Grounding runs on every prompt, before the model writes a word. Copilot checks Microsoft Graph for content relevant to your question, using the semantic index to match on meaning. What it finds is appended to your prompt, and only that combined package goes to the model. The model never reads your mailbox directly, and grounding can only surface what you personally have permission to see.
That is a real head start: no hunting for a file to paste in first. The catch is that grounding runs fresh each time, so the same vague question can land differently depending on what ranks highest that day. Naming the file, mailbox or time range turns a guess into a lookup. Copilot can also ground answers in current web results; that is web grounding.