The EU AI Office has now published three completed enforcement actions under the Act. Reading them in sequence is more instructive than reading any of them on its own.

1. The first action, dated late February. €4.2 million. A medium-sized European labour-platform provider deploying a model used to score worker eligibility for shift assignments. The cited violation: failure to maintain the documentation regime required for a high-risk-category deployment. The model itself was not the issue. The records around the model were the issue.

2. The second action, dated mid-March. €11.8 million. A US lab that had been licensing a general-purpose model into the EU market without filing the systemic-risk attestations required at its scale tier. The cited violation: structural — failure to register at the threshold the lab itself acknowledged crossing. The fine, against the lab’s quarterly revenue, was small enough that internal communications described it as a cost of operations.

3. The third action, dated last week. €0.6 million. A small Spanish startup deploying a generative model into a customer-service application without maintaining a clear opt-out mechanism for end users. The cited violation: end-user-rights compliance. The fine itself was modest. The publication of the action, which will sit on the AI Office register for the firm’s lifespan, is the sanction.

What the three actions reveal, taken together, is the shape of the enforcement architecture as it actually operates.

The Act, as drafted, was framed around constraining the largest models. The first three actions are, in volume, evenly split between the largest, the medium, and the small — by count, not by total fine value. The architecture is reaching across the full size spectrum. This is, in some respects, a feature: the Act applies to all comers and the Office is acting accordingly.

It is also, in another respect, a feature with consequences. The transactional cost of compliance is identical across the spectrum, in the sense that every regulated entity has to maintain the same documentation regime. The fixed cost of compliance is harder for the smallest entity to bear; the marginal fine for the largest entity is easier to absorb. The Act, in practice, hits hardest on the middle.

This is a known regulatory pattern. Every documentation-heavy regime hits the middle hardest. The Act is no exception, and the first three actions are evidence — not anomaly — of the pattern.

A useful number: the small startup’s fine, calculated as a percentage of revenue, is roughly four times the lab’s fine on the same metric. The Act, as enforced, is in absolute terms most painful for the entity least able to absorb it. The framing of the press release — proportionate enforcement — obscures the ratio.

Two structural observations:

— The Office has not yet brought a frontier-lab action that targets the model itself, as opposed to the structural compliance around the model. The reason is documentation access. Frontier labs, even when registering, control the data that would be needed to assess their model’s performance against the Act’s substantive criteria. The Office has not yet developed the audit machinery to evaluate the data on its own terms.

— The Office has signalled, in three public statements this quarter, that the next phase will involve substantive model assessments. The audit machinery is being built. The timeline for the first such action is not before late 2026.

The first three fines, in summary, were procedural. The next three will, if the Office holds its stated trajectory, begin to be substantive. That is a meaningful shift, and the labs paying close attention are, by all account, preparing for it.

For now: the Act is operating. It is operating in the predictable way regulatory regimes operate at this stage. The exceptional thing would be if it stopped.