On 12 October 2026 the European Commission opened a four-week consultation on a draft implementing regulation, registered as Ares(2026)9589403, that sets out how the AI Office investigates and fines certain AI systems. Feedback closes on 9 November, and adoption is planned for the fourth quarter, according to PPC Land's reading of the draft. The text implements Articles 75c and 75d of the AI Act. It covers AI systems built on a general-purpose model by the same provider, and AI integrated into designated very large online platforms and search engines.
The case for a firm procedure
The strongest argument for the Commission is that enforcement powers without procedure are a liability for everyone. Article 75a already lets the AI Office request information, order access to AI systems, appoint independent experts and carry out inspections. A regulator with those powers and no written rules on evidence, secrecy and deadlines would invite litigation and inconsistent practice. Procedural rules are also what the Court of Justice, which has unlimited jurisdiction over AI Office fines, will use to test the Office's decisions. Writing them down early is a sign of seriousness.
On several points the draft is better than the usual enforcement settlement.
What the draft gets right
According to PPC Land's summary, the addressee gets no file access before preliminary findings are notified. After that, it receives the documents cited in those findings with business secrets redacted. A second layer gives a limited number of named outside lawyers, economists and technical experts the full unredacted file. Those advisers must not be employed by the addressee, and they must tell the Commission if they join the addressee or a competitor during the case or within three years afterwards.
This adviser-only model lets a firm test the evidence against it without handing competitors or its own engineers a rival's trade secrets. The business-secrets rules are similarly measured. Suppliers may be asked to justify each confidentiality claim and provide non-confidential versions, unjustified claims may be treated as not confidential, and a supplier gets one week's notice and a chance to object before disclosure. For an industry whose main assets are model weights, training data and evaluation methods, that matters. It also protects third parties such as advertisers and publishers who hand over evidence.
Where the clock problem lies
The weak point is time. Article 75c(8) sets a five-year limitation period on the AI Office's powers. For one-off infringements it runs from the day the infringement is committed, and for continuing ones from when it ceases. The statute leaves the detail of interruption to an implementing act under Article 75d. This draft is that act.
Per PPC Land, six steps each restart the period: information requests, requests to evaluate a system, inspections, invitations to a structured dialogue, opening proceedings, and re-opening proceedings. An absolute ceiling of 10 years applies, but suspension periods, such as while a Court of Justice case is pending, extend it.
The Commission's likely answer is that long-running AI systems change constantly and that firms could otherwise run out the clock. That is fair as far as it goes. But three of the six triggers are informal or cooperative: an information request, a system-evaluation request, and an invitation to structured dialogue. Under a literal reading, a routine request could restart the period. Combined with no stated minimum reply period for preliminary findings, which PPC Land notes the draft omits compared with the parallel rules for model providers, the result is a regulator with long, flexible reach and a regulated firm with little certainty.
Why this matters for innovation
Legal certainty is a cost input for model developers and platform builders. A firm that ships a feature in 2026 could face a fine decision in 2036 under rules it cannot now fully anticipate, for a system that will have been rebuilt several times. The practical effect is to push compliance teams toward over-documentation and companies toward holding back features in the EU. That cost falls hardest on firms without large legal departments. Although the draft names no company and applies by reference to Article 75(1), the exposure is real for any provider in scope.
The fine ceilings add to the stakes. PPC Land reports that Article 99 caps most infringements at €15 million or 3% of worldwide annual turnover, and prohibited practices at €35 million or 7%. Article 75c also allows periodic penalty payments of up to 5% of average daily income or turnover per day. When penalties are that large, the rules governing when the clock stops matter as much as the penalties.
What the consultation should fix
Three amendments would keep the draft's strengths and tighten its weaknesses:
- Limit interruption to formal, binding steps. Opening proceedings, formal decisions and inspections should restart the period; voluntary dialogue and simple information requests should not.
- Add a minimum reply period for preliminary findings, matching the model-provider rules, so that adviser-only file access is usable in practice.
- Make the 10-year ceiling harder to extend, so that suspension is confined to pending court proceedings and judicial authorisation of inspections.
The PPC Land analysis also flags numbering and cross-reference errors in Articles 4, 5 and 6 of the draft. Those are drafting issues, but in a text whose central job is to define time limits, they should be corrected before adoption.
The overall picture is of a draft that takes defence rights seriously and then undercuts the benefit with an elastic clock. Industry and civil society have until 9 November to say so, and the Commission should treat that as a chance to make the AI Office's authority predictable as well as strong.