Skip to main content
Praxikon

How the Commission supervises general-purpose AI models: Articles 90 to 94 of the AI Act

··24 min read

A Dutch municipality builds a digital assistant that answers residents' questions about waste collection, parking and passports. It develops the assistant itself, but under the hood runs the language model of a large technology company, called through an API. That language model is a general-purpose AI model (Article 3, point (63)). Compliance of the model provider with its AI Act obligations is supervised not by a Dutch authority but by the European Commission: it has exclusive powers to supervise and enforce Chapter V and entrusts the implementation of these tasks to the AI Office (Article 88(1)).

How that supervision works is set out in Article 90 to Article 94. This piece goes through those articles paragraph by paragraph and then looks at them from the municipality's side: what does it notice when the Commission requests measures from its model provider, what can it do itself, and what does it record? A bank follows, and briefly education and HR. The examples are fictitious.

In brief

The route has four steps towards providers of general-purpose AI models: a qualified alert from the scientific panel to the AI Office (Article 90), followed by three powers of the Commission, exercised through the AI Office: a request for documentation and information (Article 91), an evaluation of the model, if necessary with access through APIs or the source code (Article 92), and a request for measures, up to restricting the making available on the market, withdrawing or recalling the model (Article 93). The provider has the procedural rights of Article 18 of Regulation (EU) 2019/1020 (Article 94). A provider that does not comply with a request risks a fine of up to 3% of worldwide turnover or EUR 15 million, whichever is higher (Article 101).

The Digital Omnibus on AI (Regulation (EU) 2026/1744) did not amend these articles. They are in Chapter IX and have applied since 2 August 2026; so has Article 101 (Article 113).

All these powers are directed at the provider of the model. A municipality or bank that builds the model into its own system is not a party to the procedure and has no rights under Article 94. As a downstream provider, it can lodge a complaint (Article 89(2)).

This piece explains the articles one by one. The news on the start of enforcement is in GPAI enforcement has started. When the AI Office also supervises AI systems is covered in the EU AI Office: what it does and when it supervises you.

The framework: Articles 88 and 89

Articles 88 to 94 form Chapter IX, Section 5. Article 88(1) gives the Commission supervision of Chapter V, "taking into account the procedural guarantees under Article 94". National market surveillance authorities may request the Commission to exercise these powers where that is necessary and proportionate to assist with their own tasks (Article 88(2)), without prejudice to Article 75(3).

The AI Office may take the necessary actions to monitor the effective implementation and compliance by providers of general-purpose AI models, including their adherence to approved codes of practice (Article 89(1)). Downstream providers have the right to lodge a complaint alleging an infringement of the Regulation. The complaint is duly reasoned and indicates at least the point of contact of the model provider, the facts, the provisions concerned and the reason for the complaint (Article 89(2)). A downstream provider is a provider of an AI system which integrates an AI model, also where the model is provided by another entity based on contractual relations (Article 3, point (68)). An organisation that merely uses a ready-made system from a supplier is not one. The Commission has a separate channel for these complaints, which it limits to infringements of Articles 53 to 55.

The four steps at a glance

StepWhoCondition in the textFine for non-compliance
Alert (Article 90)Scientific panel, to the AI OfficeReason to suspect a concrete identifiable risk at Union level, or that the model meets Article 51Not applicable
Documentation and information (Article 91)Commission, possibly after a structured dialogueNecessary to assess complianceArticle 101(1)(b)
Evaluation (Article 92)AI Office after consulting the Board, possibly with independent expertsInformation under Article 91 insufficient, or investigation of systemic risksArticle 101(1)(d)
Measures (Article 93)Commission, possibly after a structured dialogueNecessary and appropriateArticle 101(1)(c)

Step 1: the qualified alert (Article 90)

The scientific panel of independent experts (Article 68) may provide a qualified alert to the AI Office where it has reason to suspect that a general-purpose AI model poses concrete identifiable risk at Union level (paragraph 1, point (a)), or that a model meets the conditions referred to in Article 51 for a model with systemic risk (paragraph 1, point (b)). Point (a) covers any model, not only models already classified as having systemic risk.

