The European Union has moved from preparing its artificial intelligence rulebook to using its formal oversight powers. On 1 September 2026, the European Commission’s AI Office sent its first requests for information to more than 30 AI providers under the AI Act. The action followed the start of key enforcement powers on 2 August 2026, after a one-year compliance period for providers of the most advanced general-purpose AI models.
The development is significant, but it should be described precisely. A request for information is not, by itself, a finding that a company broke the law, nor does it mean that every recipient will face a fine or model restriction. It is an evidence-gathering measure within a broader enforcement framework that allows the Commission to request documentation, evaluate models, seek corrective measures and impose penalties when the legal conditions are met.
What happened when the EU sent enforcement requests to AI firms
The 1 September requests represent the AI Office’s first formal enforcement push against AI providers under the AI Act. More than 30 providers received requests for information, according to reporting on the Commission’s move. The requests focused on general-purpose AI model providers and were directed especially at firms that had not participated in informal compliance dialogues with the AI Office.
That context matters. The consolidated AI Act text says the AI Office may begin with a structured dialogue before issuing a request for information. The reported focus on companies outside earlier informal discussions therefore indicates that voluntary or preparatory engagement and formal evidence gathering can form different stages of the EU’s oversight process.
The central point is procedural: the EU has begun asking AI providers for information through powers granted by the AI Act. The requests open a formal oversight process; they do not establish liability on their own.
Commentary and secondary reporting have described the requests as the opening move in EU-level AI enforcement. Reports linked their expected scope to model security, external evaluations and post-market monitoring. These are areas that help an authority assess how a general-purpose model is tested, supervised and managed after it becomes available.
The number of recipients also shows that this was not framed as an isolated inquiry into one provider. At the same time, the available facts do not identify every recipient, the individual questions each company received or whether all requests were identical. It would therefore be misleading to assume that every provider faces the same issue or level of regulatory concern.
A timeline of the enforcement launch
- Late July 2026: European Commission Executive Vice-President Henna Virkkunen confirmed that the AI Office and national authorities would begin enforcing the AI Act and new transparency rules from 2 August.
- 31 July 2026: The Associated Press reported that the EU was creating a new Brussels enforcement team focused on risks including AI deepfakes, illicit imagery and hacking.
- 2 August 2026: Key transparency requirements and Commission enforcement powers entered into application. EU materials state that powers concerning the most advanced general-purpose AI models began applying after a one-year compliance period.
- 1 September 2026: The AI Office sent its first requests for information to more than 30 AI providers.
This sequence shows a relatively direct transition from the legal start of enforcement to formal information gathering. It also gives providers a clear reference point: preparation was no longer only about interpreting future obligations after 2 August 2026. The Commission could begin asking for evidence and examining whether covered providers had met requirements already in application.
Why the requests for information matter
A formal request for information gives the Commission a way to move beyond public claims and high-level policy statements. It can seek the records needed to understand how a provider approaches applicable AI Act duties. That distinction is important in a technically complex market, where a model’s public description may not reveal its evaluation process, internal controls or post-market monitoring arrangements.
The AI Act gives the Commission an enforcement toolkit rather than a single penalty mechanism. Depending on the legal and factual circumstances, the Commission can request information, conduct model evaluations, request corrective measures and impose fines. The framework also allows the AI Office to ask providers to take measures that may include restricting a model’s public availability when necessary.
- Information requests can require providers to produce relevant documentation and explanations.
- Model evaluations give the Commission a route to examine a model rather than relying solely on a provider’s account.
- Corrective measures can address identified shortcomings without treating every matter as an immediate maximum-penalty case.
- Financial penalties may apply to relevant infringements when the statutory conditions are satisfied.
- Availability restrictions may be requested if limiting public access to a model becomes necessary.
These powers make the accuracy and completeness of a provider’s response consequential. EU guidance states that the Commission can impose fines of up to 3% of global annual turnover or €15 million, whichever is higher, for relevant AI Act infringements. The availability of that ceiling does not mean it will automatically be applied to a recipient or that every deficient answer will lead to the maximum amount.
Proportionality and the specific infringement remain central to any responsible interpretation of enforcement risk. A request is the beginning of an information process, while a fine would require a separate legal basis and assessment. Companies, investors and customers should resist collapsing those stages into a single line.
Why reliable documentation is now operationally important
AI governance often produces many different records: technical reports, evaluation results, risk assessments, monitoring procedures and internal approvals. A formal request tests whether those materials are coherent, current and capable of being supplied through an accountable corporate representative.
