Guide AI Governance Microsoft 365

Microsoft Copilot: how to protect confidential data and fix permissions before rollout

Copilot works within the permissions your users already have. If those permissions have built up for years without review, confidential information becomes much easier to find. This guide covers what to check and fix before and during a Microsoft 365 Copilot rollout.

RD
Ryland Deakin
Lead Consultant, Cyvra
17 July 2026
10 min read
Updated 10 November 2026

What Copilot can reach, and why permissions decide it

Microsoft offers more than one Copilot, and they carry different risks. Copilot Chat is included with eligible Microsoft 365 business plans. By default it answers from the web and from content the user supplies, such as an uploaded or open file. Licensed Microsoft 365 Copilot is grounded in work data through Microsoft Graph: email, chats, meetings, and files in SharePoint, OneDrive and Teams. Agents and connectors can add further data sources and actions on top of either.

Licensed Copilot has no access of its own. It works with the signed-in user's permissions and does not bypass access controls. What changes is how easy content becomes to find. If someone in finance still has access to an HR folder that should have been removed years ago, Copilot could surface or summarise those files when they are relevant to a request about project costs. The fix belongs in the permissions and data protection controls, not in the AI tool.

Check what you already have

Copilot Chat may already be available to your users. Before assessing risk, list which Copilot experiences are switched on, who holds a Microsoft 365 Copilot licence, and which agents and connectors have been approved. An organisation that has only Copilot Chat does not have the same exposure as one with full work-data grounding.

Four risks to plan for

Exposure can happen inside the organisation or outside it, and the controls differ. These are the four risks we see most often before implementations.

Risk 1
Internal oversharing
Copilot shows a colleague information they can technically open but have no business need to see. Years of unreviewed permissions make this more likely.
Risk 2
Data used outside approved purposes
Working with personal data in Copilot is normal. The risk is special category data, financial records or trade secrets being used in ways your policies and lawful bases don't cover.
Risk 3
Answers shared onward
A response that includes wrongly accessible information can be pasted into an email or a Teams chat with an outside party. Internal oversharing becomes external disclosure.
Risk 4
Untrusted content and unapproved tools
Documents and emails can carry hidden instructions meant to steer Copilot's answers (prompt injection), which matters most for agents. Unapproved AI apps are a separate risk: see shadow AI risks and agentic AI security.

Copilot also brings risks of its own: answers can be wrong, and agents can take actions through the tools they are connected to. Every prompt and response is also a new record that needs retention rules.

Excessive permissions: where most of the work is

Permissions drift. Someone needs a folder for a week and keeps access for years. A project closes and its SharePoint site stays open to the whole company. After team changes and departures, access granted earlier no longer matches current roles. Review members, guests, sharing links (especially "anyone" links) and site owners, and check the retention policies that apply to old content.

A permissions review doesn't need to cover everything at once. Start with your most critical data: contracts, financial records, HR information and intellectual property. Apply least privilege, so each user only has access to what their current role requires.

If you have thousands of sites, you can use temporary controls while the review runs. Microsoft includes two in Microsoft 365 Copilot licences through SharePoint Advanced Management:

  • Restricted Content Discovery keeps a site out of organisation-wide search and Copilot answers. Permissions stay as they are, so people with access can still open the files directly. It limits discovery, not access.
  • Restricted Access Control limits a site to members of named groups, even if others had access through permissions or links. Use it where access itself must be restricted.

Labels, encryption and DLP: what each one stops

A sensitivity label does not automatically stop Copilot from using a document. A label without protection settings classifies the file and can drive other policies, but it doesn't change who can read it. A label with encryption and usage rights does restrict access, and Copilot can only use content where the user has the rights to extract it.

Microsoft Purview also has DLP policies built for Copilot. They can stop Copilot from processing files and emails that carry selected labels, block responses to prompts that contain chosen sensitive information types, and stop Copilot from running web searches for those prompts. Some of these features are still in preview, policy changes can take hours to apply, and a blocked item may still appear as a citation even though its content is not used. Test with different user profiles before you rely on them.