The alert is duly reasoned and indicates at least the provider's point of contact, the relevant facts and the reasons, and any other information the panel considers relevant (paragraph 3). Upon the alert, the Commission, through the AI Office and after having informed the Board, may exercise the powers of the Section for the purpose of assessing the matter; the AI Office informs the Board of any measure according to Articles 91 to 94 (paragraph 2). The text says "may": the alert does not oblige the Commission to act. An alert under paragraph 1, point (a), can also lead to a decision designating the model as presenting systemic risks (Article 52(4)).

A qualified alert requires a decision by at least a simple majority of the members of the panel (Article 18 of Implementing Regulation (EU) 2025/454). The AI Office processes a complete alert within two weeks and decides whether to launch measures under Articles 91 to 93; if it does not, it closes the alert (Article 19 of Implementing Regulation (EU) 2025/454). A municipality or bank cannot issue a qualified alert itself; only the panel can.

Step 2: documentation and information (Article 91)

The Commission may request the provider to provide the documentation drawn up in accordance with Article 53 and Article 55, or any additional information necessary to assess compliance (Article 91(1)). Before sending the request, the AI Office may initiate a structured dialogue with the provider (paragraph 2). Upon a duly substantiated request from the scientific panel, the Commission may also request information where that is necessary and proportionate for the panel's tasks (paragraph 3).

The request states the legal basis, the purpose, what information is required, the period for providing it and the fines under Article 101 for supplying incorrect, incomplete or misleading information (paragraph 4). The provider or its representative supplies the information. Lawyers duly authorised to act may do so for their clients; the clients remain fully responsible (paragraph 5). A provider established in a third country has, except for a model released under a free and open-source licence without systemic risk (Article 54(6)), an authorised representative in the Union, which provides information to the AI Office upon a reasoned request (Article 54(3)(c)).

On its enforcement page the Commission distinguishes a simple request for information by the AI Office, where a fine can follow for an incorrect or misleading reply, from a request issued by decision, where it can also follow for failing to reply or replying incompletely. Article 91 and Article 101(1)(b) do not make that distinction themselves; the Commission says its overview does not bind it.

Step 3: evaluating the model (Article 92)

The AI Office, after consulting the Board, may conduct evaluations of the model for two purposes: to assess the provider's compliance where the information gathered under Article 91 is insufficient (paragraph 1, point (a)), or to investigate systemic risks at Union level of models with systemic risk, in particular following a qualified alert under Article 90(1), point (a) (paragraph 1, point (b)). The Commission may appoint independent experts, including from the scientific panel, to carry out evaluations on its behalf; they must meet the criteria of Article 68(2) (paragraph 2).

For the evaluation, the Commission may request access to the model through APIs or further appropriate technical means and tools, including source code (paragraph 3). The request for access states the legal basis, the purpose and reasons, the period for providing access and the fines under Article 101 for failure to provide access (paragraph 4). Before requesting access, the AI Office may initiate a structured dialogue to gather more information on the internal testing of the model, internal safeguards for preventing systemic risks and other measures the provider has taken (paragraph 7).

The implementing acts of paragraph 6 are in Implementing Regulation (EU) 2026/1755, in force since 10 August 2026. Access may include access through APIs, to source code, to model weights and to the hosting infrastructure, and must not be subject to constraints that materially impede an appropriate evaluation (Article 2 of Implementing Regulation (EU) 2026/1755). The provider may submit reasoned observations regarding an appointed expert (Article 3 of Implementing Regulation (EU) 2026/1755).

Step 4: requesting measures (Article 93)

Where necessary and appropriate, the Commission may request providers to (Article 93(1)):

  • take appropriate measures to comply with the obligations set out in Articles 53 and 54 (point (a));
  • implement mitigation measures, where the evaluation carried out in accordance with Article 92 has given rise to serious and substantiated concern of a systemic risk at Union level (point (b));
  • restrict the making available on the market, withdraw or recall the model (point (c)).

Three things stand out. Point (b) first requires an evaluation under Article 92; a suspicion alone is not enough. Point (c) sets no condition of its own beyond "necessary and appropriate" and is not limited to models with systemic risk; in our reading it is still the heaviest tool, which proportionality reserves for cases where lighter measures do not suffice. And point (a) names Articles 53 and 54, not Article 55; in our reading a request for measures on the additional obligations for models with systemic risk runs through point (b) or point (c). An infringement of Article 55 can also be fined directly (Article 101(1)(a)). What recalling a model means is not defined: Article 3, points (16) and (17), define recall and withdrawal only for AI systems.

