Grounding
Grounding means an AI agent only states facts it can trace to real content or live system data, rather than generating from memory. It is the single most important defense against hallucination in customer support.
Retrieve or refuse
A grounded agent retrieves the relevant material before it answers. If nothing solid comes back, it does not improvise, it caps its confidence and escalates. This "retrieve or refuse" rule is what stops the classic failure where the agent invents a plausible-sounding policy because it had no real source. Retrieval strength is the strongest single predictor of a correct answer.
Citing the source
Grounding is only trustworthy if you can check it. Resolve's agent cites the content it used, so a human can verify the answer and so the agent is not "confidently wrong." When a ticket resolves from content that did not exist yet, the resolution can be drafted into a new article, but only a human-approved draft becomes groundable, so a hallucinated answer never silently becomes a cited source.
Frequently asked
What does it mean for an AI answer to be grounded?
It means every fact in the answer was retrieved from your real content or live data, not generated from the model's memory. If the agent cannot ground a claim, it refuses and escalates rather than guessing.
How is grounding different from a normal chatbot search?
A search widget shows articles and hopes the customer reads them. A grounded agent retrieves the facts, drafts a specific answer, verifies it, and cites the source, or refuses if it cannot.