Skip to main content
Praxikon

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

··21 min read

The AI Office is not a separate agency but a function of the European Commission: references in the AI Act to the AI Office are to be understood as references to the Commission (Article 3(47) AI Act). It enforces the rules for general-purpose AI models and already had powers over AI systems where the model and the system come from the same provider. The Digital Omnibus extended and specified that competence. The AI Office is now exclusively competent for two groups of AI systems: systems that a provider builds on a general-purpose AI model of its own or from its own group, and AI systems in very large online platforms and search engines (Article 75(1) AI Act). That competence applies to the providers of those systems, and to deployers only when they are also the provider or form part of the same undertaking. For a municipality, bank or employer that uses an AI system, this means: the AI Office can act against your vendor, but your own duties fall under national supervision.

Picture this: a Dutch municipality procures a writing assistant from a vendor that develops both the language model and the assistant itself. The assistant does not say that it is AI. Where does the municipality complain, and who will supervise the municipality itself?

In short

Since 21 February 2024 the AI Office has been part of the Commission's Directorate-General for Communications Networks, Content and Technology. It enforces the obligations of providers of general-purpose AI models (Articles 88 to 94 and 101 AI Act). The Digital Omnibus extended and specified its competence over AI systems: it is now exclusively competent for the AI systems in Article 75(1) AI Act, with the powers of a market surveillance authority, fines and periodic penalty payments (Articles 75a to 75d). Those powers apply since 2 August 2026. If you use an AI system that you do not provide yourself, supervision of your own duties stays national. The GDPR is not covered: for your own processing, that is your national data protection authority.

What the AI Office is

The Commission established the office by decision of 24 January 2024 (OJ C, C/2024/1459), before the AI Act itself had been adopted. The decision places the office within the administrative structure of the Directorate-General for Communications Networks, Content and Technology (DG CNECT) and entered into force on 21 February 2024. According to its preamble (recital (7)), the establishment does not affect the powers and competences of national competent authorities in the supervision of AI systems.

The AI Act builds on that. The Commission develops Union expertise and capabilities in AI through the AI Office, Member States facilitate its tasks, and the office is to be allocated adequate resources to exercise its enforcement powers (Article 64 AI Act, paragraph 3 as amended by Regulation (EU) 2026/1744). According to the Commission's page on the AI Office, it employs more than 125 staff across six units and two advisers.

Three European bodies sit alongside the office. The AI Board consists of one representative per Member State; the AI Office provides the secretariat but does not vote (Article 65(2) and (8) AI Act). The advisory forum provides technical expertise from industry, civil society and academia (Article 67 AI Act). The scientific panel of independent experts supports enforcement and can send the AI Office a qualified alert about a general-purpose AI model (Articles 68 and 90 AI Act). An overview of the whole system is in EU AI Act enforcement: the key agencies; this article is about when the AI Office is your supervisor.

What the AI Office does

TaskSourceSince
Supervising and enforcing the obligations of providers of general-purpose AI models: requesting information, evaluating models, requiring measures, fines of up to 3% of worldwide annual turnover or EUR 15 million, whichever is higherArticles 88 to 94, 101 and 111(3) AI ActObligations since 2 August 2025, enforcement since 2 August 2026; models placed on the market before 2 August 2025 by 2 August 2027
Procedural rules for evaluations and fining proceedings against model providersImplementing Regulation (EU) 2026/175510 August 2026
Encouraging and facilitating codes of practiceArticle 56 AI ActCode of practice for general-purpose AI models published on 10 July 2025
Exclusive supervision of the AI systems in Article 75(1), with its own powers, fines and periodic penalty paymentsArticles 75 and 75a to 75d AI Act2 August 2026
A template for a questionnaire, including through an automated tool, for the fundamental rights impact assessment (FRIA)Article 27(5) AI ActThe FRIA duty itself applies from 2 December 2027
A Union-level AI regulatory sandbox for the systems in Article 75(1)Article 57(3a) AI ActA power, with no deadline
Helping national authorities obtain information about a general-purpose AI model, within 30 daysArticle 75(3) AI Act2 August 2026

The code of practice is voluntary. According to the Commission, it and the AI Board have confirmed that the code is an adequate voluntary tool for providers of such models to demonstrate compliance. The Commission also issued guidelines on the scope of the obligations for providers of general-purpose AI models. They are not binding, but they set out how the Commission enforces. More on the powers against model providers is in GPAI enforcement has started.