Before a measure is requested, the AI Office may initiate a structured dialogue (paragraph 2). If, during that dialogue, the provider of a model with systemic risk offers commitments to implement mitigation measures, the Commission may, by decision, make those commitments binding and declare that there are no further grounds for action (paragraph 3). When fixing a fine, the Commission takes into account such commitments and commitments made in codes of practice under Article 56 (Article 101(1), second subparagraph).

The implementing regulation adds two things. Before opening proceedings, the Commission may, on grounds of urgency due to a risk of serious damage to health, safety requirements or other grounds relating to the public interest covered by the AI Act, order interim measures by decision, such as preventing a model from being made available on the market, based on a prima facie finding of an infringement (Article 5(3) of Implementing Regulation (EU) 2026/1755). And closed proceedings can be reopened, for example when commitments made binding are not respected (Article 6(2) of Implementing Regulation (EU) 2026/1755).

Procedural rights of the provider (Article 94)

Article 94 makes Article 18 of Regulation (EU) 2019/1020, the Market Surveillance Regulation, apply mutatis mutandis to the providers of the model, without prejudice to more specific procedural rights in the AI Act. Under that provision, a measure, decision or order states the exact grounds on which it is based. It is communicated without delay, together with information on the remedies and their time limits. Beforehand, the economic operator is given the opportunity to be heard within an appropriate period of not less than 10 working days, unless urgency makes that impossible; in that case the hearing follows as soon as possible and the measure is reviewed promptly.

That provision refers to remedies under the law of the Member State. Which remedies are available against a request for measures by the Commission is not spelled out in Article 94. In our reading, an action for annulment before the EU Courts is available if the request is intended to produce legal effects (Article 263 TFEU); whether a request under Article 93 does, the text does not say. Only for fines does the AI Act name the court: the Court of Justice has unlimited jurisdiction and may cancel, reduce or increase the fine (Article 101(5)). More specific rights are in Article 101(2) (preliminary findings and the right to be heard) and in the implementing regulation: no less than 21 days for written observations on preliminary findings, and access to the file after those findings (Articles 7 and 8 of Implementing Regulation (EU) 2026/1755).

Fines (Article 101)

The Commission may impose on a model provider a fine not exceeding 3% of its annual total worldwide turnover in the preceding financial year or EUR 15 000 000, whichever is higher, when it finds that the provider intentionally or negligently infringed the relevant provisions of the Regulation, failed to comply with a request under Article 91 or supplied incorrect, incomplete or misleading information, failed to comply with a measure requested under Article 93, or failed to give access for an evaluation under Article 92 (paragraph 1, points (a) to (d)). The amount depends on the nature, gravity and duration of the infringement. A lower cap for SMEs or small mid-cap enterprises, as Article 99(6) and (6a) provide for other fines, is not in Article 101.

The power to fine is time-barred five years after the conduct or, for continuing or repeated conduct, five years after it ceases. Any action taken by the Commission for the investigation interrupts that period, such as a request for information or access, an invitation to a structured dialogue or the opening of proceedings, but the power expires after ten years at the latest, not counting the time the case is pending before the Court of Justice (Article 10 of Implementing Regulation (EU) 2026/1755). The AI Act does not require a public register of requests or fines addressed to model providers; information on fines is communicated to the Board as appropriate (Article 101(4)).

From when?

  • The AI Act entered into force on 1 August 2024, the twentieth day after publication (Article 113, first paragraph). Entry into force is not the same as application.
  • The obligations of model providers (Chapter V) have applied since 2 August 2025, as has Chapter VII, which contains the scientific panel (Article 113, third paragraph, point (b)).
  • Articles 88 to 94 (Chapter IX) have applied since 2 August 2026, the general date of application (Article 113, second paragraph). Article 101 is in Chapter XII but is excepted from 2 August 2025, so it too has applied since 2 August 2026 (Article 113, third paragraph, point (b)).
  • Models placed on the market before 2 August 2025 must comply by 2 August 2027 (Article 111(3)). How that transition period relates to the powers in Articles 90 to 93 is not expressly settled in the text.
  • The Digital Omnibus on AI entered into force on 27 July 2026. It amended Article 113, but not Articles 88 to 94 or 101.

