
AI agents are moving beyond answering questions. They’re searching SharePoint, summarising documents, retrieving customer information, interacting with business applications, calling APIs and, increasingly, taking actions on behalf of employees.
That creates a very different data governance challenge.
A traditional application might allow an employee to open a document. An AI agent could potentially retrieve information from multiple documents, combine it, summarise it and send the result somewhere else.
The agent might be working exactly as designed. The problem could be that the underlying data is overshared, insufficiently classified or not adequately protected.
AI agents don’t remove existing information governance problems, they can amplify them.
Before deploying agents into your organisation, I recommend reviewing these five Microsoft Purview control areas.
π 1. Data Security Posture Management (DSPM): Know what your agents can see
You cannot govern what you cannot see.
Before deploying agents, establish where sensitive information exists, where it is exposed and how AI is interacting with organisational data.
Microsoft Purview Data Security Posture Management (DSPM), including its AI-related capabilities, can help organisations understand data exposure and AI usage.
Review:
- Where sensitive information is stored across Microsoft 365.
- Whether SharePoint sites and OneDrive content are overshared.
- Whether sensitive information is appropriately classified and labelled.
- Which supported AI applications and agents are being used.
- Whether AI interactions are exposing sensitive information.
- Which recommended remediation actions should be prioritised.
Pay particular attention to inherited permissions, broad access groups, stale content and documents that have never been classified.
Imagine an agent designed to help HR employees answer policy questions. If its users have access to an old SharePoint site containing confidential employee information, the agent may be able to retrieve that information too.
The agent isn’t necessarily the problem. The data access model is.
ποΈ Jim’s takeaway: Assess your information estate before connecting agents to it. Fix oversharing and data exposure before expanding AI access.
π 2. Sensitivity Labels and Information Protection: Give your data context
An AI agent needs more than permission to access information. Your organisation needs to understand the sensitivity of that information and apply the appropriate protection.
Microsoft Purview sensitivity labels help classify and protect organisational information.
Before deploying agents, review:
- Your sensitivity label taxonomy.
- Whether labels are consistently applied to important information.
- Whether appropriate encryption and access restrictions are configured.
- Whether labels are available to the supported AI workloads.
- Whether your policies restrict agents from processing particular sensitive content.
For example, you might have labels such as:
- Public
- General
- Confidential
- Highly Confidential
A document labelled Highly Confidential may require stronger controls than a document labelled General.
For supported Copilot Studio scenarios, Purview can use sensitivity labels to restrict processing of protected SharePoint knowledge sources through applicable DLP policies. Sensitivity label support also varies by agent type and knowledge source.
Do not assume that applying a sensitivity label automatically blocks every AI agent from accessing the content. The label, encryption, user permissions, agent type and applicable policy configuration all matter.
ποΈ Jim’s takeaway: Your labels need to mean something operationally. A classification scheme that is inconsistently applied or not connected to enforcement policies provides limited protection.
π 3. Data Loss Prevention (DLP): Control what agents can process and share
This is one of the most important controls to review before introducing autonomous or semi-autonomous agents.
An agent may retrieve information from one system and use it in another. It might send an email, create a Teams message, update a record or pass information to an external service.
You need to understand where DLP can inspect and restrict those activities.
Review:
- Which AI application locations are covered by your DLP policies.
- Whether sensitive information in prompts and responses is detected.
- Whether labelled content can be restricted from supported agent interactions.
- Whether external sharing and communication actions are controlled.
- Whether endpoint DLP protects users interacting with third-party AI services.
- Whether custom connectors, HTTP actions and external integrations have been assessed separately.
For example, consider an agent that can retrieve customer information from SharePoint and send an email.
You should establish whether the relevant Purview DLP policies can detect and restrict sensitive information in the supported interaction and communication paths.
For Copilot Studio agents, Microsoft documents specific DLP and sensitivity-label support, including restrictions for supported SharePoint knowledge-source scenarios. Do not assume every agent action or external connector is automatically inspected by Purview DLP.
ποΈ Jim’s takeaway: Don’t just ask what data an agent can read. Ask where that data can go next, and which controls can prevent it from leaving an approved boundary.
π 4. Insider Risk Management: Detect risky AI behaviour
Not every AI-related incident starts with a malicious employee.
An employee might accidentally provide sensitive information in a prompt. An agent might be manipulated through prompt injection. Someone might attempt to use AI to access or extract information they should not be handling.
Microsoft Purview Insider Risk Management (IRM) can help identify and investigate supported risky AI usage.
Review:
- Whether the Risky AI Usage policy template is configured.
- Whether relevant AI activities are being monitored.
- Whether prompt injection and attempts to access protected materials are included in your risk assessment.
- Whether investigation teams have appropriate permissions.
