Play 39 — AI Meeting Assistant
Real-time meeting transcription, speaker diarization, action item extraction, summary generation, and follow-up scheduling integrated with Microsoft Teams/Outlook.
Architecture
| Component | Azure Service | Purpose |
|---|---|---|
| Transcription | Azure AI Speech | Real-time speech-to-text with continuous recognition |
| Diarization | Azure AI Speech (Conversation Transcription) | Speaker identification (up to 36 speakers) |
| Summarization | Azure OpenAI (GPT-4o) | Executive summary, key decisions, topic segmentation |
| Action Items | Azure OpenAI (GPT-4o-mini) | Structured extraction: owner + task + deadline + priority |
| Integration | Microsoft Graph API | Teams bot, Outlook summary, Planner tasks, Calendar follow-ups |
| Hosting | Azure Container Apps | Scalable meeting processing runtime |
| Secrets | Azure Key Vault | Speech key, OpenAI key, Graph credentials |
How It Differs from Related Plays
| Aspect | Play 04 (Call Center Voice) | Play 33 (Voice Agent) | Play 39 (Meeting Assistant) |
|---|---|---|---|
| Focus | Customer support calls | Conversational voice bot | Multi-party meeting intelligence |
| Speakers | 2 (agent + customer) | 2 (user + bot) | 2-36 (meeting participants) |
| Diarization | Optional | N/A | Core feature (who said what) |
| Output | Sentiment + resolution | Spoken response | Summary + action items + follow-ups |
| Integration | CRM, ticketing | Voice UI | Teams, Outlook, Planner |
DevKit Structure
39-ai-meeting-assistant/
├── agent.md # Root orchestrator with handoffs
├── .github/
│ ├── copilot-instructions.md # Domain knowledge (<150 lines)
│ ├── agents/
│ │ ├── builder.agent.md # Transcription + diarization + extraction
│ │ ├── reviewer.agent.md # PII, consent, summary quality
│ │ └── tuner.agent.md # Accuracy, cost, integration tuning
│ ├── prompts/
│ │ ├── deploy.prompt.md # Deploy meeting assistant
│ │ ├── test.prompt.md # Test transcription pipeline
│ │ ├── review.prompt.md # Audit PII + consent
│ │ └── evaluate.prompt.md # Measure extraction accuracy
│ ├── skills/
│ │ ├── deploy-ai-meeting-assistant/ # Full deployment procedure
│ │ ├── evaluate-ai-meeting-assistant/ # WER, DER, action item F1, summary scoring
│ │ └── tune-ai-meeting-assistant/ # Diarization, summarization, cost tuning
│ └── instructions/
│ └── ai-meeting-assistant-patterns.instructions.md
├── config/ # TuneKit
│ ├── openai.json # Summarization model settings
│ ├── guardrails.json # PII, consent, quality thresholds
│ └── agents.json # Teams/Outlook/Planner integration
├── infra/ # Bicep IaC
│ ├── main.bicep
│ └── parameters.json
└── spec/ # SpecKit
└── fai-manifest.jsonQuick Start
# 1. Deploy infrastructure
/deploy
# 2. Test with sample meeting audio
/test
# 3. Review PII handling and consent
/review
# 4. Evaluate transcription + extraction quality
/evaluateCost
| Service | Dev | Prod | Enterprise |
|---|---|---|---|
| Azure AI Speech | $0 (Free) | $80 (Standard S0) | $300 (Standard S0) |
| Azure OpenAI | $40 (PAYG) | $280 (PAYG) | $900 (PTU) |
| Microsoft Graph API | $0 (Included) | $0 (Included) | $0 (Included) |
| Container Apps | $10 (Consumption) | $120 (Dedicated) | $350 (Dedicated HA) |
| Cosmos DB | $5 (Serverless) | $60 (800 RU/s) | $350 (4000 RU/s) |
| Blob Storage | $2 (Hot LRS) | $20 (Hot LRS) | $60 (Hot GRS) |
| Key Vault | $1 (Standard) | $3 (Standard) | $10 (Premium HSM) |
| Application Insights | $0 (Free) | $20 (Pay-per-GB) | $80 (Pay-per-GB) |
| Total | $58/mo | $583/mo | $2,050/mo |
Key Metrics
| Metric | Target | Description |
|---|---|---|
| WER | < 10% | Word Error Rate for transcription |
| DER | < 15% | Diarization Error Rate |
| Action Item F1 | > 82% | Precision + recall of extracted actions |
| Summary Faithfulness | > 4.5/5.0 | No hallucinated facts in summary |
| Groundedness | > 0.85 | FAI guardrail threshold |
| Cost per meeting hour | < $1.00 | Speech + LLM costs combined |
WAF Alignment
| Pillar | Implementation |
|---|---|
| Reliability | Continuous recognition with auto-reconnect, fallback to batch transcription |
| Security | Managed Identity for Key Vault, PII redaction, recording consent enforcement |
| Cost Optimization | GPT-4o-mini for extraction, batch mode for recorded meetings |
| Operational Excellence | Structured logging, per-meeting metrics, evaluation pipeline |
| Performance Efficiency | Real-time streaming (<500ms latency), chunked summarization |
| Responsible AI | PII detection, consent prompts, confidential content filtering |
FAI Manifest
| Field | Value |
|---|---|
| Play | 39-ai-meeting-assistant |
| Version | 1.0.0 |
| Knowledge | F1-GenAI-Foundations, O2-Agent-Coding, T2-Responsible-AI |
| WAF Pillars | security, reliability, responsible-ai, cost-optimization |
| Groundedness | ≥ 85% |
| Safety | 0 violations max |
Last updated on