Main example: the municipality with an assistant for residents' questions

The municipality develops the assistant and puts it into service under its own name. It is therefore a provider of an AI system (Article 3, point (3)) and a downstream provider (point (68)); the technology company is the provider of the model. The AI Office does not supervise the municipality itself, because model and system do not come from the same provider or undertaking (Article 75(1), point (a)). Its supervisor will be national; in the Netherlands none has yet been designated for the AI Act.

What does it notice of a request for measures? If the Commission requests the model provider to implement mitigation measures or to restrict the making available of the model, the assistant may start behaving differently, lose a function or stop working for a while. The municipality is not a party: Articles 91 to 93 are directed at the model provider, and the rights under Article 94 are its rights. The AI Act does not oblige the Commission to publish such a request, nor does it expressly oblige the model provider to inform its customers of one. The model provider must, however, keep information and documentation up to date and make it available to providers that integrate the model, so that they understand its capabilities and limitations and can comply with their own obligations (Article 53(1)(b)); the obligation does not apply to a model released under a free and open-source licence whose parameters, including the weights, are made public, unless it has systemic risk (Article 53(2)). In our reading, a measure that changes the model's behaviour or availability should be reflected in that updated information. When and how exactly, the text does not say.

What can it do itself? It can lodge a complaint if the model provider does not meet its obligations towards it, for example if updated information is not forthcoming (Article 89(2)). If the Commission opens proceedings following that complaint and intends to close them, the municipality first gets the possibility to express its views (Article 6(1) of Implementing Regulation (EU) 2026/1755). It can also agree by contract that the provider informs it of requests, commitments and changes, and prepare a route to another model.

Its own obligations remain. Since 2 August 2026, as provider, it ensures that residents are informed that they are interacting with an AI system, unless that is obvious (Article 50(1)); for the generated text, the machine-readable marking of paragraph 2 may apply; for systems placed on the market before 2 August 2026, by 2 December 2026 at the latest (Article 111(4)). See chatbots and the duty to disclose AI. It also takes measures for AI literacy (Article 4). An assistant that answers general questions is in principle not a high-risk system. If the municipality starts using it to assess whether residents are eligible for a benefit, that can change (Annex III, point 5(a)), with obligations from 2 December 2027: as provider, the high-risk requirements and registration (Article 49(1)), and as deployer, registration of the use (Article 49(3); Article 26(8)) and a fundamental rights impact assessment (Article 27(1)).

The privacy side. The assistant processes residents' questions, often with personal data. Record what role the model provider has for what comes in through the API (Article 28 GDPR if it is a processor), whether it uses that data for training, and whether a DPIA is required (Article 35 GDPR). The municipality's own processing is supervised by the Dutch Data Protection Authority (Autoriteit Persoonsgegevens), also when the model provider falls under the Commission. Where the model provider processes the data cross-border, for instance for training, the authority of its main establishment may act as lead supervisory authority (Article 56 GDPR). If it uses the data for its own purposes, such as training, it is in our reading itself a controller for that use, and the municipality needs a legal basis to disclose the data.

Example 2: the bank and the insurer

A bank builds its own assistant, on another provider's model, that helps staff summarise credit files. Like the municipality, it is a downstream provider, with the same complaint route and the same dependency: if the Commission restricts the making available of the model (Article 93(1)(c)), the process may come to a halt. In our reading, that dependency on a single model therefore belongs in the bank's outsourcing and continuity policy.

If the bank starts using the system to evaluate the creditworthiness of natural persons, it is in principle a high-risk system (Annex III, point 5(b)), with obligations from 2 December 2027: as provider, the high-risk requirements and registration (Article 49(1)), and as deployer, a fundamental rights impact assessment (Article 27(1)). The market surveillance authority is then in principle the financial supervisor, in so far as the use is directly connected with the provision of the financial service, unless the Member State, in justified circumstances and provided coordination is ensured, designates another authority (Article 74(6) and (7)). For an insurer that builds a model into risk assessment and pricing in relation to natural persons for life and health insurance, the same applies through Annex III, point 5(c), including the fundamental rights impact assessment. Supervision of the model provider stays with the Commission in every case.

