Why your AI receptionist must never make things up — what anti-hallucination means in practice

A scenario that happens more often than it should

A customer calls your showroom on a Tuesday afternoon. Your AI receptionist picks up, greets them by your business name, and the caller asks about a Carron Quantum double-ended bath — specifically whether it is in stock and what it would cost installed.

The AI answers confidently with a price that includes fitting, and says it is available for delivery the following Thursday. The customer thanks it, hangs up, drives to your showroom on Thursday morning expecting to pay that price and take delivery.

The actual price is higher. The fitting is quoted separately. Thursday is fully booked. Your team spends a long time managing a frustrated customer who has rearranged their morning based on information your AI invented. The customer leaves. They do not come back. They tell other people what happened.

What does AI hallucination actually mean?

The term hallucination sounds dramatic, but the mechanism is straightforward. AI language models generate responses by predicting what text should come next based on patterns in training data. They do not look anything up, verify facts against a live source, or distinguish between what they know and what they are guessing.

When asked a factual question the AI cannot answer accurately from its training data, it does not say “I don’t know.” It generates a response that sounds correct — a plausible price, a plausible date, a plausible stock level — and delivers it with the same confidence as a response it can actually back up.

This is not a flaw that will be fixed in the next model version. It is a fundamental property of how language models work. The solution is not a better AI — it is an AI that is architecturally prevented from guessing on facts that matter.

Why is AI hallucination a bigger problem for trade businesses?

For a consumer web app, a wrong suggestion is an annoyance. For a trade business, a wrong answer from your phone system is a broken commercial promise.

Pricing errors create expectation gaps that your staff have to close in person, with a dissatisfied customer in front of them. Availability errors send customers to your showroom or to a site survey expecting something your diary cannot support. Appointment conflicts mean an installer travels to a job that was never properly booked.

The costs are not abstract. They are in staff time spent on complaint management, in customers who do not return, and in the reputation damage of a business that told someone the wrong thing and then had to correct it face to face. A bathroom purchase is a significant decision for most households. Customers remember when the process goes wrong.

How IBIOS’s AI Receptionist will handle uncertainty (coming soon)

The IBIOS AI Receptionist (coming soon) is being built around three principles that together will prevent hallucination from reaching a caller.

First, every factual response will be grounded in live ERP data retrieved at the moment of the call. The AI will not answer pricing questions from memory or training data — it will read your actual product catalogue pricing before it speaks. It will not answer availability questions from assumptions — it will read your live stock levels and calendar. There will be no cached snapshot of your business that can go stale.

Second, every factual claim will be confidence-scored before being voiced. The AI will evaluate how certain it is based on what it retrieved. If the product was not found in the catalogue, or the price has more than one applicable variant, or the calendar data returned ambiguous results, the confidence score falls below the threshold and the response is intercepted.

Third, intercepted responses will be replaced with a graceful handoff — not a robotic refusal. The caller will be offered a callback or a transfer, with an explanation that sounds like good customer service rather than a system limitation.

What “graceful uncertainty” sounds like to the caller

The difference between a naïve AI and a well-designed one is audible in the words it chooses when it is not certain.

A naïve AI, asked what a bath costs installed, generates a price and delivers it. A well-designed AI, asked the same question, checks the live catalogue, finds that installation pricing requires a separate survey, scores the confidence on the total figure as insufficient, and says:

To the caller, this sounds like a business that takes their enquiry seriously. The AI is not refusing to help — it is being careful on their behalf. That is a better customer experience than a confident wrong answer, and it is certainly a better outcome for the business.