1,451 AI Medical Devices, Two Real Winners: The Truth About AI in Medtech

Article 1 of 5 in the Pharma vs. Medtech Series

AI already spots dangerous colon growths in real time, no AI-driven robot performs surgery without a surgeon, and one sliver of medtech already posts pharma-level margins — a map of where AI has actually landed in medtech, and where it hasn't

Prepared by Richstorm.co

Key Takeaways

• The FDA has cleared 1,451 AI/ML-enabled medical devices since 1995, with roughly 40% of that total cleared in just the last two years — but clearance proves safety, not proven patient outcomes.

• A study of 173 CE-certified European radiology AI products found outcome-level evidence stuck near 24% even as basic peer-reviewed evidence doubled, with MASAI and IDx-DR standing out as rare exceptions with real outcome proof.

• Surgical robots remain teleoperated — a human controls every motion — which is why AI is progressing there faster than in general-purpose robotics; da Vinci alone logged over 3.1 million procedures in 2025, up from 2.6 million in 2024.

• Cardiovascular devices and surgical robotics are medtech's two outliers: Abbott's Medical Devices segment ran a 33.7% margin in 2025 against an 18.2% company average, and Intuitive Surgical's moat runs on 25 years of surgeon training — both moats built on switching costs, independent of any patent or AI feature.

• AI is now layering directly onto both pre-existing advantages, raising the open question of whether a similar moat can be built anywhere else in medtech.

A note on terms: "AI" throughout this piece follows the FDA's own usage — its public device list is officially named "AI/ML-Enabled Medical Devices," tracking software that learns patterns directly from data.

Four Categories, One Confusion

"AI in medtech" is not one story. It clusters into four distinct categories, each solved by different technology and at a different stage of maturity.

Table 1: The four places AI has entered medtech.

The pattern: the further a category sits from direct patient contact, the more mature its AI is. Imaging AI reads a static scan. Implantable AI has to make a real-time decision inside a living body. That difference in stakes is also a difference in how much evidence each category has been required to produce.

Accurate Is Not the Same as Proven

Regulatory clearance measures safety and basic effectiveness. It does not measure whether a device changes what happens to a patient. Only two AI-medtech products have cleared that higher bar with real outcome data.

Two different questions are at stake. "Accuracy-level" evidence asks whether the AI correctly reads a scan compared to a known right answer — testable against old images, no patient treated any differently. "Outcome-level" evidence asks whether real patients who got the AI-assisted care actually fared better than those who didn't — provable only through an actual clinical trial. A device can pass the first test easily and still have no proof it clears the second.

Table 2: What full clinical proof looks like versus what most products have.

Among 173 CE-certified radiology AI products studied in Europe — a different, smaller population than the FDA's full device list, but a useful proxy for the sector — the share with any peer-reviewed evidence rose from 36% in 2020 to 66% in 2023. Real progress. The share with higher-level, outcome-changing evidence stayed roughly flat near 24% the entire time.

MASAI and IDx-DR show what full proof looks like. Most of the category has not gotten there.

Why Surgery Is Where AI Moves Fastest

Surgical robots sidestep AI's two hardest physical-world problems by design: a surgeon directly controls every motion, so the system never has to independently interpret an unpredictable environment, and it never leaves a stationary, powered operating room, so battery and mobility limits do not apply. That is why AI has progressed here faster than in general-purpose robotics — and why the competitive landscape just cracked open.

Table 3: The surgical robotics competitive landscape, 2026.

All five AI features in plain terms:

• Case-data analytics (da Vinci) studies data from millions of past procedures to spot patterns — comparing what's happening in the current surgery against a huge library of past ones.

• Force feedback (da Vinci) lets the surgeon actually feel how much resistance they're pushing against through the hand controls, restoring a sense of touch that early robotic systems removed entirely.

• Touch Surgery AI video analytics (Hugo) watches video from the surgery and automatically recognizes which step of the procedure is happening, flagging important moments as they occur.

• Versius Team (Versius Plus) is a live dashboard that uses data analytics to track how the robot is being used across a hospital — case volume, efficiency — in real time.

• Machine-learning collision prediction (Ottava) watches the robot's arms in real time and warns before two arms, or an arm and the patient, are about to collide inside a tight surgical space, similar to a car's collision-warning system.

Even here, autonomy has firm limits. Fully autonomous surgery is not expected in 2026, 2027, or probably by 2030, except in the most tightly constrained procedures — the field's own researchers say so plainly. What AI is doing instead is narrow, supervised augmentation: AI-assisted suturing under a surgeon ready to take over at any moment, plus vision assistance and safety-layer collision prediction.

The Money Machine Inside Medtech: Cardiovascular

Surgical robotics shows where AI is moving fastest — but its moat, where one exists, isn't about speed. Hugo, Versius Plus, and now Ottava all caught up to da Vinci's technology within about a year and a half of each other, proof that fast technical progress alone doesn't protect anyone's economics. What actually protects Intuitive Surgical is the surgeon training built up over 25 years — a different kind of moat than AI's own progress. Cardiovascular runs the same pattern for a different reason: a moat that already existed before AI arrived, for AI to reinforce.

Abbott's device business makes 33.7% operating margins — nearly double the company's own 18.2% company-wide average. The reason has nothing to do with AI — yet, and the full three-year trend behind that number is covered in the next article in this series.

Table 4: What actually builds cardiovascular's moat, and where AI fits in.

Why AI moves the first two rows and not the third: AI features get built directly into a company's own implants and its own surgical workflow, so a patient or doctor who wants those AI benefits has to stay with that company to get them — a second reason to stay, layered on top of the original one. Manufacturing and trial barriers don't move, because they're physical and regulatory, not informational: AI doesn't make a titanium alloy easier to machine to tight tolerances, and it doesn't shorten a clinical trial. Those barriers sit outside what AI is actually good at.

Cardiovascular and surgical robotics are the two parts of the sector that already proved they could escape commodity dynamics, by mechanisms that predate AI entirely. Whether AI ends up reinforcing those moats or quietly eroding them is a separate question for each, and not an obvious one — a dedicated article in this series takes on both directly.

The Scorecard

For now, medtech-AI breaks down cleanly: two products with real outcome-level proof (MASAI, IDx-DR), a large majority of products with accuracy data but not outcome data, a surgical robotics field making genuine progress inside a narrow, supervised envelope, and one cardiovascular segment already running pharma-adjacent margins for reasons that predate AI by decades.

The scale of activity across all four categories is real. Full clinical proof remains rare. Cardiovascular and surgical robotics are the two places medtech has already shown it can build a real, durable moat — and the next article in this series asks the harder question for both: does AI make each moat stronger, or does it hand a future competitor the tool to break it?

Read the full breakdown here: The Medtech Moat AI Can't Easily Break

Article 2 of 5 in the Pharma vs. Medtech Series