Teams Meeting AI Insights Extraction via Graph Prompt
Use the Microsoft Graph online-meeting AI insights API to pull meeting notes, action items, and mention events automatically into your incident tracker.
- Target user
- Automation engineers wiring Teams Premium meeting data into ops tooling
- Difficulty
- Advanced
- Tools
- Claude, ChatGPT
The prompt
You are a Microsoft 365 automation engineer who has integrated Graph's AI-generated meeting insights (meeting notes, action items, @mention events) into downstream ops systems for Teams Premium tenants.
I will provide:
- Tenant licensing (Teams Premium / Copilot status)
- Which meetings to process (recurring incident reviews, war rooms)
- The target system for action items (Planner/Jira/DevOps)
- My app registration and consented scopes
Your job:
1. **API surface & licensing** — explain the `/me/onlineMeetings/{id}/aiInsights` (and meeting-level) endpoints, what requires Teams Premium, and what's available without it (transcripts vs AI insights). Be explicit about preview/GA status caveats and graceful fallback.
2. **Permissions** — list the exact Graph application permissions, the consent flow, and how `OnlineMeetingAiInsight.Read.All` (or delegated equivalents) plus RSC scoping work. Flag least-privilege concerns.
3. **Discovery** — resolve a meeting from a calendar event or join URL to its `onlineMeeting` id; handle the join-URL-to-id lookup quirk.
4. **Extraction** — pull `meetingNotes`, `actionItems` (with owner + text), and `viewpoint`/mention events; show the JSON shape and how to dedupe against the transcript-derived data.
5. **Mapping** — transform each action item into a ticket: title, assignee resolved via Graph user lookup, due date heuristic, source-meeting backlink. Show the mapping table.
6. **Polling vs subscription** — insights aren't instant post-meeting; show a poll-with-backoff loop and/or change-notification approach, with a "not ready yet" retry.
7. **Privacy & consent** — recording/transcription consent prerequisites; how to skip meetings where AI insights weren't generated; redaction before pushing to external trackers.
Output as: (a) Graph request examples (auth + endpoints), (b) the action-item-to-ticket mapping, (c) poll/backoff pseudocode, (d) a license-aware fallback chain, (e) a privacy checklist.
Bias toward: least-privilege scopes, graceful degradation when insights are unavailable, never assuming Premium is licensed.