The U.S. Food and Drug Administration (FDA) has taken a major step in its digital transformation by rolling out agentic AI tools to all agency employees. Announced on December 1, 2025, this deployment aims to support reviewers, scientists, and investigators with complex, multi-step regulatory and administrative tasks while preserving human control over all decisions. The move follows several years of experimentation with AI and large language models (LLMs) inside the agency, culminating in the successful internal adoption of the generative AI assistant known as Elsa.
By making these new capabilities available across centers and offices, the FDA is signaling that AI is no longer a peripheral pilot but a core enabler of its public health mission. However, the tools remain strictly optional and are governed by a growing set of safeguards, including human-in-the-loop oversight, secure cloud infrastructure, and policies that prevent training on confidential submissions. Together, these measures illustrate how a major regulator is trying to harness the efficiency of agentic AI without compromising its scientific standards or data protection obligations.
Understanding what “agentic AI” means in the FDA context
In its announcement, the FDA describes agentic AI as a class of advanced systems that can plan, reason, and execute multi-step actions toward a defined goal. Unlike single-shot chatbots, these systems can chain together tasks, such as searching databases, generating summaries, drafting emails, and tracking workflows, under a unified logic. They are designed to operate as goal-oriented assistants that orchestrate multiple underlying AI models, rather than a single monolithic algorithm.
Crucially, the agency emphasizes that agentic AI operates under built-in guidelines, including explicit human oversight at key decision points. That means an AI workflow might, for example, assemble a pre-market review briefing package or flag anomalies in post-market safety data, but a human reviewer must validate the outputs before they inform any regulatory action. This model aligns with the FDA’s broader risk-based stance toward AI in medical product development, which stresses transparency, accountability, and validation of algorithmic tools.
By framing agentic AI as a support technology rather than an autonomous decision-maker, the FDA is seeking to balance innovation against public expectations of due diligence. External reporting on the rollout underscores this point: agency spokespeople have clarified that the tools are exploratory, do not replace human judgment, and cannot independently make or finalize regulatory decisions. This nuance is central to building trust among both FDA staff and the industries the agency regulates.
From Elsa to agency-wide deployment: how the FDA got here
The agentic AI launch builds directly on the FDA’s earlier deployment of Elsa, an internal LLM-based assistant introduced agency-wide in mid‑2025. Elsa was developed inside a high-security GovCloud environment to give staff a secure way to query internal documents, summarize lengthy filings, and draft technical correspondence. According to the FDA, more than 70% of employees have used Elsa, and program teams have iterated frequently on its design to better fit real-world workflows.
Feedback from this first wave of generative AI adoption helped the agency understand both the benefits and the limits of LLMs in a regulatory context. Early users encountered familiar issues such as hallucinations, inconsistent citation of sources, and sensitivity to prompt wording. In response, FDA teams focused on guardrails like strong retrieval from authoritative internal sources, narrow-scoped use cases, and mandatory human review of AI-generated outputs. These lessons laid the groundwork for a second generation of tools capable of orchestrating more complex, multi-step processes.
Agentic AI, as now deployed, is less about replacing Elsa and more about expanding what Elsa-style tools can do. Rather than simply answering a question or drafting a single memo, an agentic system might assemble data from multiple submissions, cross-check it against guidance documents, generate a structured comparison, and schedule a meeting with stakeholders, all under the supervision of a human reviewer. This progression from a stateless chatbot toward workflow-aware “agents” is at the heart of the new deployment.
Key use cases: from pre-market reviews to inspections
The FDA has highlighted several concrete domains where agentic AI is expected to provide the greatest value. One is pre-market review of drugs, biologics, and medical devices. These reviews involve vast volumes of scientific, clinical, and manufacturing data, along with complex regulatory histories for sponsors. Agentic AI can assist reviewers by organizing and summarizing prior correspondence, extracting key parameters from study reports, and assembling side‑by‑side comparisons of similar products or trials. This does not change evidentiary standards, but it can reduce the manual over associated with navigating large digital dossiers.
Another set of use cases lies in post-market surveillance and compliance. Post-market safety monitoring requires continuous analysis of adverse event reports, real-world evidence, literature, and inspection findings. Agentic systems can be tasked with periodically scanning and structuring incoming data, flagging patterns of concern for human epidemiologists and compliance officers. Similarly, AI agents can assist inspection teams by generating pre‑inspection briefings, cross‑referencing prior observations, and proposing prioritized checklists tailored to a facility’s history and risk profile.
The agency also expects substantial gains in internal efficiency for meeting management and routine administrative tasks. AI agents can help coordinate calendars across large project teams, prepare agenda materials based on latest documents, track action items, and draft follow‑up communications. While these functions are less visible than line-grabbing scientific applications, they may be critical for an agency that has faced staffing shortfalls and rising workloads. The goal is to let highly trained scientific staff focus more on substantive assessment and less on logistics and paperwork.
Security, data protection, and trust safeguards
A central concern in any regulatory AI deployment is the protection of sensitive data, including trade secrets in product submissions and confidential patient-level data in clinical studies. The FDA has repeatedly emphasized that its AI systems, including Elsa and the new agentic tools, run in a high-security GovCloud environment with strict access controls. Models are configured not to train on user inputs or on any data submitted by regulated industry, preventing inadvertent leakage of proprietary information into general-purpose model weights.
