Key Leaders Depart as the FDA’s AI Fast Track Reaches a Governance Crossroads

The U.S. Food and Drug Administration (FDA) once portrayed artificial intelligence as essential infrastructure for shortening review times. Now, however, it is not merely model capabilities that are being tested. The successive departures in May of former Commissioner Marty Makary and inaugural Chief AI Officer Jeremy Walsh, together with leadership changes in the information technology division, have left this initiative spanning drug reviews, clinical trials, and on-site inspections facing unresolved questions over who will coordinate it, how accountability will work, and how quickly it should proceed.

The core tool, Elsa, is a large language model assistant for FDA personnel that can help read, draft, and summarize documents. Practical uses include organizing industry comments on regulatory proposals, tracing a drug’s submission history across multiple years, comparing labeling, and compiling adverse events to support safety assessments. The FDA had also planned to use agentic AI for multistep work encompassing premarket review, review verification, postmarket monitoring, and inspections. Because these uses approach the core processes of regulatory judgment, there must be a clear boundary between merely reducing administrative burdens and materially influencing the interpretation of evidence.

Before the personnel changes, the FDA launched Elsa 4.0 in May and integrated more than 40 agency-wide application and submission data sources and portals into the HALO data platform. The new version added features including customizable agents, document generation, quantitative analysis, chart creation, and searches across large databases, allowing personnel to avoid manually uploading documents in every conversation. The FDA said the system was built in a FedRAMP High security environment, would not use data submitted or entered by industry to train models, and required domain experts to verify inputs, analytical processes, and the use of outputs.

These security measures, however, cannot replace performance validation. The information disclosed by the FDA to date is insufficient for outside observers to assess Elsa’s accuracy, omission rates, and hallucination risks across different review tasks. The agency has also not clearly explained how errors are recorded, which uses are prohibited, or what auditable trails should be retained when AI participates in a review. Experts who formerly served as FDA AI policy officials also believe Elsa is primarily intended to augment personnel’s work rather than replace final decision-making. Without consistent agency-level governance, however, individual centers may redevelop their own approaches, potentially resulting in differing levels of transparency.

Acting Commissioner Kyle Diamantas has said AI remains an FDA priority, but the policy direction has become more cautious. Law firm Mintz observed that the new leadership team has slowed some previously fast-tracked pilot programs and extended the comment period for a proposal to use AI to support early-stage clinical trials. The proposal envisions AI helping to collect, interpret, and report trial signals in real time, after which FDA systems would process data meeting predefined criteria. Comments submitted so far broadly support exploring this pathway but call for stronger safeguards for safety, procedural rigor, and bias.

For now, the changes primarily affect how the FDA itself uses AI and do not necessarily mean that rules governing pharmaceutical companies’ use of AI in drug development have shifted. BioPharma Dive cited former FDA policy experts as saying there is currently no indication that existing policies on sponsors’ use of AI have changed. Industry still needs clearer model-validation requirements, however, as well as an understanding of how regulators will use AI to read submission materials, so companies can prepare documents that are consistent, traceable, and suitable for combined machine and human review.

The FDA’s next choice is not simply between “accelerating” and “pausing.” Traditional guidance procedures can provide legal and procedural stability, but their pace may not keep up with model updates. Rapid pilots, meanwhile, risk bringing unresolved issues involving bias, errors, and responsibility into high-risk decisions. Whether Elsa can evolve from an efficiency tool into trusted regulatory infrastructure will depend on whether the new leadership team can publicly define the boundaries of its use, validation results, and chain of responsibility—not merely continue adding features.