On 20 July 2026 the European Commission approved the content of its Guidelines on the transparency obligations of providers and deployers of AI systems under Article 50 of the AI Act, as C(2026) 5054 final with an annex. Much of the commentary since has described the Guidelines as adopted and in application. Their actual status is narrower, and for anyone building against them the difference matters, so this piece starts there before getting to what the Guidelines say about AI agents, which is the part worth building against regardless.

Status first: approved in content, not yet adopted, and non-binding either way

What happened on 20 July 2026 is that the Commission approved the Guidelines' content. Formal adoption is a separate step, and the Commission's covering Communication, C(2026) 5054 final, states that the Guidelines apply only from that moment. As at the date of this piece, that step is pending. So the document in circulation is an approved text awaiting adoption, not an instrument in application.

Even once adopted, the Guidelines bind no one. Paragraph 5 says so in terms: they are non-binding, and authoritative interpretation of the AI Act remains with the Court of Justice. What the Guidelines offer is the Commission's considered reading of Article 50, which is the reading market surveillance authorities are most likely to start from. That makes them worth building against, and it makes conflating them with the Regulation a category error. The obligations are in Article 50. The Guidelines are a lens on them.

The binding baseline: Article 50(1)

The binding transparency duty for systems that talk to people has applied since 2 August 2026. The first sentence of Article 50(1) of the AI Act:

Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use.

Note what that text requires and what it does not. The person must be informed that they are interacting with an AI system. The Article says nothing about disclosing whose AI system, or on whose behalf it is acting. For a chatbot on a company's own website, that gap rarely matters, because the context answers the question. For an AI agent that leaves its deployer's premises, so to speak, and transacts with third parties, the gap is exactly where the interesting questions live.

Paragraph 31: the dual disclosure

That is the gap the Guidelines address for agents. Paragraph 31, in the section on the scope of Article 50(1), says:

AI agents must be designed and developed in such a way that they disclose both their artificial nature and the person on whose behalf they are acting, considering the need for transparency of the origin and the delegation of authority and accountability for the consequences of their actions.

Two disclosures, not one. The artificial nature, which is Article 50(1)'s own requirement, and the person on whose behalf the agent is acting, which is the Commission's elaboration for agents, grounded in its reading of the provision's purpose. The paragraph's stated rationale carries as much weight as its rule: transparency of origin, delegation of authority, and accountability for consequences. The Commission's position, in plain terms, is that when software acts for someone, the counterparty is entitled to know there is a someone, and who it is.

Being precise about the legal weight of each half: an agent that conceals its artificial nature from a person it interacts with has an Article 50(1) problem, subject to the obviousness carve-out. An agent that discloses its nature but not its principal has met the Article's text while falling short of the Commission's approved, not yet adopted, non-binding reading of it. Those are different exposures. A deployer deciding what to build should know which one it is managing. Most will conclude that building to paragraph 31 is cheap insurance, since the authorities most likely to ask questions are the ones most likely to hold the Commission's reading.

What a run record has to show

A design that discloses is one thing. Showing that a given interaction carried the disclosure is another, and it is the question that arrives after the fact, when a counterparty, a client's counsel or an authority asks about a specific exchange on a specific day.

Read against paragraph 31, a record of an agent's run answers that question if it shows three things. That the disclosure of artificial nature was made in the interaction, not merely available somewhere in the product. On whose behalf the agent was acting in that interaction, which for agents serving multiple users or clients can change run by run and is not derivable later from configuration. And when in the interaction both happened, because a disclosure after the transaction concludes is a different fact from one at first contact.

None of that can be reconstructed afterwards from a system prompt or a settings page. The system prompt shows what the agent was told to do. Whether a particular run did it is a fact about the run, and it is only answerable from a record made at the time by whatever instrumentation the deployment has in place.

Keep the registers apart

A closing distinction, because conflation here is becoming common in commentary. Article 50(1) is binding law in application since 2 August 2026. The Article 50 Guidelines are an approved text, pending formal adoption, non-binding on their own terms. Paragraph 31's dual-disclosure expectation lives in the second register, not the first. An analysis, a product claim or a compliance posture that cites paragraph 31 as if it were the Regulation is overstating the law; one that ignores it is understating where interpretation is heading. The defensible position states both, which is what this piece has tried to do.

For the related question of what happens on 2 December 2026, including the end of the Article 111(4) transitional for Article 50(2) marking, see our analysis of that date.

Sources: consolidated EU AI Act, CELEX 02024R1689-20260727; Commission Guidelines on the transparency obligations under Article 50, C(2026) 5054 final of 20 July 2026 with annex, via digital-strategy.ec.europa.eu; status as described, content approved, formal adoption pending, non-binding per their paragraph 5.

Truveil records the disclosure text an agent's instrumentation logs, and the capacity in which the deployer declares the agent was acting, event by event, where that instrumentation captures them.

Frequently asked questions

Are the Commission's Article 50 Guidelines binding?

No. The Guidelines state at paragraph 5 that they are non-binding, with authoritative interpretation of the AI Act reserved to the Court of Justice. Their content was approved on 20 July 2026 as C(2026) 5054 final; formal adoption, from which they apply, is pending.

What does paragraph 31 of the Article 50 Guidelines say about AI agents?

That AI agents must be designed and developed to disclose both their artificial nature and the person on whose behalf they are acting, in view of transparency of origin, delegation of authority and accountability for the consequences of their actions.

Does Article 50(1) itself require an agent to disclose who it acts for?

No. Article 50(1) requires that natural persons be informed they are interacting with an AI system, unless that is obvious in context. The duty to disclose the principal is the Commission's elaboration for AI agents in the Guidelines, not the Article's text.

Since when does Article 50 apply?

The Article 50 transparency obligations have applied since 2 August 2026. The Digital Omnibus (Regulation (EU) 2026/1744) did not defer them.