The U.S. Food and Drug Administration (FDA) is currently at a critical crossroads regarding its digital strategy, as the departure of several high-profile leaders has sparked questions about the agency’s trajectory in adopting and regulating artificial intelligence. While AI integration has become a permanent fixture within the agency’s operational framework, the path forward appears less certain following the exit of former Commissioner Marty Makary and other key technology officials. This transition marks a potential shift from a highly centralized, aggressive AI rollout toward a more traditional, perhaps fragmented, approach that emphasizes departmental autonomy and established regulatory norms.
The momentum behind AI at the FDA has been substantial. Between fiscal years 2024 and 2025, the number of AI use cases reported by the agency surged by a staggering 148%, according to data analyzed by the Bipartisan Policy Center. This rapid expansion was part of a broader federal initiative to leverage emerging technologies to improve government efficiency. However, with the recent leadership vacuum, industry experts and former agency officials are closely monitoring whether the FDA can maintain this pace or if internal bureaucratic structures will return to a decentralized model that could slow down innovation and transparency.
The Makary Era and the Centralization of AI
During his tenure, Marty Makary was a vocal proponent of a unified, agency-wide approach to artificial intelligence. His vision centered on breaking down the silos between different FDA centers—such as the Center for Drug Evaluation and Research (CDER) and the Center for Devices and Radiological Health (CDRH)—to create a cohesive technological infrastructure. The centerpiece of this effort was the launch of Elsa, an agency-wide large language model (LLM) designed to optimize performance and streamline the review process for medical products.
Elsa was not a sudden development but rather the evolution of internal experiments. It found its roots in "CDER GPT," a pilot program developed within the drug evaluation division. Under Makary’s leadership, this tool was expanded into a robust, retrieval-augmented generation (RAG) system. RAG technology is critical in a regulatory environment because it reduces the risk of "hallucinations"—instances where an AI generates plausible-sounding but factually incorrect information. By confining the LLM to a well-defined database of trusted FDA documents, Elsa allows staff to summarize years of regulatory submissions, condense industry comments on proposed rules, and assist in the drafting of internal reports without compromising factual accuracy.
In December 2024, the agency further signaled its commitment to automation by announcing plans to deploy "agentic AI." Unlike standard chatbots that respond to prompts, agentic AI systems are designed to perform complex, multi-step tasks with a degree of autonomy. The FDA intended to use these agents to support premarket reviews, conduct site inspections, and manage administrative burdens. The goal was clear: to allow human reviewers to focus on high-level decision-making while AI handled the labor-intensive documentation and data-sorting tasks that often bottleneck the approval process.
Leadership Departures and Governance Uncertainty
The departure of Makary is only one piece of a larger leadership exodus that has left the FDA’s AI governance structure in flux. Alongside the former commissioner, the agency saw the exit of Jeremy Walsh, the Chief Artificial Intelligence Officer (CAIO), and Sridhar Mantha, the acting Chief Information Officer (CIO). These roles are pivotal in implementing the White House’s Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, which mandates that federal agencies appoint leadership to oversee AI risks and benefits.
Acting Commissioner Kyle Diamantas has publicly maintained that AI remains a "top priority" for the agency. However, the absence of a permanent CAIO and CIO creates a vacuum in strategic oversight. Tala Fakhouri, currently the Chief Artificial Intelligence and Regulatory Strategy Officer at Parexel and a former FDA AI policy official, has noted that the lack of clear leadership could lead to a "fragmented" approach. In recent professional conferences, FDA representatives have increasingly spoken on behalf of their individual divisions rather than presenting a unified agency-wide stance.
This potential return to a department-specific model raises concerns about consistency. If the Center for Biologics Evaluation and Research (CBER) develops different AI standards or tools than CDER, it could create confusion for pharmaceutical sponsors who often submit applications that cross-divisional boundaries. Furthermore, a decentralized model may lack the budget and political will to maintain large-scale projects like Elsa, potentially relegating AI to a secondary tool used sporadically across different offices.
The Statistical Surge: Tracking AI Integration
The 148% increase in AI use cases reported by the Bipartisan Policy Center highlights how deeply the technology has already permeated the FDA’s daily functions. As of 2025, the FDA’s inventory of AI projects spans various categories, including:
- Premarket Review Assistance: Tools used to screen applications for completeness and to identify key data points in clinical trial results.
