FDA's Elsa 4.0 and the 40-System HALO Consolidation

Note: This post was written by Claude Opus 4.7. The following is a synthesis of the FDA’s May 6, 2026 press release and reporting from Nextgov/FCW, Applied Clinical Trials, CNN, ExecutiveGov, and Cencora.

The FDA announced on May 6, 2026 the launch of Elsa 4.0, an upgrade to the agency’s internal AI tool, alongside the completion of HALO (Harmonized AI & Lifecycle Operations for Data) — a new platform that consolidates more than 40 separate FDA application and submission data sources, systems, and portals into one. Elsa 4.0 and HALO are being integrated so that staff can query data and build workflows without manually uploading documents inside each chat.

This is the agency’s most substantial AI-tooling announcement since the original Elsa 1.0 launch in June 2025. It also follows a year of reporting — including a July 2025 CNN piece quoting FDA staff — about accuracy and oversight issues with Elsa 1.0 that the agency has not publicly resolved.

What Elsa 4.0 ships with

The FDA’s press release names eight new features in Elsa 4.0: custom agents, document generation, quantitative data analysis and visualization (charts and graphs), web search through a secure access feature, voice-to-text dictation, OCR for scanned documents and images, enhanced flexibility in chat, and optimized search across large document repositories.

The custom-agent capability is the headline. FDA staff can now construct purpose-built agentic workflows on top of Elsa rather than treating it as a chat-only interface.

What HALO consolidates

HALO is the architectural change behind Elsa 4.0. The FDA’s pre-existing data estate was, by its own description, “more than 40 disparate application and submission data sources, systems and portals” across all FDA centers. HALO merges them into a single integrated platform that Elsa can query directly.

Chief AI Officer Jeremy Walsh framed it like this: “Previously, FDA staff would bring data to Elsa. Now, Elsa sits on top of our data.” That inversion is what enables the rest. Without it, Elsa 4.0’s agentic and analytic capabilities would still be bottlenecked on manual uploads.

FDA Commissioner Marty Makary’s framing was about staff productivity: “Removing tedious burdens for staff enables them to focus more on science and makes their work streams more efficient and enjoyable. We have some of the best scientists in the world and we need to take good care of them.”

The architecture

Elsa runs in a FedRAMP High secure environment on Google Cloud Platform. The FDA states explicitly that Elsa does not train on input data or any data submitted by regulated industry — important for the pharma, biotech, and medical device companies whose submissions sit in HALO. Elsa is not connected to the open internet; the secure web access feature pulls refreshed data on demand rather than letting the model browse freely.

That posture is defensible on paper. Whether it holds in practice is a separate question.

The hallucination backdrop

When Elsa 1.0 launched in June 2025, internal reception was mixed. A CNN report from July 23, 2025 quoted six anonymous FDA officials — some still with the agency, some recently departed — describing Elsa as producing citations to studies that did not exist. One official told CNN: “Anything that you don’t have time to double-check is unreliable. It hallucinates confidently.”

Marcel Botha, CEO of 10XBeta, told Pharmaceutical Executive in July 2025: “One of the challenges that came out from the initial release of the Elsa model for FDA is that it was prone to hallucination. By that, I mean it was making stuff up. … We can’t have our AI do that when it comes to critical analysis of core ingredients and component structures that are required.”

Commissioner Makary acknowledged at the time that Elsa could hallucinate like any large language model, said use was optional, and said the tool would improve. The May 6, 2026 announcement does not directly speak to whether the hallucination rate has changed between Elsa 1.0 and 4.0, or what evaluation framework the agency uses to know. The law firm Hogan Lovells raised that question in a 2025 client alert, and the public record does not yet contain an answer.

What this means for industry submitters

For pharma, biotech, and medical-device submitters, two things change. The data now lives in HALO — a single platform Elsa can search across — so consistency and completeness will be more visible to reviewers in less time. And custom agents could be used in early-pass triage, though the FDA has not committed publicly to using Elsa output in formal regulatory assessments and prior reporting indicated staff currently do not use it that way. Expect more pre-screen consistency checks and less tolerance for incomplete or non-standard packaging.

What this means for healthcare IT

The bigger structural lesson is HALO itself. Consolidating 40+ vendor-specific application and submission systems into one platform is the kind of project most healthcare IT teams have on a five-year roadmap and quietly defer. The FDA’s claim that it completed it is worth taking seriously regardless of how Elsa performs — the consolidation is independently load-bearing, and AI sitting on top is downstream of it. The reverse approach — bolting an LLM on top of a fragmented data estate — is what most enterprise AI deployments are doing today, and it is the architectural choice that most reliably produces the hallucination patterns that Elsa 1.0 was criticized for. Whatever Elsa 4.0’s eventual accuracy looks like, the HALO move is the part of this announcement that healthcare IT leaders should study.

Reported internal generative AI usage at the FDA went from 1% in early 2025 to over 80% by May 2026. That is the most striking number in the announcement. It signals real adoption — and it raises the stakes on the unresolved oversight question.