Disclosing AI without breaking the conversation

People talking to an AI should know it. There is broad agreement on that.

Since August 2026 the EU AI Act has required it for systems that interact directly with people. The real question is a practical one: how do you say it so that nobody trips over it?

Poor disclosures come in two kinds. One hides in the small print. The other is so long that the customer forgets her question before the conversation has even started. Neither does its job.

#Why disclosure is more than a duty

Article 50(1) of the AI Act requires that people are informed when they interact with an AI system, unless that is obvious anyway. Our post The EU AI Act in customer service covers the details.

Duty aside, disclosure is useful. People who know what they are talking to phrase things differently: shorter, clearer, often with exactly the keywords the agent needs. And they know they can ask for a person.

#Timing: at the start, once, clearly

The disclosure belongs at the start of the conversation, at the latest at the first interaction. It does not need repeating in every message.

It does need to reappear when something changes: when a person takes over, or when the agent returns after a pause.

A simple three-part pattern works well.

1. Who is speaking: "This is your insurer's digital assistant." A name is fine; pretending to be a person is not.
2. What it can do: half a sentence on typical requests, such as contract questions and address changes.
3. The way to a person: "Just say 'agent' at any time and I'll put you through."

#Wording that works

Good disclosures are short, active and concrete. The table shows common versions and why the second column works better.

Test the texts with people who do not know your processes. If they cannot tell at first reading whether a person or an AI is replying, you have found a weak spot. A quick test with colleagues from other departments is often enough.

#Every channel has its own form

In web messaging, the disclosure is visible and stays visible: as the first message and as a label on the sender, such as "AI agent" next to the name. When a person takes over, the sender visibly changes too.

On the phone, there is no sender label. A voicebot therefore has to introduce itself in the greeting, before the first question. Brevity matters even more than in text: one sentence is enough, and the conversation does the rest.

The voice itself is part of the disclosure too. A synthetic voice that sounds like a specific person undermines any announcement. Choose a voice that is pleasant to listen to without imitating a human being.

Messaging apps and SMS follow the rules for text, with one twist. Conversations there often stretch over hours. When the agent comes back after a long pause, a short reminder is appropriate.

#When AI helps in the background

Not every AI talks to customers itself. Univerx Assist suggests replies that employees review, adapt and then send themselves. Here, a person is speaking, using a tool.

Whether and how you disclose this kind of support is a question for your own principles.

What matters is that the roles stay clear: the employee is responsible for what she sends. Generated content that goes out without human review deserves different treatment from reviewed suggestions.

Univerx Assist· active

Conversation

AI agent, disclosedHandover from AI agent · context complete

I received the decision letter, but the amount does not match my application.

I have the case in front of me. One moment, I am checking the calculation.

Sentiment: calmVoice

Univerx Assist

Suggested reply

The letter deviates from the application. I will propose a correction and send you the confirmation by message.

InsertAccept

Summary

Request: deviation in the letter. Documents on file. Next step: correction, confirmation by message.

#The handover as a second disclosure

The switch from agent to person is a moment the customer should notice. A simple sentence does it: "I'm now connecting you with Ms Weber from customer service." In chat, the sender changes as well.

The same applies in reverse. If an agent takes over again after the conversation, for a confirmation for instance, that should be recognisable. More on handovers in Why the handover decides.

#Define disclosures centrally

In a contact centre with many flows, disclosures easily spring up in many places, each slightly different. It is better to define, approve and version texts and rules centrally. That way you can later show which version applied when.

In Univerx, disclosure in conversation is part of the basic settings of AI agents and voicebots. It is approved together with the agent, and every action the agent takes is logged.

The guide Introducing AI agents responsibly covers the rollout as a whole, and Accessibility in customer service explains what makes customer interfaces accessible. Terms such as voicebot and EU AI Act are explained in the glossary.