What is grounding?
Grounding means tying an AI answer to a source you trust, such as a file, a database row, or a web page with a citation, instead of letting the model answer from memory alone.
In plain English
Microsoft uses 'grounded in your business data' for Copilot with a work license. Perplexity's Agent API docs say answers can be web-grounded with citations. The pattern is the same: the answer should point at something you can open. Pick the source before you pick the adjective. 'Be accurate' is a wish. 'Use only the note dated today and quote the line you used' is a rule you can check. If the note does not contain the answer, the honest draft says so. Open the source once before you forward the draft. A link you did not click is a rumor with punctuation. If two sources disagree, the draft should show both lines and stop, not average them into a third price you never charged. Averaging is how a fictional fee is born. Refuse a blended number.
How an owner would use it
If a setting says the assistant can use the open web, decide whether that is allowed for this job. For a client email, ground it in the client folder. For a public fact, open the citation.
Common mistakes
Grounding in the open web is a different choice from grounding in your folder. Owners turn on web search for a client email and get a fact about a different company with a similar name. For work about your customer, the source is the customer file. A citation you do not open is decoration. If the tool cannot show a source, do not call the answer grounded.
A worked example
A one-person PR shop
The brief is the only source
This is an illustration, not a customer result. A client update is grounded in the shared note: what ran, what is next, what is blocked. The model is told not to add impressions or quotes that are not on the note. The owner checks the three lines against the note before sending.
Related terms
Prompts and guides
Questions owners ask
What is grounding in AI?
Grounding is connecting the model's answer to sources you can check, so it is not speaking only from training memory.
How do I ground an assistant?
Point it at the folder or record for that job, ask it to name the source, and do not send the draft if the source is missing.
Is grounding the same as RAG?
RAG is one way to ground an answer: retrieve your documents, then write. Grounding is the broader rule that the claim should point at something you can open, including a row in your own system or a page you trust.
Product guides are on AI apps for business.