A Practical Guide to Preventing Data Leaks When Adopting AI

When business owners talk about AI risk, they usually picture external threats. They worry that a public model will swallow their intellectual property and regurgitate it to a competitor, or that sensitive customer records will somehow end up on the dark web.

While those concerns are understandable, they miss the most common and immediate risk facing small and mid-sized businesses today: internal oversharing.

If you give your staff an enterprise AI tool like Microsoft Copilot, it will only access what those staff members already have permission to view. The problem is that in almost every business with 10 to 250 employees, internal permissions are far too loose. Files containing executive pay, redundancy plans, disciplinary records, and financial forecasts often sit in open folders or public Teams channels where they were accidentally saved years ago.

Human employees rarely stumble across these files because nobody spends their afternoon clicking through hundreds of SharePoint subfolders. AI, however, indexes everything in seconds. If an employee asks a natural question about company performance or team salaries, a properly functioning AI tool will happily read and summarise those neglected files.

Plugging AI into an insecure file system does not create new security holes. It simply shines a spotlight on the ones you already have. Here is a practical, step-by-step approach to securing your data before you deploy AI across your business.

Separate public chatbots from enterprise AI

The first step in risk management is helping your staff distinguish between consumer AI tools and enterprise AI solutions.

When an employee uses a free consumer tool like standard ChatGPT, their prompts and uploaded documents may be used to train future foundation models. If an operations manager uploads a customer spreadsheet or a draft contract to clean up the formatting, that data has left your perimeter.

Enterprise tools work under entirely different legal and technical terms. When you use Microsoft Copilot with commercial data protection enabled, your data is ringfenced. It is not used to train the base model, it does not leave your tenant, and it is protected by the same compliance boundaries as the rest of your Microsoft 365 environment.

To manage this risk immediately: - Establish a clear policy stating that no client data, financial figures, or proprietary code may be pasted into free, public AI chatbots. - Provide a sanctioned, commercial alternative. If you ban all AI tools without offering a secure alternative, staff will simply use consumer tools on their personal phones. - Ensure all business accounts are signed in to commercial versions of these tools where data protection guarantees are legally active.

Audit your SharePoint and file structure for oversharing

Once you have addressed external tools, turn your attention inward. Enterprise AI models read your company data via the Microsoft Graph. If a document is set to "Everyone except external users" or sits in an "All Company" team, the AI considers it fair game for any employee who runs a search.

Begin with a permissions audit focused on high-risk repositories: - HR and payroll: Ensure that folders containing salary reviews, grievance procedures, medical records, and pensions are explicitly restricted. They should be stored in private, dedicated document libraries with strictly defined member lists, not inside general company folders. - Finance and accounting: Balance sheets, tax filings, and management accounts should never live on general shared drives. - Board and executive communications: Strategic plans, acquisition talks, and ownership agreements must be isolated in permission-restricted sites. - "Public" Teams channels: When staff create a new channel in Microsoft Teams, the default setting is often public to the organisation. Review existing teams and change sensitive operational channels to private.

You do not need to audit every single file your company has produced over the past decade. Focus entirely on containment: identify the 5 percent of your data that could cause commercial or reputational harm if viewed by the wrong staff member, and isolate it first.

Implement basic data classification

Relying on folder permissions alone is rarely enough, especially as businesses grow. Documents get downloaded, forwarded, and re-uploaded into the wrong places.

This is where Microsoft Purview and sensitivity labels become essential. Sensitivity labels allow you to apply rules directly to the files themselves, rather than relying solely on the folder they sit in.

A practical classification framework for an SMB should remain simple. Avoid complex, multi-tiered systems that confuse staff. Three or four tiers are usually sufficient: - Public: Information intended for external marketing, public consumption, or general release. - Internal: Standard business documents, project notes, and general communication that any employee can read. - Confidential: Commercially sensitive information, including client contracts, project pricing, and internal operational data. - Highly Restricted: Payroll, executive strategy, HR records, and legal disputes. Access is locked to specific named individuals, regardless of where the file is stored.

When enterprise AI encounters a file marked with a restrictive label, it enforces those protections. If a user asks the AI to summarise a project, the AI will ignore any restricted files the user does not have explicit permission to read.

Clean up legacy links and guest access

Small businesses often accumulate digital clutter that creates major security risks over time. Two areas deserve immediate attention: anonymous sharing links and lingering guest accounts.

In Microsoft 365, staff frequently share files using links set to "Anyone with the link can view." These links bypass authentication entirely. If one of these links was generated three years ago for a contractor, that file might still be indexed and accessible.

Take the following concrete actions: - Disable anonymous "Anyone" links at the tenant level, or set them to expire automatically after a short period (such as seven days). - Set the default sharing link type to "Specific people" rather than "Anyone in your organisation." This forces employees to consciously choose who gets access to a file. - Review external guest accounts. Former contractors, vendors, and clients often retain access to internal Teams channels long after their contracts have ended. Revoke access for any guest who has been inactive for more than 90 days.

Establish a practical roll-out schedule

Data security is not an all-or-nothing proposition. You do not need to wait until every document in your business is perfectly organised before you begin adopting AI. Doing so would mean waiting forever.

Instead, control risk through a phased deployment: - Start with a pilot group: Deploy enterprise AI licences to a controlled group of 5 to 10 employees. Choose staff who understand data handling and can provide constructive feedback. - Monitor search and output behaviour: Use the pilot phase to test what the AI surfaces. If an employee in operations notices that the AI referenced a confidential document in a response, you have identified a permission leak that needs fixing before broader rollout. - Educate your staff: Train employees not just on how to write prompts, but on basic data hygiene. Teach them how to check file permissions and why they must save files in the correct libraries.

Securing your data for AI does not require months of disruption or expensive enterprise software. It simply requires a deliberate, structured clean-up of the permissions you already have. Before you assign your first licences, run an internal audit of your high-risk files. Find out what is visible today, so you can adopt AI with confidence tomorrow.