AI hallucinations at the front desk, and how to stop them

A hallucination is when an AI produces a confident, fluent answer that is not based on anything: a price it was never given, an opening time it inferred, a policy that sounds reasonable and does not exist. At a front desk this means a customer told "yes, we take that plan" or "Saturday at 10 works" when neither is true. Four mechanisms reduce it: grounding on approved facts, an honest fallback when facts are missing, a check of the draft before it sends, and supervised approval for the first weeks. The first two do most of the work, and a product without them will eventually invent something in your name.

Why models do this

A language model is trained to produce text that reads as likely. Asked a question, it produces the most plausible continuation, and a plausible answer to "do you take Delta?" is "yes, we do". Nothing in the model's basic operation distinguishes "I know this" from "this sounds right". Fluency is not knowledge. The model is not lying; it is doing exactly what it does, and that is the wrong thing for a front desk.

What it looks like in a conversation

"Yes, we are open until 8pm on Saturdays." (You close at 5.)

"A crown is around 900 dollars." (You never loaded crown prices.)

"You can cancel up to two hours before with no charge." (Your policy is 24 hours.)

"Dr Lee is available Thursday at 10." (Dr Lee does not work Thursdays.)

Each is fluent, specific and helpful-sounding. Each creates a commitment the business did not make. The customer has no way to tell, and neither does a person skimming the log unless they know the facts.

Mechanism 1: grounding

The AI is given your approved facts and told to answer only from them. A question with no matching fact has nothing to draw on. This removes most hallucinations at the source. Our post on grounding explains it.

Mechanism 2: the honest fallback

When no fact matches, the AI says so: "I am not sure about that. Let me check with the team and get back to you." This has to be the designed behaviour, not an accident. Products that treat "I do not know" as a failure to be minimised produce hallucinations; products that treat it as the correct answer to an uncovered question do not.

Mechanism 3: a check before sending

Before sending, the draft is checked: does every claim in it (a price, a time, a policy) correspond to a fact it was given? Some products automate this; approval mode does it with a person reading the draft beside its source. Either way, it catches the cases where the model, given facts, still embroidered them.

Mechanism 4: supervised weeks

Approval mode for the first weeks. A person reads every draft. The hallucinations that survive the first three mechanisms show up here, in the drafts that get edited, and each one points at a missing or ambiguous fact. After a clean week, release categories.

The hallucination that is really a stale fact

"We are open until 8pm on Saturdays" may be a hallucination, or it may be a fact you loaded in January and changed in June. Same symptom, different fix. Grounding makes this the more common kind, and the fix is a fact owner and same-day updates.

Testing a product for it

Ask for a price you did not load. Ask about a plan, a service, a location, a policy you did not load. Ask for a slot on a closed day. Ask about a staff member's schedule. Then ask about the things you did load and compare the answers to your facts word for word. A product that passes all of these can be supervised for a week and trusted; one that invents even one answer cannot.

Frequently asked questions

Can hallucinations be reduced to zero?

With grounding, an honest fallback and a check before sending, to the point where the remaining errors are stale facts rather than inventions. That is the right target: nothing said that you did not say first.

Does a better model help?

Better models hallucinate less and still hallucinate. The mechanisms above are what make the difference at a front desk, whatever the model.

What about voice?

The same mechanisms apply to spoken replies. Voice adds the need to phrase "I am not sure" naturally, which good systems do.

What Frontiva does here

Frontiva's AI agents answer only from structured facts and approved FAQs in knowledge. When nothing matches, the question is logged as a gap instead of guessed, every reply shows the source it came from, and approval mode lets a person check each draft before it sends. See what an AI receptionist should refuse to answer for the questions that should never get a factual answer at all.