Hallucination is the catch-all term for when a language model produces fluent, confident output that is factually wrong. The hallucinated content can be subtle, a slightly-wrong policy detail, a wrong date, or egregious, a phone number that does not exist, a refund policy the company does not have.
Production-grade voice agents reduce hallucination through three layers: grounding (the agent answers only from approved knowledge sources, not from model priors), tool gating (high-stakes outputs come from deterministic tool results, not generated text), and a hallucination grader (a separate model that flags suspicious outputs before the customer hears them).
A vendor's hallucination rate matters most where the agent is asserting facts, policies, prices, availability. It matters less where the agent is summarizing what the caller said. Asking a vendor for their hallucination rate on the fact-assertion slice, with the methodology, is the right diligence question.
Vorel ships a hallucination grader that runs on every fact-assertion turn, with per-vertical calibration. Flagged turns route to a human before the response goes out.