A hallucination is when an AI model states something that sounds confident but isn’t true or isn’t supported by the information it was given.
Why it matters
A wrong price, an invented policy, or a promise the business can’t keep costs more than a missed answer, because the customer acts on it. That’s why serious deployments limit what an agent can say to what it can back up.
What it looks like
A caller asks whether a product comes with a lifetime warranty, and an agent making things up says yes. A well-grounded agent checks the business’s actual policy and gives the real answer. If they don't know, they'll transfer to a live agent.
Related terms
Guardrails · Retrieval-augmented generation · Model evaluation · Large language model