How Microsoft 365 Copilot Works

The request-to-response loop

Type "draft a status update from this week's project emails" into Copilot and four things happen in order, every time. First you enter a prompt in a Copilot experience, such as the chat or an app like Word. Second, Copilot preprocesses that prompt through a step Microsoft calls grounding: it pulls in relevant context for you and attaches it to what you typed. Third, Copilot sends this grounded prompt to a large language model (LLM), the AI that actually writes the words. Fourth, Copilot returns the generated response to the app and to you. The model is the same powerful engine whether you ask about your project or about world capitals; grounding is the step that makes the answer about your work.

Grounding is the part that personalizes the answer

Grounding is where Copilot gathers the context an answer needs. For the licensed Microsoft 365 Copilot, that context comes from your work content in Microsoft Graph^[1]: your emails, chats, documents, and meetings. Microsoft describes grounding as improving the specificity of your prompt so you get answers that are relevant and actionable for your task, and the prompt can include text from input files or other content Copilot discovers for you. Without grounding the LLM would only have general training knowledge; with it, the model writes against your actual project emails. This is the single most important idea on the page: Copilot answers from context, so an answer is only as good as the context grounding could collect.

The semantic index helps grounding find the right content

To ground well, Copilot has to find which of your thousands of files and messages are relevant. It uses semantic indexing^[2], an index built over your Microsoft Graph data that understands meaning, not just keyword matches, so a prompt about "the budget overrun" can surface the right spreadsheet even if it never uses that exact phrase. The index respects your existing security and permission boundaries while it searches. You never configure this as an end user; it runs behind grounding so that the context handed to the LLM is the most relevant content you are allowed to see.

The four-step Copilot prompt flow: prompt, grounding against Microsoft Graph, the LLM, and the returned response. Based on the Microsoft 365 Copilot architecture diagram.

What grounds your answer: web, work, and the context you control

Whether Copilot can reach your work content at all depends on which tier you are using, and that is the difference most exam questions turn on. Microsoft 365 Copilot Chat is the free, secure-chat tier and is grounded in the web only. It does not reach into your files, emails, or chats as part of the chat experience. The licensed Microsoft 365 Copilot adds work grounding: it personalizes responses with your work content in Microsoft Graph and the web together, and it requires the Microsoft 365 Copilot add-on license^[3]. So a free user asking "summarize my unread email" gets nowhere automatically, while a licensed user gets a real summary.

You can still feed work content to the free tier

Web-only does not mean work-blind. A Copilot Chat user can bring organizational content into a single chat in three ways, per Microsoft: as part of the prompt (paste it in, upload a file with the "+" button, or type "/" and pick a file from the ContextIQ menu); by using Copilot Chat side-by-side in apps like Teams and Outlook, where it is aware of content you have open; or through a pay-as-you-go agent that has access to organizational content. The difference from the licensed tier is that the free tier only sees what you hand it, one chat at a time, while the licensed tier grounds across your work content for you.

The context you choose changes the answer

Grounding always respects your permissions, so Copilot only ever uses items your account can already open; it never widens your access. Within that boundary you steer the result. The app you are in sets the working context: Copilot in Word works against the open document, Copilot in Excel against the spreadsheet. Referencing a file or person in the prompt points grounding straight at that item instead of letting it guess. Attaching or pasting content adds it to the prompt directly. Each of these narrows or sharpens what the LLM receives, which is why a vague prompt with no referenced file and a precise prompt that names the file can produce very different answers from the same question.

Web grounding serves both tiers; work-content grounding through Microsoft Graph is the licensed add-on, and in either tier you can add context yourself.

Exam-pattern recognition: why answers vary, and how to verify

Exam items at this level rarely ask you to recite architecture. They describe a business user getting a surprising result and ask why, or what to do. The signal to read for is context: what could Copilot ground against here, and was that the right context?

Common stems and the right read

"Two colleagues ask Copilot the same question and get different answers." The right answer is that Copilot grounds in the content each person can access, and permissions differ, so different context yields different answers. It is not a bug, and rewording the prompt will not align them. "A free Copilot Chat user asks Copilot to summarize their meetings and gets nothing useful." The cause is the tier: Copilot Chat is web-grounded and does not reach work content on its own, so the fix is the Microsoft 365 Copilot license (or handing it the content in the chat), not a better prompt. "Copilot in a Teams chat won't pull in a related email thread." That experience is scoped to the single chat thread; it cannot reach other chats, emails, or files, so the limitation is by design. "The answer is plausible but the user isn't sure it's right." The intended action is to check the citations: when Copilot grounds in your content it returns clickable citations to the source items, so the user opens the cited source and confirms rather than trusting the wording.

Distractors to reject

Watch for answers that blame the model's writing quality when the real issue is missing or wrong context, and for answers that claim Copilot "saw a file it shouldn't have." Copilot grounds only within your permissions and never widens access, so an answer that assumes it bypassed access controls is wrong. Likewise reject "reword the prompt" as the fix when the scenario is actually a tier limit or a permissions gap, because no rewording adds context Copilot cannot reach.

Map a surprising Copilot result to its real cause: permissions, tier, or scope, and verify any answer through its citations.

What grounds the answer: web vs work content

Aspect	Web-grounded (Copilot Chat)	Work-grounded (Microsoft 365 Copilot)
License	Free, no add-on license	Requires the Microsoft 365 Copilot add-on license
Default grounding source	The web only	Your work content in Microsoft Graph, plus the web
Reaches your files/emails/chats	Only content you hand it (upload, paste, "/" reference)	Automatically, limited to items you have permission to see
Personalized to your work	No, unless you supply the content	Yes, grounded in your emails, chats, documents, and meetings
Returns citations to your work items	Only for content you supplied	Yes, clickable citations to the grounded sources

Decision tree

Cheat sheet

Sharp facts the exam loves — give these one last read before exam day.

