When a support chatbot gives a wrong answer, the first instinct is to blame the model. But when I looked at how these failures are described in support and operations communities, a different pattern kept showing up: the model was faithfully repeating the documents it was given. The documents were the problem. Typical examples: Two FAQ pages with different refund windows (7 days vs 14 days) Two different support email addresses, both "official" A policy last reviewed two years ago ...