It can, if it's allowed to guess. AI that generates speech is built to give fluent, natural answers, and when it lacks the right information it can fill the gap with something that sounds right but isn't. A well-run AI receptionist is held to what it has actually been told about your business, and when a caller asks something outside that, it says so, takes their details and passes the question on. For your callers, an honest “I'll get someone to call you back” is far better than a confident wrong answer. And as one tribunal decision shows, the business is likely to be held responsible for what its AI says.

Why AI sometimes sounds sure and is still wrong

Generative AI predicts likely words and sentences. That's what makes it sound natural on the phone. It's also why it can produce an answer that is fluent, polite and false. This is often called “hallucination”, and Singapore's Model AI Governance Framework for Generative AI (published by IMDA and the AI Verify Foundation on 30 May 2024, checked 13 November 2026) treats it as one of the risks that came with generative AI. The framework also notes that no AI system is foolproof.

On a phone line, the risk looks like this: a caller asks something the AI hasn't been given, and instead of saying “I'm not sure”, it offers a price, a date or a policy that sounds plausible.

What a wrong answer costs the caller

Think about it from the other end of the line. A wrong answer isn't an abstract error. It's a person who:

  • turns up on a day you're closed
  • plans their budget around a price you don't charge
  • books time off for an appointment that was never confirmed
  • relies on a refund or cancellation policy you don't have

Most callers won't complain. They'll just feel let down, and they'll remember where the bad information came from. That's the real cost: not the call itself, but the trust you lose with someone who was ready to become a customer.

Who is responsible when an AI gets it wrong?

There's a well-known example from Canada. In Moffatt v. Air Canada, decided by British Columbia's Civil Resolution Tribunal on 14 February 2024 (full decision, checked 13 November 2026), a customer asked the airline's website chatbot about bereavement fares after a death in the family. The chatbot said they could apply for the reduced fare after travelling. The airline's own policy page said the opposite.

The airline argued it couldn't be held liable for what the chatbot said. The tribunal disagreed, finding the chatbot was simply part of the airline's website and that the company is “responsible for all the information on its website”. It ordered the airline to pay the difference in fare, plus interest and fees.

Three cautions about this example:

  • It's a small-claims decision in Canada, not Singapore law, and it involved a text chatbot, not a phone line.
  • It doesn't mean every AI mistake leads to a claim.
  • It does show how a tribunal can view the argument “the AI said it, not us”.

In Singapore, the Consumer Protection (Fair Trading) Act treats a misleading representation or false claim made to consumers as an unfair practice (CCCS, checked 13 November 2026). We're not aware of a Singapore case applying this to an AI phone assistant, and this isn't legal advice. But the safe assumption is simple: what your AI tells callers, your business is telling callers.

Which questions carry the most risk

Not every question is equally risky. A rough way to sort them:

Usually low risk, if your information is accurate and current:

  • Opening hours and holiday closures
  • Address, parking and how to get there
  • Which services you offer, in general terms
  • How to book or who to contact

Higher risk, and often better passed to a person:

  • Exact prices and quotes, especially for custom work
  • Confirming a booking, a slot or stock that the AI can't actually see
  • Refunds, cancellations, warranties and other policies with conditions
  • Anything medical, legal, financial or safety-related
  • Promises about outcomes, such as “that will definitely fix it”

The rule of thumb: if a wrong answer would change what the caller does next, the AI should either have the exact answer from you or pass the question on.

What a good “I don't know” sounds like

An honest answer doesn't have to sound unhelpful. Hypothetical examples of how it can sound to a caller:

  • “I don't have the price for that particular job, so I've noted your question and someone from the team will call you back with an accurate quote.”
  • “I can't confirm that slot myself. Can I take your name and number so the team can confirm it with you?”
  • “That's a question for the team rather than me. I'll pass on exactly what you've asked.”

Notice what each one does: it admits the limit, tells the caller what happens next, and makes sure the question isn't lost. Compare that with a confident guess that turns out to be wrong. The first feels like good service. The second feels like being misled, even when no one meant to mislead.

What to decide before an AI answers your phone

These are business decisions, not technical ones. Write the answers down:

  1. What may it answer on its own? List the topics where you're confident the information is correct and stable.
  2. What must it always pass on? Prices for custom work, complaints, policy exceptions and anything sensitive are common choices.
  3. How fast will someone call back? A promise to call back is only reassuring if it happens. Decide who does it and by when.
  4. Who keeps the information current? Prices, hours and policies change. Pick one person to update them and a regular time to check.
  5. How will you spot mistakes? Read the call summaries or records regularly, at least at the start, and look for answers that weren't quite right.

Our guide to whether an AI receptionist should sound human covers the related question of being honest that it's an AI in the first place.

When an AI receptionist is the wrong choice

Be honest about this before you decide:

  • If your answers change constantly. If prices, stock or availability change by the hour and the AI can't see them, most calls will end in “someone will call you back”, which may frustrate callers more than voicemail would.
  • If most calls need judgement. Custom quotes, negotiations and complex diagnoses are better handled by a person.
  • If callers are often upset or in difficulty. People who are distressed, grieving or complaining usually want a human.
  • In an emergency. Anyone describing an emergency should be told to call 995 for ambulance or fire, or 999 for police, and shouldn't be kept on the line by an AI.
  • If callers ask for a person. Some customers simply prefer to talk to someone. Make sure they can, and that someone actually calls back.

A hypothetical call, done well

This is an invented example. At 9.15pm, a caller asks a car servicing workshop whether this year's full service will cost the same as last year's. The AI has the workshop's standard price list but not the caller's history.

It gives the current published price for a standard full service, explains that it can't see past invoices, and offers to have the team check the caller's previous job and call back the next morning. It takes the caller's name and number. The owner reads the summary at 8am and calls back at 9am with the right figure.

Nothing was invented. The caller got a useful answer that night and an accurate one the next morning.

Try to catch one out yourself

Reading about this only goes so far. The quickest way to judge an AI receptionist is to ask it something it can't possibly know and listen to what it does.

We built one for our own enquiries. Her name is Penny. She tells callers she's an AI in her very first line, and when she's asked something she can't answer, she takes the caller's name, number and question so the team can follow up, with a summary of every call. You can talk to her by voice or chat in your browser at penny.pencilglow.com, with no sign-up and no need to leave your details. Ask her about something she couldn't know. Try to trip her up. Then decide what you'd want your own callers to hear.