The U.S. Food and Drug Administration (FDA) launched its first agency-level generative AI assistant, Elsa, in June 2025. Aimed at all 18,000 employees (scientific reviewers, compliance investigators, policymakers, etc.), this tool marks a significant step in the FDA’s digital transformation, aiming to enhance review efficiency, reduce repetitive paperwork, and accelerate the review process for drugs and medical devices without compromising scientific rigor.
Project Positioning and GoalsElsa’s full name was initially “Efficient Language System for Analysis,” but it is now branded simply as “Elsa.” This tool is designed for the FDA’s 18,000 employees, including scientific reviewers, compliance investigators, and policymakers, aiming to help staff complete their daily tasks more efficiently through automated processing and intelligent analysis.Elsa’s main objectives include:
- Reducing review cycles: By quickly processing large amounts of information, it shortens the review time for drugs and devices.
- Minimizing paperwork: Automatically generating meeting minutes, summarizing adverse event reports, etc., to alleviate the paperwork burden on employees.
- Demonstrating safe and controllable AI usage: Setting a benchmark for AI applications in the industry, ensuring high standards of scientific accuracy and reliability during the regulatory process.
Technical Architecture and FeaturesElsa is built on Anthropic’s Claude model and operates in a highly secure AWS GovCloud environment (IL4 security level), with physical isolation and internal network access only, ensuring data security and privacy. Key design points of the tool include:
- Document library mode: Elsa only accesses the FDA’s internal document library, avoiding the use of unreliable external information, implementing a mandatory “document library mode” + citation chain, where the system must provide traceable paragraph-level sources; otherwise, it refuses to answer, thereby reducing the occurrence of hallucinations.
- Mandatory citation mechanism: When generating content, Elsa must provide traceable citations to ensure the accuracy and reliability of the information.
Application Scenarios of Elsa
- Initial screening of clinical trial protocols: Automatically extracting points that do not comply with guidelines to save review time.
- Drug label comparison: Quickly listing differences between new and old labels, highlighting potential safety risks.
- Adverse event signal overview: Automatically generating trend summaries based on FAERS and literature databases.
- High-priority inspection lists: Optimizing inspection resource allocation based on historical compliance data.
- Internal code generation: Quickly generating executable code snippets for non-clinical databases.
- Meeting minutes and email templates: Transcribing meeting recordings into structured minutes, accompanied by a list of action items.
Current Status and ChallengesDespite the anticipation surrounding Elsa’s launch, it faces several challenges in practical application. Approximately 55% of FDA employees have logged in for trial use, but frequent users are concentrated in the Center for Drug Evaluation and Research (CDER); usage rates are lower in the Center for Devices and Radiological Health and the Center for Food Safety and Applied Nutrition. According to internal feedback, some employees express concerns about the tool’s accuracy and system integration, especially when handling complex datasets, as Elsa sometimes generates inaccurate results. In early June, there was a case of “hallucinated citations” (the model fabricated three non-existent references), triggering an internal alert; this was temporarily resolved through the mandatory citation chain feature. Jeremy Walsh, the FDA’s AI lead, emphasized that hallucination issues can be mitigated by designing more precise questions. For example, users need to clearly define the query scope and context to reduce the likelihood of the model generating incorrect information.Additionally, the rapid deployment process has raised questions about insufficient testing and training, with internal certification training for “Elsa Champions” conducted quarterly, and a draft of the “FDA Generative AI Output Verification SOP” is in progress.The FDA leadership acknowledges that Elsa is still in its early stages, and future success will depend on continuous optimization and better integration with FDA systems. As the technology matures, Elsa is expected to play a greater role in drug review and regulatory processes.Significance for the IndustryThe launch of Elsa not only improves internal processes at the FDA but also sends an important signal to the entire pharmaceutical industry. As regulatory agencies gradually adopt AI technology, pharmaceutical companies need to reassess their regulatory strategies to adapt to this change. The FDA’s implementation of the “document library lock + mandatory citation + full traceability” mechanism may become an implicit standard for future submissions of AI-assisted regulatory materials by companies.In summary, Elsa serves as the FDA’s “surgical scalpel,” providing support to regulators in a highly confidential environment, aiming to enhance work efficiency through intelligent means. Despite facing challenges, its potential is immense, and it is expected to have a profound impact on drug review and public health protection in the future.