Before You Roll Out Copilot, Fix What It Can Already See

When businesses start looking at Microsoft Copilot, the first question is usually the same: "Will it leak our data?" It is a fair worry, but it is slightly the wrong one. The better question is: "What can our people already reach that they should not?"

Copilot does not break your permissions. It exposes them.

Microsoft is clear about how Copilot works. It grounds its answers in content in Microsoft Graph, such as emails, chats and documents, and, in Microsoft's words, it "only surfaces organizational data to which individual users have at least view permissions."

That is reassuring, and it is also the catch. Copilot does not hack its way into files. It simply makes everything a user is allowed to see far easier to find. A salary spreadsheet buried in an old SharePoint folder, shared years ago with "Everyone except external users", was safe mostly because nobody could find it. Ask Copilot to "summarise recent HR changes" and that obscurity disappears.

Microsoft's own deployment guidance treats oversharing as the first job. Its Copilot privacy documentation explains why: the permission models in services such as SharePoint should help ensure "the right users or groups have the right access to the right content", and that includes access granted to people outside your organisation, for example through shared channels in Teams.

Where oversharing usually hides

In small and mid-sized businesses the pattern is rarely malicious. It is years of convenience:

• Broad sharing links. "Anyone in the organisation" links that were created for one email and never removed.
• Open Teams and SharePoint sites. Sites created as public by default, or with a company-wide group added as members.
• Leavers and movers. Staff who changed departments but kept access to their old team's files. We covered a related issue in our piece on lingering access after offboarding.
• Over-privileged accounts. Too many people with broad rights, which we discuss in too many admins in Microsoft 365.
• Unlabelled sensitive data. Contracts, payroll, board packs and client records with nothing marking them as sensitive.

A practical readiness plan

Microsoft's blueprint is organised around three pillars: remediate oversharing, set up guardrails, and meet regulations. Here is how we would translate that for a business of 20 to 500 users.

1. Find the exposure first

Start by identifying your highest-risk sites and files: finance, HR, legal, executive and client data. Microsoft's guidance relies on SharePoint Advanced Management, which it says is included with your Copilot licence, and on Microsoft Purview. Use them to review sharing, access and governance before anyone gets a licence.

2. Apply interim restrictions where needed

If a site is clearly too open and cannot be fixed this week, restrict it temporarily. A short period of tighter access is better than Copilot surfacing sensitive content on day one.

3. Fix the access, then label the data

Clean up sharing links and group memberships, and give site owners the job of confirming who really needs access. Then use sensitivity labels. Microsoft states that when data is encrypted with Purview Information Protection, Copilot honours the usage rights granted to the user, so labelling is a genuine guardrail rather than paperwork.

4. Control agents and extensions

Copilot can also use agents and connectors. Microsoft says admins can see the permissions and data access an agent requires in the Integrated apps section of the Microsoft 365 admin center and choose which agents are allowed. Treat that approval list as you would any other app approval. It ties in closely with the risks in shadow IT and apps connected to your data.

5. Decide what you will log and keep

Copilot stores a record of prompts and responses. Microsoft says admins can review that data through Content search or Microsoft Purview, and can use Purview to set retention policies for Copilot chat interactions. Decide in advance what your retention and review approach is, particularly if you handle personal information under POPIA or GDPR.

A note for European and UK customers

Microsoft states that Copilot is an EU Data Boundary service and that EU traffic stays within the boundary. It also notes that models provided by Anthropic as a subprocessor are currently excluded from the EU Data Boundary, and that admins can choose whether to use those models. If data residency matters to you, make that a conscious decision rather than a default.

The takeaway

Copilot is not a new security hole, but it is a very effective magnifying glass. A tidy permissions model and sensible labelling make it a safe productivity tool. A messy one turns every old mistake into a search result.

None of this needs to delay your AI plans. A short, structured permissions review before rollout is usually enough to change the conversation from "Is it safe?" to "Where can it help us first?" If you would like a second pair of eyes on your tenant, our AI solutions and cybersecurity teams are happy to talk it through.