1.2. What "Grounding" Means and Why Context Is Everything
💡 First Principle: A model's output is only as reliable as the information it was given to work with. "Grounding" is the practice of anchoring AI responses to specific, verifiable sources — and it is the single most important mechanism for improving Copilot accuracy.
Think of grounding like the difference between asking a knowledgeable colleague a question in the hallway ("Hey, what was the Q3 revenue?") versus asking them while they are looking at the actual Q3 report. In the hallway, they are working from memory — and memory can be wrong. With the report open, they are working from a source you can both verify.
When you provide Copilot with a document, a file, or web data as part of your prompt, that source becomes the grounding context. Copilot prioritizes the grounding source over its trained knowledge, which dramatically reduces fabrications for topics covered in that source.
Grounding sources in Microsoft 365 Copilot include:
| Source Type | Examples | How to Use |
|---|---|---|
| Files | Word documents, Excel spreadsheets, PDFs | Attach to prompt or use @mention |
| Emails and calendar | Outlook history, meeting invites | Referenced via Microsoft Graph |
| Teams conversations | Channel threads, chat history | Copilot in Teams can reference meeting content |
| SharePoint and OneDrive | Team sites, personal files | Semantic index makes these searchable |
| Web data | Current news, external pages | Bing integration (when enabled) |
The Microsoft 365 semantic index is what makes grounding powerful at scale. Instead of searching files by filename or keyword, the semantic index understands the meaning of your content — so when you ask Copilot "What did we agree on in last week's vendor meeting?" it can surface relevant content from emails, calendar, and Teams transcripts without you specifying exact search terms.
⚠️ Exam Trap: "Grounding" and "training" are different things. Grounding happens at inference time — when you run a prompt — and is specific to your query. Training happened months or years ago when the model was built and cannot be changed by users. You cannot "train" Microsoft 365 Copilot on your company data; you can only ground its responses by providing that data in context.
Reflection Question: A colleague complains that Copilot gave them incorrect financial figures in a report summary. What is the most likely explanation, and what should they do differently next time?
