Natera’s intelligent appointment scheduling with Amazon Bedrock AgentCore
Amazon Web Services Cem Onan
Natera built a voice agent on AWS that books home phlebotomy visits for oncology patients. It replaces manual scheduling calls and keeps the conversation moving while backend checks run.
Based on reporting by Amazon Web Services, Cem Onan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Natera is using Amazon Bedrock AgentCore to automate appointment booking for patients who need blood draws at home. That sounds simple until you look at the moving parts: authentication, SMS verification, vendor availability, and the need to keep the conversation calm and natural while all that work happens in the background.
The company’s old setup ran on Amazon ECS, used Twilio for phone calls, and leaned on a third-party AI provider for voice and orchestration. It worked, but the team wanted better accuracy, easier scaling, and a less awkward experience for patients who are already dealing with cancer treatment. The new system keeps the human feel of a phone call, but shifts the heavy lifting to AgentCore.
Natera’s design has three main ideas. One WebSocket handles telephony, another handles model inference, and an orchestration layer sits between them. When the agent needs to check something like appointment availability, it can send a brief filler response instead of leaving silence. And instead of forcing full verification at the start, the system starts with low trust and grants more access as the patient is authenticated.
The migration from ECS to AgentCore runtime also meant reworking how sessions and connections behave. Container-local state had to move into AgentCore memory, and long-lived WebSockets had to be adapted to an invocation-based runtime. The team used connection pooling during a call and keyed state by actor ID so conversation history could survive beyond a single container lifecycle.
AWS says the system reached 100% tool-calling accuracy in 500 end-to-end call simulations, with sub-7-second perceived latency and a cost of less than USD 0.01 per completed call. That’s the kind of number that gets attention because it’s not just about sounding smart on a demo; it’s about making a scheduled medical call feel like a normal phone call instead of a support ticket.
My take — AI-written commentary, not fact-checked reporting
This is the right way to use agents: narrow, operational, and chained to real constraints instead of vague “AI transformation” theater. Healthcare doesn’t need more chatbot poetry; it needs systems that can verify, route, and remember without making patients sit in silence. The industry keeps pretending reliability is boring, which is usually how it ends up being the only thing that matters.
Read more about this at: Amazon Web Services