Using Sensitivity Labels and Auto-Labeling to Secure Copilot Data

The requirement is to ensure that Microsoft 365 Copilot responses do not use content from project documents stored on SharePoint Site1. The core constraint is that Copilot must respect data boundaries based on document classification.

To achieve this, you must first identify the specific documents (those with the PR-XXXXXXXXX pattern) and then enforce a policy that restricts their access or usage by AI services. Microsoft Purview offers two primary mechanisms for this: Data Loss Prevention (DLP) policies and Sensitivity Labels.

A DLP policy (Option D) can detect sensitive information and prevent it from being shared externally or accessed in unauthorized ways. However, standard DLP policies are primarily designed to prevent data leakage (e.g., blocking an email with a credit card number). While newer DLP capabilities include "Copilot protection" settings that can block Copilot from accessing certain content, the most robust and intended method for controlling internal AI model behavior regarding specific content types—especially when combined with other security requirements—is often linked to how content is classified and protected at rest.

However, looking closely at the specific technical requirement: "Microsoft 365 Copilot responses must NOT use content from the project documents." This implies a need to classify the documents so that Copilot knows to exclude them. Sensitivity labels (Option A) allow you to classify content and apply protection settings. Crucially, sensitivity labels can be configured to restrict who can view the content, but more importantly, they integrate with Microsoft Purview's Information Protection to manage data governance.

Let's re-evaluate Option A vs D. DLP policies can be configured to protect Copilot interactions. According to Microsoft documentation, you can create a DLP policy that includes "Microsoft 365 Copilot" as a location and configure actions to block queries containing sensitive info. But does it specifically stop Copilot from using existing documents? Yes, if the content is flagged as sensitive and the policy blocks AI processing of that content type.

Wait, there is a nuance. Sensitivity labels are used to classify data. Auto-labeling policies can automatically apply these labels based on patterns (like the PR-XXXXXXX regex). Once labeled, sensitivity labels can enforce permissions. If the label restricts access to Group1 only, and the user accessing Copilot is not in Group1, Copilot won't see it. But the question doesn't specify user restrictions, just that Copilot shouldn't use the content.

Actually, the most direct way to prevent Copilot from indexing or using specific content types is through Sensitivity Labels combined with Auto-labeling. Why? Because Sensitivity Labels can have a setting called "Restrict sharing" or can be integrated with Purview Information Barriers. However, the specific feature to prevent Copilot from using content is often handled via DLP policies targeting Copilot.

Let's look at the comments. The comment says DLP prevents leakage but doesn't classify. This is a trap. DLP can classify and act. But the key insight is that Sensitivity Labels are the mechanism for identifying and classifying the documents (via auto-labeling) and then applying protections. If you want to ensure Copilot ignores them, you typically mark them as highly sensitive or restricted.

Correction: In modern Microsoft 365 architecture, to prevent Copilot from reading specific content, you use Sensitivity Labels. You create a label, configure it to restrict access or marking it in a way that Copilot respects (e.g., by restricting external sharing or specific permissions), and use an auto-labeling policy to find the PR-XXXXXXX pattern and apply the label. Once labeled, the label's permissions govern access. If the label is set such that only specific users can access it, Copilot (acting as the current user) will be restricted if they don't have permission. Alternatively, there is a specific setting in Sensitivity Labels to "Exclude from search results" or similar, but more commonly, DLP policies are used to block the action of Copilot querying sensitive data.

Let's look at the official guidance. To protect Copilot, you use DLP policies. You can create a policy that detects sensitive info and blocks Copilot from generating a response. However, the question asks what to create to meet the requirement for project documents. The project documents are identified by a pattern. You need to identify them first. DLP policies can detect patterns, but Sensitivity Labels are better for long-term classification and governance.

Actually, the correct answer is A. a sensitivity label and an auto-labeling policy. Here is why: You use the auto-labeling policy to scan Site1 for the "PR-XXXXXXXXX" pattern and apply a Sensitivity Label. This Sensitivity Label can then be configured to restrict access or indicate high sensitivity. While DLP can block actions, Sensitivity Labels provide the persistent classification that ensures the data is treated correctly across all services, including Copilot. Specifically, you can configure the sensitivity label to restrict who can view the content. If the label restricts viewing to a group that the Copilot user (or the context in which Copilot operates) is not part of, or if the label triggers specific Copilot restrictions (like "Do not index" or "Block AI processing" depending on the license/config), it works. Most critically, Auto-labeling is the only option here that addresses the identification aspect (the pattern matching) efficiently, minimizing admin effort. DLP policies can also detect patterns, but Sensitivity Labels are the standard for classifying and protecting content at rest and in motion, which is the foundation for Copilot compliance. The comment dismissing DLP is partially right in that DLP is reactive, whereas Labels are proactive classification. To ensure Copilot never uses them, classifying them as restricted via Labels is the best practice.

Therefore, creating a Sensitivity Label (to define the protection rules) and an Auto-labeling Policy (to automatically apply it to the PR-XXXXXXX documents) is the correct approach.