A language model on its own writes what sounds likely, based on patterns in its training text. It knows nothing about your return window, your prices or last week’s delivery delay. Grounding closes that gap: before writing, the system looks up the relevant material and writes from it.
Why it matters
Customer email is full of facts that are specific to one business and change over time. An ungrounded answer can be fluent and polite and still state the wrong price or promise a refund you do not offer. Grounding lowers that risk, and, just as important, it makes the answer checkable: if you know which page a reply relied on, you can verify it in seconds instead of rereading everything.
Grounding is not a guarantee. If the source is outdated, the answer will be too, and a system can still misread a source.
An example
Asked “do you ship to Canada?”, an ungrounded reply might guess yes. A grounded one reads your shipping page, finds Canada missing from the list, and says so, naming the page.
See also why AI answers sound right and still get the facts wrong.