Why AI hallucinations happen
The US National Institute of Standards and Technology calls this risk confabulation in its Generative AI Profile (NIST AI 600-1): confidently stated but erroneous or false content. Models produce the most likely next words, and when the facts are missing, the most likely words can still be wrong.
For customer-facing assistants, the main causes are gaps in the knowledge base, outdated or conflicting sources, and instructions that reward always having an answer. The fixes are equally practical: answer only from approved sources, say when information isn’t available, list promises the assistant must never make, and send uncertain questions to a person.
Example: an invented delivery fee
A party rental company’s chatbot is asked whether delivery to a nearby town is free. Its knowledge lists delivery fees only for the home city. An unguarded bot answers “Yes, delivery is free within 30 miles,” a policy the company never had. A guarded assistant says it doesn’t have the delivery fee for that town and asks the owner to confirm. This scenario is illustrative.
How InsertChat guards against hallucinations
InsertChat’s assistant is instructed never to invent business facts, prices, policies, availability, or commitments, and to say so when information is missing or conflicting; that setting can’t be turned off. You can list promises or decisions it must not make, and by default it asks a person for help when an answer is uncertain.
On calls, it waits for a verified result before saying anything was booked. Knowledge gaps in knowledge and answers show which real questions need better sources. See the party rental setup.
