Why Microsoft 365 Copilot Is Safer With Your Business Data

Why Microsoft 365 Copilot Is Safer With Your Business Data

    • Home
    • Why Microsoft 365 Copilot Is Safer With Your Business Data
Copilot Data Governance

Why Microsoft 365 Copilot is safer with your business data

With Copilot, your prompts, responses and files stay under your Microsoft contract, are never used to train the AI, and are only shown to people who already have permission to see them. Here's what's protected, what isn't, and what to check before rollout.

See how your data flows Book a Copilot readiness chat
0Copilot prompts or responses used to train foundation AI models
4routes where data can step outside the boundary, each with an admin setting
7simple checks to make before you roll Copilot out to your team
The real risk

Personal ChatGPT vs Microsoft 365 Copilot

The biggest risk isn't Copilot. It's staff pasting company data into personal AI accounts.

Consumer ChatGPT (Free, Plus, Pro)

A public service

Anything a user pastes in goes to OpenAI under OpenAI's consumer terms, not your company's contract.

  • Training is on by default. Each user has to turn it off themselves, and a thumbs up or down can still send that conversation for training.
  • No company permissions. Once a document is pasted in, your access controls no longer apply.
  • No admin oversight. Personal accounts sit outside your Microsoft 365 audit, retention and data loss prevention tools.
Microsoft 365 Copilot

Inside your tenant

Copilot runs inside your Microsoft 365 environment and only works with data the signed-in user can already open.

  • Never used for training. Prompts, responses and Microsoft Graph data aren't used to train foundation models.
  • Respects your permissions. Users only see what they already have access to.
  • Covered by your contract. Microsoft's Product Terms and Data Protection Addendum apply.

OpenAI's paid business plans (ChatGPT Business, Enterprise and Edu) and its API are different: OpenAI states it does not train on those by default.

How Copilot works

Four steps, every time you ask

Microsoft describes each Copilot request in the same four stages.

You type a prompt

In a Microsoft 365 app such as Word, Outlook or Teams.

Copilot grounds it

It searches Microsoft Graph (email, files, chats, meetings) using your own permissions.

The AI model answers

The grounded prompt goes to a large language model, and the answer is checked against Responsible AI filters.

It comes back to you

The response returns to the app and is stored in your Copilot history inside Microsoft 365.

See it in motion

Your data stays inside Microsoft's boundary, unless a route out is on

Follow a prompt as it moves through Copilot. Blue stays protected; amber steps outside the boundary.

Microsoft 365 service boundary · Microsoft DPA applies Your user types a prompt in Word, Outlook, Teams Copilot orchestration grounding, permission checks, Responsible AI filters AI models (Azure) Hosted by Microsoft, not trained on your data Microsoft Graph email, files, chats, meetings only what this user can see AI subprocessors OpenAI: on by default Anthropic: off in UK run by third parties web query agent request Outside the boundary Bing web search short query, IDs removed DPA does not apply Third-party agents query + relevant data vendor's own terms apply
Swipe to see the full diagram
Protected by Microsoft's DPA Third-party AI, DPA still applies Leaves the boundary, admin controlled

Everything inside the blue boundary is covered by Microsoft's Data Protection Addendum. The amber routes sit outside it, and your admins decide whether they're allowed.

Protected by default

What keeps your data safe in Copilot

As stated in Microsoft's own documentation, updated August and September 2026.

No training on your data

Prompts, responses and Microsoft Graph data aren't used to train foundation models.

Permission trimming

Copilot only surfaces data a user already has at least view access to.

Contract cover

Microsoft's Product Terms and DPA apply, including to OpenAI and Anthropic as Microsoft subprocessors.

Sensitivity labels

Content protected with Microsoft Purview labels is only used within the rights the user holds.

No human review

Copilot has opted out of the abuse monitoring that includes human review of content.

Data residency

Covered by Microsoft's data residency commitments since 1 March 2024, plus the EU Data Boundary for EU users.

Prompts stay in Microsoft 365

Prompts and responses are stored within Microsoft 365 and don't leave the boundary without your direction.

The catch

Copilot is very good at finding everything people are allowed to see. If a salary spreadsheet is shared with "Everyone", Copilot will surface it. Fix oversharing first.

Worth knowing

Where data can leave the protected boundary

Four routes take data partly outside Microsoft's normal protections. Each one has an admin setting.

Route
What leaves
What changes
Default
Admin control
Web search (Bing)
What leavesA short search query built from the prompt, with user and tenant IDs removed. Not whole files or prompts.
What changesThe DPA, EU Data Boundary and HIPAA don't apply to the query
DefaultOn
Admin control"Allow web search in Copilot" policy
Third-party agents and plugins
What leavesA query based on the prompt, chat history and data the user can access
What changesThe agent's own privacy terms apply
DefaultDepends on what's approved
Admin controlMicrosoft 365 admin center > Integrated apps
OpenAI as a Microsoft subprocessor
What leavesData needed to run GPT models, processed by OpenAI on Microsoft's behalf
What changesDPA still applies. Inside the EU Data Boundary, but outside in-country processing commitments
DefaultOn since 24 July 2026
Admin controlCopilot > Settings > AI providers operating as Microsoft subprocessors
Anthropic (Claude) as a Microsoft subprocessor
What leavesData needed to run Claude models
What changesDPA still applies. Outside the EU Data Boundary and in-country processing
DefaultOff for UK and EU
Admin controlSame setting as above
Also check flex routing. When it's switched on, Copilot can process requests outside your region during peak load. You'll find it under Copilot > Settings > Flex routing during peak load periods.
Before you roll out Copilot

Your 7-point readiness checklist

Tick them off as you go. If any feel unclear, we're happy to work through them with you.

0 of 7 complete

Sources

  1. Data, privacy, and security for Microsoft Copilot (Microsoft Learn, updated 18 Aug 2026)
  2. How does Microsoft Copilot work? (Microsoft Learn, updated 14 Sep 2026)
  3. Understanding AI functionality and models in Microsoft Online Services (Microsoft Learn, updated 10 Sep 2026)
  4. OpenAI as a subprocessor in Microsoft Online Services (Microsoft Learn, updated 18 Aug 2026)
  5. Anthropic models in Microsoft Online Services (Microsoft Learn)
  6. Data, privacy, and security for web queries in Microsoft Copilot (Microsoft Learn, updated 18 Aug 2026)
  7. How your data is used to improve model performance (OpenAI Help Centre)
  8. OpenAI will be available by default in Microsoft 365 (Wortell, on the July 2026 change and flex routing)

Want Copilot switched on safely?

We'll review your sharing, labels and Copilot settings, then tell you in plain English what's safe to switch on. Book a Copilot readiness chat with our Southampton team.

Book a Consultation

Pick a time that suits you to talk through your Copilot rollout.

  • Oversharing and permissions review
  • Web search, agent and AI provider settings checked
  • Clear, practical next steps for your team
View our live calendar