Does the EU AI Act apply to companies outside the EU?

"We're not in the EU, so this doesn't apply to us" is the most common thing companies get wrong about the EU AI Act — and it is wrong for a specific, checkable reason. Establishment is not the test. The Regulation was drafted to reach conduct, not addresses.

This article works through Article 2(1) of Regulation (EU) 2024/1689 — the routes by which a company outside the Union comes into scope, the two situations where you must appoint someone inside the Union before you sell anything, and the exclusions that genuinely do get you out.

This is an information service to help you plan, not legal advice. To see which duties attach to your own use of AI, check which obligations apply to your company.

How a non-EU company gets caught

Article 2(1) lists seven categories of person the Regulation applies to. Three of them can catch a company with no EU establishment — (a) providers, (c) the output rule, and (e) product manufacturers — and a fourth, (d), catches non-EU distributors in the Union supply chain (importers, by definition in Article 3(6), are established in the Union). Point (f) is the mirror image: it applies to the EU-established representative a non-EU provider must appoint.

Article 2(1)	Who it catches	What triggers it
(a)	Providers placing an AI system on the market or putting it into service, or placing a general-purpose AI model on the market, in the Union	Applies "irrespective of whether those providers are established or located within the Union or in a third country"
(c)	Providers and deployers located in a third country	"where the output produced by the AI system is used in the Union"
(d)	Distributors — the limb that can reach a non-EU company (importers are, by definition, established in the Union: Art. 3(6))	Being in the supply chain for the Union market
(e)	Product manufacturers placing an AI system on the market with their product under their own name or trademark	Your product carries the AI and your brand
(f)	Authorised representatives — themselves established in the Union (Art. 3(5)) — acting for non-EU providers	Accepting the mandate

Of the operator categories, only (b) — deployers "that have their place of establishment or are located within the Union" — turns on where the operator is. The routes that reach non-EU companies — (a), (c), (e) and the distributor limb of (d) — are about what you do and where the effects land.

Two of these deserve unpacking, because they are where most non-EU companies are actually caught.

Route one: you put it on the Union market

Article 2(1)(a) is the straightforward one. If you place an AI system on the Union market, put it into service there, or place a general-purpose AI model on the Union market, you are a provider in scope. The clause says so expressly: it applies whether you are established in the Union or in a third country.

"Placing on the market" is defined in Article 3(9) as the first making available of an AI system or a GPAI model on the Union market. Article 3(10) defines "making available on the market" as supply "in the course of a commercial activity, whether in return for payment or free of charge". Free tiers, betas, and open availability are supply. Price is not the test.

Route two: the output rule

Article 2(1)(c) is the one companies miss. It applies the Regulation to providers and deployers established in a third country "where the output produced by the AI system is used in the Union".

Nothing here requires the system to be sold in the EU, hosted in the EU, or marketed to Europeans. The system can run entirely on your own infrastructure, in your own country, for your own corporate purposes. If its output is used in the Union, the clause is engaged.

Recital 22 explains why it exists, and the example it gives is worth reading closely. It describes an operator established in the Union contracting a service to an operator in a third country; the third-country system processes data lawfully collected in and transferred from the Union, and returns the output to the contracting operator in the Union — all "without that AI system being placed on the market, put into service or used in the Union". The stated purpose of Article 2(1)(c) is "to prevent the circumvention of this Regulation".

One precision worth having. Recital 22 says the Regulation should apply "to the extent the output produced by those systems is intended to be used in the Union". Article 2(1)(c) says the output "is used" in the Union. Those are not identical tests, and the enacting terms are what bind — a recital aids interpretation, it does not override the article. Only the Court of Justice can settle the point authoritatively. Plan against the article.

