What exactly does my AI agent have to tell the user?

That it is an AI, before or at the point of the exchange, clearly and understandably. No specific wording is mandated. What matters is that the person cannot reasonably believe they are talking to a human: a visible notice as the conversation opens, an agent name that does not pass itself off as a colleague's first name, and an honest answer if the user asks directly. The most common failure is not a missing notice — it is a notice buried in terms and conditions nobody reads.

My agent is already in production. Do we have to rebuild it?

Almost never. In most deployments we see, compliance comes down to three things: adding explicit disclosure when the conversation opens, making sure the agent answers honestly when asked whether it is human, and documenting what it does, with which data, on what legal basis. That is hours of work, not months. The real project appears elsewhere: when nobody can state precisely which actions the agent is permitted to take, or prove afterwards what it did.

Does this also cover internal agents used only by employees?

The transparency obligation protects people interacting with the system, and your employees are people. An internal assistant handling HR or legal queries faces staff who are entitled to know who they are talking to — all the more so because the context is asymmetric: an employee who believes they are writing to a colleague in HR phrases things differently. Beyond the text of the law, this is a question of internal trust, and adoption of an internal agent depends directly on that trust.

Who is responsible — us or our provider?

The roles in the regulation — provider, deployer — carry different obligations, and this is something to settle in writing before going live, not after an incident. In practice, whoever puts the system in front of people carries the duty to inform: if you operate the agent for your customers or your staff, do not assume your provider carries that responsibility for you. What a serious provider should give you instead is the documentation, the logs and the technical mechanisms that let you demonstrate compliance.

EU AI Act: Your AI Agents Must Disclose (2026)

Substance

What applies, exactly

The obligation fits in one sentence: people interacting with an AI system must be told. No mandated wording, no prescribed format — but an obligation of result, which is that nobody can reasonably believe they are talking to a human. For a conversational agent, that is settled in the first three seconds of the exchange.

What the obligation is not: a prohibition. You can deploy an agent that answers customers, qualifies enquiries and triggers actions. What changes is that the person on the other side has a right to know what they are talking to — and you must be able to demonstrate it.

This deserves emphasis because it is consistently misread in meetings: transparency is not a commercial handicap. The deployments we see succeed announce the agent in the first line and convert better — because a user who knows they are talking to a machine writes shorter, more direct requests, which are easier to handle correctly.

Dates

The real timeline, without approximation

Date	What comes into effect	Who is concerned
2 February 2025	Prohibited AI practices and AI literacy obligations	All affected organisations
2 August 2025	Governance rules for general-purpose AI models (GPAI)	Providers of general-purpose models
2 August 2026	Transparency rules — telling a person they are interacting with an AI	Anyone deploying a system that talks to people
2 August 2026	General application date; the AI Office and national authorities become responsible for implementation, supervision and enforcement	European and national authorities

Dates from the European Commission's official pages, checked on 13 September 2026. Lumyniq, 2026.

Update, 13 September 2026: the Commission's official timeline sets Article 50 at 2 August 2026. For generative AI systems already on the market before that date, the Digital Omnibus allows until 2 December 2026 for marking generated content. Telling a person they are dealing with an AI is therefore a duty in force. The full timeline is in our EU AI Act guide for businesses.

Application

Concretely, on your agents

1. Disclosure on opening. A clear notice when the conversation starts, visible without any action from the user. Not a line in the terms, not a tooltip on hover. If your agent has a first name, add what it is: "Léa, AI assistant" rather than "Léa".

2. An honest answer to the direct question. Test it today: ask your agent whether it is human. A surprising number of production agents answer evasively, because the prompt tells them to be "warm and natural" and never addresses this case. It is the test anyone would run — a journalist, an unhappy customer, a regulator.

3. Explicit handover to a human. When the agent hands off to a person, the transition must be visible. And the reverse: a human picking up a thread the agent opened must not imply they were there all along.

4. The record. Being able to show, six months later, what the agent said and did. This is where compliance meets engineering: with no action logging, you can demonstrate nothing. We cover this in our guide to deploying an AI agent in production.

Avoid

The mistakes we see most

Confusing transparency with small print. A notice nobody reads does not discharge a duty to inform.
Fronting the agent as a fake colleague. A first name, a photo, an email signature: that is precisely what the text was aimed at.
Treating this as a legal problem. Disclosure is a product decision, made in the interface. No law firm can implement it for you.
Forgetting internal agents. Your staff are people interacting with the system.
Not knowing what the agent is allowed to do. That is the real project, and it goes beyond transparency: see our guide to AI agent security.

To do

The checklist, in order

Inventory every touchpoint where an AI speaks to a person — external and internal.
Add visible disclosure when each conversation opens.
Test "are you a human?" on every agent, and fix the prompt.
Make the handover to a human visible, in both directions.
Check that the agent's actions are logged and retrievable.
Put in writing who is provider and who is deployer, with your supplier.
Document the legal basis and what happens to the data exchanged.

If the first six are handled, you are in good shape on transparency. If the last one stops you, that is not an AI Act problem — it is a GDPR problem that existed before the agent, and the agent has just made it visible.

Want an outside read on an agent already in production? We do this scoping regularly for SMEs and mid-caps in regulated sectors — see our custom AI agents or our broader guide to the AI Act for business.

FAQ

Frequently asked questions — AI agent transparency

The European Commission states that the AI Act's transparency rules come into effect in August 2026, and that 2 August 2026 is the general application date from which the AI Office and Member State authorities are responsible for implementing, supervising and enforcing the regulation. The Commission's official AI Act Service Desk timeline sets Article 50 at 2 August 2026. For generative AI systems already on the market before that date, the Digital Omnibus allows until 2 December 2026 for machine-readable marking of generated content. The duty to tell a person they are dealing with an AI already applies.

Your AI agents must now introduce themselves

What applies, exactly

The real timeline, without approximation

Concretely, on your agents

The mistakes we see most

The checklist, in order

Frequently asked questions — AI agent transparency

Read next

Sources

Let's talk about your project

Contact details