From Sidhant Tamrkar | Product & Market Analysis
AI Office Can Fine Model Providers 3% of Turnover. What Downstream Buyers Owe
On this page
Since 2 August 2026, the European Commission can fine a general-purpose AI model provider up to 3% of worldwide turnover or €15 million, whichever is higher. That power sits in Article 101 of the EU AI Act, and it targets OpenAI, Anthropic and Google, not the SaaS company that calls their API. EU AI Act GPAI enforcement has arrived. Your share of it is smaller than the headlines suggest, and it is not zero.
Key takeaways
- Article 101 fines reach model providers, not the companies building on their models. The 3% or €15 million ceiling applies to providers of general-purpose AI models. A SaaS company that integrates a model through an API is a downstream provider of an AI system, which is a different legal role.
- A downstream company becomes a model provider only past a high bar. The Commission's 2025 guidelines use an indicative test of fine-tuning compute above one third of the original model's training compute. Prompting, retrieval and light fine-tuning sit far below it.
- Your live exposure today is Article 50, enforced nationally. Telling users they are talking to an AI and marking synthetic output carry a separate €15 million or 3% ceiling under Article 99(4), and that ceiling is lower for SMEs.
- The real downstream risk is supply, not penalty. Article 93 lets the Commission order a provider to restrict, withdraw or recall a model. If your product depends on one model, that is a continuity risk no fine schedule captures.
Reader for this post: the founder or director of a B2B software company that ships features on top of a third-party model. You decide which obligations your company carries and which ones you push back up the supply chain, so the map below is written for that decision.
What changed on 2 August 2026 for GPAI enforcement
The obligations on general-purpose AI model providers are not new. Chapter V of the AI Act, covering Articles 51 to 56, has applied since 2 August 2025. What arrived this August was the machinery to enforce them.
Article 113 of the regulation carved Article 101 out of the 2025 start date. The fining power and the investigative powers in Articles 91 to 93 all fell under the general application date of 2 August 2026. For one year the rules existed without a penalty attached. That year is over.
The first visible use of the new powers came on 14 October 2026. According to Euronews, the AI Office sent formal requests for information to OpenAI and Anthropic. The requests covered security, external evaluation and how the models are used. Commission Executive Vice-President Henna Virkkunen said the AI Office "has started enforcing these rules".
Who the AI Office is and why it matters to buyers
The AI Office is a unit inside the European Commission. For general-purpose AI models it is the enforcer, not the national regulators. That is an unusual split. National market surveillance authorities handle almost everything else in the Act, including the transparency duties most SaaS companies now carry.
So one regulator watches your model supplier and a different regulator, in each member state, watches you. Your compliance story has to work for both, and they will not necessarily ask the same questions.
What OpenAI, Anthropic and Google owe under Articles 53 and 55
Every general-purpose AI model provider carries four baseline duties under Article 53(1). Keep technical documentation to the Annex XI standard. Give downstream providers the information set out in Annex XII. Maintain a copyright policy that respects machine-readable opt-outs. Publish a sufficiently detailed summary of training content using the AI Office template.
Providers outside the EU must also appoint an authorised representative inside the Union under Article 54, by written mandate. That representative is the address the AI Office writes to.
The extra layer for systemic-risk models
A model is presumed to have high-impact capabilities, and therefore systemic risk, when its cumulative training compute exceeds 10^25 floating point operations under Article 51(2). The frontier models most SaaS products call sit in this tier.
Article 55 adds four more duties for them. Evaluate the model with state-of-the-art protocols, including adversarial testing. Assess and mitigate systemic risks at Union level. Track and report serious incidents to the AI Office without undue delay. Maintain adequate cybersecurity for the model and its physical infrastructure.
The open-source carve-out is narrower than people assume
Article 53(2) exempts models released under a free and open-source licence from the documentation duties in points (a) and (b). It does not exempt them from the copyright policy or the training summary. And the exemption vanishes entirely for a model with systemic risk.
This matters if your plan for avoiding regulatory dependency is to self-host an open-weight model. You may escape the provider's documentation trail, but you also lose the Annex XII information package you would otherwise be owed. Licence terms add their own layer, which the breakdown of open-weight model licence clauses works through clause by clause.
The Code of Practice is the compliance route most providers chose
Article 53(4) and Article 55(2) let providers demonstrate compliance through a code of practice until harmonised standards exist. The Commission's General-Purpose AI Code of Practice was published on 10 July 2025. Its signatory list includes Amazon, Anthropic, Google, Microsoft, Mistral AI and OpenAI.
