The FDA's AI-enabled device list quietly passed 1,000 — what that actually means for clinicians

The U.S. Food and Drug Administration’s running list of “AI/ML-enabled medical devices” crossed a quietly significant threshold this spring: more than 1,000 cleared products. For an authorization category that effectively did not exist a decade ago, the curve has been steep — and most of it has happened in the last five years.

It is tempting to read this as the moment AI became routine in clinical practice. The number deserves a more careful reading.

What the list actually counts

The FDA’s tracker is a public catalog of devices that have a marketing authorization — typically a 510(k) clearance, occasionally a De Novo or PMA — and that the agency has identified as using artificial intelligence or machine learning in some part of their function. A few caveats worth holding onto:

It counts authorizations, not deployments. A device with clearance is allowed onto the U.S. market. Whether any health system has actually procured, integrated, and used it is a separate question the list does not answer.
Most entries are radiology. Imaging — particularly mammography, CT, and chest X-ray triage — has dominated the list for years, and continues to. Pathology, cardiology, and ophthalmology are growing but remain a minority slice.
“AI” is a broad tent. The list mixes deep-learning models, classical statistical tools that re-branded under the AI umbrella, and rule-based decision support that uses ML at one narrow step. The agency does not currently grade authorizations by model complexity or autonomy.

The 1,000 milestone is real, but it is closer to “1,000 cleared features” than “1,000 distinct AI products clinicians touch every day.”

The shape of what’s been cleared

Reading down the list, three patterns repeat:

Workflow triage and prioritization. A large fraction of cleared products help radiologists, ED physicians, or stroke teams find work in their queue: flag the suspected large-vessel occlusion, surface the likely pulmonary embolism, mark the chest film with a pneumothorax. These tools are typically deployed as a layer on top of an existing PACS or worklist rather than as primary diagnostic claims.
Quantification. Tools that measure something a human already measures — ejection fraction, hippocampal volume, breast density — but faster, more consistently, or with fewer manual steps. The clinical claim is usually time saved or variability reduced, not a new diagnostic capability.
Image enhancement. Denoising, motion correction, low-dose reconstruction. These products often replace a step in the scanner pipeline; the radiologist may not even know an AI model is involved.

A smaller but growing slice covers things like AI-augmented cardiac monitoring patches, retinal-image screening for diabetic retinopathy, and dermatology triage. The closer the model gets to making a diagnosis rather than flagging one, the more carefully the authorization is scoped — and the more its label restricts use to a defined population, modality, and clinical context.

What the headline number doesn’t tell you

A clinician evaluating a vendor’s pitch should still ask the same four questions the 1,000-device milestone does nothing to answer:

Was the model evaluated on people who look like my patients? Most pivotal datasets remain heavily skewed toward U.S. and Western European populations and toward the specific scanner manufacturers used during the submission.
What’s the failure mode? A triage model that misses 5% of cases is doing something very different from one that flags 5% as urgent when they aren’t. “Sensitivity” and “specificity” in the FDA summary are starting points, not the whole story.
How is the model updated? The agency’s Predetermined Change Control Plan pathway is now live, which lets manufacturers commit, at submission time, to a defined set of post-market changes they intend to make. Not every cleared product uses it. Without a PCCP, meaningful model updates require a new submission.
Does it integrate? A high-performing model that requires a parallel viewer or a custom workstation usually does not survive contact with a working radiology department for long.

None of these questions are exotic. They are the same questions a sensible procurement committee has been asking for years. The 1,000-device milestone makes them more, not less, important.

What changes from here

Two things are worth watching over the next twelve months:

The post-market surveillance signal. As more cleared devices reach meaningful volume, the FDA’s MAUDE reports — and quieter institutional incident registries — will start to give a clearer picture of how these tools perform outside the controlled environment of a submission study. Expect a few high-profile retractions or label updates; they are healthy.
The shift from imaging into the rest of the hospital. EHR-embedded sepsis prediction, OR scheduling optimization, and ambient documentation tools have all started accumulating authorizations, but they have not yet hit the kind of density radiology has. When they do, the procurement and governance questions move out of the radiology department and into the office of every CMIO.

The 1,000-device milestone is best read as a marker that the regulatory machinery now handles AI submissions as routine, not as a sign that the clinical machinery has caught up. The interesting work — the integration, the post-market evidence, the equity audits — is still mostly ahead.

This is the v1 launch article on AI in Healthcare. We publish independent, source-linked reporting and analysis for clinicians, regulators, and the people building these systems. Pitch us at editorial@ai-in-healthcare.example (placeholder address until the production domain lands at AIH-7).