In addition to technical isolation, the agency is building process-based safeguards. Outputs from AI tools must be reviewed and validated by human staff before they are incorporated into official actions such as approvals, warning letters, or guidance documents. This human-in-the-loop validation is both a quality control mechanism and a legal safeguard, ensuring that final responsibility rests with trained officials whose names appear on regulatory decisions. For staff, this framework clarifies that AI is a tool they are accountable for using wisely, not a black box that absolves them of responsibility.
These internal protections complement the FDA’s external policy work on AI, such as draft guidance on the use of AI to support drug and biologic regulatory decisions and existing principles for Good Machine Learning Practice in medical device development. By aligning its internal AI experimentation with the expectations it is placing on industry, the agency seeks to avoid a double standard and demonstrate that regulators are willing to hold their own tools to the same level of scrutiny as those they review.
Agentic AI in the broader FDA AI strategy
The deployment of agentic AI across the agency is not happening in isolation; it is part of a years‑long effort to integrate AI into both the products the FDA evaluates and the processes it uses. On the policy side, the agency has released guidance to clarify how AI and machine learning models can be used in regulatory submissions, emphasizing transparency, validation, and a risk-based approach. Center-specific initiatives, such as CBER’s Artificial Intelligence Coordinating Committee, are also working to standardize how reviewers assess AI/ML technologies in biologics and vaccines.
At the same time, the FDA has begun qualifying AI-based tools for specific scientific applications. A recent example is the qualification of AIM‑NASH, an AI system that analyzes liver biopsy images to support drug development for metabolic dysfunction-associated steatohepatitis (MASH). This tool is designed to improve consistency and speed in pathology assessments for clinical trials, with AI-generated scores feeding into, but not replacing, expert interpretation. Such cases illustrate how AI can move from experimental technology to an accepted component of evidence generation in drug development.
By experimenting with AI internally while also setting standards for its external use, the FDA is creating a feedback loop: regulators who work with agentic AI in their own workflows may be better equipped to evaluate similar technologies proposed by sponsors. Conversely, scrutiny applied to industry tools helps shape internal expectations about documentation, validation, and monitoring. The agency-wide agentic AI rollout thus sits at the intersection of operational modernization and regulatory policy evolution.
Opportunities and risks for public health and industry
If implemented effectively, agentic AI could have significant positive ripple effects on public health. Faster and more efficient review workflows could shorten time-to-decision for safe and effective therapies, especially in rapidly evolving areas like oncology, gene therapies, and infectious disease. Improved post-market surveillance might detect emerging safety signals earlier, allowing more timely label changes, risk mitigation measures, or product withdrawals when necessary. For food safety and cosmetic oversight, AI-assisted inspections and data analysis may better target limited resources toward higher-risk products and facilities.
For industry, a more AI-enabled FDA could translate into clearer analytic expectations, more consistent reviews, and potentially shorter feedback cycles. As regulators become more comfortable with AI-supported methods, sponsors may find a more receptive environment for incorporating model-based analyses, synthetic control arms, and advanced image analysis into their development programs. At the same time, the presence of AI inside the agency may raise the bar on technical rigor, documentation, and explainability for tools used in submissions.
However, the expansion of AI also introduces new risks. Over-reliance on automated workflows could, if not carefully managed, obscure edge cases, minority populations, or novel safety signals that do not fit existing patterns. Biases in data or model design might propagate through large-scale workflows if not routinely audited. There is also the risk of perceived opacity: if stakeholders cannot understand how agentic systems contribute to regulatory reasoning, trust in outcomes may erode. Addressing these risks requires continuous monitoring, model governance, and clear communication about the limits of AI-derived insights.
Culture change and the Agentic AI Challenge
Technical deployment is only part of the story; an equally important dimension is cultural adoption. To encourage experimentation and skill-building, the FDA is launching a two‑month Agentic AI Challenge in conjunction with the rollout. Staff across centers are invited to design and prototype agentic workflows that solve real operational problems, with selected solutions to be showcased at the FDA’s Scientific Computing Day in January 2026. This kind of internal challenge is meant to surface bottom‑up ideas and empower domain experts to shape how AI is used in their daily work.
Such initiatives help overcome a common barrier in public sector technology projects: the gap between centrally procured tools and the nuanced needs of frontline staff. By giving reviewers, investigators, and policy analysts room to experiment, within a secured sandbox and under clear rules, the agency can discover novel, high‑value use cases that might not emerge from top-down planning alone. At the same time, the challenge format creates a feedback channel where issues of usability, reliability, and governance can be identified early.
Over time, this kind of structured experimentation could foster a new generation of “AI‑literate regulators” who understand both the power and the pitfalls of agentic systems. That internal expertise is likely to be critical as the FDA continues to engage with sponsors, patient groups, and international regulators on the role of AI in medical product lifecycles and food and device safety. The agentic AI rollout is thus as much an exercise in workforce development as it is in technical modernization.
Ultimately, the FDA’s deployment of agentic AI across the agency represents a significant inflection point in how a major health regulator uses digital tools. By shifting from isolated pilots to an agency-wide platform, the FDA is betting that AI can meaningfully augment human expertise in managing complex regulatory workloads. The success of that bet will depend on maintaining rigorous human oversight, aligning internal practices with public guidance, and continuously evaluating how AI-supported processes affect the timeliness and quality of regulatory decisions.
As agentic AI evolves and additional use cases emerge, the FDA’s experience will likely inform broader debates about responsible AI in government. Agencies around the world are watching how the FDA handles issues of security, bias, transparency, and accountability at scale. If the balance between innovation and safeguards can be maintained, the deployment of agentic AI may become a template for how public institutions can modernize without compromising their foundational obligations to equity, scientific integrity, and public trust.
