EU AI Act enforcement: Who can investigate your system?

A practical map of the AI Office, national authorities, the EDPS, GDPR overlap, investigative powers and…

The short version

EU AI Act enforcement divides cases among the AI Office, national authorities, privacy regulators and the EDPS.
National authorities can fine prohibited practices up to €35m or 7% of worldwide annual turnover.
Founders must identify provider roles, regulator jurisdiction, procedural stage and GDPR overlap before answering requests.

The expensive part of an AI Act investigation may begin with a polite email asking for information. Reply badly, or send the complaint to the wrong authority, and things can get spicy fast.

Before losing sleep over fines of up to €35 million or 7% of worldwide annual turnover for prohibited practices, answer the question most founders skip: who actually investigates you?

General-purpose AI models can bring the European Commission’s AI Office into the room. Most other systems go to national competent authorities. The European Data Protection Supervisor handles systems used by EU institutions, while privacy authorities can investigate personal data moving through the same product.

One AI service can attract several regulators, each carrying a different legal instrument. Benvenuti in Europa.

Who Brussels investigates, and why the provider matters

The EU AI Office directly supervises providers of general-purpose AI models, along with specified systems linked to those models and AI inside designated very large platforms or search engines. Most business applications remain under national supervision. Calling every product built on a foundation-model API a Brussels case would be like calling every espresso machine a Lavazza branch. Imagine a lab supplying a general-purpose model while another company wraps it inside recruitment software used to screen applicants. The AI Office supervises the underlying model at provider level, while national authorities usually supervise the independently built recruitment system and its use. Brussels can regain direct system-level responsibility when the system provider belongs to the same undertaking as the model provider, or when the AI sits inside a designated platform or search engine. If nobody at your company can draw the corporate map beside the product diagram, cancel the compliance workshop and start there.

Bar chart comparing current figures against their baselines: Time for one 1024×1024 Qwen-Image-2.1… 54 s versus 28 s, Time for one 1024×1024, 40-step image… 4 s versus 62.5 s, Artificial Analysis Intelligence Index… 38 s versus 58 s.

The Office has serious investigative tools. It can seek information, interview consenting people, require access to models for evaluations and inspect premises in relevant system cases. It may request corrective measures, including restrictions on a model’s public availability. An incorrect or misleading response creates separate sanction exposure. When the Commission sends its request through a formal decision, silence or an incomplete reply can also trigger fines. If the evidence establishes an intentional or negligent breach, the Commission may adopt a penalty decision, with the infringement’s nature, gravity and duration affecting the amount. That polite email deserves more attention than the average SaaS renewal notice.

Complaints travel through the same jurisdictional plumbing. The AI Office assesses a submission confidentially and checks whether the matter falls within its remit. When another authority owns the case, the Office may seek the complainant’s consent and refer it to the relevant national market-surveillance or fundamental-rights authority. Filing a complaint does not automatically open proceedings because the Commission retains discretion over enforcement priorities. Systems outside the Office’s assigned categories stay with national competent authorities. The chain runs from confidential assessment to possible referral, then investigation and, potentially, a finding of breach. Skip those steps in a headline and a routine information request becomes Law & Order: Chatbot Unit.

At the International AI Summit in Dublin on October 14, 2026, Commission Executive Vice-President Henna Virkkunen said the AI Office had sent information requests concerning security practices and was preparing further questions after incidents involving advanced models. The public has not seen the requests, their legal basis, their deadlines or their recipients. We also lack the company names and incident facts. Nobody knows whether the Commission will open formal proceedings, establish a breach or impose a sanction. Confident posts claiming Brussels has already convicted somebody are fan fiction with footnotes.

There is also a real legal dispute around systems escaping an R&D environment. Barry Scannell, a partner at William Fry, argues that development and testing fall outside the Act’s scope and that an accidental escape does not mean the system was placed on the market. The Commission’s reported position focuses on effects reaching the EU market. Scannell’s objection deserves a fair hearing because market impact and placement on the market are different legal concepts. No published decision or court judgment has settled the issue, so certainty currently sells at a generous markup.

Risk classification is not the jurisdiction map

Risk levels determine which obligations an authority can enforce. The system’s provider and institutional user determine who gets the file.

The famous risk pyramid helps until somebody treats it like a vendor scorecard. Classification follows intended purpose and actual use. Prohibited practices include specified forms of harmful manipulation, exploitation of vulnerabilities and social scoring. Transparency duties cover areas such as chatbot disclosure and synthetic-content labelling. General-purpose model obligations sit beside those system categories, so one colourful pyramid cannot explain the enforcement map. A provider may face model-level supervision in Brussels while its customer answers system-level questions from a national authority. The regulator follows the legal role even when the marketing page calls everyone an “AI partner,” a title with roughly the legal precision of “vibes manager.”

The AI Office’s and Member States’ national enforcement powers began applying on August 2, 2026, but only for provisions enforceable at that point. The date did not activate every obligation in one dramatic Brussels switch-flip.

Under the AI Omnibus amendments described by BSI, Annex III high-risk compliance moves to December 2, 2027, compared with the original August 2026 timetable. That gives authorities and deployers more time to build the procedures those systems require.

High-risk systems embedded in regulated Annex I products move to August 2, 2028, one year later than their previous application date. BSI’s assessment is direct:

The AI Omnibus should not be viewed as a rollback of the AI Act.

I agree, provided the extra time produces trained investigators and working evaluation infrastructure. Another last-minute scramble would be European procrastination wearing a conference lanyard.

