Who Classifies an AI System Under the EU AI Act

The common reading of the European AI regulation is that it sorts systems into risk tiers and puts the burden on whoever built them. The first half is roughly right. The second half is where organisations that only ever bought a tool discover they are holding something heavier than they expected.

The tiers are chapters, not a sidebar

The structure of Regulation (EU) 2024/1689 makes the sorting explicit. Chapter II is titled prohibited AI practices and consists of a single article listing them. Chapter III is titled high-risk AI systems and runs to dozens of articles, opening with classification rules and then setting out what compliance requires: a risk management system, data governance, technical documentation, record-keeping, transparency toward deployers, human oversight, and accuracy with robustness and cybersecurity.

Two things follow from the shape alone. Almost the entire regulatory weight sits in one tier, so the classification question is close to the whole question. And the high-risk list is not frozen: a separate article provides for amending it, meaning a system outside the tier today can be inside it later without anyone touching the system.

Provider and deployer are different roles with different duties

The regulation separates the party that places a system on the market from the party that uses it. Providers carry the heavy obligations. Deployers carry their own, lighter set, and in some circumstances a fundamental rights impact assessment.

Most buyers stop reading here and conclude they are deployers. That conclusion is correct until it is not, and the article that changes it is easy to miss.

The three ways a buyer becomes a provider

Article 25 states the conversion directly:

Any distributor, importer, deployer or other third-party shall be considered to be a provider of a high-risk AI system for the purposes of this Regulation and shall be subject to the obligations of the provider under Article 16, in any of the following circumstances:

(a) they put their name or trademark on a high-risk AI system already placed on the market or put into service, without prejudice to contractual arrangements stipulating that the obligations are otherwise allocated;

(b) they make a substantial modification to a high-risk AI system that has already been placed on the market or has already been put into service in such a way that it remains a high-risk AI system pursuant to Article 6;

(c) they modify the intended purpose of an AI system, including a general-purpose AI system, which has not been classified as high-risk and has already been placed on the market or put into service in such a way that the AI system concerned becomes a high-risk AI system in accordance with Article 6.

Regulation (EU) 2024/1689, Article 25(1)

Read those three against what organisations actually do with purchased tools.

Putting your name on it. White-labelling a supplier's system, or presenting it to clients under your own brand, is the first circumstance. This is ordinary commercial behaviour and it is squarely within the text.

Substantially modifying it. The line between configuring a system and substantially modifying it is not one a procurement team should be drawing on instinct.

Changing what it is for. The third is the sharpest. It expressly covers taking a general-purpose system that was not high-risk and changing its intended purpose so that it becomes high-risk. Nothing about the system changes. What changes is the use somebody put it to, and the party that changed the purpose becomes the provider.

The handover, and its exception

Where the conversion happens, the original provider stops being the provider of that system and is obliged to cooperate with the new one, making available information and reasonable technical access so the new provider can meet its obligations.

That sounds reassuring until the carve-out: the cooperation duty does not apply where the initial provider has clearly specified that its system is not to be changed into a high-risk system. A supplier can say so and thereby decline to hand anything over.

The practical consequence for a buyer is a contract question rather than a technical one. Whether your supplier has made that specification, and what happens to your intended use if they have, is answerable before signing and very expensive afterwards.

Why this reads differently from the US position

Worth holding both in mind, because they ask opposite questions at opposite times.

The European approach classifies before deployment and attaches duties to the role you occupy. The Federal Rules of Evidence say nothing about deployment at all and ask, after the fact and only if a dispute reaches that point, whether an opinion rests on reliable principles reliably applied.

One regime asks what you are before you act. The other asks what you can prove afterwards. An organisation operating across both is not choosing between them, and documentation built for the first is generally the documentation that answers the second.

The companion article in this section covers a voluntary framework that produces much of that documentation as a side effect. For the data these systems get pointed at, our other work on court data is the way across.