FDA builds its own agentic AI

The Food and Drug Administration has rolled out agentic artificial intelligence tools to all agency employees, expanding an internal AI push from chat-style help to systems that can handle multi-step work. (fda.gov) The agency announced the deployment on December 1, 2025, and said the new tools can support meeting management, pre-market reviews, review validation, post-market surveillance, inspections, compliance and administrative work. FDA said the system is optional for staff and operates with built-in human oversight. (fda.gov) Agentic AI is the label for software that plans and executes a sequence of steps toward a goal, rather than only answering a prompt. FDA said its version can combine multiple AI models inside governed workflows to help reviewers, scientists and investigators finish more complex tasks. (fda.gov) The rollout builds on Elsa, the FDA’s agency-wide generative AI tool launched on June 2, 2025. FDA said Elsa was designed for reading, writing and summarizing, and by December more than 70% of staff were using it voluntarily, according to internal agency data cited by the agency. (fda.gov 1) (fda.gov 2) Before Elsa went live, Commissioner Marty Makary said on May 8, 2025, that an AI-assisted scientific review pilot had cut some review tasks from days to minutes. He then set a June 30, 2025 deadline for all FDA centers to move onto a common secure generative AI system. (fda.gov) FDA has framed the project as part of a broader technology modernization plan, not a stand-alone experiment. Its fiscal 2024-2027 information technology strategy tied new AI tools to data sharing, cybersecurity, governance and a “OneFDA” push to connect work across centers. (fda.gov) The agency has also spent years using narrower forms of AI in regulatory science. FDA says its research and operations already include machine learning to target high-risk seafood imports, detect adverse events in postmarket data and improve regulatory review workflows. (fda.gov) The guardrails are a central part of the design. FDA said the tools run in a high-security GovCloud environment, keep information inside the agency, and do not train on staff inputs or on data submitted by drugmakers and other regulated companies. (fda.gov 1) (fda.gov 2) That setup leaves the FDA using AI as internal infrastructure rather than as an autonomous regulator. The agency’s public description keeps humans in the loop, makes staff use voluntary, and places the new systems inside existing review, surveillance and inspection processes. (fda.gov)