The European Commission published final guidance on July 20 for AI Act transparency obligations that start applying on August 2, 2026.
The immediate temptation is to treat this as a labelling task. The operational problem is larger. A business first needs to know where AI appears, whether it is acting as a provider or deployer, what the system produces, whether meaningful human review occurs, and how the disclosure survives the complete workflow.
The obligation follows the workflow
The Commission says providers must design certain interactive AI systems to inform people that they are interacting with AI. Providers must also add machine-readable markings for certain generated or manipulated content.
Deployers have their own obligations in specified situations. The Commission highlights deepfakes, AI-generated text on matters of public interest without human review or editorial control, and the use of emotion-recognition or biometric-categorisation systems.
That does not mean every AI-assisted output receives the same label. The applicable requirement depends on the role, system, content, context, and level of human control. That is exactly why a generic policy document is not enough.
A disclaimer cannot repair missing traceability
Disclosure becomes fragile when AI use is scattered across tools and teams.
A chatbot may be configured by one team, connected to data by another, embedded by a third, and reviewed only after a customer reports a problem. A publishing workflow may use AI during research, drafting, image production, editing, or distribution without recording which stage changed the final output.
Adding a sentence at the end cannot reconstruct that chain. The control has to sit where the work happens:
- inventory public-facing AI interactions and generated-content workflows;
- record the provider, deployer, system owner, purpose, audience, and output type;
- identify where machine-readable marking or visible disclosure is required;
- document when human review or editorial control occurs and what that review covers;
- keep the disclosure attached through exports, edits, publishing, and reuse;
- define who checks the workflow when the model, vendor, interface, or regulation changes.
Build the evidence before the deadline
The Commission's final Code of Practice on marking and labelling AI-generated content is voluntary. It gives providers and deployers a practical route for preparing for the transparency obligations, but using a tool or signing a code does not by itself prove that every workflow complies.
The useful first step is an inventory, not a banner. Once the business can see its affected systems and content paths, legal counsel can determine which obligations apply and operators can put the required disclosure, marking, review, documentation, and ownership in the right place.
This is operational guidance, not legal advice. Specific Article 50 decisions should be reviewed with qualified counsel.
Related services: AI Strategy, Implementation, Leadership