Labels on SharePoint sites and Teams cover the container. They don't label every document inside it. A Confidential label on an HR site does not mean every file in that site is kept out of Copilot.

Each control solves a different problem:

  • Permission review removes unnecessary access. It doesn't classify information.
  • Sensitivity labels identify sensitive content. On their own they may not prevent retrieval.
  • Encryption and usage rights restrict access and extraction. They must be configured and tested.
  • Copilot DLP restricts processing of selected content or prompts. The supported conditions and licences vary.
  • Restricted Content Discovery limits search and Copilot discovery for a while. It doesn't revoke permissions.
  • Restricted Access Control enforces group-based site access. It needs sound group governance.
  • User training reduces careless use. It doesn't replace technical controls.
  • Audit and monitoring detect and investigate activity. They can't prevent exposure.

UK GDPR and Copilot: what to check

Under enterprise data protection, Microsoft states that prompts, responses and the Microsoft Graph data used as context are not used to train its foundation models. That separates the business service from personal AI tools on consumer terms. Your duties as controller remain.

  • Contract: confirm which services and features Microsoft's Product Terms and Data Protection Addendum cover, including subprocessors and international transfers. The DPA supports your assessment but doesn't replace your obligations as controller.
  • Purposes and transparency: review the purposes and lawful bases for personal data processed with Copilot. Check that your privacy information still describes what you do, especially if Copilot introduces new purposes, recipients or ways of assessing people. Not every AI use needs a separate notice.
  • Data residency: Microsoft stores Copilot interaction content at rest in your tenant's data residency region. Advanced Data Residency and Multi-Geo change the commitments, and processing can take place elsewhere. The admin centre setting is a starting point. Map the actual data flows, integrations and transfers.
  • Automated decisions: UK GDPR has specific rules for decisions made solely by automated means with legal or similarly significant effects, as amended by the Data (Use and Access) Act 2025. Summarising a contract or drafting an email does not trigger them. Check the cases where Copilot output shapes decisions about people.

Whether you need a DPIA depends on what you use Copilot for. Drafting and summarising your own documents is different from managers using it to analyse individual employees. Our guide to when AI processing needs a DPIA covers the screening.

A staged rollout

Each stage should leave evidence you can show to management or a regulator. The aim is to reduce, control and monitor risk. No rollout gets to zero.

Identify the Copilot products. Copilot Chat, licensed Copilot, agents and connectors. Evidence: a product and licence inventory.
Map critical information. Sensitive sites, teams and mailboxes, and who owns them. Evidence: a data risk assessment.
Fix permissions. Remove unnecessary access, "anyone" links and overly broad groups. Evidence: access review results.
Apply interim controls. Restricted Content Discovery, Restricted Access Control or DLP where the review isn't finished. Evidence: tested restriction results.
Configure information protection. Labels, encryption and Copilot DLP by data category. Evidence: policy settings and test cases.
Confirm privacy and contracts. Purposes, the DPA, retention and cross-border flows. Evidence: a legal and supplier assessment.
Validate auditing. With auditing on, Copilot activity is recorded in Purview Audit (Standard): who, when and which resources. Prompt and response text is reached through eDiscovery or DSPM for AI. Check retention periods and who may search. Evidence: recorded test interactions.
Train and pilot. Hands-on training based on real tasks, then a pilot with representative users sized to your risk. Evidence: findings and fixes.
Expand on conditions. Approve wider rollout once critical findings are closed. Evidence: a management decision and an exception register.
Monitor changes. New sites, sharing links, agents and data sources. Evidence: a recurring governance report.

Before you allow custom agents or external connectors, list their data sources, the permissions they use, the suppliers involved and the actions they can take. An agent connected to a CRM, an ERP system or automation tools carries different risks from Copilot summarising documents.

From one engagement

In one client engagement (a single case, not a benchmark; details simplified and figures approximate), a professional services firm of roughly 100 to 150 staff ran a permissions review before Copilot adoption. It found that around a third of its SharePoint sites had at least one user with access they should no longer have had. Several of those sites held contracts with confidentiality clauses and client fee information.

