FDA deployed agentic AI agency-wide in January 2026. Here's what that means for your 510(k).

FDA named their agent Elsa. Deployed it agency-wide January 12, 2026. Not a pilot. Not a selected center. Agency-wide — premarket reviews, 510(k) processing, postmarket surveillance, inspections. The announcement was framed as an operational change, not a technology experiment. That means your next 510(k) is read by an AI reviewer before a human reviewer sees it.

The agent isn't looking for typos. It's looking for structural problems — inconsistencies between sections, missing predicate comparisons, gaps in clinical evidence summaries against FDA guidance expectations, adverse event narratives that don't match the risk analysis. Things a human reviewer takes two weeks to surface, Elsa finds in minutes.

This changes the submission calculus. Not eventually. Now.

What Elsa actually does

Elsa operates as a network of coordinated agents working across FDA's review workflow simultaneously. FDA's January 2026 announcement describes four primary functions: summarizing adverse event reports from MAUDE, flagging narrative inconsistencies within and across submission sections, surfacing missing predicate comparisons in 510(k) submissions, and identifying gaps in clinical evidence summaries against performance testing expectations in FDA guidance documents.

That fourth function has the most direct impact on 510(k)s. FDA has published guidance for hundreds of device types — specific performance benchmarks, testing standards, and clinical evidence thresholds a submission is expected to address. Elsa has access to all of it. When your performance testing section doesn't explicitly map to the guidance FDA expects, the agent flags the gap. A human reviewer doing this manually might miss it, or take three cycles of Additional Information requests to surface it. Elsa surfaces it on the first pass.

The adverse event summarization matters for predicate selection. A predicate with significant adverse event history in MAUDE is a signal that substantial equivalence claims need careful examination. Elsa correlates your predicate choice against the MAUDE record automatically. If your predicate has known field failures in the performance area where you're claiming equivalence, that surfaces before your submission reaches a human reviewer.

The narrative consistency check is the one most engineering teams underestimate. A 510(k) spans multiple sections — device description, intended use, substantial equivalence, performance testing, biocompatibility, software documentation, labeling — written at different times, often by different people, assembled at the end. Elsa reads them as a single document and flags wherever the narrative drifts. A device described in Section 2 as having three operating modes but tested in Section 10 for only two is a flag. An intended use statement mentioning one patient population but clinical data covering a different one is a flag. These inconsistencies exist in almost every hand-assembled submission. Elsa finds them all.

What happens when Elsa flags your submission

AI-flagged submissions arrive at human reviewers pre-annotated with the gaps Elsa identified. The human reviewer doesn't start with a blank slate. They start with a list of issues an agent already found, evaluating whether your submission adequately addresses those issues — or whether you'll receive an Additional Information request that stops the 510(k) clock.

i-GENTIC AI — an AI regulatory submission company led by the former Medtronic CTO — launched submission agents in February 2026 specifically because their team saw this shift inside FDA and understood the asymmetry it creates. Their early data shows a 97% first-time submission success rate for submissions built with agent assistance. That number is consistent with what you'd expect when submissions are reviewed against the same framework the reviewing agent uses.

The submissions that struggle in this environment are the ones built the old way: collected after the fact, assembled from Word documents, copy-pasted into eSTAR, with design rationale reconstructed from engineer memory weeks or months after the decisions were made. Those submissions have always had consistency problems. Now there's an agent specifically trained to find them.

How a hand-assembled submission fails an automated consistency check

Walk through the mechanics of a typical 510(k) assembly. Design work finishes — or reaches a milestone close enough to submission. A regulatory team collects documentation from engineering: the DHF, test reports, risk management file, software documentation, biocompatibility data. Engineers write summaries of decisions made months earlier. The intended use statement is drafted from product marketing copy and retrospective engineering input. The substantial equivalence argument is assembled by a regulatory professional synthesizing information from multiple sources, none of which were written with this submission in mind.

