Clinical AI is moving beyond a model that produces a recommendation for a clinician to review. ARPA-H’s ADVOCATE program is explicitly attempting something more operational: a patient-facing AI paired with supervisory agentic AI, designed to function around the clock as a digital member of a cardiovascular care team and ultimately pursue FDA authorization.
The program is designed around care delivery, not a standalone model
ARPA-H announced selected teams on Sept. 9, 2026, with up to $33.7 million committed in the first year. The program’s stated goal is to build a reliable, FDA-authorized clinical agentic-AI system for cardiovascular care.
ARPA-H frames the access problem directly: nearly half of U.S. counties lack a cardiologist. Its program design therefore combines technical performance with patient-facing interaction, supervisory logic, clinical integration, and a scalable deployment plan.
Patient-facing software changes the regulatory and safety boundary
FDA’s 2026 Clinical Decision Support guidance distinguishes certain non-device decision-support functions from software that remains subject to device regulation. FDA specifically notes that its existing digital-health policies continue to apply to software functions that meet the device definition, including functions intended for patients or caregivers.
That boundary matters for agentic systems because the deployment question is not only whether an AI can reason about clinical information. It is how authority is allocated, when humans must intervene, how outputs are monitored, how software changes are controlled, how cybersecurity is managed, and how patients understand what the system is doing.
“AI informs a clinician.”
The human user remains the primary operational actor.
“AI participates in an ongoing care workflow.”
That introduces supervision, escalation, patient interaction, monitoring, change management, and accountability as first-order system requirements.
“Human in the loop” is not a workflow specification.
Safe deployment requires concrete decision rights: what the AI may do, what it may recommend, what requires approval, what triggers escalation, what is logged, how changes are validated, and who owns the outcome when the system and the care team disagree.