The AI Office Fining Procedure Draft: Who Actually Supervises Your AI Product, And What To Keep On File

On 12 October 2026 the European Commission published a draft implementing regulation on Articles 75c and 75d of the AI Act. It sets out how the AI Office will conduct proceedings that can end in a fine against the AI systems it supervises directly: who can see the file and when, how trade secrets are handled, how long the Office has to act and how documents are exchanged. The draft is open for comment on the Have your say portal, initiative 19292, until 11.59 p.m. Brussels time on 9 November 2026. After that it goes to a committee of Member State representatives, and the Commission expects to adopt it by the end of the year. The text is a draft and its cover says it reflects only the preliminary views of Commission services, so treat every detail below as subject to change.

The legal basis is the Digital Omnibus, Regulation (EU) 2026/1744, which postponed the high risk obligations and, in the same package, added Articles 75a to 75d to the AI Act. Those articles give the AI Office the powers of a market surveillance authority, including document requests, inspections and non compliance decisions, for a narrow group of systems. The new draft should not be confused with Implementing Regulation (EU) 2026/1755, in force since 10 August, which already covers fining procedures for providers of general purpose AI models under Article 101. The new draft deals with AI systems, not models, and it fills the gap that opened when the Omnibus moved some applications under Brussels rather than national authorities.

The first question for any founder or compliance lead is whether the AI Office is your regulator at all. Under Article 75(1), the Office supervises AI systems where the same provider supplies both the general purpose model and the system built on it, and AI systems embedded in very large online platforms and search engines designated under the Digital Services Act. In practice that points at products such as ChatGPT, Gemini, Claude, Le Chat, Meta AI inside Instagram and Facebook, and AI features inside large search engines, although the Commission has not published a named list. A European startup that builds a product on the OpenAI or Anthropic API is a different case. The model and the application come from different companies, so supervision stays with the national market surveillance authority in the relevant Member State, under that country’s own procedural rules.

That distinction matters for how you write your AI Act position. Many vendor questionnaires and board papers now say something like "we are regulated by the AI Office", which is usually wrong for companies that integrate a third party model. Your AI system inventory should name the authority that would actually supervise each system, based on who supplies the model and who supplies the application, and it should record the Member States where the system is placed on the market. If you are a deployer only, using ChatGPT or Claude inside the business, you are the deployer of someone else’s system and the national authority is again the relevant one. Getting this right is a small edit, but it changes which procedural rules, deadlines and complaint routes apply to you.

For the providers that are in scope, three parts of the draft stand out. A company under investigation can see the AI Office file only after it receives preliminary findings, not before. Information a company wants kept confidential must be identified and justified by the deadline set, otherwise the Office may treat it as non confidential, and any disputed disclosure comes with a reasoned decision and at least a week of notice. And the limitation period is five years from the infringement or from when a continuing infringement ends, reset by every investigative act, including information requests, requests for access to evaluate a system, inspections, invitations to a structured dialogue and the opening of formal proceedings, with an absolute cap of ten years plus suspensions. In practice an active investigation will almost always run to the cap, so records need to survive at least that long.

The draft does not change the fine levels, which already sit in Article 75c and point to Article 99. As reported, the ceilings are 35 million euros or 7 percent of worldwide turnover for prohibited practices under Article 5, 15 million euros or 3 percent for other infringements and for breaching decisions or commitments, 7.5 million euros or 1 percent for supplying incorrect or misleading information, and periodic penalty payments of up to 5 percent of average daily turnover to compel cooperation. Larger companies face the higher of the fixed amount and the percentage, while SMEs and the new small mid cap category face the lower. One detail is worth reading twice: when the AI Office imposes the fine, the 3 percent tier covers a breach of any AI Act provision, not only those listed in Article 99. Fines go into the EU budget and decisions are published with the company name and amount.

Even if the AI Office will never write to you, the draft is a useful checklist, because national authorities have similar powers to request information and documents under their own rules. Information requests are the first step, and the 1 percent tier for misleading answers means the quality of your first response matters as much as the underlying compliance. Keep a single register of AI systems with the model provider, version, intended purpose and the Member States served. Keep dated evidence of your Article 50 transparency measures and your Article 5 screening, since those obligations have applied since 2 August 2026. Mark trade secrets in technical documentation as you write it, with a one line reason, so you are not justifying confidentiality under a deadline. And set your retention for AI governance records to at least ten years, which is longer than many teams default to for SOC 2 or ISO 27001 evidence.

Much of this already maps to work you may be doing for ISO 42001 or EN 18286, the AI quality management system standard cited in the Official Journal on 19 October. Clause 7.5 of ISO 42001 on documented information and the record keeping provisions in EN 18286 both expect controlled, retrievable records, and the AI system inventory sits at the centre of both. If you manage evidence in Vanta, Drata, Secureframe, Sprinto, Thoropass or Hyperproof, add a field for the supervising authority to each AI system record, check that retention on AI evidence is set long enough, and tag documents that contain trade secrets. Our earlier posts on the AI Office information requests to OpenAI and Anthropic and on EN 18286 cover the questions to put to your model providers.

A short list for this week. Confirm for each AI system whether the AI Office or a national authority would supervise it, and correct any policy, questionnaire answer or board paper that says otherwise. Ask your model providers how they would notify you if an AI Office proceeding affected the model or service you depend on, and record the answer in your vendor review. Review retention and confidentiality marking for AI documentation. If you are a provider in scope, or an industry body representing one, read the draft and decide whether to comment before 9 November, because the rules on file access and response times are where the text is still thin. For everyone else, the draft is not a new obligation, but it is the clearest picture yet of how an AI Act investigation will actually run.