FDA says its AI assistant, Elsa, now sits on top of the agency’s consolidated application and submission data. For an agency built around evidence, that raises a fair question: how does the system affect a device review?
Most of what circulates as the answer came from trade coverage, repeated often enough to read as settled fact. The agency’s own record is much thinner.
What FDA has confirmed about Elsa
FDA launched Elsa agency-wide on 2 June 2025. The agency said it arrived ahead of its 30 June deadline and under budget.
The first use cases were practical: summarizing adverse event reports, comparing labels, helping with clinical protocol reviews, generating code for nonclinical databases and identifying inspection priorities. FDA also said the models would not train on data submitted by regulated companies.
The architecture was less clear. FDA described a high-security GovCloud environment. The launch release did not name AWS, Deloitte, a retrieval architecture or a foundation model. Those details arrived later through trade coverage and have circulated since as though the agency confirmed them.
The public story changed on 6 May 2026. FDA announced Elsa 4.0 and HALO, a platform consolidating more than 40 application and submission sources across the agency. It said it had begun connecting Elsa to HALO, so staff can work with agency data without uploading it by hand.
Chief AI Officer Jeremy Walsh put it plainly: “Previously, FDA staff would bring data to Elsa. Now, Elsa sits on top of our data.”
Elsa 4.0 added custom agents, document generation, quantitative analysis, charts, OCR, dictation and secure web search. FDA said the system runs in a FedRAMP High Google Cloud environment and still does not train on industry data.
None of this shows that Elsa reads every 510(k), reads a submission before a person does or shapes a final decision.
What is known about Elsa in device review
The Elsa 4.0 announcement mentions medical devices once, in the boilerplate description of the agency at the bottom. It does not mention CDRH, 510(k)s or device review.
Silence proves nothing on its own. It is still the most current evidence available, and it does not support the idea that device review is where Elsa has been pushed hardest.
The clearest account of a device-specific tool is older. NBC News reported in June 2025 that two people familiar with CDRH-GPT described a beta system meant to help device reviewers. They reported failed uploads, broken queries and weak connections to internal systems. They also said it could not reach recent or paywalled literature.
That reporting describes a troubled beta in June 2025. It says nothing about the tool’s condition today or whether its functions moved into Elsa 4.0.
The same caution applies to hallucinations. CNN reported in July 2025 that current and former FDA staff had seen Elsa invent or misstate research. Commissioner Marty Makary described a narrower role, pointing to tasks such as finding studies and summarizing meetings. He also said use was optional.
Both accounts can be true. A system can save time and still produce confident nonsense, which is why the controls matter more than the demo.
Inspections are a separate claim
FDA started a one-day inspectional assessment pilot in April 2026. By late April, the agency said it had completed about 46 assessments and most ended with No Action Indicated.
The FDA announcement says facilities are selected using risk-based criteria such as product type, prior inspection outcomes and operational characteristics. Higher-risk or complex facilities are not eligible. The release does not mention Elsa or AI at all. A later Sidley analysis reported that Makary linked the pilot to AI-backed risk analysis.
FDA is using AI in operations and it is testing shorter inspections. Public evidence does not show that Elsa selects the sites.
The gap is traceability
Device companies know the rule: if software helps create or maintain a regulated electronic record, its controls need to match the risk. Under 21 CFR 11.10, covered systems require validation, access controls and record protection.
Part 11 does not automatically apply to FDA’s internal deployment in the same way. The comparison is still hard to ignore. Industry must explain what its systems do, who used them and how the record stayed trustworthy. FDA has not publicly described the equivalent controls around Elsa in enough detail for sponsors to understand its role.
The open questions are basic:
When can Elsa access a device submission?
Which review tasks may use its output?
Does a reviewer have to verify or cite that output?
Are prompts and responses retained with the decision record?
How does FDA test a model change before it reaches regulated work?
The last one has already come up. NOTUS reported in March 2026 that HHS removed enterprise access to Anthropic’s Claude. An internal FDA banner reportedly said Elsa was moving away from Anthropic and that Google’s Gemini would become its primary model. FDA’s public Elsa releases do not identify the models or explain the change controls.
That does not prove FDA skipped validation. It means the public cannot assess it.
What this changes for a submission
Very little. Nobody needs to write for an imagined robot reviewer, and the habits that help a human reviewer also help software:
the conclusion placed beside the evidence that supports it
the same device, endpoint and predicate terms throughout
tables kept machine-readable, with native text instead of scans
a record of material FDA communications, including how any AI-derived issue in review feedback was verified
I think FDA is right to automate clerical work. Reviewers should spend less time hunting through files and more time judging evidence.
FDA asks device companies to show their work. On Elsa, the agency should do the same.