Simple Data Permission Rules to Stop AI Tools from Oversharing

The nightmare scenario for any managing director rolling out an AI assistant is simple: an intern asks Microsoft Copilot a casual question about company overheads, and the system cheerfully returns a summary of executive salaries, complete with pension contributions.

This does not happen because the AI went rogue or hacked into a secure server. It happens because the AI did exactly what it was told to do. It searched every document the user had permission to read, found an unlabelled spreadsheet sitting in a forgotten folder called "General Admin", and summarised it.

AI tools like Copilot, Google Gemini, and internal search assistants do not bypass your security. Instead, they expose every security shortcut your business has taken over the past ten years. In a business of 20 to 200 people, "security by obscurity" - relying on the fact that files are buried four folders deep so nobody will stumble on them - has functioned as a makeshift governance strategy. The moment you introduce an AI engine that indexes everything instantly, that strategy collapses.

If you are evaluating AI tools, you do not need an enterprise-scale security project that takes six months and costs tens of thousands of pounds. You need four simple, practical rules to get your permissions under control before you hand out the software licences.

The "Open by Default" Reality of Growing Companies

Most small and medium businesses suffer from organic file clutter. When the company had eight employees, sharing every folder with "Everyone" was efficient. As the business grew to 40, 80, or 150 people, that legacy permission structure remained untouched.

Today, your systems likely contain three hidden risks: - Legacy SharePoint sites or Shared Drives where the default permission is set to "Everyone except external users". - Files shared via broad links, such as "Anyone in the company with this link can edit", created during a rush to meet a deadline. - Orphaned folders created by former employees who placed sensitive operational data in general collaboration spaces.

Before adding AI into the mix, you must accept a basic principle: an AI tool is a high-speed search engine. If a user can open a document manually, their AI assistant can read, summarise, and cite that document in response to any prompt. Securing your data is not an AI project; it is a fundamental access control project.

Rule 1: Eliminate Broad Sharing Links

The quickest way to introduce accidental oversharing is through loose file links. When staff share documents in Microsoft Teams or OneDrive, the default setting often allows anyone within the organisation to view the file. Over five years, this creates thousands of confidential documents floating in an accessible pool.

Take these steps to lock down link behaviour: - Change the default sharing link setting in your Microsoft 365 or Google Workspace admin console from "Organisation-wide" to "Specific people". - Force staff to explicitly name the colleagues who need access to a file, rather than using broad, company-wide access. - Run an administrative report to identify existing links set to "Anyone inside the organisation". Revoke these links systematically, starting with folders belonging to leadership, human resources, and finance.

This single policy change stops the creation of new security leaks while you work on cleaning up the past.

Rule 2: Ring-Fence Sensitive Departments into Separate Repositories

Do not rely on folder-level permissions inside a general document library. Folder-level exceptions are fragile. It takes one distracted manager clicking "Reset permissions to parent" to make a private folder visible to the entire company.

Instead, separate your data physically by department and sensitivity: - Establish strictly gated repositories for HR, executive management, finance, and legal affairs. These should be separate SharePoint sites or distinct shared drives, not subfolders within a "Company Documents" library. - Break permission inheritance completely between these sensitive sites and the rest of the company. A standard employee should not even have "read" access to the site's top-level navigation. - Ensure that membership of these groups is tied to role-based groups, not named individuals. When an employee leaves or changes roles, removing them from the "HR Team" group should instantly sever their access to every underlying document and its AI search footprint.

If a repository is inaccessible to a user, the AI tool simply acts as though those documents do not exist.

Rule 3: Apply Sensitivity Labels That Modern AI Respects

Permission groups control who can open a location, but sensitivity labels protect the documents themselves. Microsoft Purview (available in most standard business plans) allows you to tag documents as "General", "Confidential", or "Highly Confidential".

Modern workplace AI tools respect these labels when configured correctly: - Create a simple three-tier labelling system: Public, Internal, and Restricted. Avoid complex classification schemes that employees will find confusing. - Apply automated encryption to the "Restricted" label. If a document is labelled Restricted, only authorised users can decrypt it, regardless of where the file is moved or emailed. - Configure your AI assistant to respect these policies. In Microsoft Copilot, for instance, data from a labelled document inherits that protection. If a user does not have clearance for the label, Copilot will not use the document to answer queries.

Start by applying this label to your most critical files: payroll summaries, board minutes, client contracts with non-disclosure agreements, and disciplinary files.

Rule 4: Run "Can I See This?" Spot Audits

Software permissions can look correct on a dashboard while failing in practice. The only way to know if your rules work is to test them from the perspective of an ordinary user before turning on AI capabilities across the board.

Create a testing protocol: - Set up a standard, non-privileged test account that matches the profile of a junior employee. - Use the search bar in SharePoint or your internal intranet to search for obvious trigger words: "salary", "bonus", "redundancy", "acquisition", "margin", and "settlement". - If the test account can see a file in standard search results, AI will find it instantly and use it to answer questions. - Review any discovered files, identify why the permissions failed, and move them to an isolated, restricted location.

Performing this exercise for two hours will reveal 90 percent of your immediate vulnerabilities.

Moving from Cleanup to Deployment

Governance does not mean halting your AI rollout. Businesses that wait for "perfect" data cleanliness before testing tools like Copilot never deploy them at all. The goal is risk reduction, not perfection.

Begin by ring-fencing your core sensitive data - HR, finance, and leadership strategy - using the rules above. Once those repositories are confirmed dark to standard users, you can deploy your AI tools to a pilot group with confidence. The productivity gains of AI are real, but they should never come at the expense of business confidentiality. Take control of your permissions first; the rollout will be faster, safer, and far less stressful as a result.