At each step there's drift. The device description reflects how engineers describe the device this week, not how it was described in the design specification eight months ago. Performance testing reflects what was actually tested, which may have diverged from acceptance criteria in the design inputs when tests failed and were revised. The risk analysis references mitigations described differently in the FMEA than in the labeling section. The software documentation describes a version one release behind the device that will actually be submitted.

None of this is deliberate. It accumulates because the submission is a document representing engineering work — not the engineering work itself. The gap between the work and its representation grows with every week between the engineering decisions and the submission assembly.

Elsa finds all of it. Not because it understands the engineering. Because cross-referencing claims across sections of a structured document is exactly what a language model is built to do, at a thoroughness and speed no human reviewer can match.

The specific failure modes:

• Predicate comparison gaps. Your submission claims substantial equivalence in three performance areas. Testing covers two. Elsa flags the missing third before a human reviewer sees the submission. You receive an RTA.
• Intended use drift. The intended use statement was written from the product brief. Clinical performance data was collected for a slightly different patient population. The mismatch is three sentences apart in two sections. Elsa finds it.
• Software version inconsistency. Software documentation references version 2.3. Device description references version 2.4. One section was updated after a late software change; the other wasn't. Elsa flags it.
• Risk mitigation gaps. The risk management file identifies a mitigation implemented in labeling. The labeling in the submission doesn't include that warning. Elsa flags the gap between the risk file and the labeling section.
• Adverse event context. Your predicate has 47 MAUDE reports related to the same performance characteristic you're testing. Your submission doesn't address this history. Elsa surfaces it as missing predicate context.

Each of these is a flag, not an automatic rejection. But each flag is a potential RTA or Additional Information request. Each Additional Information request adds 90 to 180 days.

What a submission that passes looks like

The submission that passes Elsa's consistency check isn't a better-written version of the hand-assembled submission. It's a structurally different artifact. It doesn't drift because it wasn't assembled — it was generated.

A submission-ready engineering record has one source of truth for every claim. The device description isn't a narrative written for the submission — it's the device specification, formatted for eSTAR. The intended use isn't drafted from marketing copy — it's the user need statement that's been in the DHF since the beginning of the program. The performance testing section doesn't summarize test reports — it cites specific test results linked to the acceptance criteria the tests were written to confirm. The risk analysis isn't reconstructed — it's the live risk management file, with every control linked to its mitigation implementation.

When the device description and performance testing section say different things about operating modes, it's because they were written by different people at different times. When both are generated from the same engineering record, they can't say different things. One source.

Building toward this looks like: capturing design rationale at decision time, not after the fact. Linking test results to the acceptance criteria they confirm — bidirectional traceability throughout the program, not a matrix built at the end. Maintaining a single risk management file as the source for risk claims everywhere. Tracking software versions in one place. Using predicate MAUDE history as input to risk analysis at the beginning of the program, so by submission you've already addressed the adverse event history your predicate carries.

This isn't a documentation methodology. It's an engineering methodology. The submissions that succeed in Elsa's review environment are the ones built from programs that treated documentation as a product of engineering work — not a separate effort that happens after it.

MANKAIND

FDA is using agents to review your submission. The agents are good at finding inconsistencies across sections of a structured document. They operate on every submission, not a random sample. There's no submission that doesn't go through this review.

MANKAIND captures engineering decisions as they happen. Design inputs, risk controls, verification results, software version records, design rationale — live, linked, traceable. The platform understands the 510(k) structure. When you reach the submission window, it generates the submission from the live engineering record. Not a document assembled from scattered artifacts. A structured output from a single source of truth maintained throughout the program.

The device description and performance testing section agree on operating modes because both were generated from the same device specification. The risk analysis and labeling agree on mitigations because the labeling was generated from the risk management file. The predicate comparison addresses performance areas your guidance mapping identified at the start of the program — not areas you identified by reading the Additional Information request.

FDA is running agents on your submission. Run agents on your engineering record. The teams that don't make this shift will spend the next cycle answering Additional Information requests for inconsistencies an agent found in minutes.