Prepare data for AI
Clean up permissions, label sensitive content, retire stale files and set AI-specific data controls so Microsoft Copilot, agents and other AI apps work only with current, appropriately shared data.
The problem
AI assistants make existing content much easier to find. If a file is overshared, out of date or unlabeled, Copilot and agents can bring it into an answer for anyone who technically has access. Old and duplicated content also lowers answer quality.
Data now flows to AI through many doors: Microsoft Copilot, custom agents, Microsoft Foundry apps and third-party AI sites that staff use in a browser. Each needs a policy.12
The approach
Follow the order Microsoft uses in its foundation guidance:
- Find exposure: Purview Data Security Posture Management data risk assessments and the SharePoint Advanced Management Content Management Assessment show overshared, ownerless and inactive sites and risky links.
- Contain while you fix: Restricted Content Discovery keeps sensitive sites out of Copilot discovery; DLP for Copilot can exclude labeled files and emails and block prompts that contain specified sensitive data.
- Fix access: site sensitivity labels, removal of anonymous and company-wide links, site access reviews and ownership policies.
- Protect by default: auto-labeling and default labels. When a label applies encryption, users need both the EXTRACT and VIEW usage rights before AI apps return that content.
- Clean up: retention and deletion policies for obsolete files, and Microsoft 365 Archive for inactive sites, which prevents Copilot from processing that content.
- Watch AI use: prompts and responses go to the unified audit log; Insider Risk Management has a risky AI usage policy template; endpoint DLP can warn or block sharing of sensitive data with third-party generative AI sites.
For data outside Microsoft 365, the same principles apply: Foundry IQ honors Purview labels and synced permissions for supported sources, and Fabric data is governed in OneLake Catalog with Purview controls.
When a simpler route is enough: users on Copilot Chat without a Copilot license get mainly web grounding, so broad tenant cleanup can follow the licensed rollout rather than block it.12345
Conceptual architecture
Diagram as a list
Copilot & business apps
- AI apps: Microsoft Copilot, agents, Foundry apps, third-party AIConsumers of prepared data
Data & analytics
- SharePoint, OneDrive, Exchange and Teams contentContent AI tools can reachConnects to AI apps: Microsoft Copilot, agents, Foundry apps, third-party AI
- Retention policies and Microsoft 365 ArchiveRemove obsolete content and exclude archived content from CopilotConnects to SharePoint, OneDrive, Exchange and Teams content
Security, identity & governance
- Purview Data Security Posture ManagementRisk assessments and Activity Explorer for AI interactionsConnects to SharePoint, OneDrive, Exchange and Teams content
- SharePoint Advanced ManagementContent assessment, Restricted Content Discovery, access reviews, inactive site policiesConnects to SharePoint, OneDrive, Exchange and Teams content
- Sensitivity labels and DLP for CopilotProtect sensitive items and control AI processing
- Insider Risk Management and endpoint DLPDetect risky AI use and block pasting into third-party AI sitesConnects to AI apps: Microsoft Copilot, agents, Foundry apps, third-party AI
- Audit, eDiscovery and Compliance ManagerRecord and review AI interactions for regulatorsConnects to AI apps: Microsoft Copilot, agents, Foundry apps, third-party AI
Technologies and their roles
Microsoft Purview2
Data security and compliance for AI
DSPM, labels, DLP, Insider Risk, audit and retention cover Copilot, enterprise AI apps and other AI apps.
Microsoft 365 Copilot1
Main consumer of Microsoft 365 data
Copilot grounds in data the user can already access, so permission quality decides answer safety.
Microsoft Fabric4
Analytics data governance
OneLake Catalog and Purview controls govern data that AI and BI use.
Azure AI Search3
Permission-aware retrieval for custom AI
Through Foundry IQ, retrieval can enforce synced permissions and Purview labels.
What you need first1
- SharePoint Advanced Management (included with Copilot licenses) and Purview capabilities from at least Microsoft 365 E3
- Admin roles for the Microsoft 365, SharePoint and Purview portals
- A published sensitivity label taxonomy
- Site owners who can complete access reviews
- Decisions on how long to keep Copilot interactions and audit logs
Risks and how to reduce them
- Interim controls become permanent1
- Remove Restricted Content Discovery and other interim protections once permissions are fixed, as Microsoft's guidance says.
- Encrypted content becomes invisible to AI unexpectedly2
- Check label usage rights; AI apps need EXTRACT and VIEW rights to return encrypted data.
- Staff paste sensitive data into third-party AI sites2
- Use endpoint DLP to warn or block and the risky AI usage policy in Insider Risk Management.
- Cleanup deletes records that must be kept1
- Use retention labels and Microsoft 365 Archive to keep records while excluding them from Copilot.
Related
Sources
- Configure a secure and governed foundation for Microsoft Copilot (opens in a new tab)
- Microsoft Purview data security and compliance protections for Microsoft 365 Copilot and other generative AI apps (opens in a new tab)
- What is Foundry IQ? (opens in a new tab)
- What is Microsoft Fabric (opens in a new tab)
- Overview of Microsoft Copilot Chat (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.