Should an AI assistant tell customers it is an AI, and how?

Why pretending to be human always backfires

Customers tell an assistant things that matter to them: a health worry, a family situation, a legal deadline, their address and phone number. They share those details because they trust the business. If they later discover that the friendly person who answered was software claiming to be human, that trust breaks, and it rarely breaks only for the assistant. It breaks for the business.

There is a legal side as well. A growing number of jurisdictions require that people be told when they are dealing with an AI system, and consumer protection rules in many countries already forbid misleading customers about who or what they are dealing with. The honest path is also the safe one, and it is simply the respectful thing to do.

How to disclose well

Disclosure does not need to be heavy. Many businesses give the assistant a name and a role together, so the first message already makes things clear, and on the phone the greeting does the same. Whatever the wording, two rules never bend: the assistant never claims to be a person, and if anyone asks whether they are talking to a robot, it says yes at once, plainly, and offers a human when that is useful. Where a law requires disclosure at the very start, follow it to the letter.

Avoid both extremes. A long warning paragraph before every answer makes the assistant tiresome, while a vague phrase that leaves people guessing is not disclosure at all. Clear, short and consistent across every channel is the aim.

• Introduce the assistant by name and as the assistant of the business.
• Answer truthfully, the first time, whenever someone asks if it is an AI.
• Say what it can do for the person, and when a colleague will take over.
• Never invent a personal life, a lunch break or feelings to seem human.

Consent and the right to stop

A customer who writes to a business wants an answer to their question. That is not permission to send them promotions afterwards. Before the assistant sends messages the business starts, such as reminders or follow-ups, the customer should have agreed to them, and the record should show when and how.

Stopping must be as easy as starting. A customer who writes stop, no more messages, or the same idea in any language should be heard the first time, and the choice should apply across the business's channels, not only in the one conversation. On calls, tell people at the start when the call is recorded or transcribed, as many laws require, and respect a caller who prefers not to continue.

Collect only what the request needs

Data minimisation is a simple idea: ask only for what the task requires. Booking a haircut needs a name, a time and a contact; it does not need a date of birth. A school enquiry may need the child's age and level, but not a family's income. Every extra field is one more thing to protect and one more reason for a customer to hesitate.

Some information deserves particular care. Health details, legal matters and documents that identify a person are treated as sensitive by data protection laws in many countries, and some require prior authorisation before such data is processed, as the CNDP does in Morocco for health data. Decide at setup what the assistant may ask, how long conversations are kept, who in the team can read them, and how a customer can ask for their data to be erased.

A person is always within reach

Trust comes from knowing there is a way out. Every AI assistant should have a clear route to a person, and the customer should know it exists. When a request is handed over, the assistant should say honestly what happens next, for example that a colleague will reply in the morning, rather than promising an instant answer nobody can give.

The cases that always go to a person should be written down: urgent situations, complaints, anything medical, legal or financial beyond practical information, and any customer who asks for a human. The assistant gives no advice in those areas; it gathers the facts and passes them on.

What to ask any vendor

Before choosing an AI assistant, put these questions in writing and expect written answers. A vendor who cannot answer them clearly is telling you something important.

• Does the assistant ever claim to be human, and can we see how it answers when asked?
• Who owns the conversations and customer records, and can we export everything?
• Where is the data stored, and which outside providers process it?
• Are our customers' conversations used to train models for anyone else?
• How long is data kept, and how is a deletion request handled?
• How are consent and opt-outs recorded and applied across channels?
• What does the assistant refuse to do, and how does handover to a person work?
• What is written in the contract, and what is guaranteed?

A concrete example: Nolloi

Nolloi's assistants never pretend to be human: if asked, they say they are an AI. They give no medical, legal or financial advice, only the practical information the business approved, and they hand to the team whatever the business's rules say needs a person. The business chooses the assistant's name and tone at setup. The client owns its data and Nolloi acts as its processor, with a full export on request and 60 days of read-only access after the contract ends, all set out in a signed contract.