Article 75(1): the new exclusive competence

The original Article 75(1) already gave the AI Office powers to monitor and supervise compliance of AI systems where the model and the system are developed by the same provider. Regulation (EU) 2026/1744, the Digital Omnibus on AI, rewrote the paragraph: according to its preamble, the scope of that exclusive competence is refined, including to systems within the same undertaking (recital (31)), and it adds exceptions, the very large platforms and the office's own powers in Articles 75a to 75d. The AI Office is now exclusively competent for the supervision and enforcement of all obligations under the AI Act for two groups of AI systems.

Group (a): systems on an in-house model. AI systems based on a general-purpose AI model where the model and the system are developed by the same provider, or by providers forming part of the same undertaking. Four kinds of system are excepted and stay national:

  1. systems related to products covered by the Union harmonisation legislation in Annex I, such as machinery or medical devices;
  2. systems for critical infrastructure (point 2 of Annex III);
  3. systems provided by law enforcement authorities, border management authorities and financial institutions, insofar as they fall under Article 74(6) AI Act;
  4. systems for the administration of justice (point 8 of Annex III).

Group (b): systems in very large platforms. AI systems that constitute or are integrated into a very large online platform or very large online search engine designated under the Digital Services Act. According to the preamble of the Omnibus, the risk assessment, mitigation and audit obligations of the DSA (Articles 34, 35 and 37 of Regulation (EU) 2022/2065) are the first point of entry for those systems (recital (32) of Regulation (EU) 2026/1744).

For whom. The second subparagraph of Article 75(1) is the sentence that matters most in practice: the exclusive competence applies to the providers of those systems, and to deployers only when they are also the provider or form part of the same undertaking as the provider. The preamble says it in so many words: other deployers remain subject to national supervision (recital (31) of Regulation (EU) 2026/1744). AI systems of Union institutions and bodies are not covered either; the European Data Protection Supervisor supervises those (Article 74(9) AI Act).

What the AI Office may then do. For these systems the office has all the powers of a market surveillance authority (Article 75a(1) AI Act). It can open an investigation, on its own initiative or following a complaint under Article 85 AI Act (Article 75a(2)), request information, carry out remote or on-site inspections and order access to systems and explanations (paragraphs 3, 4 and 6). A provider can offer commitments that the office makes binding (Article 75b). If the office finds non-compliance, it can impose fines under Article 99(3) to (7) AI Act, up to EUR 35 million or 7% of worldwide annual turnover for a prohibited practice, and periodic penalty payments not exceeding 5% of the average daily income or worldwide annual turnover in the preceding financial year per day (Article 75c(4) and (5)). It publishes those decisions with the names of the parties (Article 75d(4)). Fines and penalty payments can be appealed to the Court of Justice (Article 75c(6)).

Incidents and conformity. Providers of high-risk systems under this competence report serious incidents to the AI Office instead of the national market surveillance authority; the deadlines of Article 73 AI Act apply mutatis mutandis, so at the latest 15 days, 10 days in the event of a death and 2 days for a widespread infringement or a serious disruption of critical infrastructure (Article 75(1a)). If such a high-risk system needs a third-party conformity assessment, the AI Office is responsible for it; the Commission entrusts the work to notified bodies, which then act on its behalf (Article 75(1e)).

Cooperation with the Member State. If a national market surveillance authority has well-founded reasons to suspect an infringement by a system under Article 75(1), it can ask the AI Office, through the single point of contact of Article 70(2) AI Act, to assess the matter. The office says within four months whether it will act (Article 75(2a)). Before prohibiting or recalling a system on a national market, the office first notifies the national authority (paragraph 1d).

From when. The Omnibus entered into force on 27 July 2026 and gives Article 75 no application date or transitional rule of its own. Article 75 sits in Chapter IX, which under Article 113 AI Act applies since 2 August 2026. The Commission gives the same date for the AI Office's enforcement powers. What the office can enforce depends on what already applies: the prohibitions in Article 5 (since 2 February 2025, the new prohibitions in Article 5(1), points (ba) and (bb), from 2 December 2026), the transparency obligations in Article 50 (since 2 August 2026, with a grace period until 2 December 2026 for the marking under paragraph 2 for systems placed on the market earlier, Article 111(4)) and Article 4. The requirements for high-risk systems in Annex III apply only from 2 December 2027 (Article 113, third paragraph, point (c)).