Meta is not on it. xAI signed only the Safety and Security chapter, so it must show compliance with the transparency and copyright duties by other means. If your product runs on a non-signatory's model, you are relying on a supplier that chose the harder path to proving compliance. That is worth one line in your vendor risk register.
The AI Office's four enforcement powers, and who they reach
The powers that went live in August come in a sequence. Each step escalates, and each one is addressed to the model provider rather than to its customers.
| Power | Article | What it allows | Downstream effect on you |
|---|---|---|---|
| Request documentation | Art 91 | Demand Article 53 and 55 documentation or any additional information needed to assess compliance | None directly. Your provider may ask you for usage context. |
| Evaluate the model | Art 92 | Run evaluations, including through API access or source code, where Article 91 information is insufficient | Possible findings that change your provider's model behaviour or terms |
| Request measures | Art 93 | Order compliance steps, risk mitigation, or restriction, withdrawal or recall of the model | The one that can take your model away |
| Fine | Art 101 | Up to 3% of worldwide turnover or €15 million, whichever is higher | Price and terms pressure, not liability |
Fines attach to process failures as well as substance
Article 101(1) lists four triggers. A provider can be fined for infringing the regulation itself. It can also be fined for failing to answer an Article 91 request, or supplying misleading information. Refusing an Article 93 measure is a third trigger, and denying the Commission model access for evaluation is the fourth.
Three of those four are about cooperation, not about the model. My reading is that the AI Office built itself a regime where the cheapest violation to prove is obstruction. Expect early cases, if any arrive, to look like that rather than like a ruling on whether a model is safe.
Two procedural details are worth knowing. The Commission must share preliminary findings and give the provider a chance to be heard before fining. And the Court of Justice has unlimited jurisdiction to cancel, reduce or increase any fine.
The provider-versus-deployer obligations map for downstream SaaS
This is the core of the question. The AI Act assigns duties by role, and a single company can hold several roles at once. Article 3 defines three that matter here.
A deployer, under Article 3(4), is anyone using an AI system under its own authority in a professional context. A downstream provider, under Article 3(68), is a provider of an AI system that integrates an AI model, whether the model is its own or supplied by someone else under contract. A provider of a general-purpose AI model is the company that develops the model and places it on the market.
A SaaS company that wraps a frontier model inside a product is usually a downstream provider of an AI system. When it uses that product internally, it is also a deployer. It is almost never a general-purpose AI model provider. That one distinction decides which of the following rows land on you.
| Obligation | Article | Model provider | Your company | Applies from |
|---|---|---|---|---|
| Technical documentation of the model | 53(1)(a), Annex XI | Yes | No | 2 Aug 2025 |
| Integration information for downstream providers | 53(1)(b), Annex XII | Owes it | Entitled to receive it | 2 Aug 2025 |
| Copyright policy and training content summary | 53(1)(c) and (d) | Yes | No | 2 Aug 2025 |
| Systemic-risk evaluation, incident reporting, cybersecurity | 55 | Yes, above 10^25 FLOPs | No | 2 Aug 2025 |
| Answer AI Office requests, allow evaluation | 91, 92 | Yes | No | 2 Aug 2026 |
| Tell users they are interacting with an AI | 50(1) | Only for its own products | Yes, as provider of the system | 2 Aug 2026 |
| Machine-readable marking of synthetic output | 50(2) | For its own generative products | Yes, if your system generates content | 2 Aug 2026, grace to 2 Dec 2026 for live systems |
| Label deepfakes and public-interest AI text | 50(4) | No | Yes, as deployer | 2 Aug 2026 |
| AI literacy of staff | 4 | Yes | Yes, best-efforts measures | 2 Feb 2025 |
| High-risk deployer duties, human oversight, logs | 26 | No | Only if your use case is high-risk | 2 Dec 2027 |
Article numbers refer to Regulation (EU) 2024/1689 as amended. The Article 4 text now requires measures "to their best extent" and does not require any guaranteed level of literacy. Article 50 marking grace date is per Gibson Dunn's summary of the Omnibus agreement.
Annex XII is the one thing you are owed
The most useful row in that table is the second one, because it gives you a right rather than a duty. Under Annex XII, your model provider must give you the model's intended tasks and acceptable use policies. It must also cover release date and distribution methods, architecture and parameter count, input and output modalities, licence, and context window length. Information on training, testing and validation data, including provenance and curation, rounds it out.
Most SaaS companies have never asked for this package. They should. It is the documentation you will need to show a national regulator, or an enterprise customer's procurement team, that you understood the model's limits before you shipped it. How the major providers' own terms describe these limits differs a lot, and the comparison of model provider terms sets them side by side.
