Hallucination

Language models generate plausible text; when they lack the facts they may fill the gap with something that sounds right. In a demo this looks like a quirk. On a phone with a customer it becomes a promise your business did not make. Hallucinations are reduced, not eliminated, by grounding the model in retrieved documents, constraining what it may answer and telling it to escalate when unsure.

For a growing business the practical defense is design and review. Decide which topics the agent may answer from approved sources, which it must hand to a person, and which it must never discuss (medical, legal or financial advice, for instance). Then read transcripts regularly, because the first weeks reveal the questions you did not anticipate.

Ask a vendor how the system behaves when it does not know, whether answers are restricted to approved sources, how they measure accuracy on your content, how corrections are made and how fast they take effect, and whether you can review transcripts. A vendor who claims zero hallucinations has not tested enough.