The EU AI Act for companies outside the EU

Your company has no office in the EU. The question is not whether the AI Act is strict, but whether it reaches you, in which role, and from when.

This note is for a SaaS business outside the EU, or an agency building for clients. As of 31 October 2026 we have read the consolidated AI Act (opens in a new tab) as at 27 July 2026, its Official Journal text of 12 July 2024 (opens in a new tab) for the recitals, and Regulation (EU) 2026/1744 (opens in a new tab), the Digital Omnibus on AI. Every answer names the article that decides it, and says where it does not. The decision tree runs them in order.

When the AI Act applies to a company outside the EU

If you offer an AI system to customers in the EU, or its output is used there, the Act can apply to you even with no EU office (Article 2(1)(a) and (c)). The first of those is placing an AI system on the EU market or putting it into service there, wherever the provider is established; it covers placing a general-purpose AI model there too. The second is for providers and deployers in a third country “where the output produced by the AI system is used in the Union”. Article 2(1) also reaches your EU customers in their own right as deployers, whatever you do. It reaches importers, distributors, product manufacturers, the representatives of providers outside the Union, and affected persons in the Union.

Three definitions decide the first, in Article 3(9) to (11): placing on the market is “the first making available” there; making available is supply “in the course of a commercial activity”, paid or free; putting into service is supply “for first use directly to the deployer or for own use”. The Commission’s non-binding guidelines on prohibited practices add that supply counts “regardless of the means of supply”, API and cloud included. Route one reaches, we think, any SaaS business offering an AI system to EU customers.

The second route is unamended, and reaches systems, not models. Recital 22 states the same test as output “intended to be used in the Union”, and only “to the extent” of that. The article says is used, the recital says intended, and the text does not settle which governs. Its purpose is to stop circumvention and to protect people in the Union, and its example is an EU operator contracting a high-risk task abroad, with data sent from the EU and the output returned. In its Article 50 guidelines the Commission says “unforeseeable or unauthorised downstream use” should not alone bind a provider off the EU market, while binding a deployer that foresees it.

Two questions come first. An AI system “infers, from the input it receives, how to generate outputs” (Article 3(1)), which the Commission reads as leaving out basic data processing, classical heuristics and simple prediction. Article 2 then excludes sole-purpose research, pre-release work short of real-world testing, free and open-source systems unless high-risk, prohibited or under Article 50, and use exclusively for military, defence or national security.

Provider, deployer or importer: which one you are

What you owe depends on your role: provider (it reaches users under your name) or deployer (you use it). A company outside the EU cannot be the importer of its own system (Article 3).

Roles are definitions in Article 3, not labels you choose. A provider develops an AI system or a general-purpose model, or “has an AI system or a general-purpose AI model developed”, and places it on the market or puts the system into service under its own name or trademark, paid or free. A deployer uses one under its own authority, outside personal, non-professional use. A distributor is anyone else in the chain making one available there.

An importer is located or established in the Union: Article 3(6) gives the role to a person there who places on the market a system bearing a third-country name. Who holds it, if anyone, when you sell a cloud service straight to EU customers, the texts we read do not settle.

Sell a high-risk system through an EU partner and Article 23(1) makes that importer verify your conformity assessment, documentation, CE marking, declaration and instructions, and that you appointed a representative. A distributor checks the marking, the declaration and the instructions (Article 24(1)). Under Article 22 the representative ends its mandate where it “considers or has reason to consider” that you are acting contrary to your obligations under the Regulation, and tells the market surveillance authority why. Article 22(3) makes the mandate empower it to verify your documentation, keep it ten years and be addressed in your place.

For an agency, start with the name: build a system an EU client launches under its own name and the client is the provider; launch it under yours and you are. Use AI for deliverables, deciding how yourself, and the Article 50 guidelines read you as the deployer, not the client.

Roles also move: put your name on a high-risk system already on the market, where a contract may allocate the obligations otherwise, modify one substantially, or change a system’s intended purpose so that it becomes high-risk, and Article 25(1) treats you as its provider. The initial provider then stops being that system’s provider, but must supply documentation, limitations and technical access, unless it clearly specified the system was not to be made high-risk (Article 25(2)).

