TL;DR
- Pilot Launch: The FDA has started accepting live AI- and cloud-based data feeds from AstraZeneca’s TRAVERSE and Amgen’s STREAM-SCLC oncology trials.
- Projected Gains: Officials estimate trial duration could fall 20 to 40 percent, with $120 million in annual savings funding up to 3,000 rehired scientists.
- Sponsor Deadline: A Federal Register Request for Information accepts comments until May 29, 2026, with final selection criteria due in July and pilot selections in August.
The U.S. Food and Drug Administration has announced two real-time clinical trial proof-of-concept studies, accepting direct AI- and cloud-based data feeds from AstraZeneca’s TRAVERSE study in mantle cell lymphoma and Amgen’s STREAM-SCLC study in small cell lung cancer. FDA reviewers have already received and validated real-time data signals from the TRAVERSE study through Paradigm Health, which means the pilot launches with a working data pipeline rather than a paper architecture. Speaking at the agency’s White Oak campus, FDA Commissioner Marty Makary said reviewers can now watch a fever spike or a tumor shrinking “in the cloud in real time.”
Chief Artificial Intelligence Officer Jeremy Walsh estimates the new approach could trim total trial duration by 20 to 40 percent without lowering safety standards, and Makary said annual savings of at least $120 million would fund the rehiring of up to 3,000 scientists. It is the first time U.S. regulators are accepting live trial signals from sponsors rather than waiting for end-of-phase document submissions, and the agency is presenting that savings figure as a direct funding source for reviewers it lost in the early-2025 DOGE-era staffing reductions.
Pilot Mechanics and Projected Gains
AstraZeneca’s Phase 2 TRAVERSE trial in treatment-naive mantle cell lymphoma runs with participation from MD Anderson Cancer Center and the University of Pennsylvania. Amgen will deploy the same cloud-based system for its Phase 1b STREAM-SCLC study in limited-stage small cell lung cancer. Sponsor systems push de-identified safety, dosing, and biomarker signals into a shared cloud environment, where FDA reviewers can examine them as they accrue rather than receiving locked datasets months after enrollment closes.
Early-phase studies sit at what the agency calls a drug-development bottleneck defined by high uncertainty, limited patient populations, and inefficient decision-making, with data normally moving from sites to sponsors before submission to reviewers. Walsh argued the pilot inverts that flow by letting reviewers see what is happening as it happens.
“FDA is overloaded with data, and sometimes we do not need all this data … We already have tons of data. We have been getting tons of data for decades. Can we make a decision based on less information? Can we make a decision based on signal information?”
Jeremy Walsh, FDA Chief Artificial Intelligence Officer (via RAPS)
Walsh’s 20-to-40-percent duration figure carries through to the budget: the freed capacity maps to specific scientist roles the agency wants to rehire rather than to a generic efficiency promise. Paradigm Health supplies the connective tissue, ingesting sponsor data feeds, validating them against agreed schemas, and exposing the resulting signals to reviewers without requiring sponsors to hand over locked, end-of-phase submissions.
For sponsors, the practical change is that a regulator can flag a dose-response anomaly, an emerging adverse-event pattern, or an unexpected biomarker shift inside the same week the study site records it, instead of waiting for an end-of-phase data lock. For patients enrolled in TRAVERSE or STREAM-SCLC, the architecture does not change consent, dosing, or trial design; it changes who sees the resulting signals and when. For reviewers, the workflow shifts from periodic document parsing to continuous oversight of a small, well-scoped feed.
A Phase 2 mantle cell lymphoma cohort and a Phase 1b small cell lung cancer cohort were chosen because both involve narrow patient populations, well-characterized biomarker readouts, and rapid changes in disease state, which makes them well suited to a feed that emphasizes signal density over data volume. Placing the TRAVERSE feed inside MD Anderson and Penn, two academic centers with established trial-data infrastructure, reduces the integration risk a less-equipped site might face when piping live telemetry to an external regulator.
Federal Register Request for Information and Timeline
Alongside the announcement, the FDA published a Federal Register Request for Information titled “AI-enabled optimization of early-phase clinical trials pilot program,” inviting sponsors, contract research organizations, and trial sites to propose additional studies for the broader program. RFI text outlines the data-architecture, validation, and governance expectations the agency wants prospective participants to meet.
FDA officials said the agency will accept comments until May 29, 2026, plans to publish final selection criteria in July 2026, and expects to complete pilot selections in August 2026. That schedule turns the next-cohort question into a near-term operational decision rather than a multi-year rulemaking cycle. Sponsors with active early-phase oncology, neurology, or rare-disease programs are leading next candidates, since those areas combine the small patient populations the agency cited with the kind of dense biomarker telemetry that real-time monitoring is designed to read.
A compressed schedule also pulls forward several decisions that would otherwise sit inside a sponsor’s internal Data Safety Monitoring Board cycle: which signals trip a regulator query, how a sponsor is expected to respond before the next planned analysis, and how disagreements about an interim signal are recorded for later inspection. Contract research organizations that already operate cloud-based electronic data capture stacks are likely first-wave integrators, because the RFI’s expectations map onto pipelines those vendors already maintain.
Reliability Backdrop and the Rebuilding Frame
Walsh acknowledged in the same briefing that Elsa “can hallucinate,” placing the internal tool in the same failure class as other large language models and echoing earlier reports that Elsa fabricates studies and misrepresents research data. He said over 80 percent of FDA staff now regularly use generative AI tools, up from about one percent in early 2025, a jump he tied to the agency’s earlier rollout of the internal Elsa tool.
Walsh said the idea originated in summer 2025, motivated by an FDA review process largely unchanged since the 1960s, and assembled when he stepped into the role of FDA’s first AI officer. The pilot also follows the early-2025 DOGE-era staffing cuts the rehiring plan is designed to reverse. A separate February 2026 decision shifted the FDA’s drug-application standard so that one well-controlled trial plus “confirmative evidence” can support a drug package, replacing the prior two-trial expectation.
FDA Commissioner Marty Makary tied the pilot to a longer-running drug-approval timeline he said the medical profession has accepted for decades. According to Makary, the medical profession has spent the last 50 years accepting a 10-to-12-year timeframe for new drugs to reach market, and he said patients have been receiving treatments later than they could have under a faster review system. Makary called the announcement “a milestone day” for challenging the assumption that drug approvals must take 10 to 12 years.
Investment bank Jefferies, in a same-day note to investors, framed the initiative as a step toward reducing structural inefficiencies and accelerating timelines while preserving safety, monitoring, governance, and data integrity. Pilot architects keep a clear separation between Elsa, an internal staff-facing language model, and the cloud signal monitoring deployed against TRAVERSE and STREAM-SCLC, which is built around structured trial telemetry rather than open-ended generative outputs. Where Elsa parses unstructured documents and can invent a citation, TRAVERSE and STREAM-SCLC pipelines carry typed clinical fields with known schemas, audit logs, and sponsor sign-off on every value before a reviewer sees it.
Sponsors weighing whether to expose early-phase signal data to FDA reviewers face their first concrete decision point at the May 29, 2026 comment deadline, with final selection criteria scheduled for July 2026 and pilot selections for August. Paradigm Health must keep validating the TRAVERSE and STREAM-SCLC feeds cleanly through that window for a wider cohort to inherit the workflow. The next verifiable signals will be the August pilot-selection list, the first regulator query Walsh’s team raises against a live TRAVERSE or STREAM-SCLC feed, and whether the $120 million savings figure translates into named scientist hires the agency can report against its early-2025 reductions.
