By Deborah Borfitz 

May 27, 2026 | The Food and Drug Administration (FDA) is undergoing a major step change in how it regulates and evaluates drugs, marked by a shift toward real-time clinical trials, an agency-wide rollout of generative AI tools to speed up drug application reviews, and a radical transparency push that includes public disclosure of complete response letters sent to pharmaceutical companies explaining why a drug application was rejected. Up for debate is whether these aims are within the “realm of capabilities” of health authorities, according to Kevin Bugin, Ph.D., head of global regulatory policy and intelligence at Amgen.   

Bugin, who previously served as deputy director of operations in the Office of New Drugs in the FDA's  

Center for Drug Evaluation and Research (CDER), was moderating a keynote panel discussion at last week’s SCOPE X conference featuring the reflections of former regulators and technologists with the agency. In addition to himself, these included Vid Desai, CEO of Desai Technology Consulting; M. Khair ElZarrad, Ph.D., vice president of regulatory policy at BridgeBio; and Khushboo Sharma, CEO of Accumulus Synergy, a nonprofit developing cloud-based data exchange platforms for sponsors and regulators. The conversation centered on changes in the clinical development landscape, the impact on the current regulatory system, and how regulatory policies can keep up.   

AI is materially changing how evidence and data are generated for FDA regulatory submissions, says Sharma, who like Bugin formerly held the position of deputy director of operations in CDER’s Office of New Drugs. It facilitates passive and decentralized data capture, with the information “flowing asynchronously across the country in near real time.”  

Sponsors can be running a brilliant, AI-enabled trial, and yet still be submitting their Investigational New Drug application to the FDA electronically as structured PDFs, says Sharma, which she terms “electronic paper.” The disconnect between the sophistication of data generation and regulatory submission requirements is the friction point relevant to health authorities around the world.  

“There’s a lot of change going around” in politics as well as science and technology, notes Desai, former chief information officer at the FDA. The way regulatory approvals are being handled is no match for the scale and ever-growing intricacies of submissions the agency is tasked with processing. 

By his estimation, submissions processed by the FDA doubled over the past 20 years and are likely to double again in the next two decades. “I would argue that the degree of complexity and the volume is rising faster than the ability of AI to give the agency the edge that it needs,” Desai says. The FDA recognizes the need to change from a static to a more collaborative review mode, he adds, and it’s a trend he expects will accelerate. 

Real-timing Studies 

The real-time clinical trials initiative of the FDA looks to deploy AI agents within electronic health records (EHRs) for clinical trials to directly report safety signals to regulators, says Desai. It’s a model that is likely to see further development and application in post-marketing safety surveillance, he adds.    

An agent-based, automated reporting structure could enable 95% of adverse events to be registered compared to the 3% to 5% currently being reported, Desai notes. At the agency level, robust analytics will be needed to “weed out the false positives.” A secondary concern is that under the radical transparency mindset “pretty much everything is made public as it happens,” opening the possibility of “competitors or the nut cases on social media getting access to some of those reports and misreading all kinds of nonsense into that.” 

That’s an issue that will have to be dealt with at the agency level, says Desai. Previous administrations were a “bit more thoughtful” about publicly releasing data and information, scrubbing and scrutinizing it prior to publication rather than sharing raw information, which can backfire. “There are going to be many, many false positives that we are going to have to analyze and deal with, but the agency is typically very good at doing that.” 

The onus is still on the sponsor to aggregate the data and evaluate safety signals from their clinical trials to ensure the FDA receives a comprehensive, integrated safety profile of a drug across the entire development program, says ElZarrad, who previously led the CDER Office of Medical Policy. That is best done before the agency raises a safety concern. 

It may be another decade before “spontaneous reporting” of adverse events is replaced by a direct connection to federated data, says Bugin. In Europe, this is exemplified by the DARWIN [Data Analysis and Real World Interrogation Network] EU model, established by the European Medicines Agency to provide reliable real-world evidence on the use, safety, and effectiveness of medicines and vaccines. The distributed network works by connecting to EHRs, patient registries, insurance claims, and hospital databases. 

Current State 

ElZarrad reports a deluge of AI tools are being used specifically in clinical trials and with some success in reducing failure rates oftentimes occurring for simple reasons. These include tools for summarization tasks as well as recruitment for complex conditions such as treatment-resistant depression.  

AI is now more commonly used to simulate external control arms in rare disease trials, ElZarrad notes. It’s an especially promising use case because traditional randomized controlled trials aren’t well-suited to rare diseases due to tiny patient populations, ethical issues with placebos, and vast disease heterogeneity. 

