24.4 Monitor Activities in DSPM for AI
Key Takeaways
- Wait at least one day after creating default policies before Reports populate; filter by Copilot experiences and agents, Enterprise AI apps, and Other AI apps.
- Activity explorer shows AI interaction, AI website visit, DLP rule match, and Sensitive info types events, with user, time, app category, SITs, and files referenced.
- Microsoft 365 Copilot AI interaction and SIT events require auditing; Fabric, Security Copilot, and non-Copilot apps need a collection policy with content capture for prompt and response text.
- The default Microsoft 365 data risk assessment runs weekly on the top 100 SharePoint sites by usage, with a 4-day delay the first time; custom assessment results wait at least 48 hours, do not refresh, and expire after 30 days.
- Prompt text can be missing for Microsoft Facilitator notes, users without Exchange Online mailboxes, consecutive split entries, or collection policies that do not capture content.
Monitor Activities in DSPM for AI
Quick Answer: After policies have had at least a day to collect data, use Reports for category-level graphs, Activity explorer for per-event investigation, Apps and agents for which AI apps touched sensitive data, and Data risk assessments for oversharing in SharePoint (and Fabric after setup). Prompt text appears in AI interaction events only when audit or a content-capturing collection policy applies and you have a content-viewer role.
Monitoring is the last DSPM for AI skill on the SC-401 outline, but it is how you prove the prerequisites and policies worked. Microsoft’s walkthrough is: Get started → Recommendations → Reports → Policies → Apps and agents → Activity explorer → Data risk assessments. You will be tested on where to look, how long to wait, which event type means what, and why a field is blank.
Reports: category first, then drill-through
Open Reports from Overview or the navigation pane. Wait at least a day for default policies to populate graphs. Filter using the three AI app categories Microsoft uses across DSPM for AI:
| Category | What it is for | Examples Microsoft documents |
|---|---|---|
| Copilot experiences and agents | Microsoft Copilots and agents | Microsoft 365 Copilots, Copilot Studio; also where you look after capturing Fabric and Security Copilot |
| Enterprise AI apps | Non-Copilot AI connected to the organization | ChatGPT Enterprise; Entra-registered and Foundry-connected apps |
| Other AI apps | Supported third-party generative AI sites | Sites used for Gemini and ChatGPT consumer experiences |
Enterprise-app reports that Microsoft calls out after the enterprise collection policy include total interactions over time, sensitive interactions per AI app, Insider Risk severity, and Insider risk severity per AI app. Select View details on a graph to land in Activity explorer with the same scope. If Reports are empty after 24 hours, go back to prerequisites (audit, licenses, devices, collection policy, PAYG) instead of creating a second identical one-click policy.
Keep All AI apps on Overview when you are hunting broadly, then switch categories when the question names a specific host. App ID and app host help you tell Microsoft 365 Copilots from Copilot Studio from ChatGPT Enterprise.
Activity explorer events
Activity explorer is the detailed stream. Typical columns and flyout fields include activity type, user, date and time, AI app category and app, app accessed in, sensitive information types, files referenced, and sensitive files referenced. Use the Web queries filter to find AI interactions that performed a web query; the flyout shows the search query text under Web references AI app used.
| Event | Meaning | Special requirements |
|---|---|---|
| AI interaction | The user interacted with a generative AI site or Copilot. Details include prompts and responses, except unmanaged AI apps in Edge, where text prompts only are included | Microsoft 365 Copilot and Copilot Chat need auditing. Copilot in Fabric, Security Copilot, and non-Copilot AI apps need a collection policy with content capture for prompts and responses |
| AI website visit | The user browsed to a generative AI site | Sourced from Insider Risk Management visit detection; viewing these browse-to-URL events requires IRM Analyst or Investigator |
| DLP rule match | A DLP rule matched during an AI interaction, including DLP for Microsoft 365 Copilot | You need a DLP policy that actually matched; this is not the same as a SIT observation |
| Sensitive info types | SITs were found during an AI interaction | For Microsoft 365 Copilot and Copilot Chat, auditing must be on, but no extra active policy is required |
That last row is an exam favorite: Copilot can surface Sensitive info types events from audit alone. You do not have to create DSPM for AI: Detect sensitive info added to AI sites just to see SITs in Copilot prompts. That discovery DLP policy is for browser AI sites, not for replacing Copilot audit.
Workload names in explorer match the report categories: Copilot experiences and agents, Enterprise AI apps, and Other AI apps.
Known issues you should not “fix” with random policy churn
Microsoft documents several monitoring gaps. Teach them as product behavior:
- AI interaction does not always display prompt and response text. Sometimes the text spans consecutive entries.
- Microsoft Facilitator AI-generated notes: no prompt or response is displayed, and those AI interaction events cannot be linked to Sensitive info types detected events.
- If the user does not have a mailbox hosted in Exchange Online, no prompt or response is displayed.