Cheat sheet

Sharp facts the exam loves — scan these before test day.

Copilot grounds your prompt before the model writes the answer

Copilot does not answer straight from the model's general knowledge. It first runs grounding, gathering relevant context such as your work content and adding it to your prompt, and only then sends that enriched prompt to the large language model that writes the response. Grounding is the step that makes an answer specific to your work, so a thin or generic answer usually means grounding had little context to work with, not that the model failed.

Trap Treating Copilot as if it answers purely from its training data; the personalization comes from the grounding step that pulls in your context, not from the model alone.

6 questions test this

The Copilot prompt flow is four ordered steps

A Copilot request always runs in the same order: you enter a prompt, Copilot preprocesses it with grounding against Microsoft Graph, Copilot sends the grounded prompt to the LLM, and Copilot returns the response to the app and to you. Knowing the order tells you where an answer can go wrong: the grounding step decides what context the model ever sees.

Grounding pulls your context from Microsoft Graph

For the licensed Microsoft 365 Copilot, the work context for grounding comes from Microsoft Graph, which holds your emails, chats, documents, and meetings. Grounding improves the specificity of your prompt so the answer is relevant and actionable for your task, and it can include text from files you point at as well as content Copilot discovers for you.

14 questions test this

The semantic index helps grounding find relevant content by meaning

To ground well, Copilot has to pick the right items out of all your files and messages. It uses semantic indexing over your Microsoft Graph data, which matches on meaning rather than exact keywords, so a prompt about "the budget overrun" can surface the right spreadsheet even when that phrase never appears in it. It runs automatically behind grounding and respects your permission boundaries; there is nothing for an end user to configure.

4 questions test this

Free Copilot Chat is web-grounded; the licensed Copilot adds work grounding

Microsoft 365 Copilot Chat is the free secure-chat tier and is grounded in the web only, so it does not reach your files, emails, or chats on its own. The licensed Microsoft 365 Copilot requires the add-on license and grounds in your work content in Microsoft Graph plus the web. When a task depends on your work content, the tier, not the wording of the prompt, decides whether Copilot can answer.

Trap Rewording the prompt to fix a free-tier user who gets nothing from their work content; the web-only tier cannot reach work content at all, so the fix is the license or supplying the content.

1 question tests this

A marketing analyst at Fabrikam uses Microsoft 365 Copilot chat with a work account and asks about recent industry trends. Copilot includes…

Web-only does not require a Copilot license; work grounding does

Using Microsoft 365 Copilot Chat needs no extra license because it answers from the web. Grounding answers in your work content in Microsoft Graph is the value the paid Microsoft 365 Copilot add-on unlocks. So a generic web question is fully served by the free tier, and only work-content tasks justify the license.

2 questions test this

You can still hand work content to the free Copilot Chat

Web-only does not mean work-blind. A Copilot Chat user can bring organizational content into a single chat three ways: as part of the prompt (paste it, upload with the "+" button, or type "/" to pick a file from the ContextIQ menu); by using Copilot Chat side-by-side in apps like Teams and Outlook where it is aware of content you have open; or through a pay-as-you-go agent with access to organizational content. The difference from the licensed tier is that the free tier only sees what you give it, one chat at a time.

2 questions test this

The same prompt can give different answers to different people

Because Copilot grounds in the content each person can access, the identical question can return different answers for different users when their permissions differ. This is expected behavior, not a malfunction, and rewording the prompt will not make the answers match. The variation traces to differing context, the documents and messages each account can see.

Trap Calling differing answers between two colleagues a bug; each answer is grounded in what that person can access, so different access legitimately yields different results.

2 questions test this

The active app sets Copilot's working context

Where you invoke Copilot shapes what it works against: Copilot in Word operates on the open document, Copilot in Excel on the spreadsheet, Copilot in Outlook on the mail thread. Choosing the right app is the simplest way to point Copilot at the content you mean before you even refine the prompt.

8 questions test this

Copilot returns citations so you can verify the answer

When Copilot grounds an answer in your content it returns clickable citations to the source items, so you can open the cited document or message and confirm the claim rather than trusting the wording. Checking citations is the right move whenever an answer looks plausible but you are not certain it is correct.

3 questions test this

Copilot in a Teams chat is scoped to that single thread

Copilot inside a Teams chat uses only that one chat thread as its source and cannot reach other chats, meeting transcripts, emails, or files. The scope is by design, so do not expect it to combine sources it was never given; for cross-source work you need an experience that grounds more broadly.

Trap Expecting Copilot in a Teams chat to pull in a related email or another chat; that experience reads only the current thread and cannot reference other chats or data types.

An off answer usually means missing context, not a weak model

When a Copilot answer is thin or wrong, the high-value fix is almost always to correct the context rather than the wording: confirm your tier can reach work content, reference the right file, and use the right app. The model is the same engine everywhere; what changes between a good and a poor answer is what grounding could collect.

Trap Blaming the model's writing quality for a thin answer when the real gap is that grounding had no access to the needed file or work content.

The request-to-response loop

Grounding is the part that personalizes the answer

The semantic index helps grounding find the right content

What grounds your answer: web, work, and the context you control

You can still feed work content to the free tier

The context you choose changes the answer

Exam-pattern recognition: why answers vary, and how to verify

Common stems and the right read

Distractors to reject

What grounds the answer: web vs work content

Decision tree

Cheat sheet

Also tested in

References