Databricks How the FDA Reached 85% Daily AI Use by Fixing Its Data First

Databricks says the FDA moved from scattered AI pilots to 85% daily staff adoption within roughly two months. The striking part is what came first. The agency did not begin with a better chatbot. It first consolidated data, permissions, and access across eight independently operated centers.

That distinction makes the Databricks how story more useful than another report about employees receiving an AI assistant. The FDA had already experimented with separate chatbots and data stores. Its harder problem was making sensitive information usable across organizational boundaries without weakening controls.

The primary conflict is therefore not one AI vendor against another. It is governed infrastructure against fragmented experimentation. The FDA’s reported adoption rate suggests employees use AI when it reaches trusted information and removes a specific burden. However, the public evidence still leaves important questions about measurement, accuracy, and effects on regulatory outcomes.

The FDA Turned Eight AI Islands Into One Platform

The FDA’s central change was organizational and architectural: eight centers began sharing a governed foundation instead of building isolated AI systems.

The agency’s mandate spans drugs, biologics, medical devices, food, veterinary medicine, tobacco products, and inspections. Those responsibilities sit across centers with distinct workflows, technical systems, and regulatory requirements. Each center had developed its own data resources and, in some cases, separate AI capabilities.

That independence created familiar enterprise problems. Teams duplicated technology spending, stored related information in different places, and lacked a unified view of data. A useful model in one center could not automatically reach documents or structured records controlled by another.

The scale made those divisions especially costly. Databricks says the FDA manages about one petabyte of documents, receives hundreds of gigabytes daily, and processes thousands of regulatory submissions each month. Those figures come from a vendor case study and have not been independently audited in public.

According to the FDA platform account, its technology leadership consolidated 50 to 60 data sources within three to four months. The sources came from all eight centers and entered a shared Databricks environment.

The Center for Drug Evaluation and Research, known as CDER, provided the working precedent. CDER had spent five years developing its own data platform on Databricks. Its results gave the broader organization evidence that a shared foundation could support real regulatory work.

Data exchanges that reportedly required four or five days became much faster. The platform also replaced some batch processing, where information moves on a schedule, with real-time streaming. Streaming makes new records available as they arrive instead of waiting for the next batch.

The technical consolidation alone did not settle the matter. FDA centers handle trade secrets, proprietary submissions, and other sensitive material. A common data layer would fail politically and operationally if one center feared losing control of its information.

Unity Catalog, Databricks’ centralized governance layer, became the control point. It applies permissions and tracks access across data assets. Databricks says the FDA used it to enforce granular controls, including access at the table level.

That governance layer established a practical rule. Sharing infrastructure did not mean sharing every record with every employee. Teams could connect information while preserving approval requirements and center-specific boundaries.

The agency later described the consolidated platform as HALO, short for Harmonized AI and Lifecycle Operations for Data. In May 2026, the FDA said HALO combined more than 40 application and submission sources, systems, and portals. The difference between that figure and Databricks’ 50 to 60 sources likely reflects different counting methods.

This is the first important lesson from the Databricks how account. Consolidation did not erase local responsibilities. It created a common way to locate, govern, and use information while centers retained control over protected data.

The sequence also mattered. The FDA did not ask every employee to trust a general chatbot with sensitive work. It established the information layer, demonstrated value inside CDER, and then expanded the model. Adoption followed evidence that the foundation could reduce delays without removing access controls.

That sequence challenges a common enterprise pattern. Many organizations buy an AI interface while their permissions, documents, and databases remain fragmented. Employees then encounter incomplete answers, missing context, and systems that cannot support consequential work.

A shared data foundation does not guarantee reliable AI. It does remove one predictable cause of failure: an assistant that cannot reach the right information. The FDA’s next step was placing an agency-wide interface above that foundation.

Databricks How Elsa Moved From Chatbot to Daily Work

Elsa spread because the FDA connected one interface to governed information and let employees shape it around existing procedures.

The FDA launched Elsa agency-wide on June 2, 2025. Elsa is a generative AI assistant designed for reading, writing, summarization, search, and related employee tasks. The initial Elsa launch made it available to scientific reviewers, investigators, and other staff.

The launch followed a scientific review pilot announced in May 2025. Commissioner Marty Makary ordered the agency to expand the system across all centers by June 30. The FDA released Elsa ahead of that deadline.

The agency initially listed several practical uses. Elsa could summarize adverse-event information, compare product labels, and generate code for certain databases. FDA officials also said teams were using it for clinical protocol reviews and inspection prioritization.