Article 25(4) makes the provider and its third-party supplier specify in writing “the necessary information, capabilities, technical access and other assistance” it needs. Whether a studio supplying the whole system is such a supplier is unsettled: the article speaks of what is “used or integrated in” one. Where you also host or operate it, the texts we read do not settle who is the provider. Articles 22 to 25 are in the deferred sections, so they run from the high-risk dates below.

What applies, and from when, after the Digital Omnibus

The Digital Omnibus on AI is Regulation (EU) 2026/1744 of 8 July 2026, in the Official Journal on 24 July 2026 and in force three days later, as the Commission announced (opens in a new tab). After the Digital Omnibus, the heavy high-risk rules start on 2 December 2027 for Annex III uses and 2 August 2028 for AI in products covered by Annex I. The bans and the Article 50 transparency duties already apply; two more bans start on 2 December 2026. Those high-risk rules are Sections 1 to 3 of Chapter III: the classification, the requirements, and the duties of providers, deployers and other parties. As adopted in 2024 their dates were 2 August 2026 and 2 August 2027; standards and authorities were late, the recital says.

The deferral is narrow: point (c) of Article 113 defers Sections 1 to 3 only, “with the exception of Article 6(5)”. Section 5 of the same chapter (Articles 40 to 49: standards, conformity assessment, certificates, registration), Chapter VIII (the EU database) and Chapter IX (Articles 72 to 94: monitoring, information sharing and market surveillance) are named in no point of Article 113, except Article 78, brought to 2 August 2025. The rest run from the general date: “It shall apply from 2 August 2026.” Section 5 and the database serve the high-risk rules, which do not apply yet. How they work in the meantime is open.

An Annex III use is high-risk unless the Article 6(3) derogation applies, and never where the system profiles people. Article 5’s prohibitions bind a deployer as well as a provider. Article 50 has applied since 2 August 2026. It asks a provider to disclose the AI system and mark synthetic output. A deployer discloses a deep fake and, separately, text published on “matters of public interest”, where only that duty falls away after human review, with a person holding editorial responsibility.

General-purpose AI models: what the vendor owes, what you owe

A general-purpose AI model, in Article 3(63), displays “significant generality” and performs “a wide range of distinct tasks” competently; the Commission’s indicative criterion is training compute above 10²³ floating-point operations plus the ability to generate language, images or video. Systemic risk turns on high impact capabilities, which Article 51(2) presumes above 10²⁵ floating-point operations of training compute (Article 51(1)(a)). Under Article 52(1) a provider whose model meets that condition notifies the Commission without delay, and in any event “within two weeks” of its being met or becoming known; the Commission’s guidelines say that can fall due before training ends.

Its provider is in scope once it places the model on the EU market, and the Commission treats first making one available through an API as placing it. Article 53 asks for technical documentation, documentation for the providers who integrate the model, a copyright policy honouring text-and-data-mining reservations and a public summary of training content on the AI Office template. Article 54 adds an EU representative. A model under a free and open-source licence with public parameters is spared the first two and that representative, unless it carries systemic risk, which brings Article 55’s evaluation, risk, incident and cybersecurity duties.

Call someone else’s model and you are a downstream provider (Article 3(68)). In the Commission’s guidelines the upstream actor that first makes the model available there is its provider, and the downstream actor integrating it “may be the provider of the system”. You become the model’s provider only where a modification significantly changes the model’s generality, capabilities or systemic risk. The Commission’s indicative criterion for that is compute “greater than a third of the training compute of the original model”, and your duties would then cover the modification only. Who modifies in hosted fine-tuning is unsettled, and in the guidelines control of the weights may be an important factor.

One trap is written for your position. In the Commission’s view, if a vendor clearly excluded EU distribution and use and you put its model into a system you place there, you are the model’s provider. So read the vendor’s EU terms and ask for its Article 53(1)(b) documentation.

Fines, and the lower caps for smaller companies

Member States lay down the rules; the Act lists ten factors an authority weighs; the Commission imposes the Article 101 fine.

Fines are capped lower for SMEs and, except for prohibited practices, for small mid-caps (Article 99). For them each fine in Article 99 is capped at the lower of its two figures, not the higher; the Omnibus added the small mid-cap rule. A small mid-cap is not an SME, must “employ fewer than 750 persons”, and has an annual turnover of at most EUR 150 million or a balance sheet total of at most EUR 129 million (Recommendation (EU) 2025/1099). Whether a company outside the EU counts as either is open.

