The United States government wants to build an AI that can renew a heart failure patient’s diuretic prescription without a doctor in the loop, and it wants the FDA to sign off on it within two years. On Wednesday, ARPA-H, the federal agency that funds high-risk health research, put $62.7 million behind that idea.
The program is called ADVOCATE, a strained acronym for Agentic AI-Enabled Cardiovascular Care Transformation. Its stated goal is a clinical AI system that acts as a digital member of the care team: assessing symptoms, ordering labs, adjusting medication and escalating to humans when something looks wrong. The first-year commitment is $33.7 million, with the rest subject to renegotiation as the four-year effort unfolds.
Who got the money
ARPA-H named six initial awardees across three tracks, and said it may still fund more. The patient-facing agents will be built by three companies. Atman Health is developing a voice-first system in which the patient talks to the AI and the model chooses which diagnostic questions to ask next. Tempus AI is extending its existing Olivia health app with continuous monitoring that triggers deeper analysis when a patient’s numbers shift. UpDoc is building a system in which the conversational layer is deliberately separated from clinical authority by a clinician-written rules engine that checks every proposed action against approved protocols before it happens. UpDoc’s team includes OpenAI, Microsoft and NVIDIA, and ARPA-H noted the company already holds FDA clearance for a software medical device with a patient-facing language model.
Stanford University takes the second track, a supervisory AI whose job is to watch the other agents. Its design runs suspicious outputs through three escalating stages: outlier filtering, rule-based screening and finally a deep-research auditing agent that produces a written rationale for each flagged claim. The point is an assurance layer that regulators and hospitals can inspect.
The third track is deployment. Duke University will run validation across five health systems and rural sites on both Epic and Oracle electronic records, with the American Heart Association helping with reach. Kaiser Permanente will embed the agents into Epic workflows across its 21 medical centers and more than 260 clinics, starting with shadow-mode deployments and moving to pragmatic randomized trials. Kaiser’s own research division said its award is worth up to $16 million over three years. Atman Health’s award record lists up to $7.7 million.
Why heart failure
Roughly 6.7 million Americans live with heart failure, according to figures cited by STAT, and a large share never receive optimal treatment because cardiologists are scarce outside major cities. Guideline-directed therapy for the condition is well established. The medications are known. What is missing is someone to titrate doses, respond to a weight gain or a new symptom, and keep patients on track between appointments. That gap is precisely the kind of task an always-on software agent could fill, which is why ARPA-H picked it as the first target.
The agency’s own estimate is that a working system could generate around $28 billion in annual savings across the heart failure population alone, mostly by preventing hospitalizations. That figure deserves the usual caution attached to any projection made before the product exists.
Deliberately high risk
Haider Warraich, the practicing cardiologist who manages the program, has been unusually direct about the ambition. Speaking at an NIH Collaboratory Grand Rounds session, he said the low-risk functions such as scheduling and cost estimates are welcome, but they are not the point.
“We really want to specifically focus on things that are unambiguously high-risk,” Warraich said. That includes making treatment recommendations from a patient’s record, providing a diagnosis and performing triage, and, in his words, “being able to change or renew or refill existing prescriptions, or even write new prescriptions.” The medications in scope are limited to heart failure therapy and diuretic management, but the principle is clear: this is meant to be an AI that acts, not one that merely suggests.
The contractual deadline reinforces that. Teams on the patient-facing track must submit what ARPA-H describes as a first-of-its-kind FDA authorization package within 24 months of the award. No autonomous prescribing agent has been through that process. ADVOCATE is designed to force the question of what such a review would even look like.
What could go wrong
Plenty. Language models are known to produce confident errors, and a confident error in diuretic dosing can put someone in an emergency room. That is the reason the program’s architecture puts so much weight on Stanford’s supervisory layer and on UpDoc’s rules engine sitting between the chatbot and the chart. It is also why Kaiser is starting in shadow mode, where the agent proposes but humans still act.
The other uncertainty is trust. Rural patients who already struggle to reach a cardiologist are the intended beneficiaries, and they are also the population with the least room for a bad first experience. Duke’s multi-site validation is the program’s answer to that, though the results will not arrive for years.
What is different about this week is that a federal agency has stopped treating autonomous clinical AI as a hypothetical. The money is committed, the teams are named and the clock toward an FDA filing has started. The next signal to watch is how the agency and the FDA define “safe enough” for a machine that writes prescriptions.
For more coverage of AI in health, visit Mylistingo.
