An ungrounded model answers from the internet-shaped blur it was trained on; a grounded one answers from your price list. The mechanism is usually retrieval — find the relevant fragment, put it in the context, instruct the model to use only that — plus the discipline to say 'not in the sources' instead of guessing.
Grounding is the difference between a chatbot that demos well and one you can point at customers: it is also the difference between a defensible answer and a fluent hallucination. Citations back to the source turn a claim into a checkable claim.
Related terms
RAG (retrieval-augmented generation)
Looking up relevant documents first and putting them in the prompt, so the model answers from your material instead of from memory.
Hallucination
A confident, fluent, entirely invented answer — a citation, a figure or an API that does not exist.
Knowledge base
The maintained, structured store of facts a system answers from — documents, prices, policies — that a chatbot grounds on so it can only say what is true.
The bench this belongs to
AI automationThe repetitive half of your week, handed to software that does not get bored. Inbox triage, follow-ups, reporting, data entry between tools that were never meant to talk.
