What it is
The model predicts plausible text, not guaranteed truth. When it does not know something, it sometimes "fills in" with invented details that sound believable.
Why it matters to an agency
A hallucination in a message to a client can be expensive. That is why human review at critical points and relying on real sources both matter.
How to reduce it
- Use RAG to anchor answers in your documents.
- Ask it to cite sources and to say "I don't know" when it has no data.
- Check anything sensitive before sending (prices, promises, legal details).
By the numbers
The risk concentrates in 3 kinds of data: figures, verbatim quotes and references (URLs, legal cases, papers). The practical defense has 2 layers: asking the model to cite its source or answer "I don't know", and human verification of any fact that goes out to a client. Unlike a calculation error, a hallucination sounds sure of itself: you cannot detect it by tone, only by checking against the source. Operating rule: 1 human check for every hard fact, and 0 figures without a source in messages to clients.