The EU AI Office's Role and Powers Over GPAI Providers

Almost everything in the AI Act is enforced by national authorities. General-purpose AI models are the exception: the Commission enforces Chapter V itself, through the AI Office, with its own powers, its own penalty ceiling and its own start date.

Two enforcement systems, not one

Regulation (EU) 2024/1689 splits supervision along the model-system line. High-risk AI systems go through the classic product regime: notified bodies, CE marking, and national market surveillance authorities designated under Article 70 exercising the powers in Regulation (EU) 2019/1020. General-purpose AI models go somewhere else entirely. Article 88 gives the Commission exclusive powers to supervise and enforce Chapter V, taking into account the procedural guarantees in Article 94, and the Commission exercises those powers through the AI Office, which it established by decision in January 2024 within DG CNECT. The Regulation is at EUR-Lex, and the AI Office’s remit is described on the Commission’s page.

Article 75 extends the same exclusivity one step further than people expect. Where a general-purpose AI system is based on a general-purpose AI model and both are provided by the same provider, the Commission has exclusive powers to supervise and enforce the requirements applying to that system too. A national authority investigating such a system has to hand the model question upward. In the other direction, Article 75 obliges the AI Office and national market surveillance authorities to cooperate, and a national authority that suspects a general-purpose AI system does not comply may request the Commission to exercise its powers.

The four powers, in the order they are used

Chapter IX Section 5 reads as an escalation ladder, and reading it in order is the fastest way to understand what a supervisory interaction with the AI Office would look like.

1. Information requests (Article 91). The Commission may request the provider of a general-purpose AI model to provide the documentation drawn up under Articles 53 and 55, or any additional information necessary to assess compliance. It may also request information from downstream providers about the model they have integrated. Requests must state the legal basis and purpose, specify what is required and set a reasonable period. The power to ask downstream is easy to miss and is the reason an integrator can find itself in correspondence about somebody else’s model.
2. Evaluations (Article 92). The AI Office may conduct evaluations of the model itself — to assess compliance where the information gathered under Article 91 is insufficient, or to investigate systemic risks under Article 55, particularly following a qualified report from the scientific panel. It may appoint independent experts to carry them out, and it may request access to the model through APIs or other appropriate technical means, including source code. This is the provision that makes external model evaluation a regulatory instrument rather than a research activity.
3. Required measures (Article 93). The Commission may require the provider to take appropriate measures to comply with Chapter V, to implement mitigation measures where an evaluation has raised serious and substantiated concern of a systemic risk at Union level, or to restrict the making available on the market, withdraw or recall the model. It may also accept commitments offered by the provider, including commitments to implement mitigation measures.
4. Fines (Article 101). Where those steps fail or are obstructed, the ceiling is 3% of worldwide annual total turnover in the preceding financial year, or EUR 15 million, whichever is higher.

Article 89 sits alongside as the monitoring provision: the AI Office monitors effective implementation, may take actions of its own initiative following information received, and downstream providers may lodge complaints alleging an infringement. Article 90 creates the scientific panel of independent experts, which may issue a qualified alert to the AI Office where it has reason to suspect that a model poses a concrete identifiable risk at Union level or meets the conditions for classification as a model with systemic risk.

The 3% ceiling and when it becomes usable

Article 101 lists what the 3%-or-EUR-15-million fine attaches to, and the list is broader than “breaching Chapter V”. It covers intentional or negligent infringement of the relevant provisions of the Regulation; failure to comply with a request for information under Article 91; failure to implement a measure requested under Article 93; failure to make access to the model available to the Commission or the AI Office for an evaluation under Article 92. Three of those four are procedural. A provider can be fined for how it responds to the AI Office rather than for anything about its model.

The date matters and it is not the date you would guess. Chapter V — the substantive general-purpose model obligations in Articles 53 to 55 — has applied since 2 August 2025 under Article 113. But Article 113 excludes Article 101 from that tranche, so the Commission’s power to impose these fines runs from 2 August 2026. There is a year in which the obligations bind and this particular penalty does not, which is a deliberate design choice and not an oversight. It does not make the obligations optional: Articles 91 to 93 are available in the meantime, and a required measure under Article 93 restricting availability of a model is a more serious commercial event than most fines.

Article 111(3) adds the other date worth carrying: providers of general-purpose AI models placed on the market before 2 August 2025 have until 2 August 2027 to bring them into compliance. A model released in 2024 and still being served is on a different clock from one released in 2026, which is a distinction that only exists here and is easily lost in a compliance plan built around a single deadline.

For the general penalty tiers — the 7% ceiling for prohibited practices under Article 5 and the 3% and 1% ceilings elsewhere in Article 99 — see the penalty tiers page. Article 101 is a separate provision addressed specifically to general-purpose model providers and does not sit inside the Article 99 scheme.

Procedural rights and the scientific panel

Article 94 provides that Article 18 of Regulation (EU) 2019/1020 applies mutatis mutandis to providers of general-purpose AI models, without prejudice to more specific procedural rights in the Regulation. That is the right-to-be-heard machinery from the market surveillance framework rather than a bespoke set of rights, and it means the AI Office cannot simply act: the provider is entitled to be heard, and to have the reasons for a measure stated.

Article 92(3) requires evaluations to be conducted in accordance with detailed arrangements the Commission adopts, and Article 92 as a whole is where the practical questions live that nobody has answered yet. What form does “access through APIs or other appropriate technical means” take for a frontier model. How is source code access reconciled with trade secret protection, which Article 78 obliges authorities to respect. How independent are the independent experts and how are their conflicts managed. These are open, and the answers will come from practice rather than from the text.

What this means before enforcement begins

If you provide a general-purpose AI model, the operational implication of this section is narrower than the constitutional one: the documentation you maintain under Article 53 is not filed anywhere on a schedule. It is held, and produced on request under Article 91. That changes what “ready” means. A conformity file that exists as a shared drive somebody could assemble in three weeks is not ready for a request with a stated reasonable period, and the failure-to-respond limb of Article 101 is the one that punishes it.

If you build on somebody else’s model rather than providing one, two provisions still reach you. Article 91 allows information requests to downstream providers. And Article 89 lets you complain to the AI Office about an upstream provider — a route that exists precisely because a downstream integrator depending on a model has no other leverage over its documentation. Whether that complaint route gets used in practice is one of the more interesting open questions about how this regime will actually work.