The penalties chapter has applied since 2 August 2025, except Article 101, which applies from 2 August 2026. The Commission’s enforcement page dates the enforcement powers of the AI Office and the national authorities to 2 August 2026.

Since the Omnibus, Article 75(1) makes the AI Office exclusively competent, with four exceptions, for a system built on a general-purpose model by the same provider or undertaking, and one that constitutes or is integrated into a designated very large online platform or search engine. Its powers, Articles 75a to 75d, sit in the undeferred Chapter IX, so on the text they apply from 2 August 2026, and they include the Article 99 fines and periodic penalty payments. The rest is national, under Regulation (EU) 2019/1020: build on another company’s model and yours stays there. Anyone may complain to a market surveillance authority (Article 85).

A decision tree for a SaaS company or an agency

Where the text leaves a question open the tree says so: those points are for your counsel.

A SaaS business usually answers question 4 by route one, 6 as a downstream provider; an agency turns on question 3. The dates table says when; the fines table says how much.

What to prepare, whatever the answer

None of this is compliance: it is the file we think a team keeps, whichever way the tree comes out.

1. A role record. The name each system is placed under, who places and who deploys it, the clause allocating obligations.
2. An EU-link record. Where it is offered, how, and where its outputs go.
3. A written tier assessment. A provider concluding under Article 6(3) that its Annex III system is not high-risk still documents the assessment, and 6(4) subjects that provider to the Article 49(2) registration duty. When that duty starts is open. The deferral names systems classified as high-risk. Article 6(3) says an Annex III system within its derogation is not to be considered high-risk. Article 49(2) itself is not deferred.
4. The model-vendor file. The Article 53(1)(b) documentation, the EU terms, the Code of Practice signature.
5. The Article 50 items. In the interface and the pipeline, each dated.
6. AI literacy measures. For whoever operates it. Article 4 asks for them without a guaranteed level.
7. For a high-risk provider. Your representative, its mandate’s date and scope, the importer’s checks, and the quality management standard EN 18286:2026.

Four of our pages carry the engineering: the fields we think a tool-call record needs, a permissions map for each tool, what ships with every build, what we hand over with an AI system. We document the controls; we do not certify compliance with any regulation.

Common questions

Does the EU AI Act apply to a US company with no office in the EU?

Yes: if it places an AI system or a general-purpose model on the EU market, puts a system into service there, or the output of its AI system is used in the Union.

What does “output used in the Union” mean?

Article 2(1)(c) says “is used”; recital 22 says “intended to be used in the Union”. For Article 50 alone the Commission reads out unforeseeable use.

If your product only calls a model through an API, are you a provider?

Of your own AI system, yes, as a downstream provider; of the model, only on the one-third compute test, or where your vendor clearly excluded the EU and you place the model there anyway; otherwise, in the Commission’s guidelines, the upstream actor is.

Can a company outside the EU be the importer of its own AI system?

No: the Act defines an importer as located or established in the Union.

When does a company outside the EU need an EU authorised representative?

For a high-risk system, from 2 December 2027 or 2 August 2028, and before placing a general-purpose model on the EU market. No provision we read requires one otherwise.

Did the Digital Omnibus delay the AI Act?

It moved Chapter III, Sections 1 to 3, to 2 December 2027 for Annex III uses and 2 August 2028 for Annex I products. Section 5, the database and market surveillance stand at the general date; the prohibitions since 2 February 2025. Older generative systems have until 2 December 2026 for the Article 50(2) marking, when two new prohibitions start.

If you build an AI system for an EU client, who is the provider?

Usually the client, which puts it into service under its own name; for a high-risk system, the text does not settle whether Article 25(4) reaches you as a supplier.

Review and corrections

Last reviewed 31 October 2026, against the consolidated text of Regulation (EU) 2024/1689 as at 27 July 2026, its Official Journal text of 12 July 2024 and Regulation (EU) 2026/1744. We review this note at least every three months, before each future date in the table, and on a new consolidated version or amending act, the Article 6(5) or Article 50 guidelines being adopted, a further standard referenced or a change to the Commission’s timeline.

Changes to this note. Published 1 November 2026; sources read on 31 October 2026.

Whether the Act reaches a particular system, and in which role, is a legal question, and nothing in this note is legal advice.