GraceCall

Microsoft Agents League Hackathon · Enterprise Agents track

Most agents are chatbots. GraceCall calls you.

A rental car goes overdue. No staff involved. GraceCall detects it, dials the customer, holds a real two-way conversation, and if they promise to return it — re-checks automatically and calls again if they don't. Every decision within policy. Every call logged.

---

graph TD
    classDef ms      fill:#0078D4,stroke:#005a9e,color:#fff,rx:6
    classDef svc     fill:#5c2d91,stroke:#3b1a5a,color:#fff,rx:6
    classDef groq    fill:#f97316,stroke:#c2530f,color:#fff,rx:6
    classDef pa      fill:#107c10,stroke:#0a5a0a,color:#fff,rx:6
    classDef phone   fill:#1f2937,stroke:#374151,color:#e5e7eb,rx:6

    PA["⏱️ Power Automate\nevery 1 hr"]:::pa

    subgraph msft ["  🟦  Microsoft Platform  "]
        CS["🤖 Copilot Studio\nGraceCall Agent"]:::ms
        FIQ[("📚 Foundry IQ\nPolicy · Rate Card · Agreement")]:::ms
        TOOL(["TriggerOverdueCall\ncustom action"]):::ms
        ACSSVC["📞 Azure Communication Services\nCall Automation · PSTN"]:::ms
        AZURETSS["🔊 Azure Cognitive Services\nAvaNeural TTS"]:::ms
        M365OUT["💬 Microsoft 365 Copilot\nOps queries in Teams"]:::ms
    end

    subgraph gcsvc ["  🟣  GraceCall Service · Azure Container Apps  "]
        DECIDE["decideObjective()\nrecover · extend · charge · escalate"]:::svc
        RECHECK["Re-check Scheduler\n2nd call if not returned"]:::svc
        DASH["Live Dashboard\nContainer Apps · /dashboard"]:::svc
    end

    subgraph groqtier ["  🟠  Groq — free tier  "]
        WSTT["Whisper STT\ntranscribes caller audio"]:::groq
        LLAMA["LLaMA 3.3-70B\nconversation reasoning"]:::groq
    end

    PHONE["📱 Customer Phone"]:::phone

    PA --> CS
    CS <-->|"retrieves policy"| FIQ
    CS --> TOOL --> DECIDE
    DECIDE --> ACSSVC
    ACSSVC <-->|"PSTN call"| PHONE
    PHONE -->|"caller audio"| WSTT
    WSTT --> LLAMA
    LLAMA --> AZURETSS
    AZURETSS -->|"Vera's voice"| ACSSVC
    DECIDE --> RECHECK
    RECHECK -->|"fires follow-up"| DECIDE
    DECIDE --> M365OUT
    DECIDE --> DASH

Loading

---

Why it's different

• It makes a real phone call. Not a notification. Not a chat. An actual PSTN call with a natural voice (Vera), barge-in, and two-way conversation.
• It reasons, not scripts. Same agent, two rentals, opposite decisions — booked SUV gets a recovery call, idle economy car gets an extension offer. Driven by live data + Foundry IQ policy.
• Fully autonomous loop. Detects overdue → calls → logs promise → re-checks → calls again if needed. Zero human involvement.
• Guardrails in code, not just prompts. Tools reject over-charges and over-extensions even if the model tries. Enterprise-grade, not demo-grade.
• Responsible AI built in. AI disclosure on sentence one. Do-not-call honored. No card numbers by voice. Escalates on distress. Attempt caps enforced.

---

Microsoft stack

Service	Role
Copilot Studio	Agent brain — instructions, Foundry IQ knowledge, TriggerOverdueCall action
Azure AI Foundry + Foundry IQ	Grounds every decision in retrieved policy, rate card, and rental agreement
Microsoft 365 Copilot	Ops surface — staff ask "What did Vera say to Alex?" directly in Teams
Azure Communication Services	Places the real outbound PSTN call and streams audio
Azure Cognitive Services	AvaNeural TTS renders Vera's voice
Azure Container Apps	Hosts the backend (always-on)
Power Automate	Polls for overdue rentals every hour, no human trigger needed

---

Track requirements

Criterion	How GraceCall meets it
Authored in Copilot Studio	Agent instructions + Foundry IQ knowledge + TriggerOverdueCall custom action — see copilot-studio/
Microsoft IQ layer	Foundry IQ knowledge retrieved per call; decisions cite the policy
M365 Copilot surface	Agent published to Teams; outcomes queryable by ops
Real business scenario	Rental-overage recovery with real decisions and real calls
Responsible AI	Disclosure, do-not-call, no card-by-voice, escalation, caps — enforced in code
Full agentic autonomy	End-to-end loop: detect → call → re-check → follow-up, no human in the loop

---

Works across industries

Same architecture. Swap the Foundry IQ knowledge and policy file. No code changes.

Condition detected	Agent action
Car rental overdue	Calls customer, negotiates return or extension, re-checks at promised time
Hotel checkout missed	Calls guest, offers late checkout, follows up if room not vacated
Appointment no-show	Calls patient, reschedules within policy, flags to care team
Equipment not returned	Calls site contact, arranges return, escalates if no response
Payment missed	Calls customer, confirms plan within policy, logs commitment
Delivery window missed	Calls recipient, reschedules, updates dispatch automatically

---

Built on Azure Student Subscription

No paid credits. Real constraints, real workarounds.

Constraint	Solution
Azure Voice Live not available on student tier	Bridged ACS audio stream to Groq Whisper + LLaMA over WebSocket at 24kHz PCM — no resampling
No static HTTPS URL	ngrok tunnel; CALLBACK_BASE_URL in .env
Global Standard model (no realtime tier)	Fine for orchestration; latency-sensitive conversation runs on Groq
Hard spending caps	AUTO_DIAL=0 default; re-check is opt-in; dashboard gives manual control
ACS number provisioning	Successfully provisioned a real outbound number on the student subscription

---

Run it

cp .env.example .env   # fill in keys
npm install
npm run dev            # :8080

Key env vars: ACS_CONNECTION_STRING, ACS_CALLER_ID, GROQ_API_KEY, AZURE_SPEECH_KEY, CALLBACK_BASE_URL, TRIGGER_API_KEY

npm run trigger:demo          # place a call for RNT-1001
npm run trigger:demo RNT-1002 # extend scenario

The live dashboard at https://grace-call.greenplant-d2f64cf8.eastus.azurecontainerapps.io/dashboard surfaces the full agentic loop in real-time: call transcript, decision objective, escalation status, re-check countdown, and follow-up trigger — updating as Vera conducts the conversation.

---

License

MIT. Original work for the Microsoft Agents League Hackathon.