In brief: education and HR

A university of applied sciences that builds a study assistant on another provider's model is in the same position as the municipality. If it uses the assistant to evaluate learning outcomes, it may be high-risk (Annex III, point 3(b)), with obligations from 2 December 2027; emotion recognition in education is prohibited, except for medical or safety reasons (Article 5(1)(f)). An employer that uses a ready-made recruitment platform from a supplier is not a downstream provider but a deployer. Such a platform is in principle high-risk (Annex III, point 4(a)), with obligations from 2 December 2027, and emotion recognition in the workplace is already prohibited (Article 5(1)(f)). If the supplier builds on another provider's model, the supplier, not the employer, has the complaint right under Article 89(2); the employer arranges the information it needs by contract. If the supplier builds on its own model, the AI Office supervises that system (Article 75(1), point (a)). When a change makes you a provider yourself is covered in when do you become a provider.

What you record now

  1. Which model, from whom. Per application: model name and version, the provider and its authorised representative in the Union, and whether the model was placed on the market before or after 2 August 2025.
  2. Systemic risk or not. Ask the provider whether its model has been classified as a model with systemic risk; the Commission must publish and keep up to date a list of such models (Article 52(6)).
  3. The documentation. What you received under Article 53(1)(b) and Annex XII, what you asked for and did not get, and when it was updated.
  4. Contract and fallback. Agreements on information about requests, commitments and model changes, a retest after every change, and a plan for when the model is restricted or withdrawn.
  5. Your role and route. As a downstream provider you have the complaint right under Article 89(2); as a deployer only, the route runs through your supplier.

Want to know which AI Act and GDPR obligations come into play for your application and what role your organisation has? With the case check you describe the application in a few questions.

Frequently asked questions on the supervision of general-purpose AI models

Sources

EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI); Articles 88 to 94 and 101 not amended (8 July 2026, published 24 July 2026, accessed 19 November 2026)
EUR-Lex: Treaty on the Functioning of the European Union, consolidated version, Article 263 (OJ C 202, 7 June 2016, accessed 19 November 2026)
EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 28, 35 and 56 (27 April 2016, accessed 19 November 2026)
European Commission: The enforcement framework of the AI Act (accessed 19 November 2026)

Newsletter

Every Tuesday, the AI Act week ahead in 5 minutes

A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.

Practical and short · No spam · One-click unsubscribe

Written by

Zahed Ashkara, freelance AI and privacy consultant and legal adviser

Zahed Ashkara

Freelance AI and privacy consultant and legal adviser · AI governance & safety

Zahed Ashkara is a freelance AI and privacy consultant and legal adviser for the GDPR and the EU AI Act. Organisations hire him as AI governance consultant, privacy counsel, external DPO or AI governance lead, have him carry out a DPIA, FRIA or AI Act classification, or book an in-company training. He focuses on AI governance and safety, including AI agents. He works exclusively as a freelancer, through Embed AI (trade name Praxikon). He is a Certified AI Compliance Officer (CAICO) and has been a member of the NEN standards committee Artificial Intelligence & Big Data since July 2026.

Follow on LinkedInView expert profile

Use these sources in your AI assistant too: Praxikon in Claude and other assistants

Real-World Stories

Discover how organizations apply this in practice

Related articles

How a general-purpose AI model is classified as a model with systemic risk: Article 52 of the AI Act

Article 52 of the AI Act sets out how a general-purpose AI model becomes a model with systemic risk: notification within two weeks, counter-arguments from the provider, designation by the Commission ex officio or following a qualified alert from the scientific panel, reassessment upon request and a published list. What each paragraph says, how Article 51 and Annex XIII come in, and what a municipality, bank or university of applied sciences building on such a model will notice.

EU AI ActArticle 51Article 52

The EU AI Office: what it does and when it supervises your organisation

The European Commission's AI Office enforces the rules for general-purpose AI models and is exclusively competent for two groups of AI systems, a competence that the Digital Omnibus extended and specified (Article 75(1) AI Act). For most municipalities, banks and employers it is not the supervisor of their own duties. When it is, when it is not, and what to arrange now.

EU AI ActAI OfficeArticle 75