A voice AI agent that books, confirms and cancels medical rides
A national non-emergency medical transportation provider coordinates rides for health plan members to get to their appointments, and handles millions of member calls a year. Its virtual agent was built on predefined utterances and intents, so only 5–8% of calls were resolved without an agent, and every new provider or phrase meant more maintenance. We built a voice AI agent on Amazon Nova Sonic and Amazon Bedrock. It verifies the member first, then schedules, confirms, cancels and checks rides in natural conversation. When a person is needed, it hands over to an agent in the existing Five9 contact center with a full briefing.
- Mid-teens
- containment, up from 5–8%
- >30%
- containment at peak
- 50–60K
- calls a day the platform is built for
- 4
- ride tasks handled end to end
- A medical transportation provider replaced an intent-based virtual agent with agentic voice AI on Amazon Nova Sonic and Amazon Bedrock.
- Containment rose from 5–8% to the mid-teens, with peaks above 30%.
- Members are verified first, and agents in the existing Five9 contact center get a full briefing on transfer.
- Many times more calls resolved without an agent, so call volume can grow without matching headcount.
- Added alongside the existing contact center: no big-bang replacement.
- An early production deployment of Amazon Nova Sonic.
- Agents receive only the calls that need them, with the member, intent and transcript.
- Dashboards show containment, transfers and abandonment by use case.
- Distress or low confidence always reaches a person.
- Nova Sonic speech-to-speech on Bedrock, with LiveKit Agents, LiveKit SIP and Kamailio on EKS.
- KEDA auto-scaling on concurrent calls, Redis session state, Lambda tool calls to backend APIs.
- IAM Identity Center, IRSA, Secrets Manager, TLS 1.2+ and DTLS-SRTP; Terraform throughout.
One call, from both sides
Follow a member’s call on the left and the voice agent on the right. Every call starts with an identity check, then the agent works out what the member needs, calls the ride system, and either finishes the job or hands over to a person with a full briefing. Pick a call to see each path.
- Call in
- Verify
- Understand
- Act
- Resolve
- ✓Date of birth
- ✓Member ID
- ✓Address
Member data locked
- Waiting for a request
- Call time
- 0:00
- Turns
- 0
- Actions taken
- 0
- Agent needed
- …
Illustrative call. Member, appointments and trips are fictional; the flow follows the production design: verify first, then act, and hand over to a person when it should.
The challenge
Members call to book a ride to dialysis, check where their driver is, or cancel a trip they no longer need. Most of these calls are routine, yet almost all of them reached a person: the old virtual agent resolved only 5–8% on its own.
It depended on predefined utterances mapped to intents. Every new provider, phrase or member scenario meant another manual update, so it got harder to maintain as the business grew, and staffing had to grow with call volume.
The provider wanted more than 30% of calls handled without an agent, and a natural conversation instead of a menu. It also wanted clear visibility into transfers and abandoned calls, and to start using AWS without the risk of replacing its whole contact center at once.
What made this different
Voice AI demos are easy. A voice agent that handles real members, real PHI and real volume is not.
- Mid-teens
share of calls fully handled by the AI, up from 5–8%, with peaks above 30%.
Measured on the same ride use cases before and after, from transfer tracking and transcripts.
Every caller’s identity checked before any member data is used.
Date of birth, member ID and address are a mandatory gate. Repeated failures go straight to a person.
Added alongside the existing Five9 contact center, not replacing it.
The AI sits in front of the existing contact center, so agents kept working as before while self-service grew.
An early production deployment of Amazon Nova Sonic.
A new model meant extra monitoring, retries and fallbacks in the agent, so a model hiccup never leaves a member stranded.
What changed
- Self-serviceBefore: 5–8% of calls containedAfter: Mid-teens on average, peaks above 30%
- ConversationBefore: Predefined utterances mapped to intentsAfter: Natural speech-to-speech conversation on Amazon Nova Sonic
- MaintenanceBefore: Manual updates for every new phrase or providerAfter: The model understands variations; workflows are modular
- Member dataBefore: Identity checks handled by agentsAfter: Mandatory verification (date of birth, member ID, address) before any PHI or action
- HandoffBefore: Cold transfers; members repeated themselvesAfter: Warm transfer to Five9 with member, intent summary and transcript
- VisibilityBefore: Little view of transfers, abandonment or failure pointsAfter: Grafana dashboards, transcripts and recordings for every call
- ScaleBefore: Staffing grew with call volumeAfter: Auto-scales with concurrent calls on Amazon EKS
- Contact centerBefore: Full replacement was the only modern optionAfter: AI added alongside Five9: phased, no disruption to agents
Hover or tap a tile to see what it replaced.