What this looks like in practice

Some concrete shapes, each mapped to the clause that catches it:

Situation	In scope?	Why
US SaaS company sells an AI product to customers in Germany	Yes	Art. 2(1)(a) — placing on the Union market
UK company offers a free AI tool that anyone, including EU users, can sign up for	Yes	Art. 2(1)(a) with Art. 3(10) — "free of charge" is still making available
US company runs CV-screening AI at home to shortlist candidates for its Irish subsidiary	Yes	Art. 2(1)(c) — the output is used in the Union
Indian BPO runs an AI scoring model and returns the scores to a French client	Yes	Art. 2(1)(c) — this is close to the Recital 22 example
Swiss company builds an AI system used only by its Swiss staff on Swiss matters	No	No placing on the Union market, no output used in the Union
Japanese manufacturer ships a machine with an embedded AI safety component into the EU	Yes	Art. 2(1)(e) — product manufacturer under its own trademark
US firm deploys an AI tool internally, in the US, with no EU nexus at all	No	Neither limb engaged

The pattern: an EU nexus in customers, corporate group, or downstream use of the result is enough. An EU office is not required, and the absence of one proves nothing.

Note also Article 2(1)(g): the Regulation applies to "affected persons that are located in the Union". That clause is about who holds rights under the Regulation, not who owes duties — but it signals the drafting logic throughout. The Act is oriented to where the effects land.

What gets you out

The exclusions are real, and several are narrower than they sound. From Article 2 as consolidated at 27 July 2026:

• Article 2(3) — military, defence and national security. Excluded where systems are placed on the market, put into service or used exclusively for those purposes. A third subparagraph covers systems not placed on the Union market whose output is used in the Union exclusively for those purposes. "Exclusively" is doing heavy work; dual-use does not qualify.
• Article 2(6) — scientific research and development. Systems and models "specifically developed and put into service for the sole purpose of scientific research and development", including their output.
• Article 2(8) — pre-market R&D. Research, testing or development prior to placing on the market or putting into service. But the same paragraph states expressly: "Testing in real world conditions shall not be covered by that exclusion."
• Article 2(10) — purely personal use. Disapplies deployer obligations for natural persons using AI in a purely personal non-professional activity. This is not a small-company exemption; it does not reach a company of any size.
• Article 2(12) — free and open-source licences. Excluded, unless placed on the market or put into service as a high-risk AI system, or as a system falling under Article 5 (prohibited practices) or Article 50 (transparency). The carve-out from the carve-out is where most of the interesting cases sit.

If you concluded you were out of scope on the strength of one of these, it is worth re-reading the qualifier attached to it. Four of the five have one.

The obligation people discover last: an authorised representative

This is the part that surprises non-EU providers, because it is a duty to do something before you sell, not merely a standard to meet.

• High-risk AI systems — 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."
• General-purpose AI models — Article 54(1). The same requirement, "prior to placing a general-purpose AI model on the Union market". Article 54(6) disapplies it for certain models released under a free and open-source licence.

An authorised representative is defined in Article 3(5) as a person located or established in the Union who has received and accepted a written mandate. It is a real appointment of a real party who takes on real work, not a mailbox.

Under Article 22(3) the mandate must empower the representative to verify that the EU declaration of conformity and technical documentation exist and that conformity assessment was carried out; to keep those documents available to authorities for 10 years after the system is placed on the market; to provide information on reasoned request, including access to automatically generated logs under the provider's control; to cooperate with authorities; and to handle registration duties under Article 49(1). The mandate must let the representative be addressed by authorities "in addition to or instead of the provider".

Article 22(4) adds a provision worth reading before you choose one: the authorised representative must terminate the mandate if it considers the provider to be acting contrary to its obligations, and inform the market surveillance authority. Your representative is not purely your agent.

There is no general authorised-representative requirement for non-high-risk AI systems. It attaches to high-risk systems and to GPAI models. Whether you are in the high-risk tier is a separate question — see is your AI system high-risk? — and note that most Annex III high-risk obligations now apply from 2 December 2027 rather than 2026.

