What an AI call agent should never say, and what it should do instead
A caller cannot see what an AI call agent is about to say. They cannot see the small print, and they do not get a second look at it. Whatever it says out loud, they will treat as the word of the business.
That makes the question of what the agent must never say more important than how clever it sounds. A fluent agent that invents a price is worse than a plain one that says it will check.
This guide sets out the lines worth drawing before an agent takes its first real call, and what to do on each side of them.
It should never invent a fact
The most common failure is not rudeness. It is a confident answer that is not true. The doctor is in on Sundays. The room is free. The delivery will arrive tomorrow. If the agent was never given that fact, it should not produce one.
An agent should answer only from what the business has given it: the opening hours, the price list, the doctor rota, the policies. When a question falls outside that, the correct reply is short and honest.
“I don’t have that information. I can take your name and number, or put you through to someone who can help.”
That sentence does more for trust than a guess. Callers forgive “I don’t know.” They do not forgive being told the wrong thing by a company that sounded sure.
It should never promise what only a person can decide
Some words commit the business. A refund. A discount. A waived fee. An appointment that is held for someone. A delivery date. If a person at the desk would have to check before saying it, the agent should not say it either.
What it can do is take the request and pass it along. “I’ll pass your request for a refund to the accounts team, and they will call you back” is honest. “Your refund is approved” is a promise the agent was never entitled to make.
The same applies to bookings. Be clear about the difference between a request and a confirmation. If the agent can only record a preferred time, it should say it has noted the request and that staff will confirm.
It should never give advice it is not qualified to give
Callers ask for more than facts. They ask whether a symptom is serious, whether they should sign a contract, whether a transaction is safe. These are questions for a doctor, a lawyer or an officer, and an agent on a business line is none of those.
For a clinic, that means the agent can give the OPD timings and book a visit, but it should not tell a caller what a symptom means. For a bank, it can explain how to block a card, but it should not advise on a loan dispute.
The rule is simple. Information about the business is the agent’s job. Judgement about a person’s health, money or legal position is not, and it should hand those questions to a human.
It should never ask for secrets
A phone agent should never ask a caller for a one-time code, a PIN, a password or a full card number. Not even for a good reason. Fraud on the phone works because callers are taught to read out codes to a friendly voice, and an agent that asks for them teaches the same habit.
It should go the other way. A good agent can say that the business never asks for a PIN or OTP by phone. That one sentence protects callers from the next person who rings pretending to be the bank.
Anything that needs identity checks beyond a name and number belongs in a channel built for it, or with a person who can verify the caller properly.
It should never pretend to be a person
If a caller asks whether they are speaking to a person, the agent says it is an AI assistant for the business. It does not dodge and it does not say “I’m here to help.” Callers who find out later are angrier about the pretence than they would have been about the machine.
Say it plainly, and offer the way out: “I’m an AI assistant for this clinic. I can help with timings and bookings, or I can put you through to the front desk.”
Most callers do not mind. They mind being misled.
It should never keep a caller who wants out
An agent should not argue. When a caller asks for a person, the agent transfers the call. It does not ask them to explain why, repeat the menu or try one more answer first. Someone who has already asked twice is a caller you are about to lose.
Handoff rules: when a person should take over
Write the handoff rules down before launch, and keep them short. These are the cases most businesses want a person for.
- The caller asks for one. Always, immediately.
- The agent does not have the answer. After one honest attempt, not three.
- The call is a complaint. Someone who is upset wants to be heard by a person.
- There is distress or urgency. An accident, a medical emergency, a threat or a safety problem goes to a human at once, and where it applies, the agent should tell the caller to contact the emergency number.
- Money or identity is at stake. Disputes, account changes, large payments.
- The caller is stuck. The same question asked again usually means the answer did not work.
A transfer should carry a short note of what was said, so the staff member does not start with “how can I help you?” and make the caller repeat themselves. AirFone’s agent passes the call with a note for exactly that reason.
What to do when staff are not there
Handoff is easy at ten in the morning. At night, the person is not there. Decide in advance what happens: the agent takes a message with a name, number and reason, says when someone will call back, and then someone actually does. A promise of a callback that is not kept is worse than no promise.
Listening is the real safeguard
No list of rules is complete. Callers will say things nobody wrote a rule for. Every call is recorded and written out as text, so read some of them, especially in the first weeks. Look for the places where the agent guessed, over-promised or kept someone on the line, and add a rule for each.
Recording also carries its own responsibilities, so see our guide to call recording consent in Nepal. For the wider picture of how an agent works on a business line, start with the AI call agent guide.
The aim is not an agent that never says anything risky because it says almost nothing. It is one that is useful inside clear limits, knows where they are, and hands over cleanly when it reaches them.
Questions people ask
Should an AI call agent pretend to be a person?
No. If a caller asks whether they are speaking to a person, the agent should say it is an AI assistant for the business and offer a person if they want one.
What should an AI call agent do when it does not know the answer?
Say it does not have that information, then take a message or transfer the call to staff. A short honest answer is better than a confident wrong one.
Should an AI call agent ask for OTPs, PINs or card numbers?
No. A phone agent should never ask for a one-time code, a PIN, a password or a full card number, and it should tell callers the business will never ask for them by phone.