These were bounded tasks rather than a promise that AI would make regulatory decisions. The distinction is important. Search, summarization, and document comparison can reduce preparation time while leaving judgment with qualified reviewers.

The FDA said Elsa operated inside a high-security GovCloud environment. It also said the models did not train on information submitted by regulated companies. Those protections addressed a central concern for pharmaceutical, biotechnology, food, and device businesses.

By December 2025, the agency said more than 70% of staff used Elsa voluntarily. It then deployed agentic functions, which allow software to plan and execute several connected steps under defined instructions. Staff could build workflows for meeting management, reviews, surveillance, inspections, compliance, and administrative work.

The agentic AI deployment also included a two-month internal challenge. Employees were invited to build solutions and demonstrate them during the FDA’s Scientific Computing Day. That program turned adoption into participation rather than simple software distribution.

Databricks now reports that use rose from below 1% to 85% of the FDA’s 16,000 staff within roughly two months. Its headline describes daily use, although the article does not publish the underlying analytics, date range, or definition of an active user.

That limitation matters. Opening an application once each day differs from completing a consequential workflow. An adoption percentage also says little about answer quality, time saved, or whether usage remains consistent across centers.

Still, the reported behavior offers more substance than a basic license count. Databricks says medical doctors, scientists, reviewers, and administrative staff are creating hundreds of agents each week. Employees reportedly use standard operating procedures, guidelines, and center-specific documents to ground those agents.

Grounding means supplying an AI system with approved information relevant to a request. The model can then base its response on those materials instead of relying only on general training data. This approach can improve relevance, though it cannot eliminate inaccurate outputs.

Model choice was also separated from the employee experience. Databricks says Elsa allows access to multiple models through one interface. That design reduces dependence on a single model while keeping data permissions and workflows in a shared layer.

The architecture reportedly places Model Context Protocol servers above Unity Catalog. Model Context Protocol, or MCP, is a standard way for AI systems to connect with approved tools and information. It lets an agent request governed resources without giving it unrestricted access.

For employees, those layers should remain mostly invisible. A reviewer asks a question or launches a workflow. The platform decides which approved information and tools are available under that person’s permissions.

This is where the FDA’s approach differs from a disconnected chatbot rollout. A generic assistant can draft text, but it rarely knows which internal record is authoritative. Elsa is intended to work within the agency’s information boundaries and operational context.

A similar principle applies to personal and team systems. An AI assistant becomes more useful when it works with an organized AI knowledge base, clear sources, and controlled retrieval. The FDA applied that principle at regulatory scale.

The Databricks how account therefore describes an adoption loop. Governed information makes specific workflows possible. Useful workflows attract employees. Their usage identifies new needs, which encourages teams to create additional agents.

However, rapid creation introduces a new governance problem. Hundreds of new agents each week can produce duplicated tools, uneven instructions, or poorly tested outputs. The same platform that lowers the barrier to creation must also support evaluation, ownership, and retirement.

The FDA’s initial achievement was getting employees through the door. Its longer test is whether locally created agents remain accurate, traceable, and manageable as their number grows.

The Three-Minute Review Shows Why Adoption Accelerated

The strongest evidence is not the adoption percentage but a narrow workflow that reportedly fell from days to about three minutes.

Drug reviewers sometimes need to identify starting materials, the raw inputs used during manufacturing. Relevant details can sit inside extensive regulatory submissions. Reviewers must connect those materials with products, suppliers, and manufacturers.

Databricks says this research previously required opening individual files, running keyword searches, and assembling the answer manually. The relevant information could be distributed across three to four million pages. A simple question therefore became a slow document investigation.

The team used machine learning and natural language processing to extract specific data from those documents. It tracked starting materials and relationships among products, suppliers, and manufacturers. MLflow supported the development and management of those data-processing models.

The extracted information then became accessible through Elsa. A reviewer could enter an application number and ask for associated starting materials. Databricks says the system returned a grounded answer in about three minutes, compared with days of manual research.

That claim deserves careful wording. Databricks published the case study, and the public article does not include an independent benchmark. It does not report sample size, accuracy rates, error categories, or whether every submission produced the same result.

Even with that caveat, the workflow explains why employees might return daily. It removes information hunting from a task that already has a clear question, identifiable sources, and a qualified reviewer. The AI does not need to replace the reviewer to provide value.

Venu Boppana, the FDA’s strategy and innovation leader for AI, framed success around that division of labor. He said the goal was to keep review staff from spending time searching, letting them concentrate on their core jobs.

This is a more credible adoption mechanism than requiring employees to invent uses for a general assistant. The platform meets people inside an existing process. It reduces a measurable bottleneck while preserving a human decision point.