- Whether privacy controls, including pseudonymisation, are configured.
- Whether alerts can be investigated and escalated through your existing security processes.
The objective is not to treat every unusual AI interaction as misconduct.
It is to identify activity that may indicate data theft, inappropriate use or an attempt to bypass organisational controls.
For example, an employee repeatedly attempting to use an agent to retrieve restricted information may warrant investigation, particularly when combined with other relevant risk indicators.
ποΈ Jim’s takeaway: AI governance needs to consider user behaviour as well as agent behaviour. You need visibility into how people are using agents, not simply whether an agent has been deployed.
π 5. Audit, Retention and eDiscovery: Know what happened and preserve the evidence
If an AI agent accesses sensitive information, produces an unexpected response or takes an inappropriate action, can your organisation establish what happened?
This is where audit and compliance controls become essential.
Review:
- Whether supported agent interactions are captured in the unified audit log.
- Whether prompts, responses and relevant activity records are available.
- Whether the records provide sufficient context for an investigation.
- Whether retention policies cover the relevant AI interaction data.
- Whether eDiscovery teams can search, review and preserve relevant evidence.
- Whether communication compliance policies are appropriate for supported AI interactions.
Microsoft Purview Audit can capture supported AI activity, including prompts and responses, depending on the application and integration.
Data Lifecycle Management can help manage the retention and deletion of supported AI interaction data. eDiscovery can support the preservation, collection and review of relevant information where supported.
This is particularly important for regulated organisations.
You should be able to answer questions such as:
- Which user or agent initiated the interaction?
- What information was accessed?
- What was generated?
- When did the activity happen?
- What evidence is available?
- How long must the relevant records be retained?
Remember that logging an interaction is not the same as retaining every piece of source data or recording every action taken by an external system. Validate the actual audit coverage for each agent platform.
ποΈ Jim’s takeaway: If you cannot reconstruct what an agent did, you will struggle to investigate incidents, demonstrate compliance or establish accountability.
π The important point: Purview coverage is not identical for every agent
These controls are not a universal switch that protects every AI agent.
Microsoft publishes capability and application-specific guidance for Microsoft 365 Copilot agents, Copilot Studio agents, Agent 365 and other supported AI applications.
For example, Microsoft’s Copilot Studio guidance identifies support for DSPM, auditing, data classification and sensitivity labels, with specific DLP and Insider Risk Management capabilities. Some scenarios also require pay-as-you-go billing, particularly for agents published to non-Microsoft channels.
Before deployment, validate:
- Which agent platform you are using.
- Where the agent is published.
- Which knowledge sources it accesses.
- Which connectors and tools it can call.
- Which Purview capabilities are supported for that exact scenario.
- Whether additional licensing, billing or configuration is required.
Also remember that Purview is one part of the governance model. It does not replace Microsoft Entra identity controls, least-privilege permissions, agent ownership, connector governance, secure development or operational monitoring.
π My recommended pre-deployment checklist
Before an agent is approved for production, I would want evidence that these five areas have been reviewed.
- DSPM: Sensitive data exposure and AI usage assessed.
- Information Protection: Sensitivity labels and protection policies reviewed.
- DLP: Supported agent interactions and data movement protected.
- Insider Risk Management: Risky AI usage detection and investigation considered.
- Audit and compliance: Logging, retention and eDiscovery requirements validated.
I would also require an identified business owner, a technical owner, approved knowledge sources, documented permissions and a process to disable the agent if something goes wrong.
π‘ The Jim Talks takeaway:
Organisations are rightly excited about what AI agents can do. But giving an agent access to business information is not simply another application deployment.
It introduces a new way for information to be discovered, combined, processed and acted upon.
We have spent years building governance around what employees can access and what they can share. Now we need to apply those principles to what agents can access and what they can do with it.
My framework remains:
DATA β IDENTITY β AGENT β TOOLS β ACTION β OUTCOME
Microsoft Purview plays a critical role in governing the data and compliance aspects of that journey.
But the real objective is not to deploy more controls. It is to make sure that when an AI agent is given the ability to act, the organisation understands the data it can reach, the boundaries it must respect and the evidence it leaves behind.
Don’t deploy the agent first and work out the governance afterwards. Make governance part of the deployment decision.
Microsoft resources
- Microsoft Purview β Data security and compliance for AI agents
- Microsoft Purview β Copilot Studio agent security and compliance
- Microsoft Purview β Protect sensitive data to get ready for AI
- Microsoft Purview β Data security and compliance for generative AI apps
- Microsoft Purview β Insider Risk Management
- Microsoft Purview β Audit
Always validate current licensing, feature availability and supported agent scenarios against Microsoft’s documentation before implementing controls.
Leave a Reply