AWS builds healthcare appointment voice agent on Amazon Nova 2 Sonic

AWS published a reference architecture for building a healthcare appointment voice agent powered by Amazon Nova 2 Sonic, its speech-to-speech model, running on Bedrock AgentCore. The agent authenticates patients by voice, books and manages appointments, collects pre-visit information, and escalates to human staff when a call exceeds its scope — a full conversational workflow rather than a scripted IVR menu.

The technical interest is in Nova 2 Sonic itself: a low-latency, natural speech-to-speech model is what makes a voice agent feel usable rather than robotic, and pairing it with AgentCore gives the agent the tool access and orchestration needed to actually act on calendars and records. The healthcare framing is deliberate — no-shows and routine call volume are expensive operational problems where automation has clear ROI, and voice is the natural interface for that population.

It fits AWS's broader AgentCore expansion this month, which added web search, managed knowledge bases, and policy guardrails, and reflects the agentic-AI-as-platform theme dominating the industry. Competitively, AWS is racing Google (which just shipped computer use), Microsoft, and a wave of voice-agent startups for the enterprise conversational-AI market.

The caveats in healthcare are acute: voice authentication, HIPAA compliance, and safe escalation are non-negotiable, and an agent that mishandles a medical scheduling request creates real harm and liability. AWS's emphasis on escalation-to-human and structured pre-visit collection reflects awareness of those limits, but production deployments will be judged on reliability and auditability, not demo polish.