Documentation also needs to match actual practice. A polished policy that does not correspond with an operational control may create further questions rather than resolve them. Conversely, a provider may have substantial technical safeguards but struggle to demonstrate them if responsibility for records is fragmented across engineering, legal, security and product teams.
The EU’s first requests therefore matter beyond the companies that received them. They show other providers what enforcement readiness can require in practice: not simply awareness of the AI Act, but an ability to retrieve, verify and explain evidence when the regulator asks for it.
Who is responsible for answering under Article 91
Article 91 of the AI Act addresses the supply of information to the Commission. Where the provider is a company or firm, its authorized representative must supply the requested information on the provider’s behalf. This creates a defined channel of accountability rather than allowing a corporate recipient to treat the request as an informal questionnaire with no clear owner.
The authorized representative’s role should not be confused with personally generating every technical answer. General-purpose AI documentation can depend on input from specialists in model development, evaluation, cybersecurity, legal compliance and post-market monitoring. The representative nevertheless serves as the formal point through which the provider supplies the requested material.
A practical response sequence
The precise response will depend on the language and scope of each request. Without speculating about questions that have not been published, a provider can use a disciplined process to reduce the risk of incomplete or inconsistent submissions:
- Confirm receipt and ownership. Identify the authorized representative responsible for supplying information and establish a controlled internal response process.
- Read the request narrowly and carefully. Separate each question, requested record, relevant model and applicable reporting period rather than responding with a generic compliance package.
- Map evidence to responsible teams. Determine which legal, technical, security, evaluation and monitoring functions hold the relevant source material.
- Verify factual consistency. Check that narrative explanations agree with technical records, model documentation and internal decisions.
- Identify genuine limitations. If a requested item does not exist or cannot be supplied in the expected form, the provider should not disguise that gap with vague language.
- Approve and preserve the response. Keep a clear record of what was submitted, by whom, on what basis and with which supporting materials.
This sequence is not a substitute for legal advice, and the AI Act itself remains the controlling authority. It is an operational way to treat a regulatory request as a verified evidence exercise rather than a public-relations task.
Providers should also distinguish between brevity and incompleteness. A concise answer can be effective if it directly addresses the request and points to reliable supporting material. A long response can still be inadequate if it avoids the question, mixes different model versions or relies on statements that cannot be substantiated.
The stakes are heightened by the possibility of penalties for relevant infringements, including problems involving bad or misleading replies. Because EU guidance identifies a potential maximum of 3% of global annual turnover or €15 million, whichever is higher, senior oversight of accuracy is warranted. That figure should be presented as a statutory ceiling, not a prediction about the outcome of the September requests.
How structured dialogue fits with formal enforcement
The AI Act’s framework does not require every supervisory interaction to begin with an adversarial step. The consolidated text says the AI Office may first initiate a structured dialogue before sending a request for information. Reporting on the September action likewise says the recipients were especially providers that had not joined informal compliance dialogues with the Office.
Structured dialogue and a formal information request serve related but different purposes. Dialogue can help clarify expectations, explain how a provider interprets its duties and identify where the authority needs more detail. A formal request, by contrast, uses an explicit enforcement power to require information through the legal framework.
- Early engagement can expose uncertainty before it becomes a formal dispute.
- Dialogue may help the AI Office understand a provider’s model, governance structure and evidence.
- A request for information can create a more defined record when informal engagement has not occurred or has not resolved the authority’s questions.
- Corrective measures and penalties remain separate possibilities that depend on what the evidence shows.
It would be an overstatement to conclude that participation in dialogue guarantees protection from enforcement. The available facts establish only that the first requests especially targeted firms outside informal compliance discussions. A provider that has engaged with the AI Office still needs to comply with applicable law and may still be asked to supply information.
It would also be inaccurate to portray a request as proof that a recipient refused to cooperate. A company may not have participated in informal talks for many possible reasons, none of which are established by the reported facts. The defensible conclusion is narrower: prior engagement appears to have influenced how the AI Office selected recipients for its initial information-gathering push.
What constructive engagement looks like
Constructive engagement requires more than agreeing with the regulator. A provider should be able to explain its interpretation, identify supporting evidence and state where technical uncertainty remains. If the Commission raises a concern, the provider can respond more effectively when it understands which team owns the relevant control and how a possible correction would be implemented.
