An AI agent should tell people it is an AI whenever it talks to them directly and that is not already obvious. Under the EU AI Act's transparency rules, which apply from 2 August 2026, the disclosure must be clear and come no later than the first interaction. Build it into the agent's design, not its terms of service.
Why this is now a design question
For years, telling users they were talking to a bot was a matter of good manners and brand policy. Article 50 of the EU AI Act turns it into an obligation for AI systems that interact directly with people, and the transparency rules began to apply on 2 August 2026. The practical guide to the AI Act's transparency rules sets out who holds each duty and the exceptions.
The duty to make people aware sits mainly with the provider, the organisation that builds the system, and it is framed as a design duty. That matters for agents in particular. A chatbot sits in an obvious chat window. An agent may send emails, answer phone calls, post in a ticketing system or message a supplier, and in many of those places nobody would guess a machine is involved unless they are told.
This post explains how teams can approach the requirement. It is not legal advice, and your own legal advisers should confirm how the rules apply to your systems and markets.
When disclosure is required
Direct interaction with a person
The core rule applies when an AI system is intended to interact directly with people. That covers an agent answering customers in chat, speaking to callers, replying to emails or negotiating with a counterpart's staff. It does not depend on whether the agent is high-risk. A low-risk customer service agent is still in scope.
The obvious-from-context exception
Disclosure is not needed where a reasonably well-informed, observant person would already recognise they are dealing with an AI, given the circumstances. Guidance reads this narrowly. It asks you to think about your actual audience, including people with lower digital literacy or other vulnerabilities, rather than assume everyone will work it out. If you are relying on the exception, write down why.
Content the agent generates
Separate duties cover synthetic content. Systems that generate text, images, audio or video are expected to mark the output in a machine-readable way so it can be detected as AI generated. Deployers publishing AI-generated text to inform the public on matters of public interest must also disclose it, unless a person has reviewed it and someone holds editorial responsibility.
How to build disclosure into an agent
Disclose in every channel the agent uses
List every surface where the agent reaches a person: chat, voice, email, SMS, internal tools and third-party platforms. Each needs its own disclosure that fits the medium. A short statement at the start of a call, a clear label in a chat window and a line in an email signature all serve the same purpose. Keep the wording consistent so people learn to recognise it.
Make it part of the agent, not a wrapper
Disclosure added only by a front-end template is easy to lose when the agent is connected to a new channel. Treat it as part of the agent's release, alongside its prompt, tools and policies, and test for it in your evaluation suite. A check that fails when a reply reaches a person without the disclosure is cheaper than finding the gap after launch.
Handle hand-offs and agents acting for a user
When an agent passes a conversation to a person, or a person hands it back, say so. People should not have to guess who they are talking to at each point. Take extra care when an agent acts on behalf of a named user, such as sending email from their account. The recipient should not be led to believe the user wrote it personally. The way an agent acts on behalf of a user should make that visible.
Keep a record
Log that the disclosure was shown, in which channel and in what form, along with the rest of the agent's activity. That evidence belongs in the same place as the rest of what an AI agent audit trail contains, and it supports the wider approach to compliance with non-deterministic systems.
Disclosure as part of trust
A clear disclosure does not have to make an agent feel less useful. People tend to accept automation more readily when they know what they are dealing with and how to reach a person if they need one. Pair the disclosure with an easy route to a human, which also supports human oversight without friction.
CodeDTX's AI agent development and enterprise AI integration teams design disclosure, logging and oversight into agents from the start. If you are not sure your agents identify themselves in every channel, talk to CodeDTX.
Frequently asked questions
Does the disclosure rule apply to businesses outside the EU?
It can. The AI Act applies to providers placing AI systems on the EU market and to systems whose output is used in the EU, regardless of where the provider is based. A company in India or the United States whose agent talks to customers in Europe should assume the transparency rules apply and design for them. Confirm the position for your specific products with legal advisers who know your markets.
Is a line in the terms of service enough?
Usually not. The requirement is that people are made aware, clearly and distinguishably, no later than the first interaction. A clause buried in terms that most users never read does not achieve that. Disclosure should appear where the conversation happens, in a form that suits the channel, such as a spoken statement on a call or a visible label in a chat window or message.
Does an internal agent used only by employees need to disclose?
The rule covers people in general, not only customers, so staff talking to an internal agent are in scope unless it is already obvious to them. In many internal tools the context makes it clear, but do not assume so for agents that message staff by email or chat as if they were a colleague. Workplace use can also bring separate duties, such as informing workers and their representatives.
What should an agent say when someone asks if it is human?
It should answer truthfully, every time, and never claim to be a person. Build this into the agent's instructions and test it directly, including when users phrase the question indirectly or the agent has been given a human-sounding name. A plain answer that it is an AI system, followed by an offer to connect the person to a human colleague, keeps the conversation honest and moving.