- Post-Market Surveillance: AI algorithms that monitor adverse event reporting systems to detect safety signals faster than traditional statistical methods.
- Inspectional Oversight: Systems designed to analyze historical inspection data to prioritize high-risk manufacturing facilities for on-site visits.
- Administrative Automation: LLMs used for summarizing public comments during the notice-and-comment rulemaking process.
This growth reflects a wider trend across the Department of Health and Human Services (HHS). The rapid adoption is driven by the sheer volume of data the FDA must process. With the rise of precision medicine and real-world evidence (RWE), the amount of data included in a single New Drug Application (NDA) has grown exponentially. Without AI, the agency faces a significant risk of falling behind its statutory review timelines established under the Prescription Drug User Fee Act (PDUFA).
Impact on Industry and the Need for Transparency
While the FDA’s internal use of AI is primarily intended to improve efficiency, it has direct implications for the pharmaceutical and medical device industries. One of the primary criticisms from industry stakeholders is the lack of transparency regarding how these tools are used in the decision-making process.
Fakhouri and other industry experts argue that if a reviewer is using an AI assistant to summarize a clinical study report, the sponsor should be aware of that fact. Understanding the "prompts" or the parameters used by the AI can help companies prepare their submissions more effectively. For instance, if the FDA’s AI is optimized to look for specific data labels or structured formats, sponsors can align their data packages to ensure they are interpreted correctly by the agency’s automated systems.
Moreover, there is a call for the FDA to provide clearer guidance on how drug companies themselves should validate AI tools used in clinical trials. Currently, there is a distinction between the FDA’s internal AI use and the external policies governing sponsor-led AI. While internal policies are in a state of flux, the rules for industry—such as those regarding Good Machine Learning Practices (GMLP)—have remained relatively stable. However, stability can sometimes translate to stagnation. The rapid evolution of generative AI means that a guidance document written a year ago may already be obsolete.
Rulemaking: Agility vs. Tradition
One of the most significant shifts under Acting Commissioner Diamantas is a return to traditional administrative processes. Under Makary, there were instances where policy shifts and regulatory philosophies were communicated through non-traditional channels, such as journal articles or press conferences. This "shoot-from-the-hip" style of policymaking, as described by some lawmakers, provided agility but often bypassed the rigorous public comment periods required by the Administrative Procedure Act.
Diamantas has recently moved to disavow some of the informal statements made by the previous leadership, confirming that official policy will only be developed through formal guidance documents and the federal register. While this move provides the pharmaceutical industry with the "certainty" and "predictability" it often craves, it introduces a significant challenge: speed.
The traditional FDA guidance process can take anywhere from 12 to 24 months from the draft stage to finalization. In the world of artificial intelligence, where software iterations occur in weeks or months, a two-year rulemaking cycle is an eternity. This "regulatory lag" could result in a scenario where the FDA is regulating 2025 technology with 2023 rules. To combat this, industry advocates are calling for "agile policy development"—a middle ground that involves more frequent, informal communication between regulators and developers, complemented by high-level "standing" guidance that can be updated more easily.
Future Outlook and Global Context
The FDA does not operate in a vacuum. Regulatory bodies around the world, including the European Medicines Agency (EMA) and Japan’s Pharmaceuticals and Medical Devices Agency (PMDA), are also grappling with AI integration. The EMA has already published a "Reflection Paper" on the use of AI in the medicinal product lifecycle, emphasizing a risk-based approach.
If the FDA’s internal AI strategy becomes too fragmented or slow, the United States risks losing its position as the global leader in regulatory science. The next 12 to 18 months will be a defining period for the agency. The appointment of a permanent CAIO and a confirmed Commissioner will be the first signals of whether the FDA intends to double down on Makary’s centralized vision or if it will settle into a more cautious, division-led approach.
In conclusion, while AI is undoubtedly "here to stay" at the FDA, the efficiency and transparency of its implementation remain open questions. The agency has successfully moved past the pilot phase, as evidenced by the 148% increase in use cases and the deployment of tools like Elsa. However, the true test will be whether the FDA can build a governance structure that is robust enough to ensure safety and data integrity, yet flexible enough to keep pace with the most transformative technology of the 21st century. As Tala Fakhouri noted, the technology exists to augment the work of reviewers and improve public health outcomes; what is needed now is the leadership to make that vision a transparent and consistent reality.