The process should also avoid unsupported assurances. Statements such as a model being fully safe, completely monitored or comprehensively tested are difficult to evaluate without a defined scope and evidence. Grounded communication describes what was tested, which process was followed and how the provider monitors the model after deployment, while avoiding claims that go beyond the records.
Model security, external evaluations and post-market monitoring
Secondary reporting and commentary described the first requests as focusing on model security, external evaluations and post-market monitoring. These themes align with the practical challenge facing regulators: general-purpose models can be used across many downstream contexts, so oversight cannot be limited to a single product interface or a one-time pre-release review.
Model security
Model security concerns how a provider identifies, assesses and responds to security-related risks around a model. AP’s 31 July report placed hacking among the risks receiving attention as the EU established its Brussels enforcement team. The report also highlighted deepfakes and illicit imagery, illustrating the range of misuse concerns associated with powerful AI systems.
The available facts do not establish that any named recipient enabled hacking, generated illicit material or violated a rule involving deepfakes. Those issues describe the enforcement environment and the risks highlighted at launch. They should not be turned into allegations against the more than 30 providers that received information requests.
External evaluations
External evaluation can provide a perspective distinct from a provider’s internal testing. For enforcement purposes, the important questions are likely to concern the existence, scope and results of relevant evaluation work, although the individual September requests have not been made public in the supplied facts.
A provider responding about an evaluation should be prepared to distinguish what was actually examined from what falls outside the test. It should also be able to connect any findings with decisions or corrective actions. Merely stating that an external review occurred may not explain its relevance to the model or version covered by a request.
Post-market monitoring
Post-market monitoring shifts attention from what a model was expected to do before release to what the provider learns after it is available. This is particularly important for general-purpose technology, where usage patterns and emerging risks may develop over time.
Useful monitoring evidence may sit in different parts of an organization. The key governance challenge is ensuring that relevant observations can reach decision-makers and, when required, be explained to the regulator. The first enforcement requests signal that monitoring should be treated as an ongoing compliance function, not as a document created once and archived.
Security, evaluation and monitoring form a connected cycle: assess the model, observe how risks develop, document the evidence and make corrections where the evidence supports them.
This cycle also helps explain why the Commission’s toolkit includes both evaluations and corrective measures. Information can lead to closer examination, and examination can reveal a need for changes. Restrictions on public availability are available if necessary, but the framework does not imply that restriction is the default response to every concern.
Transparency rules are now an enforcement issue
EU policy materials state that the AI Act’s transparency rules and key enforcement powers became enforceable in August 2026. Henna Virkkunen’s late-July announcement similarly confirmed that the AI Office and national authorities would begin enforcing the Act and new transparency requirements from 2 August.
Transparency in this setting is not simply a matter of publishing broad statements about responsible AI. Regulatory transparency depends on whether covered organizations can provide the information required by law and support it with reliable documentation. The September requests turn that principle into an immediate administrative task for the recipients.
The launch also involves more than one type of authority. The Commission’s AI Office has the central role described for general-purpose AI model enforcement, while national authorities are part of the wider AI Act enforcement landscape. Organizations should therefore identify which authority is acting, under which provision and in relation to which system or model before deciding how to respond.
Four distinctions that prevent misleading claims
- A request is not a violation finding. It asks for evidence; it does not prove wrongdoing.
- A maximum fine is not an expected fine. The ceiling of 3% of global annual turnover or €15 million, whichever is higher, applies to relevant infringements and is not an automatic charge.
- An available restriction is not an announced ban. The Office may request measures that include restricting public availability if necessary, but the first requests do not themselves establish that outcome.
- An enforcement focus is not an accusation. Deepfakes, illicit imagery and hacking were highlighted as areas of concern at the launch; that does not show that every recipient was linked to those harms.
These distinctions are essential for trustworthy coverage. AI regulation attracts strong political and commercial reactions, and imprecise language can distort both legal risk and public understanding. Describing the procedural stage accurately gives businesses enough reason to prepare without suggesting an enforcement outcome that has not occurred.
The same discipline applies to providers’ own communications. If a company discloses that it received a request, it should avoid implying that routine cooperation equals regulatory approval. Likewise, it should not characterize the inquiry as a penalty unless the Commission has actually taken that separate step.
What AI providers should do now
The first formal requests offer a practical signal to providers that were not among the initial recipients. Waiting for a letter before organizing evidence can increase the risk of rushed, inconsistent or incomplete responses. A more durable approach is to build a traceable compliance system around the models and obligations that fall within the AI Act.
