Natera Implements AI-Driven Appointment Scheduling Using Amazon Bedrock AgentCore

Natera, a global diagnostics company specializing in cell-free DNA testing, has introduced an AI-powered voice scheduling system to improve the patient experience for mobile phlebotomy appointments. Powered by Amazon Bedrock AgentCore, the system allows patients to book appointments through natural conversation, replacing manual call-based scheduling. The architecture employs a dual-WebSocket bridge pattern, event-driven latency masking, and a progressive trust authentication model, ensuring tool-calling accuracy of 100% during validation and a perceived latency of under seven seconds.

The existing workflow involved manual scheduling with multiple steps, including patient authentication, appointment preference collection, and coordination with vendors. While functional, this process was limited in scalability and engagement. Transitioning to Amazon Bedrock AgentCore enabled Natera to build a fully managed, scalable, and flexible voice agent capable of handling complex vendor coordination and telephony requirements beyond pre-built solutions.

The system architecture separates telephony streaming from model inference via a dual-WebSocket bridge, facilitating independent swapping of telephony providers or foundation models. Latency masking is achieved through parallel filler responses, maintaining natural conversation flow even during backend delays. Authentication is managed through a progressive trust model, starting with phone number verification and escalating to full identity verification, with conversation context seamlessly transferred between sessions.

Migration from Amazon ECS to AgentCore runtime involved decoupling orchestration logic, refactoring the agent’s entry point, and migrating session state to AgentCore memory, which offers durable, cross-session persistence. Challenges included adapting long-lived WebSocket connections and managing session continuity, addressed through connection pooling and externalized state management.

The request flow begins with Twilio establishing WebSocket connections for inbound and outbound audio streams, with the agent orchestrating interactions and tool calls. Latency masking responses are generated based on calibrated tool latency distributions, ensuring smooth conversation flow. Patient authentication employs hashed phone numbers for initial sessions, with full verification performed later, and conversation history is retrieved from AgentCore memory and external event streams for comprehensive context.

The system also integrates Amazon Bedrock Knowledge Bases to answer patient inquiries about procedures, policies, and other FAQs with over 90% accuracy. The core scheduling workflow involves retrieving appointment options, confirming details, and storing appointments in patient profiles. Compliance and safety are maintained through Amazon Bedrock Guardrails.

Early validation shows increased call duration, reduced premature call endings, and higher patient satisfaction scores, indicating improved engagement and operational efficiency. Natera plans to expand this AI-driven approach across other operations, empowering non-technical teams to deploy conversational agents rapidly.

For organizations interested in adopting similar solutions, AWS recommends deploying agents on AgentCore runtime, integrating session memory APIs, implementing event-driven latency masking, and designing progressive authentication flows to enhance user experience and security.