How the FDA Built an AI Platform That 85% of Its Staff Now Use Daily

How the FDA Built an AI Platform That 85% of Its Staff Now Use Daily | Databricks Blog

Skip to main content

How the FDA broke down center silos to launch an enterprise AI platform that reached 85% staff adoption in just two months

How medical doctors, scientists, and reviewers are building their own custom AI agents to automate workflows instantly

What it takes to cut heavy regulatory research times down from days to three minutes using a governed data foundation

“We showed the value of having a foundational data platform, and that success story became contagious.”—Venu Boppana, Strategy & Innovation Leader (AI), Office of Digital Transformation, US FDA

Almost every American interacts with the FDA before breakfast. The agency regulates the food we eat, the medicine we take, and the medical devices we rely on. Every 20 cents spent by a US consumer touches something the FDA oversees. Behind that trust is an extraordinary volume of data: a petabyte of documents, hundreds of gigabytes arriving daily, thousands of regulatory submissions flowing in every month across eight centers responsible for drugs, biologics, devices, veterinary medicine, tobacco products, food safety, and inspections. To keep pace with that demand, the FDA's Office of Digital Transformation has built ELSA, a generative AI platform available to all 16,000 FDA staff, and Halo, the governed data foundation underneath it, which runs on Databricks.

Eight centers, eight silos

The FDA's organizational structure reflects the breadth of its mandate. CDER handles drugs. CBER covers biologics. CDRH oversees devices. Each center, along with those covering veterinary medicine, tobacco, inspections, and food safety, had built its own AI capabilities independently. Separate chatbots, separate data stores, significant cost duplication, and no unified picture of the data needed to power AI effectively.

The IT leadership recognized the fragmentation and initiated a consolidation effort. Within three to four months, the team brought 50 to 60 data sources from all eight centers into a single Databricks platform. The proof point that made consolidation possible was CDER, which had already spent five years building a data platform on Databricks. Data sharing between centers that previously took four to five days was drastically sped up. Real-time data streaming replaced batch processing. When the other centers saw those results, adoption followed quickly.

Unity Catalog addressed the security concerns that initially gave some centers pause. FDA handles trade secrets and sensitive regulatory data that requires strict access controls. Unity Catalog provided the governance layer to prove that data could be contained, that assets would not be shared without proper approvals, and that granular table-level access could be enforced across the entire platform.

From chatbot to agentic AI

With the governed data foundation in place, the FDA deployed ELSA to all 16,000 staff. Users can choose from multiple models and conduct their work through a single interface. Within roughly two months of launch, adoption went from less than 1% to 85% of FDA staff.

What surprised the team was how quickly usage moved beyond simple question-and-answer. Medical doctors, scientists, and administrative staff are now creating their own agents at scale, with hundreds of new agents built per week. Staff take their standard operating procedures, regulatory guidelines, and center-specific documents, load them into workspaces within ELSA, and build agents that can answer grounded, FDA-specific questions instantly.

The architecture that makes this possible layers MCP servers on top of Unity Catalog. The combination of structured, governed data and accessible tooling turned agent creation into something any staff member can do, not just data scientists.

Answers in three minutes instead of days

The impact is concrete. One example: FDA reviewers evaluating drug applications need to understand starting materials, the raw inputs used in manufacturing. That information is buried across three to four million pages of regulatory submissions. Reviewers previously opened individual documents, ran keyword searches, and pieced together answers manually.

Using Databricks ML and NLP capabilities through MLflow, the team extracted key data assets (starting materials, product-supplier-manufacturer relationships) from millions of pages and exposed them through ELSA. Now a reviewer enters an application number, asks for the starting materials, and gets a grounded answer in about three minutes. The same task previously took days.

Scaling across all centers

The FDA is now focused on extending this model across the organization. MCP tools built for CDER are being adapted for other centers, each with its own data context and regulatory requirements. The goal is to free review staff from hunting for information so they can focus on their core expertise: evaluating whether drugs, devices, and biologics are safe and effective.

The foundation that made it all possible was not the AI itself, but the governed data platform underneath it, the consolidation that broke down silos, and the access controls that earned trust across eight independent centers.

“If we can get our review staff to not spend time searching for information and instead focus on their core job, that is where we really see success.”—Venu Boppana

Get the latest posts in your inbox

Subscribe to our blog and get the latest posts delivered to your inbox.

Sign up

View all blogs