Before you run
Data & scope
- Required tables
- Time window
- 7 days
- Microsoft Copilot data connector enabled and CopilotActivity populated; the Microsoft 365 OfficeActivity connector alone is not a substitute.
- Microsoft Purview auditing enabled for the relevant Copilot workloads, applicable product licensing, and query access to CopilotActivity.
Documentation review by Codex (AI). No live-tenant execution or detection-performance claim. Validate schema, data coverage and expected results in your workspace.
// Hunting starter. Documentation checked 2026-11-13; not tenant-tested.
let lookback = 7d;
CopilotActivity
| where TimeGenerated >= ago(lookback) and TimeGenerated <= now()
| summarize AuditRecords = count(), FirstSeen = min(TimeGenerated),
LastSeen = max(TimeGenerated), SampleRecordIds = make_set(RecordId, 3)
by OrganizationId, Workload, RecordType, AppHost
| order by AuditRecords desc, LastSeen desc
| take 200
Read-only query. Run in your own authorized workspace after checking prerequisites.
Start here to establish which Copilot audit events actually reach the workspace. The current Microsoft Copilot connector writes to CopilotActivity. This recipe intentionally uses that supported table instead of assuming Copilot interaction payloads appear in OfficeActivity. It is a coverage and activity inventory, not a misuse detector. Connector reference.
Read the result
The query groups by organization, workload, normalized record type and hosting application. It does not hard-code one interaction name, allowing you to see the types your connector actually supplies. RecordType is a documented string; LLMEventData is a separate dynamic payload that this inventory does not expand. The three record IDs provide bounded examples for a follow-up. CopilotActivity schema.
AuditRecords is a count of ingested rows, not a count of unique users, prompts, sessions or successful actions. The query neither deduplicates record IDs nor guarantees that each product emits the same event granularity. The first and last timestamps help identify a quiet or newly onboarded workload; blank host or workload values should be investigated rather than silently discarded.
Triage and tune
Compare observed groups with the Copilot products your organization expects to use and the auditing configuration for each. Inspect a sample record in context before parsing its nested data. For an unexpected host or record type, establish the owning product, users and approved use case before treating the difference as suspicious.
Microsoft documents user and administrative Copilot activity within Purview Audit; availability still depends on auditing and product use. Non-Microsoft AI auditing can have separate pay-as-you-go requirements, so do not assume every possible AI workload is included. Audit scope and billing distinctions.
Limits and validation
No rows can reflect a missing connector, a silent workload, ingestion delay or unavailable retained data. OfficeActivity being healthy does not establish CopilotActivity coverage. The output is capped at 200 groups and intentionally excludes prompt text and resource details. This query is read-only, documentation-checked and not tenant-tested. Validate the path with an approved benign Copilot interaction and its corresponding audit record before building behavioral detections.
Verify the source
Microsoft references
Measure normal activity, validate known test cases, account for ingestion delays, choose a threshold and entity mapping, and assign an owner. Results require investigation.
Follow the detection learning path →