The FDA deployed a generative AI platform called ELSA to all 16,000 of its staff. Within roughly two months, adoption climbed from less than 1% to 85% of staff using it daily. That number matters. But the story behind it is mostly about the data infrastructure built underneath the AI.
TL;DR
- -The FDA unified eight centers' separate data stores onto Databricks, then deployed ELSA, a generative AI platform, to all 16,000 staff, reaching 85% daily adoption within roughly two months.
- -Unity Catalog governance resolved security objections from centers handling trade secrets, and staff without data-science backgrounds are now building hundreds of AI agents per week.
- -CIOs evaluating internal AI rollouts should audit whether fragmented data stores and access-control gaps are limiting adoption before deploying a front-end AI tool.
Eight Centers, Eight Separate Data Stores
The FDA's structure reflects the breadth of its mandate. Eight centers covering drugs, biologics, devices, veterinary medicine, tobacco, food safety, and inspections had each built AI capabilities independently. Each ran separate chatbots and separate data stores. There was no unified data picture and significant cost duplication.
One center made consolidation credible. CDER had already spent five years building a data platform on Databricks. Other centers saw the results. Data sharing that previously took four to five days was drastically sped up. Real-time streaming replaced batch processing. Buy-in followed.
Within three to four months, the team brought 50 to 60 data sources from all eight centers into a single Databricks platform.
Governance Was the Gating Factor
The FDA handles trade secrets and sensitive regulatory submissions. Security concerns gave some centers pause at first. Unity Catalog provided the governance layer that resolved those objections. It enforced granular table-level access controls. It ensured data assets were not shared without proper approvals. That demonstration of containment earned cross-center trust.
Venu Boppana leads AI strategy and innovation at the FDA's Office of Digital Transformation. He described the effect: "We showed the value of having a foundational data platform, and that success story became contagious."
Staff Started Building Agents Without Being Asked
Once the governed foundation was in place, ELSA went live. Staff can choose from multiple models and work through a single interface. Usage moved beyond simple question-and-answer faster than the team expected.
Medical doctors, scientists, and administrative staff are now building hundreds of new agents per week. They load standard operating procedures, regulatory guidelines, and center-specific documents into ELSA workspaces. The agents answer grounded, FDA-specific questions. Staff without data-science backgrounds are building them.
The architecture layers MCP servers on top of Unity Catalog. That combination makes agent creation accessible across the organization.
Days Reduced to Three Minutes
One result is specific and measurable. Drug reviewers evaluating applications need to identify starting materials. Those are the raw manufacturing inputs. That information is buried across three to four million pages of regulatory submissions. Reviewers previously searched documents manually.
Using Databricks ML and NLP capabilities through MLflow, the team extracted key data assets and exposed them through ELSA. A reviewer now enters an application number and asks for starting materials. The same task that previously took days now takes about three minutes. That is the clearest performance claim in the evidence.
What CIOs Should Evaluate
This account comes from a single Databricks blog post by Molly Just-Behr, published July 23, 2026. The source does not include independent verification, cost figures, or detail on failure modes during rollout. Those are material gaps for any organization considering a similar path.
With that scope stated, Metrotechs analysis points to three audit questions worth running before scaling internal AI.
Map your data fragmentation first. The FDA's adoption speed followed consolidation, not the AI launch. If data lives in departmental silos with inconsistent access controls, a front-end AI tool will expose that fragmentation rather than fix it.
Test governance with your most skeptical stakeholder. The FDA's cross-center buy-in came after Unity Catalog demonstrated containment to centers with the highest security threshold. Identify that team in your organization and start there.
Watch whether non-technical staff build agents without prompting. The FDA source frames bottom-up agent creation as a surprise outcome, not a planned feature. If your rollout produces only IT-driven use cases, the enablement model may need revision.
The source does not state whether the FDA's architecture or tooling choices transfer without modification. Treat the sequence as a hypothesis worth testing: governed data first, AI layer second, broad enablement third.
Sources and supporting resources