Data Protection and Governance for Microsoft 365 and Copilot · Topic 3 of 6
How Copilot Accesses and Protects Data
AB-900: how Microsoft 365 Copilot uses Microsoft Graph and permissions, which controls protect data, and the Microsoft responsible AI principles.
- 5 min
- Includes a diagram
- Copilot only uses content the signed in user can access.
- Microsoft Graph and the semantic index supply the organizational context.
- Prompts and organizational data are not used to train the foundation models.
What it is
Microsoft 365 Copilot answers questions using your organization's own work content, such as email, chats, meetings and files. It does not open a back door to that content. It can only use what the signed in user is already allowed to see. Microsoft documentation now also uses the shorter name Microsoft Copilot, but the behavior described here is the same.
The key phrase for the exam is grounding: Copilot adds relevant organizational data to your prompt before the model writes a response.
How it works
How Copilot accesses data
Copilot is an orchestration of three parts: a large language model, your content in Microsoft Graph, and the Microsoft 365 apps you work in. When you ask a question, Copilot retrieves relevant content and combines it with your working context, such as the meeting you are in. The model then writes a response.
Prompts, responses and data accessed through Microsoft Graph are not used to train the foundation models. Data stays within the Microsoft 365 service boundary and is protected with the same commitments as your other Microsoft 365 content.
How Microsoft Graph influences responses
Microsoft Graph is the layer that connects your Microsoft 365 data: mail, calendar, chats, files, meetings and people. Copilot uses it to find content and relationships. The semantic index helps find content by meaning, not only by keywords, and it honors the same user access boundary.
| Graph effect | Result |
|---|---|
| More relevant content in Graph | Richer, more accurate answers |
| Content the user cannot access | Never used in the answer for that user |
| Graph connectors | Bring in external data, shown only to users allowed to see it |
| Stale or messy content | Can appear in answers, so clean up matters |
Permissions and other controls
The first control is the existing permission model. If a user cannot open a SharePoint file, Copilot cannot use it for that user. This also means oversharing becomes visible. A file shared too widely before Copilot can now surface in answers.
| Layer | Control | Effect on Copilot |
|---|---|---|
| Microsoft 365 | Site and file permissions, restricted access control, restricted content discovery | Limit what Copilot and search can reach |
| Microsoft Purview | Sensitivity labels with encryption, DLP for Copilot, retention, auditing, DSPM for AI | Restrict, record and govern use |
| Microsoft Defender | Detects risky AI app use and threats, and can block risky apps through Defender for Cloud Apps | Protects against unsafe tools and attacks |
| Copilot service | Filters for harmful content and defenses against prompt injection | Reduce unsafe responses |
Administrators can use DLP for the Microsoft 365 Copilot location to keep Copilot from summarizing items with chosen sensitivity labels. Interactions can be audited, searched with eDiscovery and retained.
Responsible AI principles
Microsoft's responsible AI principles are fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability. They guide how Copilot is built and used.
| Principle | Meaning in plain words |
|---|---|
| Fairness | Treat people equitably |
| Reliability and safety | Work as intended and avoid harm |
| Privacy and security | Protect data and respect privacy |
| Inclusiveness | Work for people with different abilities |
| Transparency | Be understandable, such as citing sources |
| Accountability | People remain responsible for AI systems |
Copilot reflects these by showing citations, warning that responses can be wrong, and expecting the user to review the output.
A worked example
A new hire asks Copilot for the salary review plan. Copilot finds no answer, because the file sits in a site only HR can open. Later an HR file is found to be shared with Everyone except external users. Copilot now returns it to anyone who asks. The fix is not in Copilot. The administrator tightens permissions, uses a label with a DLP rule for Copilot, or uses SharePoint restricted content discovery.
How the exam asks about this
- "A user cannot see a file. Can Copilot show it to that user?" No, Copilot respects the user's permissions.
- "Which service connects emails, files and chats so Copilot can use them?" Microsoft Graph.
- "Which responsible AI principle covers citing sources and explaining limits?" Transparency.
What to remember for the exam
- Copilot only uses content the signed in user can access.
- Microsoft Graph and the semantic index supply the organizational context.
- Prompts and organizational data are not used to train the foundation models.
- Purview, Defender and Microsoft 365 permissions all add protection.
- The six responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability.
Copilot does not create new access. It exposes existing oversharing faster. "Copilot ignored permissions" is wrong in any scenario.
Common mistakes
Assuming Copilot can read everything in the tenant. It is limited to each user's own access.
Thinking data is used to train the model. Organizational data and prompts are not used to train the foundation models.
Blaming Copilot for oversharing. The cause is permissions that were too broad. Fix the permissions and labels.
Check yourself
Why does Copilot never show a user a file they cannot open?
It honors the user's existing Microsoft 365 permissions.
What does Microsoft Graph provide to Copilot?
The organizational content and relationships, such as mail, files, chats and meetings, used to ground responses.
Name the six responsible AI principles.
Fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability.
Topic 3 of 6 in Data Protection and Governance for Microsoft 365 and Copilot
Test yourself on Data Protection and Governance for Microsoft 365 and Copilot
The full Microsoft AB-900 question bank has 499 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Microsoft. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.