AI Office and AI Act Service Desk

What is the AI Office?

The AI Office is the European Commission's own AI regulator. A Commission decision of 24 January 2024 set it up inside DG CNECT, the department for communication networks, content and technology. It has more than 125 people across six units, among them one on regulation and compliance and one on AI safety.

The AI Act Service Desk is the front door the same Commission built for everyone else. It went live on 8 October 2025, and it is where a company can put a question about the regulation to a team of experts working with the AI Office.

Together they are the European floor of the AI Act: one body that supervises the largest model providers itself, one counter where everyone else can ask what the rules mean. The floor you are far more likely to meet is the national one, which has its own entry on AI Act enforcement in Belgium.

What the AI Office supervises on its own

Article 88 gives the Commission exclusive powers to supervise and enforce Chapter V of the AI Act, the chapter with the obligations of providers of general-purpose AI models, and the Commission hands that work to the AI Office. Exclusive is the load-bearing word: no national authority in any member state can open a file against OpenAI, Google, Anthropic or Mistral about the model itself. The Digital Omnibus on AI, in force since 27 July 2026, added certain AI systems a model provider builds on its own model, and AI systems inside the very large online platforms and search engines the Digital Services Act covers.

The second half of the job is writing rather than policing. The AI Office runs the process behind the codes of practice and prepares the Commission guidelines on questions like the prohibited practices and the definition of an AI system. Those texts carry no article number, but a national inspector reads them as the yardstick.

The AI Office versus a national market surveillance authority

Ask both the same question, which systems does it supervise, and the split is clean.

The AI Office supervises general-purpose AI models and the companies that make them, anywhere in Europe. Its counterparts are a handful of large model providers, and it can request documentation, evaluate a model itself and fine the provider up to 15 million euros or 3 percent of worldwide turnover. It will never look at the chatbot on your website.

A national market surveillance authority supervises AI systems on the market or in use in its own country: risk class, technical documentation, human oversight, labels, registration. Its counterparts are every company in that country, from a hospital group to a bakery, and it has nothing to say about how a model was trained.

The line runs along the model, not along company size. If you built it or bought it and put it in front of customers or staff, it is national.

The other bodies whose names turn up

The AI Board has one representative per member state, with the AI Office as secretariat and the European Data Protection Supervisor there as an observer. It coordinates joint market surveillance work and issues opinions, so that 27 authorities read the regulation the same way.

The scientific panel is the group of independent experts in Article 68, and the Commission named 60 of them on 1 June 2026. Its sharpest instrument is the qualified alert in Article 90: when it suspects a general-purpose model carries a concrete risk at Union level, it flags that to the AI Office, which decides whether to investigate.

The advisory forum is the stakeholder body, with members from industry, start-ups, SMEs, civil society and academia, and with the Fundamental Rights Agency, ENISA, CEN, CENELEC and ETSI as permanent members. It was named the same day and advises the Board and the Commission.

What the AI Act Service Desk gives you today

Four things sit on the platform. The AI Act Explorer is the full text with annexes and recitals in a readable interface. The compliance checker is a questionnaire that walks you through scope and answers, indicatively, whether you are a provider or a deployer and what follows from that; in September 2026 it still carries a beta label. The timeline lists the application dates, and the question form reaches the expert team behind the desk.

Two limits before you lean on it. The platform was promised in all 24 official languages by early 2026, and today it runs in English, French, German, Spanish, Italian and Polish, so a Flemish company reads its own regulation in a second language. And the desk gives information, not permission: an answer does not bind a national authority, does not settle a dispute with a vendor and is not a defence in court.

Both the tools and the rules under them are still moving: the omnibus rewrote dates and widened the AI Office's powers within two years of the Act, and the scientific panel and the advisory forum only got their members in mid-2026. Write the date next to any answer you file.

Where a Belgian company takes its question

Take a Hasselt software company that sells a portal to accounting firms, with a chatbot on top of a commercial model. Three questions land on the same morning and go three different ways.

Does the chatbot have to say it is a chatbot? That is Article 50 and your own obligation as the provider of that system. Questions about your own obligations go to the national authority, and until Belgium has one, to the Service Desk form.

A competitor's AI-generated product photos carry no label. A complaint about a system is an Article 85 route to the market surveillance authority of your own country, never to the AI Office. Belgium has not passed the law that gives one the powers, so today that leaves the courts, and the Gegevensbeschermingsautoriteit where personal data are involved.

Our model vendor will not confirm it published the summary of its training content. That obligation is the vendor's and the AI Office polices it, so nobody opens a case on your behalf. Your lever is the contract: ask in writing at procurement and put the answer in your AI register.