Does the EU AI Act apply to you if you are outside the EU?

Yes, quite often, even if your company has never had an office, a bank account or a signed contract in Europe. The EU AI Act decides who it covers by looking at the AI system and where its output ends up, so a firm in Toronto, Austin or Brisbane can sit inside it.

I think the alarm is out of proportion for most readers abroad: coverage usually means a short list of duties, not a compliance programme. The duty likelier to hurt a small firm sits in privacy law, which is why whether a pasted client record counts as a breach has a guide of its own.

Does the EU AI Act apply to non-EU companies? Article 2’s three routes

Article 2(1) lists seven groups the Regulation applies to, and three settle the question for a business abroad. Point (a) covers any provider that puts an AI system or a general-purpose model on sale in the Union, whatever country it calls home. Point (b) covers deployers established or located in the Union.

Point (c) is the long arm. It reaches providers and deployers in a third country whenever the output produced by the AI system is used in the Union. Recital 22 gives the reason: without it, an EU company could hand a high-risk task to a contractor abroad and import only the result.

That recital speaks of output “intended” for use in the Union, but the Regulation never says where incidental use ends. Nobody can safely plan on that line, as I read it. Ask the question separately for each tool and each client, and note the answer beside that tool in your staff AI rules.

Deployer vs provider: the Article 3 definitions behind the quiz question

A quiz statement doing the rounds says a deployer is any person or organisation that develops or commissions an AI system and places it on the market under its own name or trademark. True or false? False. That sentence paraphrases Article 3(3), which defines the provider.

Article 3(4) is shorter. A deployer is any person, company or public body using an AI system under its authority, unless the use is personal and non-professional. The Commission’s own illustration: whoever develops a CV-screening tool is its provider, and a bank using that tool is its deployer.

Roles can change hands. Under Article 25(1), a deployer that brands a high-risk system, substantially modifies one, or turns another system to a high-risk use inherits the provider’s duties. A thin app sold under your brand on someone else’s model worries me more: every request that app sends is now part of your product.

Who has to appoint an EU authorised representative?

Only some providers do, and the AI Act asks it of no deployer. Article 22 tells a provider established outside the Union to appoint one, by written mandate, before making a high-risk AI system available on the EU market. Article 54 sets the same rule for providers of general-purpose AI models.

Article 3(5) requires the representative to be located or established in the Union. It keeps the documentation for ten years, answers authorities’ reasoned requests and can be addressed instead of the provider. Should it conclude the provider is in breach, Articles 22(4) and 54(5) oblige it to walk away and report its reasons.

Article 54(6) spares open-source model providers who publish weights, architecture and usage information, unless the model carries systemic risk. I distrust any representative that promises to ask no questions; the Regulation expects it to read your file. IT providers who set these tools up for clients have a programme of their own with us.

What a business outside Europe owes today: literacy, bans and labels

If Article 2 catches you, Articles 4, 5 and 50 bind you now. Of the Article 5 bans, the one an office could meet is software reading the emotions of staff or students, allowed only for medical or safety reasons. Article 50, on AI notices and labels, has applied since 2 August 2026.

Article 4 reaches every provider and deployer. Regulation (EU) 2026/1744 replaced it in July 2026: providers and deployers must take measures that help their staff, and others running AI for them, build AI literacy. The new text adds that no one has to guarantee a particular level in any individual.

That rewrite reads to me as relief, not repeal. The duty survives, with proportion written in, and the Commission must publish practical examples. A deployer labels deepfakes and AI text published to inform the public; the text duty lapses once someone with editorial responsibility has reviewed it.

EU AI Act timeline after Regulation (EU) 2026/1744

The AI Act entered into force on 1 August 2024 and applies in stages under Article 113. The Digital Omnibus on AI, Regulation (EU) 2026/1744 of 8 July 2026, appeared in the Official Journal on 24 July and took effect three days later, on 27 July 2026.

It moved the high-risk dates and added two bans from 2 December 2026: AI that generates non-consensual intimate imagery of an identifiable person, and AI that generates child sexual abuse material. It left Article 2(1), the general-purpose model rules and the August 2026 start of Article 50 where they were.