Transparency is your live obligation, and it is not new ground
Article 50 applied from the same 2 August date, but it is enforced by national authorities, not the AI Office. A support chatbot built on a frontier model makes you the provider of that system for Article 50(1) purposes. The full duty list, and the marking problem for generated content, is covered in the EU AI Act transparency checklist, so it is not repeated here.
When a downstream company becomes a GPAI model provider itself
There is one route by which the Article 101 regime can reach a company like yours. If you modify a general-purpose AI model enough, you become the provider of a new model.
The Commission's guidelines on general-purpose AI providers, published on 18 July 2025, give an indicative test. A downstream modifier is presumed to be a new provider when the training compute used for the modification exceeds one third of the original model's training compute. If the original model carries systemic risk, the base becomes one third of the 10^25 FLOP presumption threshold. Where the original compute is unknown, the general thresholds stand in for it.
What the one-third test means in practice
For a frontier model trained with well over 10^25 FLOPs, one third of the training compute is a very large budget. A typical enterprise fine-tune on a few million tokens of support transcripts is many orders of magnitude below it. Retrieval-augmented generation, prompt engineering and tool use involve no training compute at all.
Even past the threshold, the guidelines limit the new provider's duties to the modification. You would supplement the original provider's documentation rather than rebuild it. That is a meaningful softening, and it is one reason I think most SaaS companies are spending compliance attention in the wrong place.
Which EU AI Act fines can actually reach a downstream SaaS company
Article 101 cannot reach you unless you become a model provider. The penalties that can reach you sit in Article 99, imposed by national authorities under each member state's rules.
Article 99(4) sets a ceiling of €15 million or 3% of worldwide turnover for breaches of listed operator duties. The list includes deployer obligations under Article 26 and transparency obligations under Article 50. Article 99(5) sets €7.5 million or 1% for giving incorrect or misleading information to authorities.
The SME rule flips the arithmetic
Article 99(6) says that for SMEs and start-ups, each fine is capped at whichever of the percentage or the fixed amount is lower. For a company with €20 million of turnover, 3% is €600,000, and that becomes the ceiling. No equivalent rule exists in Article 101, because the model providers it targets are not small.
The AI literacy duty in Article 4 does not appear in the Article 99(4) list at all. That does not make it optional. It does mean the direct fine exposure most vendors warn you about is thinner than their sales decks imply.
Where this map is weakest
A map that tells SaaS buyers their exposure is limited should state the ways it could be wrong. There are four.
First, the one-third compute test is indicative, not statutory. Commentators have pointed out that a significant change in a model's generality or capability can occur well below that threshold. The guidelines leave room for that judgement. A company doing heavy domain fine-tuning should not treat the number as a safe harbour.
Second, the enforcement record is close to empty. As of mid-October, the only reported step is the information requests to OpenAI and Anthropic, and the detail comes from press reporting rather than a published Commission decision. The Parliament Magazine reported the AI Office had around 40 staff in July 2026, against a target of 160 by 2030. How it prioritises is a forecast, not a fact.
Third, there is a live dispute over testing. An unnamed Commission official told Euronews the Act becomes enforceable when a model escapes a test environment and affects the EU market. A practitioner quoted in the same piece argued the Act does not apply to development and testing at all. That question will be settled by a case, not by this post.
Fourth, high-risk classification can pull you into a much heavier regime. If your feature is used for hiring, credit scoring or another Annex III use, Article 25 can make you the provider of a high-risk system from December 2027. This map assumes a general-purpose SaaS feature. It is wrong for a recruitment screening product.
What to ask your model provider for in writing
If the AI Act puts most of the regulatory weight on your supplier, your job is to make sure the supplier's compliance actually flows to you. That happens in contracts, not in policy documents.
| Ask | Article basis | Weak version | Version that works |
|---|---|---|---|
| Integration documentation | 53(1)(b), Annex XII | Link to public model card | Named delivery of the Annex XII package, updated on each model version |
| Notice of regulatory action | 91, 93 | Nothing in the contract | Notice within a set period of any Article 93 measure affecting EU availability |
| Deprecation and withdrawal | 93(1)(c) | Standard deprecation policy | Minimum migration window if a model is restricted or withdrawn in the EU |
| Compliance route | 53(4), 55(2) | General compliance warranty | Statement of whether the provider relies on the Code of Practice, and which chapters |
None of these asks requires you to become a compliance expert. They require your supplier to tell you what it already has to produce for the regulator. The broader data clauses that sit alongside them are covered in the guide to AI clauses in a data processing agreement, and the full vendor review sequence is in the AI agent procurement checklist.