Establish clear governance
Responsibility should be visible across legal, technical and executive functions. The authorized representative needs a reliable route to obtain information from the teams that develop, test, secure and monitor the model. Escalation paths should also be clear if a response uncovers a gap or conflicting records.
Maintain a model-level evidence map
Providers should know which documentation relates to which model and version. An evidence map can connect evaluation records, monitoring processes and corrective decisions without assuming that one generic policy proves compliance for every model.
The map should be designed for verification rather than volume. Its purpose is to help the organization locate source material and explain why it is relevant. This can reduce the chance that a formal response mixes outdated documentation with current practices.
Test response readiness
A provider can conduct an internal exercise based on the powers the Commission now holds. The exercise need not guess the exact contents of a future request. It can instead test whether the organization can identify an accountable representative, retrieve supporting records and reconcile technical and legal explanations.
- Select a covered model and define the version under review.
- Locate the relevant security, evaluation and monitoring records.
- Ask responsible teams to explain the evidence in consistent terms.
- Check whether claimed controls are operating as documented.
- Record unresolved gaps and assign corrective ownership.
- Confirm that the authorized representative can access an approved response package.
Prepare for more than document production
The Commission can conduct evaluations and request corrective measures as well as seek information. Readiness should therefore extend beyond producing files. Providers need a process for assessing a regulatory concern, deciding what change is required and documenting how that change is implemented.
They should also understand the possibility of a request to restrict public availability if necessary. This power makes deployment and distribution controls relevant to enforcement planning. The supplied facts do not show that such a restriction was imposed through the first requests, but the authority exists within the framework.
Keep statements accurate and supportable
Accuracy is the most immediate lesson from the potential penalties associated with relevant infringements and misleading replies. Every statement submitted to the Commission should be reviewed against source evidence. If an answer depends on assumptions, estimates or a limited test scope, those limitations should be made clear rather than hidden.
Providers should avoid treating this as a one-time legal drafting exercise. Engineering and operational teams often understand the evidence, while legal and compliance teams understand the request and its consequences. A credible submission depends on coordination between both sides.
What the first enforcement push means for the AI market
The September requests mark a transition from rulemaking and compliance preparation to active supervision. For general-purpose AI providers operating in or serving the EU market, regulatory capability now includes responding to evidence requests and potential model evaluations, not merely following policy debates.
The move may also change how customers and business partners assess providers. A trustworthy compliance posture is increasingly tied to the ability to explain governance, testing and monitoring with evidence. However, receipt of a request should not be used as a simple measure of product quality or legal compliance because the initial selection reportedly emphasized providers that had not participated in informal dialogues.
Investors and procurement teams should apply the same caution. The existence of an information request is material context, but its significance depends on the request, the provider’s response and any later Commission action. Without those details, claims that a recipient has either been cleared or condemned would be premature.
Signals to watch as enforcement develops
- Whether the Commission publicly explains the categories of information it sought from providers.
- Whether information requests lead to model evaluations or structured corrective measures.
- How the AI Office uses dialogue alongside formal enforcement powers.
- Whether any case results in a fine or a request to restrict public model availability.
- How EU-level action for general-purpose AI models interacts with enforcement by national authorities.
These are developments to monitor, not outcomes established by the first round of requests. The most reliable analysis will continue to separate official Commission actions, EU legal text, reported facts and outside commentary. That source discipline is particularly important when enforcement is new and procedural details are still emerging.
The first push also demonstrates why compliance should be treated as an operating capability. The AI Office can request information, evaluate models, demand corrective measures and impose fines where legally justified. Providers that can connect governance claims to current evidence will be better positioned to respond accurately than those relying on general assurances.
The EU sends enforcement requests to AI firms at a decisive point in the AI Act’s implementation. The 1 September 2026 requests to more than 30 providers followed the 2 August start of key enforcement and transparency powers, with reporting indicating a particular focus on general-purpose AI providers outside earlier informal compliance dialogues. The action is best understood as the opening of formal evidence gathering, not as a collective finding of wrongdoing.
For providers, the practical message is clear: know who is authorized to answer, maintain reliable model documentation, verify every statement and be prepared for evaluation or corrective action if the evidence warrants it. For readers, customers and market observers, the equally important lesson is to distinguish requests from rulings and statutory maximums from actual penalties. That grounded approach reflects both the seriousness of the EU’s new powers and the procedural fairness required when describing their use.
