What the European Commission published
On 20 July, the European Commission published guidelines for providers and deployers of certain AI systems under Article 50 of the EU AI Act. The transparency obligations start applying on 2 August 2026. The guidelines are meant to clarify how organizations can meet those obligations in practice.
Article 50 covers several different situations. People may need to be told when they are interacting with an AI system. Providers of systems that generate synthetic audio, images, video or text must support detection of that output in a machine-readable format. Deployers face disclosure duties for deepfakes and, in defined circumstances, AI-generated text used to inform the public on matters of public interest. Separate notice requirements apply to emotion-recognition and biometric-categorisation systems.
The important distinction: not every AI use has the same duty
This is not a blanket rule that every AI-assisted sentence needs a label. The obligation depends on the system, the output, the purpose and whether the organization acts as provider or deployer. Article 50 also contains specific exceptions and adjusted disclosure rules, including where text has undergone human review or editorial control and a person or organization holds editorial responsibility.
That makes a generic disclaimer a weak response. A footer cannot determine which workflow produced an output, whether that output is customer-facing or public-interest information, who reviewed it, or which disclosure belongs in the channel where a person encounters it.
Transparency is a workflow requirement
For companies using AI in customer service, document production, knowledge work or communications, transparency has to be designed into the flow of work. The system needs to know what it is producing, for which audience, through which channel and under whose responsibility.
In practice, that can mean preserving provenance, adding machine-readable markers where required and technically applicable, presenting a clear notice in a chat or voice interface, and recording whether human review took place before content was released. Those controls belong beside identity, permissions, logging and approval gates — not in a policy document disconnected from execution.
The channel is part of the AI operating layer
A useful way to look at enterprise AI is as an operating layer across data, methods, tools, governance, channels and agents. The new guidelines make the channel layer concrete. A disclosure only works when it reaches the person at the right moment: in the interface, document, media item or publication process where the AI output appears.
This also means that buying an AI tool does not finish the job. The tool may generate the output, but the organization still needs to connect that output to its own review process, content systems, customer channels and evidence trail. Compliance and operational design become the same implementation problem.
A practical check before 2 August
Start with an inventory of live AI workflows, not an inventory of models. For each workflow, identify who encounters the output, whether the organization is provider or deployer in that context, whether the output is synthetic media or public-facing text, and who has editorial responsibility.
Then test the actual route to production. Can the system preserve provenance? Can it place the right notice in the right channel? Is human review visible and attributable? Can the organization reconstruct what happened later? If those answers depend on memory or manual convention, the control is not operational yet.
Laava's perspective
The Commission's publication reinforces a practical point: responsible AI is not a separate compliance track beside implementation. Governance has to travel with the work. Companies moving from isolated experiments to production AI need an operating layer that connects data, methods, tools, controls and channels around real processes.
The sensible first step is therefore not a broad compliance programme built on assumptions. Pick a real workflow, classify the obligations that genuinely apply, and implement the required disclosure, review and audit trail in the system itself. That produces both evidence and a repeatable pattern for the next workflow.