Designing a multi-agent AI platform that guides clinic staff through every patient touchpoint — from scheduling and billing to post-operative follow-up and insurance appeals.
Every patient interaction required staff to manually pull records, verify insurance, coordinate between providers, and document outcomes across disconnected systems. High cognitive load led to errors, delays, and patient dissatisfaction — especially at surgical handoff points.
The patient sits at the center of every decision. Four clinical roles coordinate around them, with AI agents handing off context at every step — forming a closed loop that always returns value back to the patient.
Two end-to-end calls powered by Deepgram. Aura-2 plays the dialogue while Nova-3 Medical transcripts drive the right-panel checklist; HITL popups pause audio at every decision point so you drive the agents yourself.
Each persona runs on its own bundle of agents inside the CareAssist platform. Click any use case to see the screen that persona works in.
The core design challenge was establishing appropriate AI boundaries. Each agent needed clear escalation paths — moments where the system explicitly deferred to human expertise rather than proceeding autonomously. This was especially critical for clinical and billing decisions, where errors carry real consequences for patients.
The real-time coaching layer became the key design mechanism for building staff trust — giving humans visibility into what the agent was doing and why, at every step.