Yes, at the start of every conversation and whenever a customer asks. In the EU, Article 50 of the AI Act requires that people are told when they are interacting with an AI system, unless it is obvious. Elsewhere it is good practice. The agent should never use a human name or persona that suggests a person is typing.
Why disclose, beyond the legal duty?
Because customers find out. They notice the speed, the phrasing, the way the agent handles an unexpected question. A customer who discovers halfway through that they were talking to an AI, after being led to think otherwise, stops trusting not just the agent but everything it told them. Disclosure at the start costs one line and avoids that.
Disclosure also changes how customers use the agent. Someone who knows they are talking to an AI asks more directly, and knows to ask for a person when the matter is sensitive. That makes escalation work better.
What should the disclosure say?
Three things, briefly: that this is an AI, what it can help with, and how to reach a person. For example: "I'm the AI assistant for this account. I can help with orders, returns and delivery. You can ask for a person at any time."
What it should avoid:
- Human names with no qualifier. "Hi, I'm Sarah" implies a person.
- Vague terms. "Virtual assistant" may not tell every customer what they need to know; "AI assistant" is clearer.
- Disclosure buried in terms. It must be in the conversation, where the customer sees or hears it.
When must the agent repeat it?
| Moment | What the agent does |
|---|---|
| Start of every conversation | States it is an AI, in every channel |
| Customer asks "am I talking to a real person?" | Answers plainly: no, and offers a person |
| Before a handover | Says a person is taking over |
| After a handover back to the agent | Says it is the AI again |
| In emails it sends without review | Signs as the AI assistant, not as a named staff member |
| On a voice call | Says it in the opening line, before collecting information |
The question "are you a bot?" must never be dodged. An evasive answer to a direct question is worse than no disclosure at all.
Does disclosure invite customers to test the agent?
Some, yes. Customers who know they are talking to an AI may try to trick it, argue with it or be rude to it. That is a reason to build proper limits, not a reason to hide what it is. The controls are covered in how to stop customers from tricking a support agent, and handling rudeness in handling abusive customers with an AI agent.
Disclosure also makes refusals easier to accept. A customer who knows the agent is an AI with set limits understands "I can't approve that, but a colleague can" better than if they thought a person was refusing. The wider point about explaining refusals is in whether a support agent can tell a customer why a request was refused.
An illustrative example
Take a utility's chat agent. A customer opens the chat and the agent's first message says it is an AI assistant, lists what it can help with and says a person is available on request.
Midway through, the customer writes: "Is this a real person? I want to talk about a scam call I got." The agent answers: "No, I'm an AI assistant. I can take the details of the scam call now, or connect you to a person. Which would you prefer?" The customer chooses the person. The agent hands over with a note. How scam reports are handled from there is set out in handling scam reports from customers with an AI agent.
The customer was never misled and got what they asked for.
What stays with people on the support team?
Choosing the disclosure wording, approving it per channel and language, and reviewing whether customers understand it. People also decide how the agent presents itself: a plain assistant label, a product name, a tone. That choice should never stray into impersonation. Disclosure is about honesty, not about negotiating with the customer over who they are talking to, a theme taken up in a support agent should not negotiate.
What controls keep it consistent?
- Fixed opening lines per channel and language, tested before launch.
- A test case for the direct question in every language, so "are you human?" always gets a plain answer.
- Email signatures that name the AI, not a staff member.
- Disclosure recorded in each conversation's log, so you can show it was given.
Where Pryme fits
In Pryme, disclosure is part of the platform guardrails for customer-facing agents, so it cannot be removed by a brief or a customer setting. Each conversation's record shows the disclosure given. More is in the AI support agent that knows when to say no.
Questions people ask
Do we need to disclose if a person reviews every reply before it is sent?
The reply is still written by an AI. Many businesses say so, for example "drafted by our AI assistant and checked by our team", which is accurate and builds trust.
Can the agent have a name?
Yes, if it is clearly presented as an AI, such as "Ava, our AI assistant". A name alone, without that label, can mislead.
Does disclosure lower customer satisfaction?
Measure it for your own customers. What tends to hurt satisfaction is discovering the truth later, not being told at the start.