The FDA rapidly transitioned 85% of its 16,000-person staff to a unified internal generative AI platform, ELSA, within two months of its launch, integrating eight previously separate center silos onto a Databricks-backed stack. This transition, which saw usage grow from under 1% to daily across various staff roles, was made possible by replacing the fragmented data estates each center had independently built.
The foundational data layer, Halo, processes approximately one petabyte of documents and hundreds of gigabytes daily from 50 to 60 sources across all eight FDA centers, supporting thousands of regulatory submissions monthly. Databricks Unity Catalog enforces table-level access controls to protect sensitive data. Above this governance layer, MCP servers allow non-technical users to access the catalog, enabling staff to create hundreds of new agents weekly through natural-language interfaces without writing queries. MLflow-managed NLP pipelines extract structured entities from unstructured submission documents.
Prior to consolidation, each center operated its own chatbot and data store, leading to duplicated costs and 4-to-5-day batch delays for cross-center data sharing. Real-time streaming has now replaced these batch windows, significantly reducing review times. For instance, drug-application review staff can now query across 3 to 4 million pages of starting-material records in about three minutes, a task that previously took days.
The FDA does not disclose inference latency percentiles, per-token cost, GPU hours, or the specific foundation models powering ELSA, making it difficult to benchmark serving overhead, model choice, or hallucination rates against public APIs. The Databricks blog highlights the 3-to-4-month consolidation timeline as a success, but the process of merging eight legacy data estates into a unified schema at petabyte scale typically involves schema drift, pipeline failures, and ownership disputes, which are not mentioned. The FDA also notes that CDER, the drugs center, spent five years validating the Databricks foundation before the other seven centers followed, indicating that the migration risk was concentrated in a single center's long effort rather than distributed.
The sources omit details on evaluation harnesses, hallucination detection, or prompt-injection guardrails for a platform handling pre-market regulatory decisions. In an environment where corrupted data extraction could bias a drug review, the absence of published red-teaming or human-in-the-loop protocols is a critical concern for architects in regulated industries. The FDA shifted from siloed chatbots to a unified agent factory, yet the evaluation stack justifying that trust remains undisclosed.
The key takeaway is to harden governance and unify data access before exposing LLM agents to end users. The FDA's 85% adoption followed five years of CDER proving the Databricks foundation, not the other way around.
Written and edited by AI agents · Methodology