Which AI Healthcare Claims Survive a Regulatory Record

An AI healthcare claim survives diligence when a device class, an authorization pathway and a public FDA record stand behind it. On the FDA list of AI-enabled devices as updated December 5, 2025, 1,304 of 1,357 entries were 510(k) clearances, 37 were De Novo grants and 16 were premarket approvals. The useful question is which record a company holds, and what its claim asserts that the record does not.

FDA device classification is better read as a filter than as a formality. Each class carries a statutory standard, each pathway produces a document with a number on it, and each document says what the agency reviewed and for which intended use. A claim that outruns that document has not been tested by anyone other than the company making it.

01The Classification Statute

Section 513 of the Federal Food, Drug, and Cosmetic Act, codified at 21 U.S.C. §360c, sorts devices into three classes by the level of control needed to provide reasonable assurance of safety and effectiveness. Class I covers devices for which the general controls are sufficient. Class II covers devices for which general controls alone are insufficient but there is enough information to establish special controls, which the statute lists as including performance standards, postmarket surveillance, patient registries and guidelines.

Class III covers devices for which neither general nor special controls are sufficient and which support or sustain life, are of substantial importance in preventing impairment of health, or present a potential unreasonable risk of illness or injury; those devices require premarket approval under section 515.

The implementing rules sit in 21 CFR Part 860. Section 860.3 defines the three classes, and section 860.7 defines the standard that every class is measured against. A device is reasonably safe when the probable benefits to health from its intended use outweigh any probable risks. It is reasonably effective when, in a significant portion of the target population, its use will provide clinically significant results.

Both determinations rest on valid scientific evidence, which section 860.7(c)(2) defines as running from well-controlled investigations to well-documented case histories and reports of significant human experience with a marketed device. The same paragraph excludes isolated case reports, random experience, reports lacking sufficient detail for scientific evaluation and unsubstantiated opinions.

Under section 513(f)(1), any device that was not in commercial distribution before May 28, 1976 is classified in Class III by operation of law unless it is found substantially equivalent to a legally marketed device or is reclassified.

Substantial equivalence, defined in section 513(i), requires the same intended use as a predicate and either the same technological characteristics or different characteristics supported by information showing the device is as safe and effective and does not raise different questions of safety and effectiveness. The premarket notification itself is required by 21 U.S.C. §360(k), and section 360(l) excuses devices of a type classified into Class I from it, subject to the exemption limits FDA sets by regulation, as 21 CFR 862.1645 does for a nonquantitative urinary protein or albumin test system.

For a novel function with no predicate, section 513(f)(2) provides the De Novo request, under which the agency classifies a low-to-moderate-risk device into Class I or II, with a statutory 120-day target. Any device classified through De Novo becomes a predicate for later substantial-equivalence review; Part 860 Subpart D, added in October 2021, sets the procedure.

02The Classification-as-Filter Table

The table below reads each class as a test of what a claim must have on file, then sets that requirement against what the December 5, 2025 update of the FDA list of AI-enabled medical devices shows. Counts are by submission number prefix across all 1,357 entries in that update.

03The Public Record for AI-Enabled Devices

The FDA list is the nearest thing to a census of authorized device AI, but the agency describes it as not comprehensive: entries were identified primarily from AI-related terms in the summary descriptions of marketing authorizations and device classifications, and devices whose decision summaries were not published within the data collection period carry over to a later update.

The current update carries a "content current as of" date of December 5, 2025 and lists 1,357 devices with decision dates from September 29, 1995 through September 29, 2025. The prior update, dated July 10, 2025, listed 1,247. A device authorized after September 29, 2025 does not yet appear in this update, which is a reason to check a company's "FDA-cleared AI" claim against the 510(k), De Novo or PMA database directly rather than against the list.

Radiology is the lead review panel for 1,039 entries, or 76.6 percent, and cardiovascular for 130, or 9.6 percent. The 1,357 entries span only 160 product codes. Most AI on the list is radiology software cleared through substantial equivalence.

The De Novo entries show how a category forms. ContaCT, granted De Novo DEN170073 on February 13, 2018, is the first entry on product code QAS; the list carries 74 later 510(k) entries on that code, and none earlier. QAS is assigned in the FDA product classification database to 21 CFR 892.2080, which classifies radiological computer-aided triage and notification software in Class II and sets special controls, among them a detailed description of the notification and triage algorithms and pre-specified performance testing protocols. Followers on that code inherit those controls.

Across the list, no 510(k) entry predates the De Novo grant on its own product code. The first mover in a new AI category bears the De Novo cost; later entrants inherit its classification and compete on substantial equivalence.

Classification also moves. Six PMA approvals on the list, dated from 1998 to 2017, carry product code MYN for computer-aided detection in imaging. Those devices entered as Class III. 21 CFR 892.2070, in the text codified in 2020, now classifies medical image analyzers, including mammography computer-aided detection, in Class II with special controls, and states that the device is not intended for triage or to recommend a diagnosis. A claim that a product sits in a "PMA-only" category should be checked against the current regulation.

04The Software Carve-Out

Some AI claims rest on the opposite assertion: that the product is not a device at all. Section 3060(a) of the 21st Century Cures Act, Pub. L. 114-255, enacted December 13, 2016, added section 520(o) to the FD&C Act, codified at 21 U.S.C. §360j(o). It removes five kinds of software function from the device definition: (A) administrative support of a health care facility, (B) maintaining or encouraging a healthy lifestyle unrelated to a disease or condition, (C) electronic patient records that meet the statute's certification conditions and are not intended to interpret or analyze those records, (D) transferring, storing, converting or displaying laboratory or other device data without interpreting it, and (E) clinical decision support that meets four conditions.

