EU AI Office Enforcement Puts AI Data Vendors Under Scrutiny
The EU AI Office is sending formal information requests to general-purpose AI model providers under the AI Act. Here is what the shift from monitoring to inquiry means for Europe's AI data supply chain.

The AI Office moves from monitoring to inquiry
The European Commission's AI Office gained formal enforcement powers over general-purpose AI model providers on August 2, 2026, and by autumn it had moved past passive monitoring into active information requests. Providers are being asked to supply documentation on cybersecurity practices, safety testing, and copyright compliance, the three areas the AI Office has signaled it will scrutinize first under the general-purpose AI rules. The Commission's preferred first step is what it calls technical compliance dialogues, structured conversations with providers and deployers that happen before any formal proceeding is opened.
For a Dublin-based AI Office with a relatively small enforcement staff relative to the scale of the market it oversees, information requests are the cheapest way to establish a paper trail. Every provider that receives one has to produce real documentation, not marketing claims, about how its model was trained, tested, and secured. That documentation obligation is where the pressure flows downstream to anyone who supplied training data, evaluation data, or red-teaming services to that provider.
Meta's decision not to sign the EU's voluntary GPAI Code of Practice has been watched closely as a test case. Declining to sign does not exempt a provider from the underlying legal obligations or from enforcement action, and non-signatories are expected to face more direct scrutiny of their compliance evidence rather than less.
What the information requests actually ask for
Under Article 53 of the AI Act, general-purpose AI model providers must maintain and, on request, disclose a sufficiently detailed summary of the content used to train the model. That summary has to go beyond a vague description of data sources; the Commission's template calls for information on the types of data, where it originated, and whether copyrighted or personal data was involved. Providers that cannot produce this with confidence are the ones most exposed when a request lands.
The requests extend to risk management documentation for systemic-risk models, meaning any provider whose model crosses the compute threshold the AI Act uses to define systemic risk has to show evidence of adversarial testing, incident reporting procedures, and model evaluation results, not just a policy document describing intentions. Fines for non-compliance with the general-purpose AI obligations can reach 3% of global annual turnover under the AI Office's dedicated GPAI powers, with broader AI Act violations carrying penalties of up to 35 million euros or 7% of global turnover, whichever is higher.
Why training data provenance is the weak point
Most AI model providers can describe their training pipeline in general terms. Far fewer can trace a specific batch of annotated data back to the vendor that produced it, the quality checks it passed, and the consent or licensing basis under which it was collected. That traceability gap is exactly what an information request is designed to expose, and it is also where outsourced annotation programs built on anonymous marketplace labor tend to fail first.
A vendor that cannot name who labeled a dataset, under what contract, and with what quality sign-off leaves its enterprise customer unable to answer the AI Office's questions when they arrive. Corpshore AI structures its annotation programs around named, directly managed annotator teams with documented training records and sign-off chains precisely because that documentation is now a regulatory requirement, not a nice-to-have.
What this means for European AI buyers right now
Enterprise AI teams in Germany, France, the Netherlands, and the wider EEA should treat the current enforcement posture as the floor, not the ceiling. The AI Office has said it expects enforcement intensity to build gradually, which means the providers and vendors that get their documentation in order now, while scrutiny is still at the information-request stage, avoid the far more disruptive experience of assembling it retroactively under a formal investigation timeline.
This also changes vendor selection criteria. A procurement team evaluating annotation or RLHF vendors should now ask directly whether the vendor can produce dataset provenance records, worker contract status, and quality audit trails on short notice, because that is precisely the kind of evidence an AI Office inquiry will eventually ask the buyer's own model provider to produce.
Building a defensible data supply chain
The practical response is not a one-time compliance exercise but an operating model change: every annotation project should generate its own audit trail as a byproduct of normal work, not as a separate reporting layer bolted on afterward. That means documented annotator qualifications, versioned guidelines, inter-annotator agreement scores, and a clear record of which humans touched which records.
Corpshore AI, which manages annotation programs across 35-plus languages and 15,000-plus operational seats, builds this documentation into its delivery process by default. Contact Corpshore AI at corpshore.ai/start-a-pilot to launch a European data program with AI Act-ready documentation built in from day one, deployable in under 14 days.
Frequently asked questions
- What is the EU AI Office and when did its enforcement powers start?
- The AI Office is the European Commission body responsible for overseeing general-purpose AI models under the EU AI Act. Its enforcement powers over GPAI providers took effect on August 2, 2026, allowing it to request information, open investigations, and issue fines.
- What happens if a company does not respond to an AI Office information request?
- Failing to respond or providing incomplete information can escalate a technical compliance dialogue into a formal investigation. Fines for non-compliance with general-purpose AI obligations can reach 3% of global annual turnover, and broader AI Act penalties can reach 35 million euros or 7% of turnover.
- Does declining to sign the GPAI Code of Practice exempt a company from enforcement?
- No. The Code of Practice is voluntary, but the underlying legal obligations in the AI Act are not. Meta's decision not to sign has been treated as a test of whether non-signatories face the same or greater scrutiny, and early signals suggest they do.
- What training data documentation does the AI Act actually require?
- Providers of general-purpose AI models must maintain a sufficiently detailed summary of training content, covering data types, origin, and whether copyrighted or personal data was used. This has to be detailed enough to survive a direct Commission request, not a general description.
- How can a company prepare its AI data supply chain for EU scrutiny?
- Start by confirming every annotation and training data vendor can produce provenance records: who labeled the data, under what contract, and with what quality sign-off. Corpshore AI builds this documentation into every program by default across its 35-plus language annotation operations.
- How do I start an AI Act-ready data program with Corpshore AI?
- Contact Corpshore AI at corpshore.ai/start-a-pilot to scope an annotation or evaluation program with documentation built for EU AI Act requirements. Programs can typically be scoped and launched in under 14 days.
Published by
Corpshore AI
https://corpshore.ai