When it concerns you, and when it does not

Your situationWho supervises your duties under the AI ActSource
You provide a general-purpose AI model yourselfThe AI OfficeArticle 88 AI Act
You build an AI system on your own model or on a model from your own groupThe AI Office, unless one of the four exceptions appliesArticle 75(1)(a)
Your AI system is (part of) a designated very large online platform or search engineThe AI OfficeArticle 75(1)(b)
You build an AI system on another company's model, for example through an APIThe national market surveillance authority; the AI Office only for your complaint about the model providerArticle 75(1) and Article 89(2)
You use a vendor's AI system and are not the providerThe national market surveillance authority, even if your vendor falls under the AI OfficeArticle 75(1), second subparagraph
Personal data are involvedFor your own processing, your national data protection authority; for cross-border processing by a vendor, possibly the lead supervisory authority of its main establishmentArticles 51, 55 and 56 GDPR

Most organisations sit in the last three rows. If you build your own AI system on another company's language model and put it into service under your own name, you are the provider of that system, but the model and the system then do not come from the same provider: Article 75(1)(a) does not apply. A borderline case is the organisation that modifies an open model so substantially that it becomes the model's provider itself; the Commission's guidelines put that threshold at a significant modification. On our reading, the model and the system then do come from the same provider.

In the Netherlands no market surveillance authority for the AI Act has been designated yet. On 20 April 2026 the Dutch government proposed that existing supervisors oversee AI within their own domain, with the Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) where there is no supervisor yet, and the AP and the Dutch Authority for Digital Infrastructure (RDI) in a coordinating role (Rijksoverheid, in Dutch). The Dutch government plans the AI Act Implementation Act for the second quarter of 2027 (Planning letter on digital affairs, Parliamentary Paper 26643, no. 1554, in Dutch). See also AI Act supervision in the Netherlands.

Three examples

The examples and the vendors are invented.

Municipality. A municipality procures a writing assistant from vendor A, which develops the underlying language model itself. That system falls under Article 75(1)(a). The AI Office supervises vendor A, for example on the duty to inform people that they are interacting with an AI system (Article 50(1) AI Act) and on the machine-readable marking of generated text (paragraph 2). If the assistant does not say that it is AI, the municipality can complain through the Commission's complaints tool for systems under the AI Office. The municipality's own duties stay national: measures for AI literacy (Article 4 AI Act) and, if it uses the assistant to publish text informing residents on matters of public interest, the disclosure that the text was generated by AI, unless a human has reviewed it and someone holds editorial responsibility (Article 50(4) AI Act). The DPIA and the processing agreement under the GDPR fall to the data protection authority.

If the municipality builds its own chatbot on vendor B's model, through an API and under its own name, it is the provider of that chatbot, the duty in Article 50(1) rests on it, and it falls under national supervision. If it uses a system to assess whether residents are eligible for benefits (point 5(a) of Annex III), its deployer duties (Article 26 AI Act) and the FRIA (Article 27 AI Act) apply from 2 December 2027. For that FRIA the AI Office is developing a questionnaire template (Article 27(5)); the municipality notifies the results to the market surveillance authority (paragraph 3), on our reading the national one, because the AI Office's competence does not reach it as a deployer. If the system was placed on the market or put into service before 2 December 2027, the high-risk duties apply only if its design changes significantly from that date; because it is intended to be used by a public authority, the provider and the municipality must in any case comply by 2 August 2030 (Article 111(2) AI Act). See the FRIA template.

Bank. A bank uses a system that assesses the creditworthiness of individuals. That is a high-risk system under point 5(b) of Annex III; the duties apply from 2 December 2027 and the bank then carries out a FRIA. For high-risk systems placed on the market, put into service or used by financial institutions, the market surveillance authority is the national authority responsible for the financial supervision of those institutions, insofar as the use is directly connected with the financial service (Article 74(6) AI Act). If the bank builds the system itself on a model from its own group, it would fall under group (a), but exception (iii) keeps it under national supervision. If the bank buys the system from a vendor that builds it on its own model, the text is less clear: exception (iii) refers to systems provided by financial institutions, and the vendor is not one. For the bank's own duties it makes no difference: it is a deployer, not a provider, so supervision stays national. Its customer service chatbot on a vendor's in-house model is usually not a high-risk system; there the AI Office supervises the vendor for Article 50. See from scoring algorithm to transparent credit decision.