The mechanism also clarifies why the data platform mattered. An advanced language model alone would not know which application, supplier, or manufacturing document contained the authoritative answer. It needed extracted records, relationships, permissions, and identifiers.

Search quality depended on earlier data engineering. Teams had to process millions of pages, define important entities, and expose those assets safely. The visible three-minute response sat above years of less visible platform work.

This pattern creates pressure for enterprise AI programs built around demonstrations. A polished conversation interface can impress executives, but employees abandon it when answers lack internal context. Daily use requires repeated value inside ordinary work.

It also creates pressure for isolated departmental tools. A center-specific assistant can move quickly because its scope remains narrow. However, it can reproduce silos when another department needs related information or shared governance.

The FDA kept CDER’s working platform as a starting point, then adapted its capabilities across centers. That approach balanced local proof with organization-wide infrastructure. It avoided waiting for a perfect enterprise design before demonstrating a useful workflow.

The historical sequence matters. CDER’s five-year platform effort preceded the rapid agency-wide adoption story. The two-month adoption figure did not emerge from two months of total development. It followed years of data preparation and a concentrated consolidation phase.

That nuance should shape how other organizations interpret the result. The story does not show that any institution can deploy an assistant and reach 85% use within weeks. It shows that rapid adoption can follow a mature data foundation and a credible internal precedent.

The workflow also illustrates the boundary between assistance and adjudication. Locating starting-material information is not the same as deciding whether a drug is safe or effective. Faster evidence retrieval can support review, but it does not validate the final regulatory conclusion.

FDA policy places human subject-matter experts throughout the process. The agency says staff verify inputs, analytical steps, and how outputs are used. That safeguard is essential because an apparently grounded answer can still omit a relevant document or misread a relationship.

The best reading of the three-minute example is therefore limited but meaningful. The platform reportedly compresses a research task, makes supporting information easier to reach, and returns time to specialists. It does not establish that AI can independently perform scientific review.

What the 85% Adoption Claim Does Not Tell Us

High usage is evidence of access and perceived utility, but it is not evidence of regulatory accuracy or better public-health outcomes.

The 85% figure comes from Databricks’ July 2026 account. The company supplies the underlying data platform, so the case study also supports its commercial narrative. Readers should treat the number as a reported customer metric rather than an independent finding.

The FDA has published earlier adoption data that supports the direction of travel. In December 2025, the agency said more than 70% of staff voluntarily used Elsa. That official disclosure makes broad adoption plausible, but it does not independently confirm 85% daily use.

Several measurement details remain absent. Neither the FDA nor Databricks has publicly explained the minimum activity required to count a daily user. They also have not disclosed adoption by center, job type, or workflow.

Uneven usage could matter. Administrative employees may use summarization frequently, while scientific reviewers may rely on Elsa for fewer, higher-stakes tasks. One agency-wide percentage can hide those differences.

The public sources also lack a system-wide accuracy evaluation. The three-minute starting-material example provides speed, but no precision, recall, or reviewer-correction rate. Those measures would reveal whether faster retrieval also remains dependable.

Governance claims require similar scrutiny. The FDA says Elsa does not train on user inputs or regulated-industry submissions. It also says the system runs in a FedRAMP High environment, a federal security standard for high-impact cloud data.

Those are important controls, but they do not resolve every risk. Authorized users can still request inappropriate information. Agents can combine permitted data in unexpected ways, and generated summaries can omit qualifications found in source documents.

Rapid agent creation expands that challenge. Each custom agent can include different instructions, tools, and document sets. Without consistent testing, two agents answering similar questions might produce conflicting results.

The FDA’s human-review requirement limits the consequences of those differences. Yet human oversight works only when reviewers understand the system’s limits and can inspect supporting evidence. Automation bias can lead people to accept confident answers too readily.

External experts raised these concerns before the broad rollout. Eric Topol of Scripps Research told Axios that the concept was sound, but the limited detail and perceived rush were concerning. He also questioned which models and training inputs supported the system.

Former FDA Commissioner Robert Califf expressed enthusiasm tempered by caution about the timeline. Industry representatives welcomed efforts to make review more efficient but emphasized protection for proprietary submissions. These reactions identified the core tradeoff before Elsa reached broad adoption.

The rollout concerns have not disappeared because usage increased. Adoption can make evaluation more urgent, since more employees and workflows now depend on the platform.

There is also a distinction between employee efficiency and regulatory throughput. A reviewer can find information faster without shortening the total review cycle. Other scientific, legal, manufacturing, and communication steps may remain unchanged.

