Of all the obligations in the EU AI Act, Article 50 is the one that hits agentic systems fastest and most broadly. Not because it is particularly strict — but because it touches almost every agent that communicates with people or produces content. It has applied since 2 August 2026; for systems already on the market before then, the marking grace period ends on 2 December 2026.

What Article 50 requires

Two things, at heart. First: anyone interacting with an AI system must be told — unless it is obvious from the circumstances. Second: content generated by AI must be recognisable as such, in machine-readable form, not just as a footnote. Both sound harmless until you apply them to an agent that answers hundreds of e-mails a day, produces documents and writes proposals into systems.

Where agents run into it

  • The agent replies to the supplier. The “where is my money” e-mail is no longer answered by Ms Huber but by the accounts-payable agent. The recipient must be able to tell.
  • The agent produces the document. Quote draft, contract summary, audit report — everything it generates that leaves the building needs a marking.
  • The agent prepares the decision. Article 50 does not apply directly here, but its sibling obligation does: if the decision affects people (credit, hiring), the system slides into the high-risk class with rules of its own.

How to build it properly

The most common mistake is to treat Article 50 as a task for the business team: “please add a notice to every e-mail.” That does not scale and gets forgotten. Solved cleanly, it is a platform feature: the agent's communication channel attaches the disclosure automatically; the document system stamps agent-generated files with machine-readable metadata; the audit trail records when which marking was delivered. Then compliance is not a behaviour people have to perform anew every day, but a property of the system — testable and provable.

What to do before 2 December

  • Inventory: which systems generate content or talk to people? Agents, chatbots, generators — list them all.
  • Channels: for each system, decide where the disclosure lives — e-mail signature, chat header, document footer, metadata.
  • Marking: choose a technical method for machine-readable marking and build it in centrally — not per use case.
  • Evidence: document in the audit trail that and how marking happened. That is the file you present if it ever matters.

Dates reflect the AI Act as amended by the Digital Omnibus. This is an engineering perspective, not legal advice — when it gets specific, we bring the lawyers to the table.