Voice AI for NEMT call centers: 75% call-handling efficiency, 98% scheduling accuracy

Operational context

Non-emergency medical transportation call centers serving phone-first patients and caregivers who need ride lookup, scheduling, cancellation, ETA checks, and reminders: with actions executed in live dispatch systems.

The bottleneck

NEMT call centers handle repetitive but time-sensitive calls. Many riders need phone access; automation had to execute real actions through transportation systems like TripMaster: not just answer questions.

• Manual agents spent hours on repetitive ride-status and scheduling requests
• Transportation data lived in vendor systems requiring controlled API tool-calling
• Errors in schedules or cancellations could affect healthcare appointment access

System limitations

Legacy call workflows could not scale with volume without adding headcount. Integrations with dispatch vendors required careful abstraction: each NEMT operator could use different transportation backends.

• No unified voice layer connected to TripMaster and third-party dispatch APIs
• Human agents were the only path for write actions (schedule, cancel, update)
• Limited observability on call outcomes and scheduling accuracy

AI opportunity

High-volume, rule-bound phone workflows are ideal for voice agents with tool-calling: when human escalation and audit trails are designed in from the start.

• Automate ride lookup, scheduling, cancellation, ETA, reminders, and alerts via phone
• Route exceptions to human coordinators with full transcript and context
• Measure scheduling accuracy and call efficiency as operator KPIs

Architecture changes

The architecture separates telephony, orchestration, model routing, and dispatch integrations: so voice flows can evolve without rewriting core transport logic.

• Twilio telephony ingress with call logging and recording policies
• .NET API orchestration layer for tool routing and validation
• LLM tool-calling mapped to TripMaster and vendor abstraction endpoints
• Fallback paths when API calls fail or confidence drops below threshold

Human-in-the-loop controls

Healthcare coordination requires human takeover when the agent is uncertain or when prohibited actions are requested.

• Warm transfer to live coordinators with transcript and attempted tool-call context
• Confidence thresholds before write actions hit dispatch systems
• Call logging and QA sampling for patient-facing language accuracy
• Kill switches per workflow during rollout

Integrations

• TripMaster dispatch APIs
• Twilio voice telephony
• Third-party NEMT vendor abstraction layer
• .NET backend orchestration services

Buyer lessons

• Phone-first operational AI must tool-call into dispatch: not sit beside it as a FAQ bot
• Vendor abstraction matters when NEMT operators use different transportation backends
• Measure scheduling accuracy and handle time, not model benchmark scores
• Human escalation is a feature, not a failure mode