On 2 August 2026 the EU AI Act stopped being a compliance project and became a law somebody can act on. The Commission's own notice is plain about what changed: the Article 50 transparency rules took effect, and three sets of bodies acquired the power to investigate, request documents, evaluate systems and impose fines.
A month later the obvious question is who has been fined. The answer is nobody, publicly, and that answer is far less interesting than the one underneath it.
The more useful question is who can actually act on you, and there the first month produced a genuinely important finding. Enforcement is live everywhere and evenly resourced nowhere. Member States were required to designate their competent authorities and single points of contact by 2 August 2025. As of March 2026, according to the European Parliament's own research service, 8 of the 27 had done so.
That gap is the story. It means identical conduct carries materially different risk depending on which member state notices it.
1. What actually switched on
Three bodies, split by what is being supervised rather than by how serious the conduct is.
National market surveillance authorities handle AI systems on their own market. This is where almost every business will meet the Act. Their powers are the ordinary market surveillance set: request documentation, request data, inspect systems, and in relevant cases reach source code.
The European AI Office handles general purpose AI models and systems built on them, which for most readers means it supervises your model provider rather than you.
The European Data Protection Supervisor covers EU institutions, bodies and agencies. If you are a business, this one is not about you, and it is regularly misreported as though it were.
The penalty ceiling for Article 50 breaches is 15 million euro or 3 percent of total worldwide annual turnover, with proportionality provisions for small and medium businesses, and a separate ceiling of up to 750,000 euro for EU institutions. Note what that is not: the 35 million euro or 7 percent tier belongs to the prohibited practices in Article 5, not to transparency. We set the tiers out in full in the EU AI Act 2026 guide, and the distinction matters because most coverage quotes the largest number available regardless of which obligation is in question.
The Commission also published three things alongside the rules that are worth knowing about: guidelines on the transparency obligations, a code of practice on marking AI generated content, and a set of icons for labelling. The code of practice is voluntary. The Article 50 obligations underneath it are not, and signing the former does not discharge the latter.
2. Why no fines yet is the expected result
Market surveillance is an ex post regime. Nothing about it produces a fine in week three.
The sequence is: a complaint or a market sweep brings a system to an authority's attention, the authority requests documentation, it evaluates what comes back, and only then does it reach a finding and consider a sanction. Each of those steps takes weeks at best. Add that most authorities are simultaneously being stood up, hiring, and working out their own procedures, and the gap between powers existing and cases appearing is measured in quarters.
So the absence of a first fine says nothing about whether the rules bite. What it does mean is that the businesses acting now are doing so before there is an example to react to, which is the more comfortable position to be in.
3. The uneven map is the actual risk
Here is the part we would want a client to understand.
The Act is enforced through a decentralised model for AI systems, with national authorities acting on their own market, and a centralised model for general purpose models, with the AI Office acting alone. The European Parliament's research service notes the obvious consequence: the decentralised pattern dominates, and it produces uneven enforcement across the Union.
With 8 of 27 single points of contact designated a year after the deadline, that is not a theoretical concern. In practice it means three things.
Your exposure is geographic, not just categorical. The same chatbot, deployed identically, sits under a well resourced authority in one member state and a not yet designated one in another. That is a fact about where your users are, which is not usually how compliance risk gets modelled.
The gap will close unevenly and without warning. An authority that designates and staffs in Q4 may open with a sweep of a visible sector. Being compliant before that is cheap; being compliant during it is not.
Do not read slow enforcement as weak law. GDPR looked dormant for roughly eighteen months and then did not. The regime here is younger, the authorities are newer, and the obligations that are live are simpler to check than most of GDPR's. Article 50 in particular is unusually easy to test from the outside: a regulator can open your chatbot and see whether it disclosed.
4. What an authority actually asks for
Strip the process down and a market surveillance request resolves to five questions.
What is the system. What was it intended to do. Which role do you occupy, provider or deployer. How was the required disclosure delivered, and when. What did a human review, and who was it.
Every one of those is a record, and the difference between a comfortable response and a bad month is whether those records are a byproduct of running the system or something your team reconstructs under time pressure. Disclosure is the one that cannot be recreated afterwards: either you logged that it was delivered at the moment it was delivered, or you are asking an authority to take your word for it.
This is exactly the argument we made from the engineering side in Engineering for the EU AI Act, and enforcement starting has not changed it. A transparency duty is a property of the system, not a policy document. It is engineered in, or it does not exist. The pleasant consequence is that the same append only trail that answers a regulator is the one that answers a security incident, which we set out in the agent security paper. Build it once, for either reason, and it serves both.
For anyone running agents that correspond with people, the human review record is the one that gets tested first, and the discipline around it is the subject of Governed autonomy. A review that took four seconds is not a review, and a log that says a human approved something without recording who or when is not evidence.
5. What we would do this quarter
Five things, and none of them are a policy document.
- Map your users by member state. You cannot reason about uneven enforcement without knowing which authorities you are actually exposed to.
- Test your own disclosure the way a regulator would. Open the chatbot. Is it obvious, at first contact, that it is not a person. If it takes three messages, that fails.
- Check that disclosure is logged, not just implemented. These are different things and only one of them is evidence.
- Write down which role you occupy for each system, provider or deployer, and why. This single line determines which obligations attach and it is the first thing anyone will ask.
- Nominate the person who answers. A request arriving with no named owner burns a week before anyone reads it.
None of that is expensive. All of it is much cheaper now than during a sweep, which is the entire argument for doing it in a quiet quarter.
6. Sources
- European Commission, safer and more transparent AI, 2 August 2026, for what took effect and which bodies gained powers.
- European Parliamentary Research Service, enforcement of the AI Act, March 2026, for the designation figures and the fragmentation analysis.
- Regulation (EU) 2024/1689, the AI Act itself.
- European Commission, guidelines on the transparency obligations and the regulatory framework overview.
- AI Act Service Desk, the official implementation resource.
This article is general information and not legal advice. The designation figures are those published in March 2026 and more member states will have designated since; check the current position for the countries you operate in. No publicly reported enforcement decision existed at the time of writing, and we have not inferred one.
This guide sits in our Operations research, alongside the governance and operations threads. The engineering companion is in Engineering, and the full library is at Research.