AI Explained
Microsoft Copilot, product by product
Microsoft's family of AI assistants: how one answers from your work data, how to start, and which one is for which job.
The map
One question, from the app to the answer
Copilot is not a model. It is the service around one: it looks up what you are allowed to see, hands that to a model with your question, and checks what comes back. Play the diagram. The steps are the ones Microsoft documents [1] [2].
Press Play to walk through it, or Listen to have it read to you.
Read all the steps
- You ask in plain words inside an app such as Word, Teams or the Copilot chat.
- Who you are comes first. Copilot can only reach content you already have permission to open.
- Grounding: Copilot looks in Microsoft Graph, your mail, files, chats and meetings, for what is relevant. This is retrieval, the R in RAG.
- If web search is allowed, Copilot also sends a short generated query to Bing, not your whole prompt.
- Your question and what was found go to a large language model, which writes the answer.
- Safety checks run, and the prompt and the answer are logged so they can be audited later.
- The answer comes back into the app you asked from.
- Connectors bring in data from systems outside Microsoft 365. Agents package instructions, knowledge and actions for one job on top of this same engine.
Start using it
From sign-in to a repeatable agent
Seven steps. The first two matter most: where to open it, and what it can see.
Open it
Sign in with your work account at copilot.cloud.microsoft, in the Microsoft Copilot app, in Edge, or inside Outlook and Teams [3].
Know what it can see
The basic tier answers from the web and from what you upload or have open. Only the paid add-on searches your organization's mail, files and meetings by itself [3]. Check which one you have before you trust an answer about your own data.
Ask with four parts
Say the goal, the context, the source to use and the format you want. For example: summarize last week's project emails from the finance team as five bullets for my manager.
Pick the mode
Leave the model on Auto, or choose a quick response for routine questions and deeper thinking for hard ones [3].
Read before you forward
Open the files an answer drew on before you send it to anyone. An answer is only as good as what was found.
Make it repeatable
When the same job keeps coming back, turn it into a declarative agent with your instructions, knowledge and actions in Agent Builder, or in Copilot Studio when it needs more [4] [5].
If you write code
In GitHub Copilot, start with inline suggestions and chat, move to agent mode for a task you watch, and hand a well-described task to the cloud agent when you want a pull request back [6].
The family
Product by product
As Microsoft's pages described them on the date in the sources. Microsoft renames and regroups these products often: follow the links for the current names and plans.
Microsoft Copilot Chat
Chat for work, signed in with a work account. The basic tier answers from the web and from what you upload or have open, and does not search your organization's data by itself [3]. Prompts and responses are not used to train the foundation models [2].
Microsoft 365 Copilot
The paid add-on, which Microsoft's page calls Premium. It grounds answers automatically in your organization's data through Microsoft Graph and Work IQ, and brings the full Copilot experience into Word, Excel, PowerPoint and OneNote [3].
Work IQ
Microsoft describes it as a workplace intelligence layer that lets agents access and reason over an organization's data, context and tools [7].
Copilot connectors
They bring in data from outside Microsoft 365. A synced connector indexes the data into Microsoft Graph. A federated connector uses the Model Context Protocol to fetch it at question time without indexing it [8].
Agents
A declarative agent is Copilot configured for one scenario with your instructions, knowledge and actions. It runs on Copilot's orchestrator and models and needs no hosting. A custom engine agent brings its own orchestrator and models, needs hosting, and leaves compliance and security to you [4].
Copilot Studio
A graphical, low-code studio for building and managing agents and workflows. An agent follows your instructions, draws on the knowledge you connect and uses tools to act [5]. Those tools can come from an MCP server [9].
GitHub Copilot
For developers: inline suggestions and chat, agent mode that proposes file changes and terminal commands for your approval, a cloud agent that works on a branch and opens a pull request, and a command-line version. It takes custom instructions, agent skills and MCP servers [6].
Microsoft Foundry
The developer platform on Azure, formerly Azure AI Foundry. It offers a catalog of more than 10,000 models from Microsoft, OpenAI, Anthropic, Meta and others, a service for building and hosting agents, and identity, access control, content filters and network isolation around them [10].
Microsoft Security Copilot
For security teams: a generative AI assistant for incident response, threat hunting and posture management, in its own portal or inside Defender, Sentinel, Intune, Entra and Purview. Its agents act only within configured permissions, and Microsoft says the models are not trained on customer data [11].
The models behind it
By default Copilot chooses a model for each prompt, and a person can ask for a quick answer or deeper thinking. Microsoft's page names Anthropic's models as available in some licensed experiences, not to everyone by default [3].
This center is independent and is not affiliated with Microsoft or GitHub. The product names belong to their owners.
Which one?
Pick the right product for the job
Choose what you are trying to do. The guidance is this center's own, built from the descriptions above.
Before you switch it on
Four things that go wrong
It finds what people could always open
Copilot can only reach content the person already has access to [1]. That is the protection and also the problem: a file shared too widely years ago is now one question away.
Hidden instructions in what it reads
A document, an email or a web page can carry text written to steer the model. Treat everything an assistant reads as untrusted, and test for it.
An agent that can act
Reading is one risk, acting is another. Give an agent its own identity and the least access the job needs, and keep a person approving anything that cannot be undone.
A confident wrong answer
Microsoft's own page for Security Copilot says responses can be inaccurate, incomplete or outdated, and that generated code should not go to production without validation, testing and review [11].
Sources
Where this comes from
Read 2026-11-15, unless a source gives its own date. A number in brackets points to the page a statement was read on. Sentences with no number, the small demonstrations and the tools are this center's own explanations.
- Microsoft Learn: Service components for Copilot (ASD Blueprint for Secure Cloud) Read 2026-11-16.
- Microsoft Learn: Microsoft Copilot Chat, privacy and protections Read 2026-11-16.
- Microsoft Learn: What is Microsoft Copilot? Read 2026-11-16.
- Microsoft Learn: Agents for Microsoft 365 Copilot Read 2026-11-16.
- Microsoft Learn: Copilot Studio overview Read 2026-11-16.
- GitHub Docs: GitHub Copilot features Read 2026-11-16.
- Microsoft Learn: Work IQ overview Read 2026-11-16.
- Microsoft Learn: Copilot connectors overview Read 2026-11-16.
- Microsoft Learn: Extend your agent with Model Context Protocol (Copilot Studio) Read 2026-11-16.
- Microsoft Learn: What is Microsoft Foundry? Read 2026-11-16.
- Microsoft Learn: Responsible AI FAQ for Microsoft Security Copilot Read 2026-11-16.