Choosing the right Copilot — a practical guide for everyday users and a technical deep dive for IT pros and power users.
Everyone reads this. Which Copilot you have, where it lives, and which one to open.
Microsoft ships several things called Copilot. Before choosing between them, it is worth knowing that the first question is not which one should I open — it is which one am I entitled to. That single fact determines whether the rest of this guide applies to you at all.
| If this is you… | You have… | And it can see… |
|---|---|---|
| You signed in with a personal account | Consumer Copilot | The public web. None of your work data. |
| You signed in for work, no Copilot add-on | Copilot Chat | The web, plus files you paste or upload into the chat yourself. |
| Your organization bought you the add-on | Microsoft Copilot (licensed) | Your email, files, meetings and chats — everything your account can already open. |
This is the question the rest of the guide depends on, and the one most often answered wrongly. The three products share a name, share an icon family, and increasingly share a front door. They do not share what they can see.
copilot.microsoft.com.m365.cloud.microsoft.Until recently this guide told you to look for the "365" in the name. That test no longer works. Microsoft has been dropping "365" from the branding, and its own documentation now uses Microsoft Copilot for both the licensed work product and the free consumer app. The name alone will not tell you which one you are in.
Two things still will: the account you signed in with, and the address it lives at.
The consumer assistant. A personal account signs in here. It is not connected to your organization's data and never sees your email, files or meetings. Fine for personal tasks — but do not paste work content into it.
The work assistant. Your Microsoft Entra work or school account signs in here, and this is the surface described throughout this guide. Copilot Chat lives at m365copilot.com and redirects into the same experience — what differs is your license, not the address.
In August 2026 Microsoft merged what had been two separate chat experiences — web chat and work chat — into a single surface, with one toggle governing where answers come from. If you learned Copilot before that change, this is the most important thing to re-learn.
This control only exists where work grounding is possible, which makes it the quickest way to confirm which Copilot you are in — see Section 2.
Copilot is not one place. Once entitlement and grounding are settled, the remaining question is where to invoke it — and the honest answer is that the surface matters less than it once did, because they share an orchestrator and a permission model. What still differs is how much context arrives for free.
| Surface | Reach for it when | Needs the license? |
|---|---|---|
| Copilot in Teams | Work is happening right now — a live meeting, an active thread. Context arrives free. | Yes |
| The Copilot app (web or desktop) | You are orienting rather than executing — exploring, planning, searching broadly. | No — Chat tier works |
| Copilot in Word, Excel, PowerPoint, Outlook | The artifact is the point. Drafting, editing and analyzing in place. | Yes, for full editing |
| Copilot Search | You want to find and scan source material yourself rather than read a synthesis. | Yes |
| Copilot Notebooks | A project you return to repeatedly. Pin the sources once; stop re-explaining context. | Yes |
| Copilot Pages | The output needs to be worked on with other people. | No |
| Agents and the Agent Store | The same shaped task recurs and you want it packaged. | Partly — see below |
Agents built only on instructions and public websites are free; agents that reach tenant data are metered and off by default for Copilot Chat users. The Copilot Agent Knowledge Reference covers how to build one and what each knowledge source can actually hold.
The reason there are two Copilots isn't about features — both can answer questions, write, summarize, and brainstorm. The difference is what you're trying to do when you open them. Microsoft designed each one around a different mental state.
Think of this as a smart teammate sitting next to you while a meeting is happening or a conversation is live. It already knows the chat thread, transcript, and attached files — and skips straight to helping you act.
Best for: Summarizing meetings · Drafting replies · Generating action items · Catching up on threads
Think of this as a quiet research room where nothing is assumed. You can search broadly across your email, files, meetings, and chats without already knowing exactly what you're looking for.
Best for: Finding documents · Researching across projects · Planning · Exploring before you dive in
A useful mental test: do you already know what task you're in? If yes, open Teams. If you're still figuring out what needs to happen, open the Copilot app.
Meeting Copilot is the most common source of "it worked last week and now it doesn't." The behavior is not random. It is governed by a setting the meeting organiser chose, often without realising what it decided.
| Organiser's Copilot setting | During the meeting | After it ends |
|---|---|---|
| On, with transcript saved | Works | Works — Recap and chat |
| Only during the meeting | Works | Gone entirely |
| Off | Unavailable | Unavailable |
This is separate from whether Copilot can see the surrounding chat, files and thread, which it can regardless. The transcript setting governs spoken content only.
Not sure which to open? Match your situation to one of these common cases.
You joined a 45-minute Teams meeting 20 minutes late
Ask "Catch me up" — if transcription is on, Copilot summarizes what you missed in seconds. If the organiser chose "only during the meeting," this works live and disappears afterwards. See Section 7.
You need to find a proposal from last quarter
Search or chat across all your files and emails without knowing exactly where it was saved.
You want to draft a reply to a long channel thread
Copilot reads the thread and drafts a reply in context — no copy-pasting required.
