AI Law. Decoded.

AI Law. Decoded.

When the EU AI Act Applies to Non-EU Companies

The answer to give an EU customer, and the reasoning behind it.

Silvia Stepitova's avatar
Silvia Stepitova
Nov 11, 2026
∙ Paid

In April, I wrote that there was no official guidance yet on what separates AI output meant for the EU from output that just ends up there. In July, the European Commission said something about it in its guidelines on transparency, the ones about labeling AI content.

It’s a partial answer, written for one article of the AI Act, and it’s not binding. It’s also the only official word so far on the question that decides whether a company outside the EU has to care at all.

That question usually arrives as a line in an EU customer’s due diligence form: does the EU AI Act apply to your organization, and if not, why not? “We’re not in the EU” isn’t the right answer.

The AI Act doesn’t only ask where you are. It asks where your AI system goes and where its output lands, which is also why it follows some companies home. Article 2 lists seven routes in, and for a company outside the EU, two of them matter most.

The first is supply. You sell an AI system into the EU, offer it there as a service, or roll it out to your own staff in an EU office. That route has been fairly clear from the start.

The second is output. Your system runs somewhere else, but what it produces is used in the EU. That route hasn’t been clear at all, and it’s the one the Commission has now started on, in two halves. A provider outside the EU that doesn’t supply its system to the EU shouldn’t be pulled under the transparency rules just because the output reached the EU incidentally, unforeseeably or without its authorization. A company outside the EU that uses an AI system is in where it foresees its output being used in the EU, for example by directing it there or by posting deepfakes on the open internet.

Both halves apply to transparency only. Within that, they point the same way: whether the transparency rules reach you depends on what you knew and did about where your output goes, and not only on where your servers sit. For the rest of the output route, it’s the most useful signal so far.

So the answer needs evidence, not just a conclusion. Where your customers are, what your terms say, and what you do to keep output out of the EU or to send it there. It also needs to be given per legal entity, because a group can be caught through one company and not another.

For most companies, a “yes” means two things today: taking measures to build AI literacy among the people working with it, and making sure nothing they run is on the banned list, which grows on December 2, 2026. The heavier obligations come with the high-risk rules in December 2027, and so does the representative inside the EU for non-EU providers of high-risk systems. A provider of a general-purpose AI model based outside the EU needs a representative already.


Below in the article: the seven routes and the exclusions, worked through one fictional group with five companies in five countries. Then when a company outside the EU needs a representative inside it (and when, on the text, it doesn’t), and the one-page record you file when someone asks how you know.

The decision tree and the record come as a one-page PDF you can download at the end.


User's avatar

Continue reading this post for free, courtesy of Silvia Stepitova.

Or purchase a paid subscription.
© 2026 AI Law. Decoded. · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for writing, podcasting, and media founders.