ARPA-H Bets on AI to Speed Up Clinical Trials
ARPA-H's SURPASS program and three companion efforts, worth up to about $100M, aim to use AI to make clinical trials faster and leaner.

The US government is betting that artificial intelligence can shorten one of medicine's slowest processes: the clinical trial. The Advanced Research Projects Agency for Health (ARPA-H) has opened a program called SURPASS that seeks AI platforms for adaptive trials, and it announced three companion initiatives, with awards of up to roughly $100 million combined, aimed at trial sites, data infrastructure, and patient-initiated enrollment. The programs do not test any drug, but they could shape how new treatments are evaluated for years.
What was announced
As reported by Nextgov, the Department of Health and Human Services unveiled four efforts at once, with SURPASS the centerpiece. ARPA-H's program page expands the name as Simulation-augmented, Real-time Platform Adaptive Seamless Trials. It describes perpetual, adaptive platform trials that combine predictive computational models, shared infrastructure, common control groups, and real-time analysis, with the aim of evaluating drugs and biologics faster, at lower cost and with fewer participants than conventional designs.
The page frames the ambition as a question: whether trials that now take more than a decade could be cut to under four years. HHS context cited by Nextgov explains the motivation: clinical development of drugs and biologics often takes over a decade, costs an average of up to $2 billion, and fails more than 90% of the time. Those are HHS's figures, not independent estimates, but they describe a widely discussed problem.
The three technical areas
SURPASS is organized around three technical areas, according to both sources:
A phaseless design engine that simulates trials so researchers can run faster studies with fewer patients. The traditional separation of phases would give way to continuous, adaptive designs.
A continuous inference engine that supports real-time or on-demand analysis, so decisions such as dropping an arm or adjusting enrollment can be made as data arrives.
An agentic AI operations layer to automate trial startup, speed up data collection and support other procedures that currently consume months of administrative effort.
Program manager Daria Fedyukina said, in remarks reported by Nextgov, that the work is meant to let teams learn in real time and generate evidence faster with less operational burden.
Three companion initiatives
Nextgov reports three related ARPA-H efforts. Unlike SURPASS, each already has a named prime awardee:
STACK aims to create more domestic trial sites and lift enrollment by using AI to activate sites faster. Evidence Health is the prime awardee, with up to $41.18 million.
COMMONS targets nationwide data infrastructure, including consent architecture and regulatory-grade data access, covering more than 40 million individuals and 3 million linked records. Evidence Health is again the prime, with up to $49.9 million.
CINCH is a patient-initiated model in which cancer patients contribute their own data to help identify suitable trials. Courage Health is the prime awardee, with up to $8.95 million.
The University of Texas at Austin's Dell Medical School is partnering with Evidence Health on STACK and COMMONS, preparing trial sites and developing a national research data resource, the report says. The dollar amounts are ceilings ("up to"), not confirmed final payments.
Together they cover three stages of the same pipeline. STACK addresses where trials can run, COMMONS addresses what data they can draw on, and CINCH addresses how patients find them. SURPASS then asks for the methodological layer on top. The design suggests ARPA-H sees AI less as a single product than as connective tissue across a fragmented system.
How it fits earlier federal steps
The announcement builds on earlier moves, according to Nextgov. Operation TrialBlazer, launched in June 2026, is an HHS-wide effort that supports trials through pilot programs, public-private partnerships, and patient connections. In April, the Food and Drug Administration started a pilot that uses AI and cloud computing to monitor trial data in real time; an FDA official estimated it could cut overall trial times by up to 40% without compromising patient safety. That figure is an official's estimate for a pilot, and it has not been demonstrated at scale in the sources reviewed.
Timeline for applicants
SURPASS is the only one of the four efforts open to outside proposals. The program page lists an optional informational webinar on October 15, an optional proposers' day on November 6, a required solution summary due November 30, 2026, and full proposals due January 22, 2027. Submitting the summary is mandatory to proceed. Any selected platform is therefore unlikely to be operating in trials before 2027 at the earliest, and likely later.
Why it matters
For healthcare AI, trial operations may be a more tractable target than diagnosis. The tasks in question, such as site activation, data cleaning, eligibility matching, and statistical monitoring, are repetitive, document-heavy, and measurable. An agentic layer that cuts startup time does not need to make a clinical judgment, which sidesteps some of the validation burden attached to tools that influence individual patient care. If the programs work, faster trials would reach patients, especially in areas such as oncology where enrollment is a persistent bottleneck.
The government's role is also significant. Trial infrastructure is expensive to build and has few commercial champions, because the benefits spread across sponsors, sites and patients. Public funding for shared data resources and common control groups could lower costs for smaller sponsors, who cannot afford to build their own.
Open questions
Several issues remain unaddressed in the material reviewed. Adaptive and simulation-based designs depend on regulators accepting evidence generated with them, and the sources do not say how the FDA will treat results from a phaseless design. Privacy and consent are central to a program linking millions of records; the COMMONS description mentions consent architecture, but details are not public in the coverage. The possibility that simulation models embed biases from historical trial populations, which have often underrepresented women and minority groups, is a standard concern for this kind of work, though the announcement as reported does not discuss it.
There is also a procurement point. Evidence Health holds the prime role on two of the three companion efforts, so a large share of the near-term money flows through one organization. That is not unusual for a fast-moving agency, but it means the success of two programs depends on a single contractor's execution.
What to watch next
Watch the November 30 summary deadline and which teams emerge, since the mix of pharmaceutical, academic and software organizations will indicate how the field sees the opportunity. Watch for FDA guidance on adaptive platform trials and on the acceptable use of AI-generated evidence. And watch whether the first operating results from the STACK and COMMONS work show measurable gains in site activation time and enrollment. Those metrics, rather than the funding totals, will show whether AI can make trials faster in practice.
Frequently Asked Questions
What is SURPASS?
An ARPA-H program seeking AI platforms for perpetual, adaptive platform trials, with a phaseless design engine, a continuous inference engine and an agentic operations layer.
What are STACK, COMMONS and CINCH?
Companion ARPA-H efforts on trial site activation (up to $41.18M), national data infrastructure (up to $49.9M) and patient-initiated cancer trial matching (up to $8.95M).
When are SURPASS proposals due?
A solution summary is due November 30, 2026 and full proposals January 22, 2027, per ARPA-H.
Related articles
Healthcare & Wellness AIONC Funds Open-Source Guardrails for Clinical AI Agents
ONC awarded MedStar about $1 million to build open-source middleware that gates and verifies AI agents handling clinical follow-ups. The work is planned, not yet proven.
