Play 54 — AI Customer Support V2
Advanced AI customer support — multi-channel (chat, email, voice), intent classification with slot filling, knowledge-grounded responses via AI Search, sentiment-aware escalation with priority routing, session memory in Cosmos DB, CSAT tracking, and channel-specific formatting (markdown/HTML/SSML).
Architecture
Full architecture details:
architecture.md
How It Differs from Related Plays
| Aspect | Play 05 (IT Ticket Resolution) | Play 54 (Customer Support V2) | Play 08 (Copilot Studio Bot) |
|---|---|---|---|
| Audience | Internal IT staff | External customers | Internal/external |
| Channels | Ticketing system | Chat + email + voice | Teams/web |
| Intent | IT issue classification | Customer intent + sentiment + slots | Topic-based routing |
| Escalation | SLA-based routing | Sentiment-aware + confidence-based | Topic fallback |
| Grounding | IT knowledge base | Product KB + policies + troubleshooting | Copilot Studio topics |
| Metrics | MTTR, resolution rate | CSAT, auto-resolution, first-contact | User satisfaction |
DevKit Structure
54-ai-customer-support-v2/
├── agent.md # Root orchestrator with handoffs
├── .github/
│ ├── copilot-instructions.md # Domain knowledge (<150 lines)
│ ├── agents/
│ │ ├── builder.agent.md # Multi-channel + intent + KB
│ │ ├── reviewer.agent.md # Response quality + escalation
│ │ └── tuner.agent.md # Intent + CSAT + cost
│ ├── prompts/ + skills/ + instructions/
├── config/ # TuneKit
│ ├── openai.json # Intent model + response model
│ ├── guardrails.json # Escalation rules, sentiment thresholds
│ └── agents.json # KB config, channel templates
├── infra/ + spec/Key Metrics
| Metric | Target | Description |
|---|---|---|
| Intent Accuracy | > 90% | Correct intent classification |
| Auto-Resolution | > 60% | Resolved without human agent |
| CSAT Score | > 4.0/5.0 | Customer satisfaction |
| Escalation Rate | 15-25% | Sweet spot: not too many, not too few |
| KB Grounding | > 95% | Responses from knowledge base only |
| Cost per Conversation | < $0.05 | ~3 turns average |
Cost Estimate
| Service | Dev | Prod | Enterprise |
|---|---|---|---|
| Azure OpenAI | $80 | $700 | $2,800 |
| Azure AI Search | $0 | $250 | $1,000 |
| Azure Communication Services | $10 | $80 | $300 |
| Cosmos DB | $5 | $75 | $350 |
| Container Apps | $10 | $100 | $400 |
| Azure Service Bus | $5 | $25 | $100 |
| Key Vault | $1 | $5 | $15 |
| Application Insights | $0 | $30 | $100 |
| Total | $111 | $1,265 | $5,065 |
Detailed breakdown with SKUs and optimization tips:
cost.json· Azure Pricing Calculator
WAF Alignment
| Pillar | Implementation |
|---|---|
| Reliability | Session memory, KB fallback, graceful “I don’t know” |
| Security | PII handling, Key Vault, no customer data in logs |
| Cost Optimization | gpt-4o-mini for intent, gpt-4o for response, KB caching |
| Operational Excellence | CSAT tracking, escalation analytics, resolution rate |
| Performance Efficiency | <1s response time (chat), parallel intent+KB search |
| Responsible AI | Sentiment-aware tone, no hallucinated policies, human escalation |
FAI Manifest
| Field | Value |
|---|---|
| Play | 54-ai-customer-support-v2 |
| Version | 1.0.0 |
| Knowledge | R2-RAG-Architecture, O2-Agent-Coding, T2-Responsible-AI, T3-Production-Patterns |
| WAF Pillars | security, reliability, cost-optimization, responsible-ai, performance-efficiency |
| Groundedness | ≥ 85% |
| Safety | 0 violations max |
Last updated on