My view is that the withdrawal clause is the one to fight for. Fines against your provider change its margins. An Article 93 withdrawal order changes whether your product works in Europe on Monday morning.
Frequently asked questions
Does the EU AI Act GPAI enforcement apply to companies using the OpenAI API?
Not directly. The Commission's powers under Articles 91 to 93 and its fines under Article 101 apply to providers of general-purpose AI models, such as OpenAI, Anthropic and Google. A company that integrates a model through an API is a downstream provider of an AI system. It carries its own duties, such as Article 50 transparency, but those are enforced by national authorities under Article 99.
What are the AI Office fines for general-purpose AI providers?
Article 101 lets the Commission fine providers of general-purpose AI models up to 3% of annual worldwide turnover or €15 million, whichever is higher. Fines can follow an infringement of the regulation, a failure to answer an information request, refusal of a requested measure, or denial of access for model evaluation. The power applies from 2 August 2026, and the Court of Justice can review any fine.
When does a downstream provider become a general purpose AI model provider?
Under the Commission's July 2025 guidelines, a company modifying a model is presumed to become the provider of a new model when its modification uses more than one third of the original model's training compute. The test is indicative rather than statutory. Even then, the new provider's duties are limited to the modification, and it can supplement the original provider's documentation instead of rewriting it.
What information must a GPAI provider give downstream providers?
Annex XII sets the minimum. It covers the model's intended tasks, acceptable use policies, release date and distribution methods, architecture and parameter count, input and output modalities and formats, licence, and integration requirements such as context window length. It also covers information on training, testing and validation data, including type, provenance and curation methodology. Downstream providers are entitled to receive it under Article 53(1)(b).
What EU AI Act compliance obligations apply to SaaS companies in 2026?
For most SaaS companies building on third-party models, the live obligations are Article 50 transparency, which applied from 2 August 2026, and Article 4 AI literacy, which has applied since February 2025. Article 50 requires telling users they are interacting with an AI and marking synthetic output. High-risk deployer duties under Article 26 apply only to Annex III use cases, from 2 December 2027.
Can the EU force an AI model off the market?
Yes. Article 93 lets the Commission request that a general-purpose AI model provider restrict the making available of a model, or withdraw or recall it. That power sits alongside requests for compliance steps and mitigation measures, and refusing it is a separate ground for a fine under Article 101. For a company whose product depends on a single model, this is the most material downstream risk in the regime.
Where to start this week
Write to your primary model provider and ask for its Annex XII information package by name. Note whether it arrives as a structured document or as a link to a marketing page. The answer tells you how ready your supplier is for the AI Office, and it costs you one email.
Then open your master agreement with that provider and search for the word "withdraw". If nothing covers a regulator-ordered withdrawal in the EU, add it to the renewal agenda now, while the AI Office is still sending letters rather than orders.
Related in this series
The transparency duties that now apply to your own product are mapped in the EU AI Act transparency checklist. For documenting which models sit inside your product, see the AI bill of materials disclosure template.
References
- Regulation (EU) 2024/1689, Article 101, Fines for providers of general-purpose AI models, consolidated text. Used for the fine ceiling and the four triggers. Articles 3, 4, 25, 26, 50, 51, 53, 54, 55, 91, 92, 93, 111 and 113 consulted on the same site.
- Regulation (EU) 2024/1689, Article 53, Obligations for providers of general-purpose AI models, with Annex XII. Used for provider duties and the open-source exemption.
- Regulation (EU) 2024/1689, Article 99, Penalties. Used for operator fine ceilings and the SME rule.
- Euronews, Luca Bertuzzi, EU's AI rules face first enforcement test over lab incidents, 14 October 2026. Used for the information requests and the testing-phase dispute.
- European Commission, The General-Purpose AI Code of Practice, last updated 7 October 2026. Used for signatories.
- WilmerHale, European Commission Issues Guidelines for Providers of General-Purpose AI Models, 24 July 2025. Used for the one-third compute test. Should be upgraded to the Commission guidelines text.
- Gibson Dunn, EU AI Act Omnibus Agreement, 27 May 2026. Used for amended high-risk dates, the Article 4 change and the marking grace period.
- The Parliament Magazine, Europe gets ready to police frontier AI, 31 July 2026. Used for AI Office staffing.
Weakest point in the source base: the one-third compute test is cited through a law firm summary rather than the Commission's guidelines document, and the first enforcement step rests on a single press report. This post is analysis, not legal advice, and reflects the regulation as amended as of 16 October 2026.
Related reading