity. Under Article 22, the representative must verify that EU declaration of conformity and technical documentation have been drawn up; must hold a copy of these documents available to national authorities; must cooperate with authorities; and must terminate the mandate if the provider acts against AI Act obligations.

Existing GDPR Article 27 representatives are not automatically AI Act representatives — the roles are distinct and may need separate appointments. Many organisations are choosing to consolidate the roles where possible, but the legal mandates require care.

Risk classification: the first practical step

Before any compliance work, every AI system within scope must be classified. The AI Act recognises four tiers: prohibited practices, high-risk systems, limited-risk systems requiring transparency, and minimal-risk systems with voluntary codes.

High-risk classification is the practical pinch point. Annex III lists eight categories of high-risk AI: biometrics, critical infrastructure, education and vocational training, employment and HR, essential services, law enforcement, migration, and administration of justice and democratic processes. Annex I extends high-risk to safety components of regulated products covered by EU harmonisation legislation — medical devices, machinery, vehicles, toys, and others.

Organisations frequently underestimate Annex III scope. AI used for screening job applications is high-risk. AI used for evaluating creditworthiness in essential financial services is high-risk. AI used for emotion recognition outside narrow medical or safety contexts is high-risk. AI used to allocate access to public benefits or services is high-risk.

For each high-risk system, providers face a substantial compliance package: risk management system, data governance, technical documentation, automatic record-keeping, transparency to deployers, human oversight, accuracy and robustness, conformity assessment, and post-market monitoring. Deployers face a lighter but still material set of obligations centred on use, human oversight, log retention, and — for public bodies and some private actors — fundamental rights impact assessment.

GPAI: a parallel regime

General-purpose AI models — foundation models including most current large language models — face a parallel obligation regime that has been applicable since 2 August 2025. GPAI providers must publish training-data summaries, respect EU copyright opt-outs (Directive (EU) 2019/790), maintain technical documentation, and make information available downstream to providers integrating the model into AI systems.

For models posing “systemic risk” — defined initially as those trained with more than 10^25 FLOPs — additional obligations apply: model evaluations including adversarial testing, systemic-risk assessment, incident reporting to the European AI Office, cybersecurity measures, and a designated point of contact within the EU.

Non-EU GPAI providers serving the EU market — OpenAI, Anthropic, Google, Meta, and others — have been working through the GPAI Code of Practice under Article 56 of the AI Act, an operational compliance pathway agreed with the European AI Office. Organisations integrating GPAI models should understand which Code commitments their upstream providers have made.

Penalties scale with severity

The AI Act’s penalty structure is designed to incentivise compliance at the corporate-strategy level. Up to €35 million or 7% of worldwide annual turnover — whichever is higher — for prohibited-practice violations. Up to €15 million or 3% for breaches of high-risk obligations. Up to €7.5 million or 1% for providing incorrect, incomplete, or misleading information to authorities.

Penalty calculation considers the nature and severity of the breach, the intent, the size of the company, and previous compliance history. SMEs and start-ups face proportionally adjusted maximums but are not exempt. The penalty regime applies regardless of where the responsible organisation is established — extraterritorial scope flows through to enforcement.

Building a compliance programme

The practical path forward for non-EU organisations rests on a five-step programme that AIPIA recommends to international members.

Step one is inventory. Every AI system used or sold in the EU, with the team responsible, deployment dates, and intended use cases. The inventory feeds risk classification under Article 6 and Annex III.

Step two is role mapping. For each system, identify whether the organisation acts as provider, deployer, importer, or distributor. The same organisation may hold different roles for different systems. Outsourced relationships need review — third-party AI suppliers’ AI Act obligations affect the buyer’s compliance posture.

Step three is gap analysis. For high-risk systems, compare current documentation, governance, oversight, and accuracy practices to AI Act requirements. Most organisations have meaningful gaps in technical documentation completeness, log retention, and post-market monitoring even where ISO 42001 or equivalent frameworks are already in place.

Step four is implementation. Build or contract the missing pieces. Engage external assessors for conformity assessment where required. Appoint authorised representatives. Establish internal AI governance with clear accountability — not a side-of-desk activity but a named function reporting to senior management.

Step five is monitoring. The AI Act’s post-market monitoring obligations are continuous. Incidents must be logged and reported; performance must be tracked against accuracy and robustness targets; relevant changes to systems must trigger re-assessment.

What AIPIA provides

AIPIA’s working group on AI Act compliance publishes practitioner-level guidance, runs member briefings within ten working days of European Commission and AI Office releases, and delivers training programmes — including the European Digital Credential in AI Act compliance. For US, UK, Gulf, and broader international members, the credential provides verifiable evidence of AI Act fluency that increasingly appears in EU procurement and partner-vetting processes.

Compliance is achievable, but it requires structured work and clear ownership. Organisations that started in 2024 are largely on track for August 2026. Those starting in mid-2026 have less margin but can still meet the deadline with disciplined execution.