Far more calls resolved without an agent.
Containment was measured on the same member use cases before and after, from the old virtual agent’s reporting and the new dashboards, transcripts and transfer tracking.
- Calls resolved without an agentBefore5–8%AfterMid-teens
- Utterance and intent upkeepBeforeManual, constantAfterNot needed
- View of transfers and abandonmentBeforeLimitedAfterEvery call, on dashboards
Containment on ride scheduling, confirmation, cancellation and status calls. Target: consistently above 30%.
How a member call is handled now
The member just talks. The agent checks who they are, does what they ask, and passes the call to a person, fully briefed, when it should.
- 01 · CallA member callsThe call arrives over SIP, is balanced by Kamailio and bridged to the voice agent by LiveKit.
- 02 · VerifyWho is calling?Date of birth, member ID and address are checked before any member data is touched.
- 03 · UnderstandNova Sonic listens and repliesSpeech in, reasoning, speech out, in one model, with natural turn-taking.
- 04 · ActTools do the workThe agent calls the provider’s APIs to schedule, confirm, cancel or look up a ride.
- Ride scheduled
- Ride confirmed
- Ride canceled
- Status given
- Warm transfer
- 05 · HandoffA person, if neededLow confidence, distress, or anything out of scope goes to a Five9 agent with the member, intent and transcript.
How we ran it
- 1RequirementsVolume, latency, HIPAA, Five9 handoff, scaling and visibility, agreed with the provider’s operations leaders.
- 2Choose the modelAmazon Nova Sonic: speech-to-text, reasoning and speech in one model, low latency, HIPAA-eligible. Self-hosted open models, custom SageMaker models and third-party voice platforms were rejected.
- 3Design the conversationGreeting and caller context, member verification, then schedule, confirm, cancel and status flows, each with escalation paths.
- 4Build the platformLiveKit and Kamailio on EKS across Availability Zones, Redis for shared session state, Lambda to the provider’s APIs, all in Terraform.
- 5TestFlow, integration, security and load testing at production volumes, plus accents, background noise and edge cases, then UAT with business experts.
- 6Tune in productionPrompt refinement, escalation thresholds and analytics reviews with the provider, release by release.
The outcome
- Self-service containment up from 5–8% to the mid-teens, with peaks above 30%
- Rides scheduled, confirmed, canceled and checked in natural conversation
- Members verified before the agent can see PHI or act
- Warm transfers to the existing Five9 contact center with a full briefing
- An architecture designed and load-tested for 50,000–60,000 calls a day
- Transcripts, recordings and real-time dashboards for every interaction
- AI
- Amazon Nova Sonic
- Amazon Bedrock
- Tool calling to backend APIs
- Voice
- LiveKit Agents and SIP
- Kamailio SIP load balancer
- WebRTC (DTLS-SRTP)
- Compute
- Amazon EKS with KEDA auto-scaling
- AWS Lambda
- Amazon ElastiCache for Redis
- Data
- Amazon DynamoDB call records
- Amazon S3
- Monitoring
- Amazon CloudWatch
- Prometheus
- Grafana
- Security
- IAM Identity Center
- IAM roles per pod
- AWS Secrets Manager
- Private subnets, TLS 1.2+
- Contact center
- Five9 (existing)
- Warm transfer
Three lessons from an early production deployment of a new voice model. New AI services need more monitoring, retries and fallbacks than ordinary software. Call flows keep changing once the business sees what is possible, so the conversation has to be modular and easy to release. And getting the agent to sound right takes continuous prompt tuning and escalation tuning, reviewed with the business every cycle.
A partner AWS has validated, on exactly this kind of work.
- AWS Advanced Tier Services PartnerRecognized by AWS for proven delivery, certified people and customer outcomes.
- AWS Migration Services CompetencyValidated by AWS for moving and modernizing workloads on AWS at scale.
- Amazon Connect Delivery PartnerValidated by AWS to deliver Amazon Connect, including migrations from Avaya, Genesys, Cisco and other platforms.
- AWS AI Services CompetencyValidated by AWS for building AI into real products, including the contact center.
- AWS Experience-Based Acceleration (EBA) partnerWe run EBA engagements with AWS to speed up migrations with hands-on, time-boxed sprints.
- ISO/IEC 27001 certifiedOur information security management system is independently certified to the international standard.
- SOC 2 Type IIIndependent auditors have tested our security controls over time, not just at a single point.
More work
All case studies →- Amazon Connect
- Amazon Lex
- AWS Lambda
- Amazon Connect
- Contact flows
- AWS Lambda
- Amazon EKS
- Amazon Kinesis
- Amazon DynamoDB








