What is an AI hallucination?

Hallucination is not a bug in the ordinary sense. These systems predict likely text, and a likely-sounding answer is often a true one and sometimes is not.

Confidence is the dangerous part. There is no hesitation in the delivery to warn anybody, which is exactly why people believe it.

Phone calls make it worse than writing does. Nothing is on screen to check, the caller cannot scroll back, and the answer is gone the moment it is said.

Grounding reduces it. A system restricted to answering from supplied material invents far less than one answering from general knowledge.

Say a caller asks your appliance repair shop's assistant, "Is the repair under warranty if it breaks again?" Your brief says nothing about warranties. A model left to fill that gap produces the most typical-sounding reply: "Yes, all repairs come with a 90-day warranty on parts and labor." It's fluent, specific and made up. Ninety days is simply what sentences like that usually say.

Six weeks later the dishwasher fails again. Your customer calls back and says your office promised 90 days. There's a recording, and he's right that it was said. You now either honor a warranty you never offered or argue with a man who has the promise on tape. A grounded assistant given the same question says, "I don't have that information, so let me get someone who does." That reply is less impressive, and it costs you nothing.

One case is worth knowing. In 2024 a Canadian tribunal ordered Air Canada to pay a passenger after its website chatbot described a bereavement refund policy that didn't exist. The airline argued that the chatbot was responsible for its own words. That argument failed, and the tribunal held the company to what its bot had said. It was a text chatbot and a small sum, but the reasoning travels. Your customers are entitled to rely on what your business tells them, whatever does the telling.

So test for it on purpose. Write ten questions your brief doesn't cover, such as a senior discount or a price for an unusual job. Call your line and ask each one. Scoring is simple: a handover or an "I don't know" passes, and any confident specific fails. Repeat the test after every big edit to your brief, since new text can open new gaps.

GreetKeeper answers from the brief you write, and hands over when a caller asks something outside it rather than improvising.

The transcript is the check. You can read what was said on every call, which is how an unwanted answer gets found in week one rather than month six.

We publish no accuracy figure for this. The mechanism is what we can describe honestly; a percentage would be invented.