Call center ai · Production

Clarus Care builds a generative AI-powered healthcare contact center prototype with Amazon Bedrock, Amazon Connect, and Amazon Lex

The problem

Healthcare practices managing high volumes of patient calls face long hold times, frustrated patients, and overwhelmed staff, while Clarus's legacy menu-driven IVR forced patients through rigid menu options that limited resolution of complex multi-intent needs and caused communication bottlenecks that could delay critical care coordination.

First attempt

The legacy IVR required patients to navigate rigid menus and relied on rigid name matching and extension numbers, making it unable to handle natural name variations or complex multi-intent requests in a single interaction.

Workflow diagram · grounded in source
1
Patient initiates contact
trigger
“A patient initiates contact through either a phone call or web chat interface.”
2
Connect routes contact
routing
“Connect processes the initial contact and routes it through a configured contact flow.”
3
Urgency assessment via Bedrock
ai_action
“the system immediately evaluates whether the situation requires urgent attention using Bedrock APIs. This first step makes sure that emergency cases are quickly identified and routed appropriately. The system uses a focused prompt that a…”
4
Multi-intent detection via Claude 3.5 Sonnet
ai_action
“Intent extraction uses Anthropic's Claude 3.5 Sonnet through Bedrock for detailed analysis that can identify multiple intents from natural language, making sure patients do not need to repeat information.”
5
Route appointment intents to scheduling module
routing
“If an 'appointment' intent is detected in the main handler, processing is passed to the scheduling module.”
6
Information collection via Nova Pro
ai_action
“Information collection employs a faster model, Amazon Nova Pro, through Bedrock for structured data extraction while maintaining conversational tone.”
7
Response generation via Nova Lite
output
“Response generation utilizes a smaller model, Nova Lite, through Bedrock to create low-latency, natural, and empathetic responses based on the conversation state.”
8
Smart transfer to staff
human_review
“Smart transfers to staff are initiated when urgent cases are detected or when patients request to speak with providers.”
9
Analytics pipeline processing
feedback_loop
“Conversation data is processed through an analytics pipeline for monitoring and reporting”
Reported outcome

Clarus Care developed a prototype generative AI contact center capable of handling multiple patient intents per call through voice and web chat channels, with smart transfer capabilities and an analytics pipeline for performance monitoring, providing a scalable foundation for their growing customer base.

Reported metrics
Patient calls handled annually15 million
Client retention rate99%
Users served across specialtiesover 16,000
Backend processing latency target<3 seconds
Show all 5 reported metrics
patient calls handled annually15 million
client retention rate99%
users served across specialtiesover 16,000
backend processing latency target<3 seconds
availability SLA requirement99.99%
Reported stack
Amazon BedrockAmazon ConnectAmazon LexAmazon NovaAnthropic's Claude 3.5 SonnetAmazon Nova ProAmazon Nova SonicAmazon CloudFrontAmazon S3Converse API
Source
https://aws.amazon.com/blogs/machine-learning/how-clarus-care-uses-amazon-bedrock-to-deliver-conversational-contact-center-interactions?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Clarus Care developed a prototype generative AI contact center capable of handling multiple patient intents per call through voice and web chat channels, with smart transfer capabilities and an analytics pipeline for…

What tools did this team use?

Amazon Bedrock, Amazon Connect, Amazon Lex, Amazon Nova, Anthropic's Claude 3.5 Sonnet, Amazon Nova Pro, Amazon Nova Sonic, Amazon CloudFront, Amazon S3, Converse API.

What results were reported?

Patient calls handled annually: 15 million; Client retention rate: 99%; Users served across specialties: over 16,000; Backend processing latency target: <3 seconds (source-reported, not independently verified).

What failed first in this deployment?

The legacy IVR required patients to navigate rigid menus and relied on rigid name matching and extension numbers, making it unable to handle natural name variations or complex multi-intent requests in a single interac…

How is this call center ai AI workflow structured?

Patient initiates contact → Connect routes contact → Urgency assessment via Bedrock → Multi-intent detection via Claude 3.5 Sonnet → Route appointment intents to scheduling module → Information collection via Nova Pro → Response generation via Nova Lite → Smart transfer to staff → Analytics pipeline processing.

Related call center ai cases