TensorBlue
AI_VOICE_RECEPTION

GenAI
Receptionist

Real-time conversational reception for hospitals & clinics

Built on PIORS framework with SFMSS simulation. Outperforms GPT-4o on reception tasks with HIS integration and safety guardrails.

Scheduling
Multilingual
HIS Integration
PHI-Secure
Emergency Routing
CALLS_TODAY
247
FCR_RATE
78.4%
AVG_LATENCY
1.2s
PHI_INCIDENTS
0

Front desks and call centers in hospitals juggle complexity that legacy IVRs cannot handle

Generic chatbots can't integrate safely with hospital information systems or uphold escalation, audit, and PHI controls.

📞
High Call Volumes
Wait times & abandoned calls
🌍
Multilingual Patients
Language barriers
🔀
Complex Routing
Scheduling, referrals, billing, prep
🔒
Strict Privacy
PHI protection requirements
Real-time conversational reception with verifiable safety rails, audit trails, and HIS connectors
PIORS
Personalized Intelligent Outpatient Reception
Based on LLM with Multi-Agent Medical Scenario Simulation
GPT-4o
Outperformed
15 Experts
Clinical evaluation
Nov 2024
Research validated
SFMSS
Service-Flow-aware Medical Scenario Simulation

Multi-agent simulation generates patient/receptionist/service-flow dialogues for supervision and stress tests. Published in NAACL 2025 Findings.

HIS Integration
Hospital Information System Connectivity

PIORS is designed to collaborate with HIS for patient-specific answers and actions — demographics, appointment lookups, scheduling with strict RBAC.

arXiv:2411.13902 • NAACL 2025
Architecture
Voice → Tools → HIS
Telephony
SIP/PSTN → WebRTC gateway
1
Voice I/O
Streaming ASR/TTS with barge-in
2
Intent & Entity
Low-latency NLU + medical lexicons
3
LLM Core + Tools
PIORS + SFMSS policy constraints
4
HIS Connectors
Read demographics, write scheduling
5
Escalation
Route to human with transcript + intent
6
TOOLING_LAYER
Scheduler
FAQ RAG
SMS/Email
HIS Connector

Conversation Flow

1
Greeting
→
2
Identify
→
3
Intent Disambiguation
→
4
Action
→
5
Confirm
→
6
Closure
SFMSS encodes allowed service-flow transitions to constrain LLM behavior
EXAMPLE_POLICIES
Emergency Detection
"I'm not a clinician, but I can quickly connect you. If you think this could be an emergency, please call your local emergency number now."
Identity Verification
"Before I discuss appointments, I'll verify your date of birth. You can also say 'human' anytime to speak with staff."
Scheduling Confirmation
"I found Dr. Rao on Tue 10:40. Shall I book it and text you the prep instructions?"

Safety Guardrails

Non-Negotiable Controls
1
Identity & PHI
Verify caller using multi-factor cues (DOB + last name + OTP)
All PHI stays on-prem/VPC
Redact transcripts for analytics
2
Clinical Boundaries
No medical advice, triage, or interpretation
Reception scope only — route clinical matters to nurse line
Refusal templates for clinical advice
3
Emergency Handling
Emergency phrases (chest pain, bleeding) → immediate escalation
Safety filters on prompt & output
Instruct to call local emergency number
4
Audit & Logging
Log: audio hash, ASR text, intents, tool calls
Immutable decision log for every operation
Retention aligned to policy (who/what/when)
ZERO PHI INCIDENTS • BY DESIGN

Deployment

Shadow → Canary → Scale
PHASE
1
Shadow
ACTIONS
Record + simulate decisions • Zero customer impact
DURATION
2 weeks
PHASE
2
Canary
ACTIONS
10–20% calls in business hours • Staff barge-in hotkey
DURATION
4 weeks
PHASE
3
Scale
ACTIONS
After KPI gates pass (FCR, routing, safety)
DURATION
Ongoing
INTENT_CATALOG
Book/Cancel Appointments
Prep Info & Directions
Insurance Questions
Refills Routing
Billing Inquiries
General Clinic Info

Production KPIs

First-Call Resolution (FCR)
No human needed
≥ 70–80%
Routing accuracy
Correct department
≥ 95%
Scheduling success
Eligible requests completed
≥ 90%
Avg turn latency (voice)
TTFB end-to-end
≤ 1.0–1.5s
Escalation rate
With retraining
Trending ↓
PHI incidents
By policy
0
TRAINING_LOOP
1
Use SFMSS to generate challenging dialogues
2
Weekly hard-negative mining from real calls
3
Labeler resolves ambiguous cases
4
Retrain prompt/rules based on feedback
Operations
Tracing
•
Red-team tests
•
Adversarial fuzzing

Conversational reception
that's clinically aligned

Built on PIORS + SFMSS. Provably better than GPT-4o on outpatient reception tasks with HIS integration and safety guardrails.

70–80%
First-Call Resolution
≥95%
Routing Accuracy
<1.5s
Voice Latency
ZERO
PHI Incidents
Deploy AI ReceptionBook Demo Call
FRAMEWORK
PIORS + SFMSS
VALIDATION
Outperforms GPT-4o
DEPLOYMENT
On-Prem/VPC

Frequently Asked Questions

How is a GenAI receptionist different from a phone tree?

Natural language understanding (multilingual, accented), context across the whole conversation, real-time scheduling and CRM updates, and human escalation that includes full transcript and intent.

Can it handle inbound and outbound calls?

Yes - inbound for reception, support, and triage; outbound for appointment reminders, lead qualification, and renewals.

Related TensorBlue Resources

Voice AI Agents

Voice agents and workflow automation.

AI Chatbot Development

Web and messaging conversational AI.