Operate AI Insight Packs

Prerequisites

insights.read; insights.write for definitions/state; insights.run for execution.
Accessible source modules and a supported workload, subscription, or tenant scope.
An active AI provider for interview/generation/refinement and scheduler/connectors for recurring external delivery.

Open /insights; sections include Today, Library, Recent Runs, and Schedule.

Open Library and start from a template or select Generate with AI.
In the Goal → interview → Generate flow, define audience, operational question, sources, scope, lookback, materiality, and output.
Review the generated definition; do not accept invented thresholds or escalation logic.
Use Preview/Sample/Review as offered. A sample is illustrative; select a workload for a real read-only preview.
Save only after validation issues are resolved.
Run once on demand with notification disabled.
Confirm routine evidence yields Nothing notable and material evidence yields Notable or Urgent.

Expected result: A reusable, validated pack exists and produces a persisted digest from accessible sources.

Verification: Open the run and compare source counts/rows, scope, lookback, verdict, and materiality-gate explanation.

In Today, review urgent, notable, and quiet counts per scope.
Open unread material cards; opening marks the digest read.
Inspect evidence bundles and flags before accepting the narrative.
Acknowledge when triage ownership is established; mark false positive when the verdict is demonstrably wrong.
Use Recent Runs to filter by verdict, pack, notification, and day; mark all read only after review.
Use Re-run now when current evidence is required. Closing the dialog does not cancel the background run.
Create a case or open linked Change Explorer/Radar context where the digest offers those actions.

Expected result: Material digests have explicit read/acknowledged/false-positive state and an appropriate handoff.

Verification: Reopen the digest and confirm state, actor/time, scope, and linked record.

Open Run / Schedule or the Schedule tab.
Select workload/scope, cadence, time, time zone, notification behavior, and lookback.
Verify the next-run preview, then enable the schedule.
Review Timeline and Coverage Matrix for watcher gaps and next-seven-day occurrences.
Pause/resume the schedule or use Edit to open Scheduled Tasks for cadence changes.
Snooze a noisy pack to suppress notifications temporarily; note that snoozed packs still run and retain records.
Review health/noise and false-positive rates; narrow sources or raise materiality, then test again.
When retiring a watcher, disable/remove its schedule as well as disabling the pack.

Expected result: The intended scope runs at the displayed local/UTC cadence and only material results request notification.

Verification: Confirm next-run time, scheduler history, resulting digest, and connector delivery log.

Expected result: A portable snapshot of that run is downloaded.

Verification: Compare report title, scope, timestamp, verdict, and findings with the in-app run.

Packs observe source data and do not directly change Azure.
AI text and thresholds require human review; Nothing notable does not prove health.
Snooze suppresses notification, not execution.
External delivery may expose operational metadata; minimize sensitive text.
Disable a bad schedule immediately, correct the pack, and test on demand before re-enabling.

Symptom	Resolution
Run has no evidence	Check scope, adapters, source scans, connection, and lookback.
Too noisy	Narrow sources/filters, raise materiality, test, or snooze while tuning.
Schedule did not run	Check pack and schedule enabled state, time zone, next run, and scheduler history.
No notification	Verify verdict gate, notify setting, and connector delivery health.
AI degraded	Use deterministic rows/counts; check provider before regenerating the definition.