AI hallucinations: why they happen and how to keep them in check on the phone

In its risk profile dedicated to generative AI, published in July 2024, NIST, the US National Institute of Standards and Technology, uses the term confabulation: the production of erroneous or false content stated with confidence, colloquially called hallucinations, which may mislead users. The problem is not just the error but the tone: nothing in the answer signals that it is doubtful.

On the phone, a hallucination can take very concrete forms: announcing opening hours that do not exist, quoting a made-up price, confirming that a service is offered when it is not, or stating that an appointment is booked when it is not. For the caller, the agent’s word is the company’s word.

A language model is trained to predict the most likely continuation of a text. It does not consult a database of verified facts: it generates words that statistically resemble a good answer. When the information requested is missing, rare or ambiguous in its data, it may produce a plausible but false answer rather than admit that it does not know.

A paper published in September 2025 by researchers from OpenAI and Georgia Tech, Why Language Models Hallucinate, puts forward a complementary explanation: common evaluation methods reward the right answer and give no credit for admitting uncertainty. Under these conditions, guessing becomes a rational strategy for the model, just as it does for a student who takes a chance in an exam rather than leave a question blank. The authors propose changing how evaluations are scored so that they stop penalising abstention.

The first safeguard is to restrict answers to the information provided by the business: website, documents, fact sheet. The agent does not draw on general knowledge to answer a question about your opening hours or prices; if the information is not in what it has been given, it says so and suggests another route. This approach reduces the risk of invention without eliminating it.

The second safeguard is repeating back: before recording a request or confirming a slot, the agent repeats what it has understood and asks for confirmation. The third is transfer: for any out-of-scope question or sensitive situation, the call goes to a person or a message is taken for a callback. The fourth is the readable transcript, which lets you check afterwards what was said and correct the reference information.

An agent that answers from your documents cannot be more reliable than those documents. Contradictory opening hours on two pages of your website, an out-of-date price in a PDF or a discontinued service that is still mentioned will produce wrong answers, without the model having hallucinated in the strict sense.

So set up a single, up-to-date source of truth, written in plain language: what you do, what you do not do, the cases where the call must be transferred. Also list the questions the agent must not answer, such as a diagnosis, legal advice or a firm price estimate, and the standard response to give in those cases.

Be wary of any pitch promising zero hallucinations. On the basis of published knowledge to date, no language model offers that guarantee, and the researchers cited above describe hallucinations as a consequence of training and evaluation methods. A serious provider talks about reducing the risk, explains its safeguards and gives you the means to check what was said.

Ask precise questions: what does the agent say when information is missing? Can it confirm an appointment without actually entering it in the calendar? Can you read back every call? The answers to these questions tell you more than an accuracy rate announced without any methodology.

Alyas answers the phone based on the information you provide, whether that is your website, a text or a PDF. It transfers to the number you have chosen, collects the caller’s details and places a transcript and a summary of each call in your customer area so that you can check its answers. The ten-day free trial lets you put it to the test on your own questions.