National competent authorities supervise the bulk of commercial and public-sector AI. Each Member State designates authorities for notification and market surveillance, plus a single point of contact. The law expects adequate technical resources and expertise covering cybersecurity, personal-data protection and fundamental rights. Member States must also assess that capacity periodically. We still do not know how actual staffing, funding or technical capability compare across the Union. A common rulebook will feel wildly different if one authority can inspect a frontier model while another has two lawyers and a heroic spreadsheet. This is where deeper EU coordination earns its keep: local authorities know their markets, while shared European teams can provide the expensive technical evaluation.

A towering dark industrial machine looms beside frail folded spreadsheet paper on a cobalt field.

The penalty ceilings explain why capable enforcement matters. Prohibited practices can attract up to €35 million or 7% of worldwide annual turnover, whichever is higher for a large undertaking. Other listed breaches, including transparency failures, can reach €15 million or 3%.

Supplying incorrect, incomplete or misleading information to a national authority can cost up to €7.5 million or 1% of worldwide turnover. SMEs and start-ups generally face the lower of the fixed ceiling or turnover percentage. Authorities must consider the circumstances of each case, including intent and cooperation. The maximum is a ceiling, never a prix fixe menu.

Passing GDPR does not close the AI Act file

Data-protection compliance covers data-protection duties. It does not give the rest of an AI product a diplomatic passport.

Take an AI hiring tool. A privacy authority may investigate candidate data and the legal basis for processing it, while a market-surveillance authority examines the system’s intended purpose, provider documentation and applicable AI Act duties. Lawfully processed data cannot rescue a prohibited practice. An AI Act classification cannot supply a missing GDPR legal basis. Separate files emerge from the same feature because each regulator is testing a different duty. The overlap becomes especially important when training data or model outputs contain personal data. Founders who put “GDPR compliant” on a sales deck and consider the product cleared have confused one inspection lane for the entire airport.

National AI authorities need data-protection expertise, while privacy regulators keep their existing powers. The European Data Protection Supervisor has a separate AI Act role for systems used by EU institutions. A private employer may hear from its data-protection authority about candidate records while a market-surveillance authority investigates the screening system. The model provider could meanwhile receive questions from the AI Office. Same feature, three legal relationships and enough correspondence to ruin everybody’s aperitivo.

A useful complaint identifies the model provider and system provider, shows who deployed the product and explains where the alleged harm occurred. It should also say whether personal data was involved. AI Office staff assess confidentiality and jurisdiction before deciding whether referral is appropriate, and the complainant’s consent may be needed before the file moves. Referral remains discretionary and guarantees neither proceedings nor compensation. Please do not make a regulator reverse-engineer your corporate org chart. Nobody enjoys that diagram, including the person who made it.

I used to file nearly every European AI investigation under one mental label. That was lazy. Now I check the authority, legal basis, procedural stage and available remedy before treating a case as precedent. As of October 16, 2026, one European monitor still recorded zero published AI Act enforcement actions, even though Virkkunen had discussed information requests two days earlier. Both claims can coexist because confidential information gathering may never produce a published proceeding.

Codes are useless without technical muscle

For model providers, the EU AI Act Code of Practice offers a recognised implementation route. Participation is voluntary, while the underlying legal obligations follow their own timetable.

The general-purpose AI code addresses model obligations, while a separate transparency code supports marking and labelling duties for synthetic content. These codes turn broad requirements into technical practices that companies can document and test. An investigator can compare a system’s behaviour with the legal duty and the provider’s claimed method. Signing a code helps explain the compliance approach. A signature cannot save a label that disappears when someone downloads the file. Compliance lives in the implementation, which has less glamour than a Brussels launch event and considerably more value.

Europe’s sovereignty argument gets very concrete here. We need independent compute and expertise to preserve evidence, evaluate models and verify corrective measures. The AI Office’s Scientific Panel brings together independent specialists in model evaluation and systemic risk, though it advises authorities and does not issue fines. I want shared European evaluation labs that national regulators can call directly. Rebuilding frontier-model capability in every Member State would be fantastically expensive and extremely on-brand.

When Mistral announced its Munich hub, Bavarian State Minister Dr. Florian Herrmann made the political case in the company’s “Hallo, Deutschland!” announcement:

Technological sovereignty is political sovereignty. That is why we need strong European AI champions that take a leading role in international competition. Mistral’s decision to choose Munich sends a strong signal for our position as a high-tech location. Bienvenue à Munich, bienvenue en Bavière!

A European passport grants no compliance shortcut. Building models and infrastructure here still gives Europe greater control over evidence, security decisions and the expertise regulators need when something goes wrong. I support a federal model where national regulators keep sector knowledge and local complaint handling, while EU-level teams conduct model evaluations that smaller Member States cannot sensibly duplicate.

By 2028, I expect Brussels to have centralised much more of that capability. If Europe still has one AI law filtered through 27 wildly different levels of technical capacity, then the founder who cannot name their regulator today has already scheduled tomorrow’s surprise.

Frequently asked questions

Who enforces the EU AI Act?

The European Commission’s AI Office supervises providers of general-purpose AI models and certain linked systems. National competent authorities handle most commercial and public-sector AI systems. The European Data Protection Supervisor covers systems used by EU institutions, while privacy regulators retain authority over personal-data processing.

What are the penalties for breaking the EU AI Act?

Prohibited practices can attract penalties up to €35m or 7% of worldwide annual turnover, whichever is higher for a large undertaking. Other listed breaches can reach €15m or 3%, while incorrect, incomplete or misleading information can cost up to €7.5m or 1%.

Does GDPR-compliant AI automatically comply with the EU AI Act?

No. GDPR-compliant AI satisfies data-protection duties, but the AI Act separately governs prohibited practices, transparency, intended purpose and provider documentation. A privacy authority and a market-surveillance authority can investigate the same product because each regulator examines a different legal duty.