What the FDA's 85% Adoption Rate Teaches Enterprises About Custom AI

When a US federal agency gets 85% of its 16,000 staff using an internal AI platform daily within two months of launch, that’s not a pilot — that’s a blueprint. The FDA just proved that enterprise AI adoption isn’t a technology problem; it’s a data foundation problem, and most companies are still trying to solve it in the wrong order.

The agency’s Office of Digital Transformation built ELSA, a generative AI platform now available to every FDA employee, on top of Halo — a governed data platform running on Databricks. According to Venu Boppana, Strategy & Innovation Leader (AI) at the Office of Digital Transformation, adoption jumped from less than 1% to 85% within roughly two months. That’s a number worth studying, because it shows exactly why so many enterprise AI initiatives stall.

Why Consolidating Data Silos Had to Come First

The FDA regulates food, drugs, medical devices, biologics, veterinary medicine, tobacco, and inspections across eight independent centers — each of which had built its own AI capabilities, chatbots, and data stores. Before consolidation, sharing data between centers took four to five days. The IT leadership brought 50 to 60 data sources from all eight centers into a single Databricks platform in three to four months, replacing batch processing with real-time streaming.

This matters because most enterprises confuse “deploying AI” with “having AI-ready data.” The FDA’s team didn’t lead with a flashy chatbot; they led with plumbing. CDER, the drug evaluation center, had already spent five years building on Databricks — and that proof point is what made the other seven centers agree to consolidate.

If you’re a mid-sized enterprise running separate CRMs, ERPs, and product databases across business units, you know this problem. The ROI story that unlocks cross-functional buy-in usually comes from one unit’s early success — not a top-down mandate. Building enterprise data pipelines and reliable API integration is the boring work that makes the exciting work possible.

The prediction: over the next 18 months, the CIOs who win with AI will be the ones who spent 2024–2025 quietly consolidating data governance, not the ones who bought the most model licenses.

How Unity Catalog Solved the Trust Problem

The FDA handles trade secrets and sensitive regulatory data. Any consolidation effort would have died on arrival without airtight access controls. Unity Catalog provided the governance layer that proved data could be contained, assets would not be shared without proper approvals, and granular table-level access could be enforced across the entire platform.

This is the piece that non-technical buyers routinely underestimate. Security and governance aren’t just compliance checkboxes — they’re the political currency that lets you convince a skeptical business unit to hand over its data. Without table-level access controls, the FDA’s eight centers would have kept their silos, and ELSA would never have been built.

Imagine you’re the head of a bank considering an internal AI assistant across retail, wealth management, and commercial lending divisions. Each division has legitimate reasons to distrust a shared data layer. The Unity Catalog approach — provable, granular, auditable access — is what turns “no” into “yes.” For regulated industries, this is especially relevant to fintech and banking software that has to pass audits.

Our take: governance tooling is going to become the highest-leverage line item in enterprise AI budgets, and vendors that can’t prove auditability at the row level will get quietly dropped from RFPs.

When Employees Start Building Their Own Agents

Here’s what leadership usually misses about AI rollouts: within two months of ELSA’s launch, medical doctors, scientists, and administrative staff started creating their own agents — hundreds per week. Not data scientists. Regular staff loading their SOPs, regulatory guidelines, and center-specific documents into ELSA workspaces and building grounded, FDA-specific agents.

This happened because the architecture layered MCP servers on top of Unity Catalog. That combination — governed data plus accessible tooling — turned agent creation into something anyone could do. The traditional bottleneck of “we need three data scientists to build a use case” disappeared.

If you’re running a professional services firm, this means your consultants could be building their own client-specific research agents by end of quarter — provided you’ve done the underlying data work. That distinction matters, and it’s worth understanding when AI agents actually beat traditional automation before you commit to an architecture.

Prediction: within a year, “citizen agent development” will be a standard KPI in enterprise AI programs, and the vendors selling closed, developer-only agent platforms will lose to the ones that expose MCP-compatible tooling to end users.

The Three-Minute Answer That Justifies the Whole Program

The concrete win: FDA reviewers evaluating drug applications need to identify starting materials buried across three to four million pages of regulatory submissions. Reviewers used to open individual documents, run keyword searches, and manually piece together answers over days. Using Databricks ML and NLP through MLflow, the team extracted starting materials and product-supplier-manufacturer relationships and exposed them through ELSA. Now a reviewer enters an application number and gets a grounded answer in about three minutes.

The business case writes itself. Days become minutes. Highly paid experts stop doing document archaeology and start doing the judgment work only they can do. As Boppana put it, “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.”

For any enterprise buyer, this is the template for justifying custom AI investment: pick one workflow where senior staff waste hours on information retrieval, quantify the current cost, and rebuild it as a grounded agent. That’s the story that gets budget approved for the next 20 use cases.

FAQ

Q: What is a governed data platform, and why does it matter for enterprise AI? A: A governed data platform consolidates data from multiple sources with unified access controls, lineage tracking, and audit capabilities. It matters because AI agents can only be as trustworthy as the data they’re grounded in — without governance, you can’t safely expose sensitive data to LLMs across a large organization.

Q: How is ELSA different from just deploying ChatGPT internally? A: ELSA is built on top of Halo, the FDA’s governed data foundation, and uses MCP servers layered over Unity Catalog to let staff build agents grounded in specific FDA documents and SOPs. A generic ChatGPT deployment can’t access internal regulatory submissions with proper access controls, and it can’t be customized by non-technical staff to answer center-specific questions.

Q: Can a smaller enterprise replicate the FDA’s approach? A: Yes, the principles scale down. The pattern — consolidate data sources, establish governance, deploy a shared AI interface, then let teams build their own agents — works for a 500-person company as well as a 16,000-person agency. The technology stack may differ, but the sequencing does not.

Key Takeaways

• Enterprises that skip the data consolidation step will hit an adoption ceiling long before they hit ELSA’s 85% number, regardless of which foundation model they license.
• Governance tooling like Unity Catalog is not a compliance line item — it’s the political mechanism that unlocks cross-departmental data sharing, and buyers should evaluate AI platforms on governance depth first.
• Expect “citizen agent development” to become the dominant enterprise AI pattern within 12 months, making MCP compatibility a required feature rather than a nice-to-have.
• The strongest ROI case for custom AI comes from a single high-value workflow (like the FDA’s three-minute starting-materials lookup) — pick that battle before proposing an enterprise-wide platform.
• Success stories inside your own organization travel faster than any external case study; find your CDER-equivalent early adopter and fund them disproportionately.