Fixing it took a few weeks. Without the review, any employee with that leftover access could have had those files summarised in answers about client work. After the review, permissions reflected current roles, sensitive documents were labelled, and users knew what to expect from Copilot.

Where to start

How long a readiness assessment takes depends on the volume of content, how tangled your permissions are, the tools you have and how mature your governance is. Set the timeline after an initial diagnosis, and fix the highest risks before you widen access.

If Copilot is already active, start with the data access governance reports in the SharePoint admin centre. They show which sites are shared with "Everyone", broad groups or through "anyone" links. That number tells you whether you have an urgent problem or a maintenance project.

Copilot can amplify the permission and data governance problems you already have, and it adds AI risks of its own. A safe rollout deals with both. For licensing, technical prerequisites and configuration, see our Copilot implementation guide. If you want to talk through the state of your environment, get in touch with our team.

Frequently asked questions

Can Microsoft Copilot expose confidential company data?

It can make overshared data easier to find. Licensed Microsoft 365 Copilot works within each user's existing permissions and does not bypass access controls, so if someone can open a confidential file they shouldn't have access to, Copilot could surface or summarise it in an answer. A user can then pass that answer on to people outside the organisation. The fix is in permissions, labels and data protection policies.

Does a Confidential label stop Copilot from using a document?

Not on its own. A label without protection settings classifies the file but doesn't change who can read it. A label with encryption and usage rights restricts access, and Purview DLP policies for Copilot can stop it from processing files with selected labels. Test both with different user profiles before you rely on them.

Does Microsoft train its AI models on our documents?

For Microsoft 365 Copilot and Copilot Chat under enterprise data protection, Microsoft states that prompts, responses and the Microsoft Graph data used as context are not used to train its foundation models. Prompts and responses are still stored, subject to your retention policies, and can be searched for audit and eDiscovery.

Can we switch Copilot on before every permission is fixed?

Some organisations do, with conditions. Review the most critical data first, use Restricted Content Discovery or Restricted Access Control on sensitive sites while the wider review runs, and start with a small pilot group. Widen access once the critical findings are closed.

How does Microsoft Copilot affect GDPR compliance?

Your organisation remains the controller for personal data processed with Copilot. Check the purposes and lawful bases, whether your privacy information still describes what you do, how long prompts and responses are kept, where data is stored and processed, and whether any use needs a DPIA. Microsoft's Data Protection Addendum supports that assessment but doesn't replace it.

What is the difference between Microsoft Copilot and ChatGPT for business use?

Microsoft 365 Copilot is built into Microsoft 365 and can work with your organisation's data in Microsoft Graph, within each user's permissions. ChatGPT business plans offer their own data controls and integrations. The risk depends on the plan, the contract, the data connected and how the tool is configured, so compare those for each tool your staff use.

Can Copilot activity be audited?

Yes. With auditing turned on, Purview Audit records Copilot interactions, including who used it, when, and which files or sites were referenced. The audit record doesn't hold the prompt or response text. To review content, use eDiscovery or Data Security Posture Management for AI, within your retention settings.

How does Cyvra help with Copilot governance?

We review permissions on your most critical data, set out labels, DLP and interim restrictions, and check the GDPR points with your DPO. The result is a rollout plan with documented controls, the risks that remain, and who owns each one.

Sources

Ryland Deakin
About the author
Lead Consultant, Cyvra · CISM · CompTIA Security+ · MCP

Ryland has delivered cybersecurity, compliance, and IT management programmes for regulated organisations across the UK and the Netherlands for over 20 years. View full profile

Work with Cyvra

Ready to adopt Copilot safely?

We review permissions, labels and DLP before and during Copilot adoption, check the UK GDPR points, and leave you with documented controls and owners.

Disclaimer: This article is for informational purposes only and does not constitute legal, regulatory, or professional advice. Cyvra makes no warranty as to the accuracy or completeness of this content. Readers should seek independent professional advice appropriate to their specific circumstances.