Setting Up Microsoft 365 Permissions to Stop AI Data Leaks

When an employee opens Microsoft Copilot and asks, "What are the salaries of the leadership team?" or "Who is on the redundancy shortlist?", they should receive a blank response. In far too many small and mid-sized businesses, they will instead receive a neatly formatted table pulled directly from an Excel spreadsheet saved four years ago.

This is not an AI glitch, nor is it a data breach in the traditional sense. Microsoft Copilot does not bypass your security boundaries. It operates strictly within the context of the user running the prompt. If a user can see a file through SharePoint or OneDrive, Copilot can read it, summarise it, and present it in seconds.

For the past decade, most businesses operated under a regime of security through obscurity. Files were technically open to the whole company, but buried six folders deep in a SharePoint site with an unhelpful name. Nobody found them because nobody went looking.

Copilot changes that equation permanently. Obscurity is dead. If you are evaluating a rollout of Copilot or similar workplace AI tools, sorting out your Microsoft 365 permissions is not an optional IT chore - it is the single most urgent task on your roadmap.

Why AI Turns Hidden Messes into Active Crises

Microsoft uses a principle called security trimming. When an employee asks Copilot a question, the system queries the Microsoft Graph - the underlying fabric that links emails, chats, documents, and calendar events. The Graph returns results only from data sources where that specific user has at least "Read" access.

The problem in most companies with 10 to 250 staff is not that permissions are broken. The problem is that permissions are far too generous.

Over years of day-to-day business, convenience regularly trumps governance. An administrator sets up a company-wide SharePoint site so staff can share templates, but later someone drops an acquisition dossier into a subfolder. A manager creates a Microsoft Team for a specific project, leaves the privacy setting as "Public" instead of "Private", and uploads contractor rate cards.

Without AI, those files sit quietly in the dark. With AI, a single vague question - such as "What are our project margins for this quarter?" - triggers a search that pulls data across every location the user can touch. The user does not even need to know the document exists. AI connects the dots and surfaces the exact figures.

The Three Dangerous Defaults in Your Tenant

Before you can fix the problem, you need to know where the exposure sits. In almost every SMB environment, permissions problems stem from three distinct patterns.

• The "Everyone except external users" group. This built-in tenant group is often used by staff to share files quickly without having to manage individual names. If sensitive folders inherit permissions from a parent folder shared with this group, your entire payroll, strategy, or legal history is open to every employee.
• Public Teams channels. When someone creates a Team in Microsoft Teams, the default setting in many tenants is "Public". That does not mean open to the internet, but it does mean open to every licensed user in your organisation. Files stored in the General channel of a public team are fully indexed by Copilot for all staff.
• Stale sharing links. When employees share a document using the "People in your organisation with the link" setting, that document becomes discoverable to anyone who performs a broad search across the tenant. Over several years, hundreds of these active links accumulate, creating a massive surface area of exposed data.

Evaluating Your Remediation Options

Fixing these issues requires a conscious choice about how your business manages risk against administrative burden. Small and medium businesses generally evaluate three distinct routes.

### Option 1: The Enterprise Lockdown (Microsoft Purview E5) Microsoft offers sophisticated tooling under its Purview suite, including automated data classification, machine-learning-driven sensitivity labels, and complex data loss prevention (DLP) rules.

For an enterprise with thousands of staff, this is essential. For an organisation of 60 people, it is often an expensive distraction. It requires premium licensing, months of configuration, and dedicated staff to monitor alerts. If you try to implement enterprise-grade DLP overnight, you will likely stall your AI rollout indefinitely.

### Option 2: Restricted SharePoint Search Recognising this dilemma, Microsoft introduced a feature called Restricted SharePoint Search. This allows IT administrators to temporarily limit Copilot to searching a curated list of up to 100 approved SharePoint sites, alongside users' personal OneDrive files and chats.

This is an effective stopgap. It allows you to roll out Copilot to an early-adopter group without waiting to audit thousands of historic sites. However, it is fundamentally a blunt instrument. It prevents AI data leaks by blinding the AI to vast portions of legitimate company knowledge, reducing the practical value of the tool.

### Option 3: Pragmatic Hygiene and Group-Based Access The most balanced path for SMBs is targeted remediation. Rather than trying to classify every document created since 2018, you focus exclusively on ring-fencing high-risk data repositories - finance, HR, executive leadership, and legal - and establishing strict site-level permissions.

This approach delivers roughly 90 percent of the security benefit of enterprise tooling with a fraction of the complexity, leaving your operational knowledge accessible to Copilot where it adds genuine value.

A Realistic Four-Step Permission Reset

If you want to deploy Copilot safely without ground-up re-architecture, use this sequential process.

• Step 1: Run a permissions report. Use the SharePoint admin centre or PowerShell scripts to identify all sites where "Everyone except external users" holds member or visitor rights. Export this list and highlight any sites containing confidential business data.
• Step 2: Sever inheritance on sensitive repositories. Remove broad organisation groups from HR, payroll, legal, and board-level sites. Enforce a rule that these sites must contain only explicit, named security groups managed by your IT partner or administrator.
• Step 3: Audit Microsoft Teams privacy. Review every Team in your tenant. Convert any public Teams containing commercial, client, or personnel data into private Teams immediately.
• Step 4: Implement basic sensitivity labels. You do not need automated tagging to start. Deploy three simple, manually applied labels: Public, Internal, and Confidential. Configure the "Confidential" label so that documents tagged with it cannot be accessed outside designated groups, regardless of where they are saved.

Maintaining Control Without Slowing Down Work

Fixing permissions once is straightforward; keeping them clean requires a regular operational rhythm.

Do not ask your staff to become compliance experts. If sharing a document securely takes four extra clicks, they will find an insecure workaround. Instead, educate team leaders on the concept of the "blast radius" - the idea that every file they save should only be visible to people who genuinely need to see it to do their jobs.

Schedule a quarterly review of group memberships and site ownership. When employees leave the business, ensure their access is revoked cleanly, and periodically review which SharePoint sites allow guest access.

Copilot provides an extraordinary return on investment when deployed on top of a well-ordered file system. It exposes bad habits, but it does not create them. By addressing your permission structure now, you protect your commercial privacy while building the foundation for an AI-enabled business that operates with speed and confidence.

If you are unsure where your sensitive data currently sits across Microsoft 365, do not guess. Run a tenant-wide permissions audit before you purchase your first Copilot license.