Employer. An employer uses a recruitment platform that analyses applications and ranks candidates (point 4(a) of Annex III). The vendor builds it on its own language model. There is no exception for employment and workers management, so the AI Office supervises the vendor, now for Article 5 and, where it applies, Article 50, and from 2 December 2027 for the high-risk duties. The employer stays under national supervision for its own duties: AI literacy now, and from 2 December 2027 human oversight and informing workers' representatives and affected workers in advance, among others (Article 26(2) and (7) AI Act). A prohibited practice illustrates the point. Emotion recognition in the workplace is prohibited, except for medical or safety reasons (Article 5(1)(f) AI Act); if the vendor supplies that feature, the AI Office acts against the vendor, but the employer that uses it answers to the national supervisor. See AI in recruitment and selection.

If you integrate another company's AI model

If you build your own AI system on another company's model, you are a downstream provider (Article 3(68) AI Act). The model provider must give you information and documentation that lets you understand the capabilities and limitations of the model and comply with your own obligations, containing at least the elements in Annex XII (Article 53(1)(b) AI Act). That duty does not apply to providers of models released under a free and open-source licence whose parameters, including the weights, are made publicly available, unless the model has systemic risk (Article 53(2) AI Act). If you do not receive it, you can lodge a reasoned complaint with the AI Office (Article 89(2) AI Act); the Commission runs a separate complaints channel for downstream providers. If a national supervisor cannot conclude an investigation of your high-risk system because information about the model is missing, it can enforce access through the AI Office (Article 75(3)). For your own system, supervision stays national. Whether a modification makes you a provider yourself is covered in when do you become a provider.

What the AI Office is not

Not a data protection authority. Supervision of the GDPR lies with independent national supervisory authorities (Article 51 GDPR), in the Netherlands the Autoriteit Persoonsgegevens. For your own processing, such as a DPIA, the legal basis for training or a data breach involving an AI application, that is your data protection authority, even if your vendor falls under the AI Office. If a vendor processes personal data across borders, the supervisory authority of its main or single establishment may act as lead supervisory authority (Article 56 GDPR).

Not your national market surveillance authority. For all AI systems outside Article 75(1), national authorities supervise. Member States must designate them (Article 70(1) AI Act); the Netherlands has not done so yet.

Not the supervisor of all prohibited practices. The AI Office enforces Article 5 only for the systems in Article 75(1). For all other systems that enforcement lies with the national market surveillance authorities, which report annually to the Commission on the use of prohibited practices (Article 74(2) AI Act).

What to do now

  1. Ask your vendor, per AI system, whether it builds the system on a general-purpose AI model of its own or of its group, and whether it considers itself covered by Article 75(1). Record the answer in your register.
  2. Provide in the contract that the vendor informs you about investigations, commitments and decisions of the AI Office that affect your system; the office publishes decisions on commitments, non-compliance and fines (Article 75d(4)).
  3. Arrange your own duties now: measures for AI literacy (Article 4) and, where they apply, the disclosures in Article 50(3) and (4). From 2 December 2027 the duties in Article 26 are added for Annex III and, for bodies governed by public law, private entities providing public services and for creditworthiness assessment and life and health insurance (points 5(b) and (c) of Annex III), the FRIA (Article 27(1)). If a system was placed on the market or put into service earlier, those duties apply only after a significant change in its design, and for systems intended for public authorities by 2 August 2030 at the latest (Article 111(2)).
  4. Know the complaint route: about a system under the AI Office through the Commission's complaints tool, about a model provider through Article 89(2), about everything else to the national supervisor once designated.
  5. Follow the AI Office's publications, such as the FRIA questionnaire template. If you are unsure whether your system is high risk, start with the case check.

Frequently asked questions about the EU AI Office

Sources

EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 51, 55 and 56 (27 April 2016, accessed 17 November 2026)
European Commission: European AI Office (accessed 17 November 2026)
European Commission: The enforcement framework of the AI Act (accessed 17 November 2026)
European Commission: The General-Purpose AI Code of Practice (code published 10 July 2025, accessed 17 November 2026)
Rijksoverheid (Dutch government): Kabinet zet stap met toezicht op Europese AI-regels (in Dutch) (20 April 2026, accessed 17 November 2026)
House of Representatives of the Netherlands: Planningsbrief Digitale Zaken, Parliamentary Paper 26643, no. 1554 (in Dutch) (accessed 17 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