Cleared, but rarely reimbursed

The Food and Drug Administration has cleared 1,614 medical devices that use artificial intelligence, and it clears more each year. Only 78 of them, fewer than one in 20, have ever had a way to be paid for each use: a billing code, a Medicare laboratory code or an inpatient add-on payment. Today 73 still do.

Few are paid per use

FDA AI-enabled devices by payment route, one square per device

• Per-use route today (73)
• Route has ended (5)
• Never had one (1,536)

A device counts only when a CMS rule, a Medicare contractor or its maker names it for a code, so 78 is a floor.

Show the numbers

Payment route	Devices	Share
Per-use route in force	73	4.5%
Per-use route ended	5	0.3%
Never had a per-use route	1,536	95.2%
All devices	1,614	100%

Sources: FDA; CMS; Medicare contractors; makers' coding guidesHealthcare AI Insights

Where insurers do pay, the money is concentrated. Medicare's payments to doctors for AI codes reached $10.3m in 2024. Analysis of coronary CT scans took 96% of it.

Everything else is paid for some other way, mostly by hospitals and imaging groups out of their own budgets. Even AI built into scanners is usually sold separately: where makers' documents say how they sell such features, 154 of 195 are options, licences or upgrades rather than part of the machine's price.

Who foots the bill

FDA AI-enabled devices by who pays, % of total

• 4.8%Paid per use: a billing code, a Medicare lab code or an inpatient add-on
• 4%No extra payment: rides on a service insurers already pay for
• 15%Inside a hardware price: scanner features, medtech systems, consumer devices
• 76%From a buyer's budget: standalone software, scanner options and upgrades, per-case services

The three unpaid groups rest partly on inference from who filed each device. The hardware share is an upper bound: scanner features whose packaging could not be checked stay counted as hardware.

Sources: FDA; CMS; makers' documentsHealthcare AI Insights

The pattern holds abroad. Across 36 payment arrangements in America, China, Germany and Britain, per-use payment appears only for AI that produces a result of its own, such as a measurement or a risk score, and for AI that treats. Insurers, in short, pay for answers, not for help.

Why? One hypothesis is a verification tax. AI that promises efficiency is hard to price while a clinician still checks every case, because the time it saves is hard to see. That may explain why AI scribes have spread fastest: their mistakes are cheap to fix. A later article in the series tests the idea.