The FDA is testing a bigger idea than another government AI tool.

Through a new pilot program, the agency wants to monitor clinical trial data in real time using cloud infrastructure and AI. Instead of waiting for companies to finish trials, assemble enormous application packages, and send regulators millions of pages after the fact, FDA reviewers would be able to see predefined clinical endpoints and other signals as the study runs.

That is a serious shift. It turns drug review from a batch process into something closer to a live data pipeline.

FDA Commissioner Marty Makary described the pilot as the agency’s first real-time clinical trial and said it challenges the assumption that it has to take 10 to 12 years for a new drug to come to market.

The promise is speed. The risk is trust.

The real target is dead time

Makary said about 45 percent of the time between a Phase 1 clinical trial and a company’s FDA application submission can be dead time filled with paperwork and tedious administrative tasks.

That is the obvious place for AI and cloud systems to matter.

Drug approval timelines are not slow only because the science is hard. They are also slow because clinical evidence moves through layers of formatting, transfer, documentation, review, and interpretation. Some of that bureaucracy protects patients. Some of it is just process drag.

The FDA’s pilot is aimed at separating the two.

If regulators can see trial signals as they happen, the review process can start earlier. If AI can help sort, summarize, compare, and flag clinical data, human reviewers can spend more time on judgment and less time digging through paperwork.

That does not make drug review easy. It makes the pipeline less stale.

AI as regulatory infrastructure

The important part is that FDA is not only using AI to summarize documents.

The agency is trying to wire AI into the regulatory workflow itself. A direct cloud feed from a clinical trial would let regulators observe patient events, endpoints, adverse signals, and relevant trial data in real time.

That turns AI from an office productivity layer into infrastructure for decision-making.

FDA Chief Artificial Intelligence Officer Jeremy Walsh said the agency wants to reach regulatory decisions faster without compromising safety. He also said the goal is to reimagine what information the agency needs and when it needs it.

That is the correct frame. The biggest gain may not come from making old paperwork faster. It may come from changing when regulators receive evidence and how they interact with it.

AstraZeneca and Amgen are first in

The pilot will start with two clinical trials from AstraZeneca and Amgen.

That matters because this is not just an internal FDA demo. It requires the agency and companies to coordinate around data standards, access, cloud infrastructure, endpoint definitions, model-assisted analysis, and review practices.

That coordination is where the hard questions live.

Who controls the data feed? How are endpoints defined? What happens if AI flags something ambiguous? How do reviewers validate model output? What gets logged? What becomes part of the official record? How do sponsors, regulators, and patients trust the same pipeline?

Real-time review can only work if the system is auditable.

Speed cannot become opacity

The danger in any AI-enabled regulatory system is not simply that the model makes a mistake. Humans make mistakes too.

The deeper danger is that speed becomes opacity.

If AI accelerates review but nobody can explain why a signal mattered, why a data point was weighted, why a trial anomaly was dismissed, or how an endpoint was interpreted, the system will lose credibility. FDA cannot treat model output like magic.

Clinical trial review needs traceability. It needs records. It needs human accountability. It needs clear boundaries between automated assistance and regulatory judgment.

That is not anti-AI. It is how AI becomes usable in high-stakes infrastructure.

The agency is already changing internally

The pilot also fits into a broader FDA modernization push.

According to the Government Executive report, FDA has consolidated 40 separate application intake systems into one system. It has also consolidated three data monitoring systems and seven adverse event reporting systems into single systems, while reducing duplicate software licenses across agency centers.

Makary said those changes should save at least $120 million per year, with savings going back into science, new technologies, and rehiring scientists.

FDA has also moved aggressively on generative AI adoption. Walsh said regular generative AI use inside the agency rose from about 1 percent of the workforce in early 2025 to more than 80 percent, with some centers exceeding 90 percent. One major internal tool is Elsa, a large-language-model system that helps employees read, write, and summarize reports.

The agency uses both Google Gemini and Anthropic Claude models.

That internal adoption matters because real-time clinical trials will not work if the agency itself still operates like a paper-era institution. You cannot plug a live cloud trial into a slow, fragmented review machine and expect the whole system to move faster.

The best version of this is faster science

The best version of FDA’s pilot is not weaker oversight. It is better-timed oversight.

If regulators can monitor clinical evidence earlier, detect problems sooner, reduce dead time, and make decisions with cleaner data infrastructure, patients could get useful therapies faster while unsafe or weak candidates are caught earlier.

That is the pro-acceleration case.

The point is not to let companies sprint past regulators. The point is to make the regulator technologically capable enough to keep up with modern science.

Medicine is already becoming more data-intensive. Trials are more complex. Devices and therapeutics are more computational. AI systems are entering discovery, trial design, patient matching, monitoring, documentation, and review. A regulator that cannot operate at that speed becomes the bottleneck by default.

FDA is trying not to be that bottleneck.

The hard part is trust

Real-time clinical trials will not be judged only by speed.

They will be judged by whether the data is reliable, whether the models are validated, whether regulators can audit the process, whether companies can participate without gaming the system, and whether patients believe safety is still the priority.

That is the real test.

AI can compress paperwork. Cloud systems can move data faster. Real-time monitoring can cut dead zones out of the review process. But the system only works if the decision trail remains clear.

The FDA’s pilot is worth watching because it shows where public-sector AI gets serious. Not chatbots. Not press-release automation. Regulatory infrastructure.

If it works, clinical trials stop moving like paperwork and start moving like live science.

That could matter more than another model demo.

Sources