You're preparing for a client you haven't spoken to in months
Ask Copilot to surface recent emails, shared files, and meeting notes involving that client across all of M365.
You want action items from yesterday's standup
Copilot pulls the saved transcript and generates a structured action item list. This only works if the meeting was transcribed or recorded — otherwise there is nothing left to read.
You don't know what you need — you just know there's something to figure out
Start here. The app's broader, exploratory design is built for open-ended starting points.
Keep this handy for your first few weeks. The choice becomes instinctive quickly.
| Signal | Teams Copilot | M365 App |
|---|---|---|
| You're inside a Teams chat or channel | ✓ | |
| You're in or just finished a meeting | ✓ | |
| You need to search across all your files | ✓ | |
| You want to catch up on a project | ✓ | |
| You're drafting a response to a message | ✓ | |
| You don't know what to search for yet | ✓ | |
| You need action items from a call | ✓ | |
| You need to research across many sources | ✓ | |
| Speed is the priority | ✓ | |
| Exploration is the priority | ✓ |
Optional. Read this when an answer in Part One surprised you and you want to know why.
For a long time the honest answer here was "same brain, different assumptions" — one model, one orchestrator, one Graph, and only the pre-loaded context differing between surfaces. That is no longer accurate. The orchestration layer and the permission model are still shared. The model is not.
Since 2026 the underlying model is partly a user-visible choice. Anthropic models can be selected alongside OpenAI ones for some tasks, and Word's default moved to Sonnet 5 with a higher default reasoning level in August 2026. Two surfaces can now answer the same prompt with different models at different reasoning effort — so "it gave me a better answer in Word" is sometimes literally true rather than a perception.
What differs is the pre-prompt context — the information Copilot receives before it ever reads your message. This pre-loaded context shapes everything: how broadly Copilot searches, how quickly it synthesizes, and what it assumes you're trying to accomplish.
Every Copilot interaction begins with a system prompt — a behind-the-scenes instruction set that tells the model who you are, what you have access to, and what it should prioritize. In Teams, that system prompt is automatically populated with rich contextual data from your current session. In the M365 app, it is not.
When you open Copilot inside a Teams chat, channel, or meeting, the orchestration layer pre-loads the following before processing your prompt:
This dense pre-loading lets Copilot skip the discovery phase entirely. It already knows what's relevant. Retrieval is narrow and fast — optimized for synthesis speed over breadth.
When you open the Copilot app, no session-specific context is pre-loaded. The system prompt gives Copilot access to your full M365 tenant data but does not narrow the scope. Copilot must infer relevance from your prompt alone.
Microsoft 365 Copilot uses semantic search over the Microsoft Graph — an index of your organization's emails, chats, files, calendar events, and SharePoint content. The critical architectural difference between the two surfaces is how aggressively retrieval is scoped before your prompt is processed.
| Retrieval Dimension | Copilot in Teams | Copilot App |
|---|---|---|
| Search scope | Narrow — current thread / meeting | Broad — full Microsoft Graph |
| Source diversity | Low — current context prioritized | High — mail, files, chats, calendar |
| Pre-filtering | Aggressive — time window + thread | Minimal — relies on your prompt |
| Synthesis timing | Immediate | After exploration (on request) |
| Result surface | Composed summary or draft | Search results + optional summary |
| Latency profile | Lower (smaller retrieval surface) | Higher (broader search) |
| Best when you know… | Exactly what task you're in | Nothing yet — still discovering |
Microsoft Graph is the unified API powering Copilot's access to your organizational data — indexing Exchange, SharePoint, OneDrive, Teams, and other M365 services. Both Copilot surfaces query the same Graph with identical permissions. Security and compliance policies apply equally.
The difference is how the query is scoped: Teams Copilot appends thread-specific filters automatically; the Copilot app sends broader semantic queries and lets results determine scope.
This part of the guide describes internal behavior, and honesty about the boundary is what makes the rest of it trustworthy. Some of what follows is documented by Microsoft. Some is careful inference from observed behavior. They deserve to be told apart.
Each Copilot surface is purpose-built for a different cognitive mode: doing (executing a known task) or thinking (figuring out what to do next).
| Design Dimension | Copilot in Teams | Copilot App |
|---|---|---|
| Assumed cognitive state | Execution | Orientation / Discovery |
| Assumed urgency | High | Low to moderate |
| Retrieval strategy | Narrow and fast | Broad and thorough |
| Synthesis timing | Immediate | On demand |
| UI metaphor | Teammate in the room | Quiet research library |
| Primary failure mode | Misses outside-context info | Can feel slow for urgent tasks |
| Primary UX signal | Speed and completeness | Breadth and control |
Every product behavior, licensing gate and limit in this guide traces to one of the pages below, each retrieved and verified on 27 August 2026. Microsoft changes this surface frequently — re-verify before repeating any of it to a customer.