Should You Tell Customers They Are Talking to an AI Agent?

Yes. Tell customers when they are talking to an AI agent, at the start of the conversation and again whenever the answer matters to a decision. Disclosure protects trust, sets accurate expectations about what the agent can do, and in a growing number of jurisdictions it is a legal expectation, so it is safer to treat it as a design requirement than as a risk to manage later.

Key takeaways

• Disclose at the first message in every channel, in plain words, and answer truthfully whenever a customer sincerely asks whether they are speaking to a person.
• Hiding the agent does not remove the difference between an agent and a person. It only moves the moment of discovery to a worse time, usually a complaint.
• Good disclosure has three parts: what the agent is, what it can do, and how to reach a human.
• A handoff to a person is part of honest disclosure. The person should receive a summary so the customer never repeats the story.
• Rules vary by country, sector and channel, so write your policy once, check it with counsel for each market, and keep a record of the wording you used.

Why does disclosure matter even when nobody asks?

Customers form expectations from the first line of a conversation. If they believe they are writing to a person, they assume judgment, discretion, the ability to bend a rule and some memory of who they are. If they are writing to an agent, they adjust: they keep requests short, they check important answers, and they expect a human to be available for exceptions. Those expectations are not a weakness to engineer around. They are how people decide how much to rely on what they are told.

The common worry is that disclosure makes people leave. In practice the risk sits elsewhere. A customer who finds out later that the friendly reply was automated tends to reread everything that came before it with suspicion, including the correct parts. That discovery usually happens at the worst moment: a billing dispute, a delayed delivery, a sensitive question. At that point the problem is no longer the answer, it is the relationship.

Disclosure also protects your own team. Staff who take over from an agent need to know what the customer was told and under what framing. If the customer thought they were speaking to a colleague, the person picking up inherits a misunderstanding they did not create.

Finally, disclosure keeps the agent honest in scope. When the first message says what the agent can do (bookings, order status, common questions), it gives you a natural boundary. Requests outside that boundary become an expected handoff instead of a failure.

What do the rules say about telling customers?

This is general orientation, not legal advice. The direction of travel is consistent: people should be able to tell when they are dealing with an automated system.

• The EU Artificial Intelligence Act includes transparency obligations for systems that interact directly with people, generally expecting that users are informed they are interacting with AI unless that is obvious from the context. The [EU AI Act overview](https://artificialintelligenceact.eu/) is a useful starting point for reading what applies to you.
• In Brazil, the data protection law builds on transparency, clear information to the data subject, and the ability to request review of decisions made only by automated processing. The LGPD checklist for AI agents goes through the points a company should verify, and the national authority, the [ANPD](https://www.gov.br/anpd/pt-br), publishes guidance.
• Consumer protection and advertising rules in many places treat misleading customers about who they are dealing with as a deceptive practice, even without an AI-specific law.
• Some regions and sectors (health, finance, public services) add their own requirements, and a few jurisdictions have rules aimed specifically at automated accounts that communicate with consumers.
• Messaging platforms, app stores and telephony providers can have their own terms about automated accounts. Read them for each channel you use.

Two practical habits make this manageable. First, write one disclosure policy that satisfies the strictest market you serve, instead of maintaining a different behavior per country. Second, keep a record of the exact wording shown in each channel and when it changed. If a question ever arises, you can show what a customer saw.

The broader principles are worth knowing as a design guide. The [OECD AI Principles](https://oecd.ai/en/ai-principles) and the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) both treat transparency and human oversight as core properties of trustworthy systems.

When should the agent disclose? A decision guide

Use this guide per channel and per type of conversation. When in doubt, disclose.

• The agent writes or speaks to a customer directly, in your company's name: disclose in the first message, and show it in the channel profile or greeting too.
• The conversation is by voice: say it in the opening seconds of the call, before any request is handled.
• A customer asks, in any words, whether they are talking to a person or a machine: answer truthfully and immediately, then offer the choice of continuing or reaching a colleague.
• The agent can change something for the customer (an appointment, an address, an order): disclose, and state the limits of what it can change without a person approving it.
• The topic is health, money, legal rights, safety or a complaint: disclose prominently and offer a human early, not only after a failure.
• The agent drafts a message and a named employee reviews and sends it: the employee is the sender and answers for it. Say in your privacy notice and public information that AI assists your team, and never let the message imply a personal act that did not happen.
• The agent works only for employees: customer disclosure does not apply, but tell staff clearly what it is. The tradeoffs are covered in internal vs customer-facing agents.

The line that should never move is the sincere question. An agent configured to deny being an agent, or to dodge the question, turns a manageable disclosure into a deception.

What wording works in practice?

Good wording is short, plain and useful. It names the agent for what it is, says what it can do, and gives a way out to a person. It does not apologize for being automated and does not oversell.

An opening message might read:

Hi, I'm the virtual assistant for [Company]. I'm an AI, and I can help with bookings, order status and common questions.

