How do you stop an AI support agent from hallucinating and breaking customer trust?

Ground the agent in your real docs, explicitly bound what it may make claims about, label it clearly as AI, and give it a fast exit to a human. The failure cases are instructive: a court made Air Canada honour a refund policy its chatbot invented, and Cursor's support bot fabricated a login policy that triggered public cancellations. Before launch, test against hundreds of real past conversations, then monitor accuracy and hallucination rate (aim under 2%) as an ongoing operations job, not a one-time setup.

A quick orientation. The real value is below: resources worth your time, from people who've actually done it.