For a firm abroad, 2 December 2027 is the date I’d plan around, because that is when the Annex III rules for hiring, credit and education begin. Anyone still selling an August 2026 high-risk deadline is quoting the 2024 text, and I’d treat the rest of their pitch with the same caution.

Can the EU fine a firm based abroad? Article 99 in numbers

In law, yes. Article 99 sets ceilings for any operator in scope, wherever it sits, and Member States’ national authorities enforce them. The ceiling for a banned practice is EUR 35 million or 7% of total worldwide annual turnover; for most other duties, Articles 26 and 50 included, EUR 15 million or 3%.

A large company is exposed to whichever figure is greater. For SMEs, start-ups included, Article 99(6) flips that to whichever is lower, though whether a firm abroad fits the EU’s SME definition is still an open question. Paragraph 6a, from the omnibus, gives small mid-caps the lower figure in the two lower bands only.

Collecting from a firm with no EU assets is a separate problem, but I see that gap as luck, not protection. The likelier pressure arrives through EU clients writing these articles into contracts, and through the questions underwriters now put about staff and AI.

GDPR and the AI Act abroad: two reach tests that don’t line up

The GDPR has a long arm too, and it works differently. In the EDPB’s reading of Article 3, it reaches a firm outside the Union that offers goods or services to people in the EU or monitors their behaviour there. The AI Act asks a different question: where is the output used?

So the two can split. A Toronto firm sending AI market reports to a Paris company, with no personal data on anyone in Europe, can sit inside the AI Act and outside the GDPR. Article 2(7), as amended, still keeps the GDPR intact, bar narrow exceptions such as the new Article 4a on bias testing.

Home law keeps running too. The FTC said in 2024 there is “no AI exemption from the laws on the books”; Canada’s privacy commissioners and Australia’s OAIC say much the same about privacy law. Running a small firm abroad, I’d fix what goes into each prompt before reading any AI Act article.

A written rule rarely stops the late-evening paste on deadline day. That needs a control acting at the moment text leaves the screen.

Before choosing one, it helps to know the less obvious ways client data reaches a model.

Before December 2027: three moves for a firm outside Europe

Outside the EU, I’d keep the whole job to three moves that fit on a card. Test every AI tool against Article 2 and write your role beside it, then file the literacy evidence. If anything you use touches hiring, credit or another Annex III area, ask the vendor now who carries the provider’s duties.

The AI Act won’t stop a client’s file reaching a chatbot. Nonimo replaces client names, identifiers and case numbers with labels before the text goes to the AI tool, and the mapping that restores them stays with you.

Teams putting that in place across several departments can start from Nonimo for organisations.

This page explains the Regulation in general terms; a lawyer who knows your contracts should confirm your own role. Other questions about AI and client data are answered across our guide library.

Sources

• Regulation (EU) 2024/1689, the AI Act (EUR-Lex): Articles 2, 3(3) to (5), 4, 5, 22, 25(1), 50, 54, 99, 111 and 113; recital 22 on output used in the Union.
• Regulation (EU) 2026/1744, Digital Omnibus on AI (EUR-Lex): Article 1(2)(b), (5), (7), (38), (39) and (40); in force on the third day after publication on 24 July 2026.
• European Commission, AI Act Service Desk implementation timeline: the dates up to 2 August 2028.
• European Commission, Navigating the AI Act: the CV-screening provider and bank deployer example; omnibus in force on 27 July 2026.
• EDPB, Guidelines 3/2018 on the territorial scope of the GDPR, version 2.1: establishment and targeting criteria, assessed per processing activity; the Article 27 representative.
• FTC, Operation AI Comply, 25 September 2024: no AI exemption from existing law.
• Office of the Privacy Commissioner of Canada, principles for generative AI: generative AI works within existing privacy law.
• OAIC, guidance on privacy and commercially available AI products: privacy obligations apply to personal information put into AI systems.

Accessed 10 November 2026.