A hallucination is the model failure mode of choice: instead of saying "I do not know", it produces fluent, well-structured nonsense. The cause is structural — LLMs are trained to predict likely text, not true text, so when the right answer is rare in the training data or absent from context, they fill in something plausible.
For agencies, hallucinations are the headline risk in any client-facing AI workflow. A made-up statistic in a pitch deck, a hallucinated case-study quote, or an invented competitor name can blow trust in a single sentence.
The fix is not "better models" — it is constraint. Ground the model in retrieved documents (see RAG), force it to cite sources, evaluate its outputs against known answers, and design the workflow so a human reviews anything that ends up in front of a client.