Purview + Copilot: Data Governance Before AI Deployment

The Fundamental Problem: Copilot Inherits Your Permissions Debt

Every organization accumulates permissions debt over time. SharePoint sites with "Everyone except external users" access. Teams channels where the entire company was added during a project kickoff and never removed. OneDrive folders shared with links that never expire. File shares migrated to SharePoint Online with flattened permissions.

Before Copilot, this permissions debt was low-risk because users rarely discovered content outside their normal workflows. Copilot changes the equation. When a user asks Copilot "summarize our company's restructuring plans," Copilot searches across every SharePoint site, Teams channel, OneDrive folder, and Exchange mailbox that user has access to — and surfaces relevant content regardless of whether the user knew it existed.

This is not a Copilot bug. It is working as designed. The problem is that most organizations have never audited what their users can actually access because, until now, it did not matter much operationally. Copilot makes it matter.

What Breaks If You Skip Governance

Real scenarios EPC Group has encountered in ungoverned Copilot deployments:

1. Executive compensation exposure. A mid-level manager asked Copilot to "find salary benchmarking data." Copilot surfaced an HR SharePoint site containing the entire executive compensation structure. The site had been shared with "All Employees" during an HR system migration and never locked down.
2. M&A document leak. An employee asked Copilot about "strategic partnerships." Copilot referenced a confidential acquisition target document stored in a Teams channel that the employee had been added to months earlier for an unrelated project.
3. Patient data in summaries. A healthcare organization deployed Copilot in Teams. A non-clinical staff member asked Copilot to summarize recent meeting notes. Copilot included PHI from a clinical meeting transcript that had been stored in a shared Teams channel, creating a HIPAA violation.
4. Terminated employee data access. Copilot surfaced documents from a shared mailbox belonging to a terminated employee. The mailbox had not been deprovisioned, and the current user had inherited delegate access through an old distribution group.
5. Stale content confusion. A sales team member asked Copilot about pricing. Copilot confidently cited a 3-year-old pricing document because retention policies had not been applied and the outdated document had higher relevance signals than the current version.

The Purview Governance Stack for Copilot

Microsoft Purview provides the governance control plane that makes Copilot safe for enterprise deployment. Here is the stack, in implementation order:

Layer 1: Permissions Audit and Remediation

Before touching Purview, fix the permissions foundation:

• Audit all SharePoint Online sites for "Everyone" and "Everyone except external users" permissions
• Review all Teams channels for membership that does not match current business need
• Identify OneDrive sharing links that are org-wide or do not expire
• Remove direct user permissions and replace with Azure AD security groups
• Review and remediate shared mailbox delegate access
• Implement SharePoint access reviews on a quarterly cadence going forward

Layer 2: Sensitivity Labels (Classification)

Sensitivity labels tell Copilot (and users) what kind of data a document contains:

• Public: No restrictions. Copilot can reference freely.
• Internal: Available to all employees via Copilot. No external sharing.
• Confidential: Restricted to specific groups. Copilot only surfaces to authorized users.
• Highly Confidential: Encryption applied. Copilot cannot index encrypted content (effectively excluded from Copilot responses).

Auto-labeling policies can classify documents at scale using trainable classifiers (detect PII, PHI, financial data) or keyword patterns. EPC Group recommends starting with manual labeling for the most sensitive content, then expanding to auto-labeling after the taxonomy is validated.

Layer 3: Data Loss Prevention (DLP)

DLP policies prevent Copilot from surfacing or copying content that violates organizational rules:

• Block Copilot from referencing documents with "Highly Confidential" sensitivity labels
• Prevent copy/paste of content containing SSNs, credit card numbers, or medical record numbers from Copilot responses
• Alert compliance teams when Copilot surfaces content from restricted SharePoint sites
• Apply DLP policies to Copilot interactions in Teams chat, Word, Excel, and PowerPoint

Layer 4: Data Lifecycle Management (Retention)

Retention policies ensure Copilot does not surface content that should have been deleted or archived:

• Apply retention labels to content based on creation date, last modified date, or event triggers
• Auto-delete drafts and working documents after project completion
• Archive historical versions that should not appear in current Copilot responses
• Retain regulated content for the required period (HIPAA: 6 years, SOX: 7 years, GDPR: varies)

Layer 5: Adaptive Protection

Purview Adaptive Protection dynamically adjusts DLP policies based on user behavior:

• Users who access an unusual volume of sensitive content through Copilot get stricter DLP enforcement automatically
• Risk levels adjust based on insider risk signals (mass downloads, access pattern anomalies)
• No manual intervention required — policies adapt in near-real-time

Layer 6: Audit and eDiscovery

Copilot interactions must be auditable for compliance:

• Copilot prompts and responses are captured in the unified audit log
• eDiscovery searches can include Copilot interactions (critical for litigation hold)
• Communication compliance policies can scan Copilot outputs for regulatory violations
• Audit retention should be configured for your regulatory requirement (minimum 1 year, recommended 7 years for regulated industries)

Pre-Copilot Governance Checklist

Use this checklist before enabling Copilot licenses for any user group:

Implementation Roadmap

EPC Group's recommended phased approach for Copilot governance readiness:

• Phase 1 (Weeks 1-4): Assessment. Permissions audit, content inventory, sensitivity classification gap analysis, current governance maturity evaluation.
• Phase 2 (Weeks 5-8): Foundation. Permissions remediation, sensitivity label taxonomy deployment, manual labeling of critical content, DLP policy creation.
• Phase 3 (Weeks 9-12): Automation. Auto-labeling policy deployment, retention policy configuration, Adaptive Protection enablement, audit logging verification.
• Phase 4 (Weeks 13-16): Pilot. Enable Copilot for 50-100 pilot users. Monitor access patterns, review audit logs, collect user feedback, tune DLP policies based on false positives.
• Phase 5 (Weeks 17-20): Rollout. Expand Copilot to broader user groups in waves. Continue monitoring. Adjust governance controls based on pilot learnings.

The Cost of Getting This Wrong

Organizations that deploy Copilot without governance face predictable costs:

• Data exposure incident response: $50,000 to $200,000 per incident (investigation, remediation, notification if PII/PHI involved)
• Compliance audit findings: $25,000 to $100,000 for remediation, plus potential regulatory penalties
• Copilot license waste: $30/user/month ($360/user/year) for licenses that get revoked after an incident until governance is fixed
• Organizational trust damage: Incalculable. When employees lose trust in AI tools due to a highly visible data exposure, adoption of all AI initiatives suffers for years

Compare this to the cost of proactive governance: $50,000 to $200,000 for a comprehensive Purview + Copilot readiness engagement, implemented once, before deployment. The ROI is not close.

Frequently Asked Questions