The FDA has linked AI to faster access to treatments, but public evidence has not established that causal chain. A careful analysis should not convert time saved on one research task into claims about faster approvals or better patient outcomes.

The agency’s May 2026 release provides a clearer view of its intended safeguards. It says experts verify inputs, analytical processes, and implementation of outputs. The same release states that Elsa does not directly connect to the internet, despite offering controlled web search.

Elsa 4.0 also added document generation, quantitative analysis, visualization, optical character recognition, dictation, and custom agents. The Elsa 4.0 update significantly widened the system’s scope.

Broader capability increases both usefulness and exposure. Document generation can save time, but it creates review obligations. Quantitative analysis can surface patterns, but a mistaken assumption can influence the resulting chart.

The FDA now needs evidence beyond logins. Useful measures include correction rates, citations opened, tasks completed, reviewer time saved, and incidents involving inappropriate access. Outcome measures should also distinguish administrative work from scientific review.

None of these uncertainties invalidates the platform’s reported adoption. They define what the adoption number can support. It shows that employees found reasons to return, not that every output was correct or every workflow improved.

That distinction protects the story from becoming vendor promotion. The FDA appears to have solved an important access and infrastructure problem. It still must prove that rapid use can coexist with consistent evaluation across thousands of employees and hundreds of new agents.

Three Signals Will Show Whether the FDA Model Holds

The next phase will test whether the FDA can convert broad experimentation into measurable, governed, and repeatable work across every center.

The first signal is adoption reporting with better definitions. The FDA should disclose active-use criteria, center-level participation, and the share of employees completing meaningful workflows. Stable use after the initial rollout would strengthen the case for durable value.

A decline would not necessarily mean failure. It could reflect the retirement of novelty-driven activity while high-value workflows remain. The important measure is whether repeat use concentrates around tasks with verified time or quality improvements.

The second signal is formal evaluation for custom agents. Hundreds of new agents each week create an inventory problem as well as an innovation opportunity. The agency needs visible standards for testing, ownership, source freshness, permissions, and retirement.

Agent evaluations should cover more than answer fluency. They should test citation accuracy, missing evidence, permission enforcement, reproducibility, and behavior when source documents conflict. Results should be compared before and after model or workflow changes.

Clear evaluation practices would strengthen the governed-platform thesis. Serious incidents, inconsistent answers, or uncontrolled duplication would weaken it. They would show that centralized data alone cannot govern decentralized agent creation.

The third signal is HALO’s expansion beyond the CDER precedent. The FDA says it is adapting tools from the drug center for other centers with different data and regulatory needs. Successful reuse would show that the architecture supports institutional variety.

This will be difficult because food inspections, medical-device reviews, biologics, and veterinary medicine do not share identical processes. A common platform must support different evidence types without forcing every center into one workflow.

The HALO consolidation positions Elsa above agency data rather than requiring users to upload documents into each chat. That change should reduce friction and provide more consistent context.

It also raises the stakes for permissions and data quality. An incorrect access rule can affect many workflows when a central assistant sits above shared systems. A stale or duplicated record can spread confusion faster than it did inside one center.

For enterprise buyers, this Databricks how case offers a demanding benchmark. The visible assistant was only the final layer. Years of data work, a successful center-level precedent, rapid source consolidation, and strict access controls came first.

The story also offers a warning against copying the headline without copying the sequence. Buying multiple models does not create trusted context. Opening an agent builder does not establish evaluation. Counting users does not measure better decisions.

Knowledge workers should care because the same mechanism applies at smaller scales. AI provides more value when it can reach approved documents, preserve source boundaries, and support a defined task. It becomes less trustworthy when context remains scattered or unverified.

Developers should watch how the FDA manages MCP-connected tools and custom agents. The difficult engineering work now sits in permissions, evaluation, observability, and change control. Model selection is only one part of that system.

Regulated enterprises should watch for public evidence linking usage to quality. The strongest future disclosure would pair time savings with error rates and human-correction data. That combination would show whether efficiency and rigor moved together.

The FDA’s reported 85% daily adoption is notable because government technology rollouts rarely earn such rapid voluntary use. Yet the most transferable lesson is quieter. People returned after the platform connected AI to governed information and recognizable work.

The question for the next year is whether that foundation can withstand scale. If the FDA publishes meaningful evaluations, controls its agent inventory, and repeats CDER’s results elsewhere, its approach will become a credible model.

If those signals remain private, the adoption claim will stay impressive but incomplete. Organizations studying databricks how the FDA built Elsa should focus on the evidence beneath the percentage. Their own AI programs should begin with trusted data, explicit permissions, and one workflow worth repeating.