IND Consulting — Outbound AI Voice Agent

📋 Outbound Call Worklist

🗄 System of Record

✓ Tasks / Notes Created

No tasks yet — complete a call.

☎ Call

ℹ What this does — and how we'd build it for real

What the AI agent does (capabilities)

Places outbound calls proactively to patients and doctors' offices — it initiates the call, it doesn't wait for someone to call in.
Natural two-way conversation driven by an LLM — greets warmly, states its purpose, asks and answers, and adapts to whatever is said (no rigid phone-menu or fixed script).
Interruptible (barge-in) — the person can cut in any time and the agent yields and responds, like a real call.
Human-sounding voice, with a distinct male/female voice for each speaker.
Four DME workflows: monthly insurance-change check (CA 30-day rule) · medical-review reminder that books the appointment directly · appointment scheduling with the doctor's office · chart-notes & CMN requests from the doctor's office.
Captures structured data mid-call — carrier, policy/group #, effective date, at-risk medical group, appointment date/time, fax timeframe.
Writes a task/note back into the system of record with a summary + the captured fields; escalates to a human only for clinical questions, disputes, or when asked.
Multi-tenant — one agent; each provider onboards as a config pack (their systems, phone numbers & rules). Owens runs on BrightTree.

How we build it for real (Azure / cloud / open-source)

Telephony (real dialing): Azure Communication Services (Call Automation) · Twilio / Amazon Connect · open-source Asterisk / FreeSWITCH + SIP — or the client's Talkdesk voice APIs.
Real-time voice loop (turn-taking, barge-in, silence detection): Azure Voice Live API · LiveKit Agents / Pipecat (open-source) · Vapi / Retell.
Speech-to-Text: Azure AI Speech (streaming) · Deepgram / Amazon Transcribe · open-source Whisper / faster-whisper (self-hosted for PHI control).
LLM brain: Azure OpenAI (GPT-4o) with tool/function calling · AWS Bedrock (Claude) / Google Vertex · open-source Llama 3 / Mistral on-prem; orchestrated with Semantic Kernel / LangGraph.
Text-to-Speech: Azure Neural TTS (incl. a branded Custom Neural Voice) · ElevenLabs / Amazon Polly · open-source Piper / Coqui.
Knowledge & rules: Azure AI Search / pgvector to ground on policy docs (RAG) + a deterministic rule engine for the CA 30-day rule, responsible-payer (DOFR), and scheduling policy.
System write-back: BrightTree · HDMS · DocFlow · Talkdesk · Pulse/Prochant via REST / HL7 / FHIR APIs — or RPA where no API exists; glued with Azure Logic Apps / Functions + Service Bus.
Data & compliance: Azure SQL / Cosmos DB (call state, tasks), Blob Storage (recordings), Key Vault, Private Link, HIPAA / BAA, PHI redaction, full audit trail, human review queue.
Run & monitor: Azure Container Apps / AKS · Application Insights + an evaluation harness · GitHub Actions CI/CD.

This demo → production, layer by layer

Layer	In this demo	In production
Dialing	Simulated in the browser	Azure Communication Services / Twilio / Talkdesk
Listening (STT)	Browser Web Speech API	Azure AI Speech / Whisper (streaming)
Brain (LLM)	Live LLM, dynamic each turn	Azure OpenAI (GPT-4o) + orchestration & rules
Voice (TTS)	Neural voice	Azure Neural TTS / ElevenLabs (custom voice)
Barge-in / turn-taking	JS on interim transcripts	Azure Voice Live / LiveKit / Pipecat
System of record	Mock JSON panel	BrightTree / HDMS / DocFlow APIs (HL7 / FHIR / RPA)
Data & compliance	None (demo only)	Azure SQL + Key Vault + HIPAA/BAA + audit trail

🗄 BrightTree — Patient / Order Records · Owens

records

Live BrightTree records — a completed call writes back to the matching row; changed cells show was → now.

📝 Write-back log

No writes yet — run a call in the Call Console (Auto Demo), and this dataset updates automatically.

🏗️ Reference architecture & data flow — outbound voice agent, end to end

Data flow

1Trigger — the scheduler + BrightTree / payer feeds surface records that need action (CA 30-day re-verify, review overdue, missing CMN).
2Decide — the Orchestrator loads the record, consults the Rules engine + Knowledge/RAG, and picks the call type & goal.
3Dial — it instructs Telephony to place the outbound call.
4Answer — the callee (patient or MD office) picks up; audio streams both ways.
5Listen — caller speech → STT → text.
6Converse — the LLM decides the next line & captures fields → TTS → speech back; repeats each turn, with barge-in.
7Hand-off — on call end, transcript + captured fields go to the Extractor / Validator.
8Write-back — validated results update BrightTree, create a task/note in DocFlow, book the appointment, or request documents.
9Escalate — anything needing a person goes to the human review queue; every step is logged for audit & monitoring.

How it decides what to call

Reads the CRM/EMR + payer feeds and finds records that need action — insurance re-verify due under the CA 30-day rule, medical review overdue, or a missing/expiring CMN.
Classifies each record into a call type: patient insurance check, patient reminder, MD-office appointment, or MD-office records/CMN.
Routes patient vs doctor-office calls, picks the right number, and loads only the context it needs (HIPAA-minimal).

What happens after the call

Extracts structured fields from the conversation (carrier, IDs, dates, appointment slot, fax ETA).
Validates against the rules, then writes back to the system of record and creates a task/note for staff.
Escalates to a human only when needed; logs a full audit trail and (optionally) the recording.