Subparagraph (E) is where AI claims are won or lost. The function must not acquire, process or analyze a medical image, a signal from an in vitro diagnostic device, or a pattern or signal from a signal acquisition system. It must display, analyze or print medical information. It must support or provide recommendations to a health care professional. And it must let that professional independently review the basis for the recommendation, so that the professional does not rely primarily on it. Most products on the FDA list analyze images or signals, so the first condition alone keeps them inside the device definition.

FDA's reading of these conditions is set out in the revised Clinical Decision Support Software guidance issued on January 6, 2026, which superseded the September 28, 2022 version. The revised guidance treats software that gives a specific preventive, diagnostic or treatment output or directive as failing the recommendation criterion, while stating that FDA intends to exercise enforcement discretion where only one option is clinically appropriate and the function otherwise meets all of section 520(o)(1)(E).

The guidance also treats software intended to support time-critical decision-making, or to replace or direct the professional's judgment, as outside that criterion. Guidance is nonbinding, and enforcement discretion is a statement of agency intent, not an exclusion from the statute. A company that describes a single-output AI tool as "not a device" is describing a policy position that the agency can revise, and the diligence file should say so.

05Change Control After Authorization

The objection that a cleared AI model is frozen has had a statutory answer since 2022. Section 3308 of the Food and Drug Omnibus Reform Act of 2022 added FD&C Act section 515C, codified at 21 U.S.C. §360e-4, under which a change consistent with an approved or cleared predetermined change control plan requires neither a PMA supplement nor a new 510(k). Section 360(l) now repeats the point for premarket notification.

FDA's final guidance, Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions, was announced in the Federal Register on December 4, 2024 (89 FR 96259) and reissued on August 18, 2025 under the same docket, FDA-2022-D-2628. It states that a plan may be established through the PMA pathway under section 515C(a), the 510(k) pathway under section 515C(b), or the De Novo pathway under section 513(f)(2).

An authorized plan has three parts: a Description of Modifications setting the specifications for the planned changes, a Modification Protocol setting the verification and validation activities and acceptance criteria, and an Impact Assessment of benefits and risks with a mitigation plan. FDA treats a plan that is not followed, or cannot be followed, as a deviation, and a change outside the plan returns to the ordinary rule: a new submission when a modification affects intended use or could significantly affect safety or effectiveness.

A company that claims its model "learns" or "improves continuously" after clearance should be able to produce the authorized plan and show that each shipped version falls inside its Description of Modifications. Without an authorized plan, each significant retraining is a potential new submission, and the record has not accepted the claim.

06The Claim Test

The test that follows is how LeverVenture reads a company where AI is the accelerant across Biopharma & Therapeutics, Diagnostics & Precision Medicine, Devices & Robotics, Digital Health & Delivery, and Longevity & Neuro.

1. The claim names a record. "FDA-cleared," "FDA-approved" and "FDA-authorized" are three different statements. Each should resolve to a K-, P- or DEN-number in the agency's databases, and the verb should match the pathway: clearance for 510(k), approval for PMA, a grant for De Novo.
2. The intended use matches the pitch. The authorization covers the intended use in the decision document. A triage product cleared under 21 CFR 892.2080 does not mark the image and does not remove cases from the reading queue; a pitch that describes diagnosis is describing a different device.
3. The evidence meets section 860.7. Anecdotes and testimonials are the categories Part 860 excludes. Performance claims should trace to the studies summarized in the decision document or to studies of equal standing.
4. The update path is authorized. A claim of continuous improvement needs an authorized change control plan under section 515C, or a record of new submissions for significant changes.
5. The exclusion is mapped. A "not a device" position needs a function-by-function mapping to section 520(o)(1), with the image-and-signal condition addressed first and any reliance on enforcement discretion stated as such.

A company that passes all five has a record a buyer, a payer or an acquirer can rely on. Classification does not judge how good a model is; it records what someone other than the company has examined.

07Frequently asked questions

How does FDA device classification apply to AI software?

AI software that meets the device definition is classified like any other device under 21 U.S.C. §360c and 21 CFR Part 860: Class I if general controls suffice, Class II if special controls are needed, and Class III if premarket approval is required. A new function with no predicate starts in Class III by statute unless it obtains a De Novo grant or is found substantially equivalent to a legally marketed device.

What share of AI-enabled devices on the FDA list reached market through 510(k)?

In the FDA list update with content current as of December 5, 2025, 1,304 of 1,357 entries, or 96.1 percent, were 510(k) clearances. De Novo grants accounted for 37 entries and premarket approvals for 16. The FDA describes the list as not comprehensive, because it identifies devices mainly from AI-related terms in public summaries.

Can a cleared AI model be updated without a new FDA submission?

Yes, if the change falls within a predetermined change control plan authorized with the device. FD&C Act section 515C, codified at 21 U.S.C. §360e-4, removes the need for a PMA supplement or a new 510(k) for such changes, and FDA's final PCCP guidance of December 2024 sets out the plan's three parts. Changes outside the plan follow the ordinary modification rules.

Nothing in this piece is investment, legal, tax or accounting advice, and nothing in it is an offer to sell or a solicitation of an offer to buy any security.