News · 13 August 2026
The AI Act enforcement phase begins: your model supplier's paperwork is now your concern
The EU's Digital Omnibus on AI entered into force at the end of July 2026, and on 2 August 2026 a major set of AI Act provisions began to apply. The Commission's AI Office gained enforcement powers over providers of general-purpose AI (GPAI) models, including the ability to request documentation, run evaluations and impose fines. Transparency obligations also applied: people should be told when they interact with a chatbot, deepfakes must be labelled, and AI-generated content should carry machine-readable marks, with a later date, 2 December 2026, for systems already on the market. For these transparency breaches, fines can reach €15 million or 3% of worldwide annual turnover. High-risk obligations were pushed back, to December 2027 and August 2028 depending on the category.
The supply-chain angle
Most organisations do not train frontier models. They build products on top of someone else's model, whether through an API or open weights. The Act distinguishes the model provider from the downstream provider who integrates it into a system. Each has duties, and they depend on one another.
If you are a downstream provider, you need information from your upstream supplier to meet your own obligations: what the model can and cannot do, what data categories it was trained on, known limitations and the security measures around it. GPAI providers must document and share certain information, and publish a summary of training content. Your job is to ask for it and to keep it.
A due-diligence checklist
- Request the model documentation and any summary of training data. Record the version and date.
- Check the licence and usage terms for restrictions on your use case.
- Test the model yourself on your task. Supplier benchmarks may not resemble your data.
- Define your own transparency measures: disclosure to users, labelling of generated content and logging.
- Plan for upstream change. Model versions change. Pin versions where you can, and re-evaluate when they move.
- Have a fallback. If a supplier withdraws or changes a model, how do you continue?
Where explainability fits
Even when a model is a black box to you, the system around it need not be. Retrieval that cites sources, constrained outputs, confidence reporting and human review points can make a product inspectable even if the underlying model is not.
Enforcement turns documentation from an optional good habit into a business risk. A supplier file kept up to date is cheap insurance.
Written in August 2026; details reflect the situation at that time.