An answer you can trace.
Grounding is the requirement that a model's answer be traceable to material you trust. In support, that material is usually a help center article, an SOP, a contract clause, or a ticket your team already resolved.
The point is inspection. A reviewer should be able to see what the sentence was based on. If they cannot, the answer may still be right, but you cannot tell, and you cannot fix the source.
Confidence is not a citation.
Grounding is not the same as a confident tone. Models can sound sure about a policy you do not have. Confidence is a style. A citation is a check.
Grounding is also not the whole job. A cited draft that nobody writes back into the ticket has informed a person. It has not finished the record. Retrieval and the write are different steps.
Support answers get audited.
A wrong refund policy, a wrong entitlement, or a wrong change window is a customer event, not a demo glitch. Grounding is how you make that failure visible before it ships, and how you explain it after.
It is also how a team improves the source. If the cited article is wrong, the fix is the article, not a longer prompt. If no article exists, the honest result is a handoff, not a guess.
A decorative link is not grounding.
Some products attach a link the model did not use, or a passage that does not say what the reply claims. Ask to see the span, not only the title of the article.
Grounding does not set your autonomy. You can ground every draft and still require a person to send it. You can also ground a reply the product sends alone. The source and the permission to send are separate controls.