- Sensitive info types detected does not display the user risk level.
- If a collection policy does not capture content, no prompt or response is displayed. The network one-click policy is the example that starts this way.
Also remember the permission overlay from the previous section: even when capture works, you still need Content Explorer Content Viewer or Data Security AI Content Viewer to read the text.
Two ingestion pipelines confuse operators. Microsoft 365 Copilot interactions continue to arrive from the unified audit log even if you disable a DSPM collection policy. Fabric, Security Copilot, enterprise AI, and unmanaged AI prompt storage depends on collection policies (and content capture). A privacy request to “stop collecting AI prompts” is therefore at least two switches, not one.
Apps and agents, and the Policies status board
Apps and agents is a dashboard of AI apps and agents in the organization. For each agent you can see sensitive data it accessed and which Purview policies protect it. Use it when the question is “which agent is over-sharing?” rather than “what did this user paste at 14:32?”
The Policies page is monitoring, not authoring. Confirm the one-click policies are on, watch PendingDeletion, and jump to Communication Compliance to remediate unethical-behavior matches. If an IRM or Communication Compliance policy is missing from the list for a view-only user, that is a permissions issue, not a failed deployment.
Data risk assessments
Generative AI amplifies oversharing: Copilot can surface stale, over-permissioned, or unlabeled content quickly. Data risk assessments identify and help remediate that problem.
Use the Microsoft 365 tab for SharePoint and OneDrive oversharing insights, and the Fabric tab for Fabric workspaces after the one-time Entra/Fabric setup.
Default assessment (Microsoft 365): runs weekly for the top 100 SharePoint sites based on usage. It appears at the top of the page with totals such as items found, sensitive data detected, and links sharing data with anyone. The first default assessment has a 4-day delay before results display. Return regularly; do not treat a one-time glance as the program.
Custom assessments: choose users and data sources and run on demand (documented as preview). After a custom assessment runs, wait at least 48 hours. Results do not update again; you need a new assessment to see changes. Custom assessments expire after 30 days. Use Duplicate to recreate the same selections. Item-level scanning (SharePoint only, after Entra app authentication) flags items with external or anonymous sharing links and shows sensitivity label and owner. Remediation actions Microsoft documents:
- Resolve — you decide the item is not an oversharing risk
- Apply sensitivity label — unlabeled or wrongly labeled items
- Notify — email the site owner (not customizable)
- Remove sharing link — use sparingly; legitimate users can lose access until the owner issues a tighter link
Published Microsoft 365 limits: 200,000 items per location (default and custom); file counts may be inaccurate above 100,000 files per location; OneDrive is not supported for item-level scanning; maximum 10 SharePoint sites for item-level scanning. Both default and custom assessments support Export to Excel, CSV, JSON, or TSV.
Open a site (or Fabric workspace) flyout and use the tabs:
- Identify (Microsoft 365) — how much data was scanned for SITs; start an on-demand classification scan if needed
- Protect — recommended remediations, which launch other products: DLP that prevents Copilot from summarizing selected labels; SharePoint Restricted Content Discovery to exempt sites from Microsoft 365 Copilot; auto-labeling for unlabeled sensitive files; retention policies for content not accessed in at least 3 years
- Monitor — counts for anyone / organization / specific people / external sharing; Start a SharePoint site access review uses SharePoint data access governance reports. Fabric Monitor can send you to Microsoft Entra access reviews
Fabric default assessment, once configured, also runs weekly for the top 100 Fabric workspaces based on usage. Protect recommendations there include Fabric DLP examples (block Anyone-with-the-link access when SITs are found, apply a label, attach a policy tip) and a default sensitivity label for new and edited Fabric items.
Scenario: empty Reports after Copilot rollout
Northwind assigned Microsoft 365 Copilot licenses yesterday and opened DSPM Reports immediately. Graphs are empty. Microsoft tells you to wait at least a day. If they are still empty after that, confirm Purview Audit is on — it is default for new tenants but may have been turned off. Do not onboard every device unless the missing data is third-party site visits.
Scenario: Fabric prompts still blank in explorer
Reports show Copilot in Fabric activity after you took Secure interactions for Microsoft Copilot experiences, but AI interaction flyouts have no text. Check three layers: the collection policy exists and is not PendingDeletion, content capture is selected with classifiers set to All, and the investigator has Content Explorer Content Viewer or Data Security AI Content Viewer. If the user has no Exchange Online mailbox, Microsoft documents that prompt text will not display even when capture is on.
Official references: Using DSPM for AI, Activity explorer events.
You need prompt and response text in Activity explorer for Copilot in Fabric. Which event type carries that content, and what capture path does Microsoft document?
What scope does the default Microsoft 365 data risk assessment in DSPM for AI use, and how often does it run?
A custom data risk assessment completed, and 30 days later the team wants updated oversharing results for the same sites. What does Microsoft document as the way to see changes?
You've completed this section
Continue exploring other exams