If you would rather speak to a person, just say so at any time and I will pass this on with the details, so you will not need to repeat yourself.

When a customer sincerely asks:

Yes, I'm an AI agent, not a person. I can keep helping, or I can bring in a colleague. Which would you prefer?

On a phone call:

You have reached the automated assistant for [Company]. I can help with appointments and questions, or connect you to our team.

A few rules of thumb:

• Put the disclosure where the customer will see it before they invest effort, not in a footer or terms page alone.
• A name and a friendly tone are fine, as long as the AI label is visible next to them. Avoid invented human biographies, photos of people, or small touches meant to imitate a person, such as fake typing pauses.
• Avoid vague labels like AI-powered as the only signal. Say that the replies come from an AI agent.
• Repeat the disclosure when the context changes: a new session after a long gap, a switch from a person back to the agent, or a move from a general question to an action on the customer's account.
• Write the wording in the customer's language and keep it consistent across email, chat and messaging. If one agent serves several channels, the same statement should appear in each, as described in one agent across email, chat and WhatsApp.

How should the handoff to a human work?

Disclosure without a real path to a person is a notice, not honest contact. The handoff is where customers judge whether you meant it.

Define the triggers in advance instead of leaving them to the agent's judgment in the moment:

• The customer asks for a person, in any wording.
• The agent fails to resolve the same issue after a small, fixed number of attempts.
• The customer shows frustration, distress or urgency.
• The topic involves money decisions, legal rights, safety or health advice.
• The request is outside the agent's defined scope, or its confidence in the answer is low.

Then define what the person receives. A good handoff template includes the customer's goal in their own words, what the agent already did or said, the data it verified, what it could not do, and any promise made. The customer should hear who will follow up and, only if you can honor it, when. If staff are unavailable until the next working day, say so plainly instead of implying an instant reply.

An illustrative scenario: a small dental clinic uses an agent on WhatsApp for appointments. A patient writes in the evening to move a visit. The agent opens by stating it is the clinic's AI assistant, offers new slots, and confirms the change. Then the patient adds that their jaw still hurts after a procedure and asks if that is normal. The agent does not give clinical advice. It says it is an AI and cannot assess symptoms, tells the patient the clinic's reception will read the message when the office opens, and gives the emergency instructions the clinic has approved for urgent pain. It flags the conversation to staff with a summary: the rescheduled slot, the exact symptom described, and the fact that the patient was told to expect a call. The patient repeats nothing when reception calls.

Notice what made it honest: the agent said what it was, stayed inside its boundary, and passed on the context. The same logic applies to every sector, from hotels to schools, only the triggers change.

How do you put this into practice?

A workable plan for a small team:

• List every channel where the agent speaks to customers, including voice, email, chat widgets and messaging apps.
• Write one disclosure line per channel and one standard answer to the question of whether the customer is talking to a person.
• Decide the handoff triggers, name the role that receives each type of handoff, and set the summary format.
• Test the wording and handoffs before launch. Running the agent in shadow mode lets your team review real conversations before the agent has authority.
• Log whether the disclosure was shown and when a handoff happened, so you can audit both.
• Read a sample of transcripts regularly, and treat any customer comment like I did not know this was automated as an incident to investigate.
• Review the policy when you add a channel, enter a new market, or change what the agent is allowed to do.

Disclosure also belongs in the question of who answers for mistakes. Telling a customer that an agent is involved does not move responsibility onto them. Your company still answers for what the agent says and does, which is the subject of who is accountable when an AI agent makes a mistake.

How Be Human handles this: in a first engagement, disclosure wording and handoff triggers are defined together with the agent's role, its tools and its approval boundaries, so they are part of the agent's operating context and not a banner added at the end. Managed operations then include recurring evaluations that check the agent still discloses correctly, escalates on the agreed triggers, and passes a usable summary to the person who takes over. Changes to wording or boundaries are made in a controlled way and reviewed by the client.

Frequently asked questions

Do I need to disclose if the agent only drafts replies and a person sends them? The sender is a person who has reviewed the message and answers for it, so a per-message label is usually unnecessary. Still, be open in your privacy notice and public information that AI assists your team, and check the rules of your market, since some require disclosure whenever AI generates the content customers read.

Will disclosing make customers abandon the conversation? Some will ask for a person sooner, and that is a signal about where your agent should hand off, not a reason to hide it. Customers tend to accept an agent that is clear about its scope, quick to help and easy to escape from. What damages engagement is an agent that pretends and then fails.

Can I give the agent a human-sounding name? A name is fine if the AI label is shown next to it and the agent answers truthfully when asked what it is. Avoid personal histories, human photos or behavior designed to make people believe a person is typing. The test is simple: would a customer feel misled if they saw the whole picture?

Does disclosing the agent shift responsibility to the customer? No. Telling customers they are speaking to an AI sets expectations, but it does not reduce your responsibility for accuracy, data handling or the actions the agent takes. Keep clear ownership inside your company, with a named person responsible for the agent's behavior and for reviewing its escalations.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.