Change management is one of the major challenges facing the FDA as it seeks to absorb the vast amount of evidence being generated, says Sharma, who questions the readiness of reviewers and staff to do things differently. In her experience at the agency five years ago, they were highly resistant to adopting new technologies.  

This new way forward, as recently announced, includes proof-of-concept trials that will report endpoints and data signals to the agency in real time. One will be conducted by Amgen, a phase 1b trial in patients with limited-stage small cell lung carcinoma. 

The surge in data coming into the agency means “pretty much every function within the FDA” is going to require much better, AI-powered tools, says Desai. The first generation of that thinking is exemplified by the agency’s use of the generative AI chatbot Elsa to streamline and accelerate regulatory reviews, and its related work with the HALO (harmonized AI and lifecycle operations) data platform unifying disparate application and submission databases across all agency centers. Desai expects this will evolve into the development of role-based AI tools, including for people currently doing code work at the agency manually.   

Massive challenges facing the FDA need to be met with significant changes at both the personnel and training levels, says ElZarrad, noting the “artificial divide” between the divisions for devices, biologics, and drugs. One of his best personal experiences at the agency was when people were assigned to dual positions within both CDER and CDRH (Center for Devices and Radiological Health). Cross-training expertise centers “will need to be normalized” in the reimagined FDA, he adds. 

The way physicians are trained also needs an update, given that the FDA has been rapidly expanding its acceptance of real-world evidence mined from EHRs and wearables. Currently, they’re handed a “hierarchy of evidence” with randomized controlled trials sitting alone at the top, he says. “Automatically, we are biasing them, [indicating] everything else is substandard.” 

What’s Needed and Not 

In terms of regulatory policy and guidance, a flurry of activity from legislators and the FDA has given way to a quiet period beyond some general principles about how AI can be used in drug development provided it is trustworthy, used responsibly, and has audit trails, says Bugin. These are the same sort of attributes previously referenced in the machine learning space.  

A new generation of guidance is needed, says ElZarrad, as is a new way of developing it. AI guidance that has been produced for drugs and biologics can’t keep up with the pace of scientific advancements, which is why he favors the idea of focusing on the principles rather than the context of use as is done with ICH and GCP guidelines. 

This will need to be accompanied by customer support from the FDA, he adds. As it is, it can be difficult to know who to go to and when to expect answers when companies have questions.  

As for transparency around regulatory decision-making, ElZarrad says he doesn’t believe the agency goes far enough. On critical decisions, the FDA needs to be able to publicly explain the required timeline to properly evaluate complex submissions. 

What’s not needed are executive orders like the one recently issued requiring the FDA to get rid of 10 regulations to pass a new one, says Desai. The tactic aimed to reduce regulations but instead resulted in “little to no regulations being passed.” 

“What we do need,” he continues, “is for Congress to change the way it funds the agency.” Currently, the FDA is issued one-year funds that it must spend in that timeframe. Given recent appropriation battles, it can take half a year for a bill to pass, leaving the agency with as little as three months to spend that one-year allocation. 

The Wish List  

All the easy things that can be done in one year by the FDA have been done, Desai says. “All the problems that remain would take multiple years of focused strategy and investment to fix, and Congress needs to find a way of funding the agency to deal with those bigger issues.” 

On Sharma’s wish list is an investment in “people over platform.” As she notes, “You can’t train a generation of reviewers overnight.” 

Bugin answers his own magic-wand question with a call for preparing the FDA for “the reality of data abundance.” The agency is presently set up for data scarcity, a fast-disappearing period when the data necessary to assess the risks and benefits of a product were hard to collect. 

Big data could be leveraged in the pursuit of getting needed medicines to patients as quickly as possible, says Bugin. AI could potentially help structure that data to make it valuable and actionable by finding signals within all the noise.    

“But the FDA has got to be ready for it ... [and also] make sure that [its] wishes are being shared across the entire healthcare ecosystem,” Begin says. “The data is not coming from silos so much anymore; it’s going to come from data being connected across healthcare systems.” What’s needed are EHRs with the right kinds of structures in place, AI literacy training, and oversight mechanisms that allow engagement between industry, technology users, and regulators, he concludes. 

AI models also need to strike “a balance between predictability and proportionality,” says Sharma, meaning they produce results that are scientifically credible and the required validation, transparency, and complexity align with their risk level and intended use. Prioritizing the novelty of a methodology introduces liabilities as well as limits real-world utility.  

Desai proposes a system for the EHR environment that’s analogous to credit card fraud alerts sent via text when an unusual activity is detected. “We better have a system like that in place if we continue to go down this deregulation path ... [and] we’re taking more risks,” he says.