ARPA-H Is Spending $62.7M to Build FDA-Authorized Autonomous Clinical AI. The Bigger Story Is the Infrastructure Around It.

ARPA-H is funding a four-year effort to build and validate autonomous clinical AI for cardiovascular care — including supervisory AI, health-system deployment, FDA authorization and reimbursement pathways.

ARPA-H announced the selected teams for its four-year, $62.7 million ADVOCATE program, with up to $33.7 million committed during the first year. The stated goal is unusually ambitious: create a reliable, FDA-authorized clinical agentic AI system capable of operating around the clock as a digital member of the care team.

Heart failure is the initial use case. But the architecture being built could matter far beyond cardiology.

This is not another clinical copilot

Most healthcare AI products still sit somewhere on the spectrum of summarization, prediction, documentation or decision support. ADVOCATE is attempting something materially different.

The patient-facing agents are intended to autonomously support heart-failure patients between visits and escalate cases to clinicians when necessary. Three companies were selected to build these systems: Atman Health, Tempus AI and Updoc.

Atman is developing a voice-first system that combines an evidence-based clinical decision engine with large language models. Tempus plans to extend its Olivia patient application with continuous monitoring and deeper clinical analysis. Updoc is separating conversational intelligence from clinical authority, using a clinician-built rules system to validate proposed actions before execution.

That separation between reasoning and permission to act may become an important design pattern for autonomous healthcare AI. It also echoes the execution-and-guardrail architecture emerging across companies tracked on the RCAI Market Map.

Then ARPA-H added an AI supervisor

Stanford University will develop a separate supervisory AI system designed to continuously monitor the clinical agents after deployment. The system is intended to identify unsafe recommendations and behavior outside expected parameters using multiple layers of screening and auditing.

The architecture is notable: AI agent → supervisory AI → human escalation → independent evaluation.

That is much closer to an operating architecture for autonomy than a traditional software product.

Kaiser and Duke are solving the deployment problem

Building an autonomous clinical agent in a controlled environment is one challenge. Putting it into real healthcare workflows is another.

Duke University will test ADVOCATE agents across five health systems and rural locations using both Epic and Oracle Health/Cerner environments. Kaiser Permanente will deploy across 21 medical centers and more than 260 clinics, embedding the agents into Epic workflows for heart-failure patients.

The deployments are expected to include shadow-mode testing and pragmatic randomized clinical trials. Becker's Hospital Review highlighted Kaiser and Duke's role in the $62.7 million program.

This resembles the enterprise deployment problem RCAI has been tracking in R1's Phare OS rollout with UF Health: the model is only one layer. Integration, workflow access, oversight and operating governance determine whether AI can actually function inside a health system.

FDA authorization is being designed alongside the technology

ARPA-H says it will work with the FDA throughout the program to develop a regulatory framework for this new category of patient-facing clinical AI. The patient-facing teams are expected to submit a first-of-its-kind FDA authorization package within 24 months of contract award.

The program also includes shared evaluation standards, interoperability requirements and reimbursement pathways. Johns Hopkins University's Applied Physics Laboratory will independently evaluate technical performance and clinical outcomes.

So the program is not simply build an AI agent. It is attempting to create the full chain: build → validate → supervise → integrate → regulate → reimburse → scale.

The RCM implications are easy to miss

ADVOCATE is a clinical program, not a revenue-cycle initiative. But the operating requirements look increasingly familiar.

Autonomous execution. Rules-based constraints. Continuous monitoring. Exception escalation. Human oversight. System interoperability. Auditability. Measurable outcomes.

Those same characteristics are appearing across the agentic RCM platforms tracked by RCAI. Stedi's push toward a headless RCM execution layer is one example: infrastructure beneath the agent becomes as important as the agent itself.

If healthcare can establish acceptable frameworks for an AI agent to participate in medication management and clinical decision-making, the threshold for allowing agents to autonomously work eligibility, prior authorization, claims, denials, coding and payment workflows begins to look different.

RCAI Take

The most important part of ADVOCATE is not the $62.7 million. It is the architecture.

ARPA-H has separated the problem into patient-facing autonomous agents, independent supervisory AI, health-system deployment infrastructure, external evaluation, FDA authorization and reimbursement.

The next phase of healthcare AI will not be determined only by which model reasons best. It will depend on which systems can reason, act, monitor themselves, escalate exceptions, integrate with existing infrastructure and prove that their actions are safe and economically valuable.

That is the same transition RCAI is watching across revenue cycle. ADVOCATE may become one of healthcare's clearest early blueprints for what production-grade autonomy actually requires.