FDA's ELSA AI platform reaches 85% staff adoption in two months; governed data and agents reduce drug review from days to 3 minutes

The FDA's Office of Digital Transformation launched ELSA, a generative AI platform available to all 16,000 staff, reaching 85% adoption within two months of deployment. ELSA runs on Databricks with Halo as its governed data foundation, consolidating 50-60 data sources from eight fragmented FDA centers (drugs/CDER, biologics/CBER, devices/CDRH, veterinary, tobacco, food, inspections) that previously maintained separate data stores and chatbots. Within roughly two months, usage moved from <1% to 85% as staff moved beyond Q&A into agentic AI: hundreds of new agents are built per week by medical doctors, scientists, and administrative staff who load standard operating procedures, guidelines, and center-specific documents into ELSA workspaces.

The architecture layers Model Context Protocol (MCP) servers on top of Unity Catalog, turning agent creation into an accessible workflow for non-data-scientists. Concrete impact: FDA reviewers evaluating drug applications previously spent days searching through 3-4 million pages of regulatory submissions to understand starting materials and supply chains for manufacturing. Using Databricks ML/NLP and MLflow extraction, the same queries now return grounded answers in ~3 minutes. This acceleration frees review staff from information hunting to focus on core expertise: safety/efficacy evaluation. CDER, having spent five years building on Databricks, proved the consolidation case—data sharing between centers that took 4-5 days dropped to real-time, spurring adoption across all eight centers.

For enterprise healthcare and regulated-industry architects, the FDA playbook demonstrates that AI adoption velocity depends primarily on governance+consolidation, not model capability. Unity Catalog's table-level access controls and data containment earned trust across eight independent centers skeptical of data sharing. The agentic AI layer unlocked by MCP servers+governed data created the adoption curve, while concrete use-case acceleration (3 minutes vs. days) proved ROI to skeptical regulatory staff. This suggests that LLM+data-governance stack maturity now drives regulatory vertical penetration more than frontier model releases.