AI and agents
Hallucination
A hallucination is output from a language model that reads as confident and plausible but is false or unsupported by the information it was given.
How it works in practice
Language models generate likely text. When the right answer is not in front of them, they can produce a likely-sounding one anyway: an invented policy clause, a made-up order number, a citation that does not exist. It happens more when the question goes beyond the supplied documents or asks for exact figures.
You reduce it rather than eliminate it. Give the model the source material (RAG), ask it to quote or cite what it used, tell it to say when it does not know, validate structured output against your own data, and keep a person on anything that leaves the building. Evals tell you how often it still happens.
A worked example
A quoting assistant drafts prices for a building supplies firm. Prices come only from the price book through a lookup tool, never from the model’s text, and the draft is rejected automatically if any line has no matching product code.
Examples in this glossary describe typical builds for companies of 1 to 1,000 people. They are not named client results.















