People must be told they are dealing with an AI system by their first interaction, clearly and accessibly, under Article 50. In chat, say it in the first message and keep a visible label. On a call, say it in the greeting. In email, put it at the top of the first reply, not in a footer.
What does Article 50 require for disclosure?
The EU AI Act's Article 50 obliges providers to design AI systems that interact directly with people so those people are informed they are interacting with an AI system. The exception is where this is obvious to a reasonably well-informed, observant and circumspect person, taking the circumstances into account.
The information must be given in a clear and distinguishable manner at the latest at the time of the first interaction or exposure, and must meet applicable accessibility requirements. Separate duties cover marking AI-generated content, discussed in Article 50 marking of AI-generated content, and the overall picture is in Article 50 transparency for AI agents.
Who is responsible for getting it right?
The provider must build disclosure into the system. The deployer, the firm putting the agent in front of customers, controls the channel, the greeting, the branding and the hand-offs, so in practice the deployer decides whether the disclosure is visible and well timed. A provider's disclosure that the deployer hides behind a custom avatar and a human first name does not achieve the purpose.
Do not lean on the "obvious" exception. A fluent agent with a friendly name and a natural voice is precisely the case where it is not obvious.
How should disclosure work in each channel?
Each channel has a different "first interaction", and each has its own trap.
Chat. Disclose before the customer types, in the chat header or opening message, and keep a persistent label beside the agent's messages. Example: "You're chatting with our AI assistant. It can check your account and answer questions. Type 'person' at any time to reach a member of the team." The trap is a header that scrolls away and a later hand-off to a person that nobody announces.
Voice. Disclose in the greeting, before the agent asks for any information. Example: "Hello, you're speaking with an automated AI assistant from the bank. I can help with cards and payments, or put you through to a colleague." For outbound calls, the greeting is the only safe moment; a customer who answers expects a person. The trap is a synthetic voice so natural that a short disclosure gets missed, so keep it at a steady pace and repeat it if the caller asks who they are speaking to. Synthetic audio may also fall under the separate marking duty.
Email. Put the disclosure at the top of the agent's first reply. Example: "This reply was written by our AI assistant. If anything here is wrong or you'd like a person to respond, reply 'person'." The trap is a footer disclaimer below the signature that nobody reads, which does not meet "clear and distinguishable".
Switching between agent and person. Announce every switch in both directions. If a person takes over a chat, say so; if an agent picks up a thread a person started, say that too. One agent working across channels needs a consistent identity and disclosure, a design question covered in whether one AI support agent can work across email, chat and phone.
How do you meet the disclosure duty and prove it?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Inform at first interaction | Disclosure fixed into the opening of every channel flow; cannot be removed by template edits | Channel configuration and template locks |
| Clear and distinguishable | Disclosure in the body or greeting, in plain words, not in a footer or terms page | Screenshots and call scripts per channel |
| Accessible | Disclosure works with screen readers, in supported languages and at a clear speaking pace | Accessibility test results |
| Not obscured by persona | Agent names and avatars reviewed so they do not imply a person | Persona approval record |
| Switches announced | Hand-off messages for agent-to-person and person-to-agent transfers | Transcripts showing hand-off messages |
| Proof it was given | Disclosure recorded as part of each conversation | Conversation records with disclosure line |
The last row is the one most firms miss. If a customer later claims they thought they were speaking to a person, you need the transcript or call record showing the words were said.
What will a supervisor ask?
- Show us the first thing a customer sees or hears in each channel.
- Do any agents use human names or voices that could suggest a person?
- How do you tell a customer that a person has taken over, or that an agent has?
- Can you prove the disclosure was given in this specific conversation?
- Is the disclosure available in every language the agent supports?
What does Article 50 disclosure not cover?
It does not require a specific form of words. It does not require disclosure where a person writes or speaks and the agent only assists in the background. It does not give customers a right to a human under the Act itself, although offering one is good practice and may be required by other rules or your own service standards. And it does not address whether the agent may do what it does; disclosure does not authorise an action. For how the duties fit the rest of the regime, see AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
Pryme files each conversation's disclosure as part of the append-only record, alongside the agent's actions and verdicts, so you can show for any customer contact what they were told and when. One agent identity can operate across channels while keeping a consistent disclosure in each.
Questions people ask
Is a disclosure in the terms and conditions enough?
No. Terms are not the first interaction and are not clear and distinguishable in the moment. The disclosure has to appear in the conversation itself.
Can the AI agent have a name?
Yes, if the name and avatar do not suggest a person and the disclosure is given. "Our AI assistant, Max" is safer than "Max from the support team".
Does an interactive voice menu need the disclosure?
A fixed menu with recorded options is not an AI system interacting with the caller. A voice agent that understands free speech and responds is, and should disclose in its greeting.