Your AI agent is ready to answer customers, and someone asks the simple question: do we tell them? In most cases the answer is yes, and the better question is how. A notice that is vague, late or buried creates exactly the distrust you wanted to avoid. A short, honest one costs almost nothing and makes the human alternative feel like a service and not like a failure.
This article is for the manager responsible for customer communication. It assumes the legal questions have been put to your lawyer and focuses on the practical side: wording, timing and the handover to a person.
1. Why tell customers at all
There are two reasons, one legal and one practical. On the legal side, the EU AI Act contains a transparency rule for systems that interact directly with people: users should be informed that they are dealing with an AI system, unless that is obvious from the circumstances (AI Act, Article 50). Whether and when it applies to you depends on your role and your use case, so confirm it with counsel. If personal data is collected, the GDPR also requires you to give people clear information about how it is used (GDPR Art. 13).
The practical reason is simpler. Customers who find out later that "Anna from support" was software tend to feel misled, even when the answers were correct. Customers who were told at the start usually judge the agent by whether it solved the problem. For what happens when communication goes wrong, see our post-mortem of an agent communication mistake.
2. Say it first, in the same place
The notice belongs at the start of the conversation, in the channel the customer is already using. Not in the terms and conditions, not on a separate page. If the agent writes emails, the first line says so. If it answers in a chat window, the first message says so. If it answers calls, the first sentence says so.
Good wording is short and concrete:
- Chat: "Hi, I'm an AI assistant for [Company]. I can help with orders, invoices and delivery questions. If you prefer a person, just say so and I'll pass you to our team."
- Email footer: "This reply was prepared by our AI assistant. A colleague reads anything involving payments or complaints. To reach a person directly, reply with the word PERSON."
- Phone: "You are speaking with an automated assistant. Say 'agent' at any time to be connected to a colleague."
Avoid human names and photos for the agent, avoid phrases such as "our team will look at this" when no person will, and avoid jargon like "LLM-powered". Customers do not need the technology, they need to know what they are talking to and what to do next.
3. Make the human option easy to use
A human alternative that exists on paper but takes six clicks to find is not an alternative. Test it with these questions:
- Can the customer ask for a person in their own words, and does the agent recognise it every time?
- Does one request suffice, or does the agent try to talk the customer out of it?
- Is the route available in every channel the agent works in?
- Is it clear what happens next and when?
For some decisions the human route is also a legal matter. The GDPR gives people the right to obtain human intervention, to express their view and to contest a decision that is based solely on automated processing and has legal or similarly significant effects on them (GDPR Art. 22). If your agent touches decisions like refunds, credit or account closure, design the human review into the process from the start, as described in human approval and audit trails.
4. Promise only what you can staff
The human route needs people with time to answer. Decide the response times before you launch and say them to the customer: "A colleague will reply within one working day" is better than "as soon as possible". If the team is available only in office hours, say that too, and let the agent still collect the details so the case is ready when the person starts.
Estimate the load honestly. If a share of customers asks for a person in the first weeks, the team needs capacity for that, not only for the cases the agent cannot solve. Plan a simple rule for peaks, for example a longer stated response time shown automatically when the queue passes a set size.
5. Hand over the case, not just the customer
The most irritating moment is the handover that starts from zero. When a customer moves to a person, the person should see the conversation, what the customer wants, what has been checked, and what the agent was unsure about. The customer should not be asked to explain again. A short summary at the top of the case is enough.
Decide in advance when the agent hands over without being asked: an angry or distressed customer, a complaint, a request involving legal or financial consequences, a question the agent has answered twice without success, and any case where the agent's confidence is low. Write these triggers down in the runbook, and tell the customer when it happens: "I'm passing this to a colleague, who will see everything we've discussed."
6. Wording that works in difficult moments
Some situations need extra care:
- A mistake by the agent. Say what happened, what is being done and who will follow up. Do not blame "the system".
- Questions about being a bot. If asked directly, answer directly, every time.
- Vulnerable customers. If a customer seems confused or upset, offer a person without being asked.
- Sensitive topics. For health, money problems or bereavement, consider routing straight to a person.
7. Review it every month
Treat the notice and the handover as part of the product. Each month look at a small set of numbers: how many customers ask for a person, how long they wait, how many cases the agent hands over on its own, and how many customers repeat their problem after handover. Read a sample of conversations, too. If almost nobody asks for a person, check whether the option is visible. If many do, ask what the agent is failing to do.
Update the wording when the agent's abilities change. A notice that says "I can help with invoices" is wrong the day the agent starts handling returns.
A short checklist
- Tell customers at first contact, in the same channel, in plain words.
- Do not give the agent a human name or photo.
- Let customers ask for a person in their own words, and accept it the first time.
- State the response time of the human route and staff it.
- Send the conversation history and a summary with the handover.
- Write the automatic handover triggers in the runbook.
- Review requests for a person, waiting times and repeated explanations every month.
Telling customers the truth takes one sentence. Keeping the promise behind it, a person who can really help, is the part that needs planning, and it is also the part that builds trust.