Being a non-EU company does not reduce the fines

The penalty tiers in Article 99 attach to conduct, not to geography, and there is no discount for distance:

Infringement	Ceiling (whichever is higher)
Prohibited practices (Art. 5)	€35 000 000 or 7 % of total worldwide annual turnover
Obligations of providers (Art. 16), authorised representatives (Art. 22), importers, distributors, deployers, or the Article 50 transparency duties	€15 000 000 or 3 % of total worldwide annual turnover
Incorrect, incomplete or misleading information to notified bodies or national competent authorities	€7 500 000 or 1 % of total worldwide annual turnover

Two details matter for a non-EU group. First, the base is total worldwide annual turnover, not EU turnover — the reach of the fine is not limited to the market that brought you into scope. Second, Article 99(4)(b) makes failures by your authorised representative under Article 22 a fining ground in their own right. Appointing someone does not transfer the risk: the representative can be fined for its own failures under Article 22, and you remain fully exposed on your own obligations — including the duty to have a compliant representative in place at all. For the full picture see EU AI Act fines.

Article 99(6) caps fines for SMEs and start-ups at the lower of the amount or percentage. Article 99(6a), inserted by the Digital Omnibus, does the same for small mid-cap companies — but only for the paragraph 4 and 5 tiers. The prohibitions tier in paragraph 3 is not included in 99(6a).

Watch the trap in Article 25

If you are a non-EU company that fine-tunes, rebrands or repurposes someone else's system, Article 25(1) can make you the provider — with the provider's obligations under Article 16 — where you put your name or trademark on a high-risk system, make a substantial modification to one, or modify the intended purpose of a system such that it becomes high-risk.

Article 25(2) has provided since 2024 that when that happens, the initial provider "shall no longer be considered to be a provider of that specific AI system", while remaining obliged to cooperate and provide the information and technical access the new provider needs — cooperation duties the Digital Omnibus sharpened. The role transfers rather than being shared. If you are the one who rebranded it, the role has transferred to you — and a new Article 99(4)(da) makes breaches of Article 25(2) and (4) a fining ground.

Did the Digital Omnibus change any of this?

No — and this is worth stating plainly, because a lot of scope commentary was rewritten after the Omnibus on the assumption that it was.

Reading the consolidated text of the AI Act as at 27 July 2026, the Omnibus amendment markers in Article 2 appear further down — at paragraph 7 and a new paragraph 13, among others. They do not appear at paragraph 1. Regulation (EU) 2026/1744 left the list of who the Regulation applies to untouched. The definitions of provider, deployer, authorised representative, importer, distributor and operator in Article 3(3) to (8) are likewise unamended.

One genuinely new item is worth diarising: Article 2(13), inserted by the Omnibus, lets specific requirements in Articles 9 to 15 and 17 to 25 be limited for certain Annex I Section A high-risk systems where other Union harmonisation legislation gives equivalent or higher protection — and requires the Commission to adopt delegated acts specifying this by 2 August 2027. That is a dated, checkable milestone that does not yet exist in most published timelines. It sits alongside the other dates in the EU AI Act timeline.

What to do if you are outside the EU

1. Stop asking where you are established. Ask instead: do we place anything on the Union market, and is the output of any of our AI used in the Union? Those two questions decide it.
2. Trace your outputs, not just your sales. Group companies, EU subsidiaries, and EU clients of your services are all routes by which output lands in the Union. This is the limb most companies have never mapped.
3. Check the free tier. "Free of charge" is expressly within "making available on the market" under Article 3(10).
4. Re-read the exclusion you are relying on. Real-world testing is outside the Article 2(8) exclusion; "exclusively" governs Article 2(3); "sole purpose" governs Article 2(6); the open-source exclusion in Article 2(12) falls away for high-risk, Article 5 and Article 50 systems.
5. If you are high-risk or a GPAI model provider, start the authorised-representative appointment early. It is a contracting exercise with a party who must be willing to hold your documentation for ten years and to terminate on you if you are non-compliant. That takes longer than people plan for.
6. Check the Article 50 transparency duties separately. They bite on chatbots and generated content regardless of risk tier, and they reach non-EU deployers whose output is used in the Union — see Article 50: what you must disclose.

Scope is the question where being wrong is most expensive, because it is upstream of everything else: if you conclude you are out, you never look at the rest. If you would rather be told when something moves than re-read the Official Journal yourself, join the waitlist — or work through which obligations apply to you now.

The official text is Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744; article references above are to the consolidated version as at 27 July 2026. This article is an information service to help you orient — it is not legal advice, and you should confirm the position against the official sources before acting.