Phase 1: The money · Article 2 of 3
From America to China, insurers pay for AI's answers, not its help
Insurers pay per use mainly when AI produces a result of its own, such as a risk score, and then they pay for the result, not the AI. Help with work they already fund rarely earns anything extra: China's price rules write that split down, and America's codes and Medicare's rates fit it.
The question: What makes insurers pay for AI?
By Chi-way Li · Oct 20th 2026
Article 1 found that only 5.0% of FDA-authorized AI devices are paid per use; three in four are bought from buyers' own budgets. This article asks what earns per-use payment.
Two questions
Two questions sort the AI that insurers pay for per use from the rest:
- Does it produce a result of its own, such as a measurement or a risk score?
- Is its work already part of what an existing fee pays a clinician to do?
Payment follows a yes, then a no. A lung-nodule cancer-risk score passes: the CT's fee does not cover the score, which had 6,832 Medicare hospital claims in nine months of 2025. Colonoscopy AI fails both and earns nothing extra in America or China.
What the cases show
Our 42 payment cases agree. Of the 17 paid per use, 9 pay for a result, 5 for a treatment such as a therapy app, and only 3 for help, each an exception below. In 17 of the 20 cases where AI helps with existing work, a hospital's or a government's budget pays.
It is a pattern from cases we chose, 36 of them American or Chinese, not a law. And a result is necessary, not enough: 437 of 476 FDA-listed imaging AI devices that measure or quantify are still bought from budgets.
Per-use payment mostly goes to results and treatment
How each type of AI is paid for, by country: 42 payment cases in America, China, Germany and Britain
Own code: an insurer pays per use through the AI's own price item or code. Patients or others: per-use payment by patients or third parties. Budget: a hospital or government buys it, including central purchase. In device price: the AI comes inside the price of the device.
Source: Cleared to Paid payment cases from pricing rules, payment data and filingsHealthcare AI Insights
What the claims show
Medicare's billing data agree. Every AI code billed more than 1,000 times by doctors in 2024, or paid separately that often in hospitals in the first nine months of 2025, pays for a result, led by FFR-CT, which estimates from a CT whether a narrowed heart artery limits blood flow. Help codes stay small: chest X-ray computer-aided detection was billed 722 times in 2024.
Two codes take most outpatient AI claims
US hospital outpatient claims for AI billing codes, January to September 2025, by product type
24 AI billing codes had claims, grouped into 15 product types.
Show the numbers
| Product type | Codes | Claims |
|---|---|---|
| FFR-CT (coronary flow from CT) | 75580 | 19,352 |
| Lung nodule CT | 0722T, 0721T | 6,832 |
| Brain MRI volumes | 0866T, 0865T | 2,669 |
| Liver and organ MRI | 0649T, 0648T, 0698T, 0724T, 0697T | 1,950 |
| Coronary plaque from CT | 0625T | 1,191 |
| Autonomous diabetic eye screening | 92229 | 1,142 |
| 3D heart model for cardiac procedure planning | C9793 | 430 |
| Echo AI for heart failure | 0932T | 428 |
| CT bone strength test | 0556T, 0555T | 341 |
| 3D anatomy model for surgical planning | C8001 | 246 |
| ECG simulation to locate arrhythmia | 0897T | 101 |
| ECG AI | 0764T, 0765T | 60 |
| Liver ablation planning | 0944T | 47 |
| Lung disease subtype from CT | 0877T, 0878T | 20 |
| Heart chamber volumes from chest CT | G0183 | 1 |
Source: CMS preliminary outpatient rate-setting file (2025 claims)Healthcare AI Insights
Lab tests follow the same rule. Medicare's lab fee schedule prices eight FDA-listed AI devices at $200 to $4,500 per test in 2026, and Medicare covers at least five machine-learning tests outside FDA's AI list, such as Veracyte's Decipher ($3,873) and Guardant's Shield ($1,495).
What the rules say
The American Medical Association's coding system sorts AI work into three kinds:
- Assistive: pointing out data for a clinician.
- Augmentative: turning data into a new measure or score.
- Autonomous: reaching a conclusion itself.
By our reading, every heavily used code is augmentative or autonomous. Medicare's 2027 hospital-rate proposal agrees: 50 software codes, FFR-CT and the lung-nodule score among them, would get a new payment status; chest X-ray detection stays folded into the exam.
China's national price guides, since late 2024, price a service by what it produces. AI-assisted radiology, pathology, endoscopy and lab tests earn no separate fee, while FFR during an angiogram has one price whether it comes from a pressure wire or from software. National insurance does not cover screening, so patients pay for Burning Rock's OverC, the one Chinese AI lab test we found.
The exceptions
Most exceptions ended or are paid by patients.
- Medicare's new-technology add-on paid hospitals up to $1,040 per stay for stroke-triage software, for two years only.
- RadNet, an American imaging chain, charges patients $40 for an AI mammogram read.
- Medicare's extra payment for older, pre-deep-learning computer-aided detection in mammograms ended in 2017, when new screening codes folded it into the exam's price; a study of 625,625 exams had found no accuracy gain.
- Chinese provinces are folding separate fees for computer-aided cervical cytology, up to 200 yuan a case, into the diagnosis fee.
Japan is the lasting exception we found: since 2024 it adds about ¥600 to a polyp removal done with detection AI.
Who keeps the gain
Why do the two questions work? Our analysis, not a finding: payment lasts when whoever pays also keeps the gain. Insurers gain from results that replace a costlier test, hospitals from help that saves staff time, scanner owners from faster scans, and governments and other long-term budget holders from screening. A rule, such as a coverage law, a quality score or a program budget, often decides who pays first. The next article follows the money.
Where payment would last
Our analysis, not a finding: outlined cells mark who would pay for each type of AI if payment followed the gain; chips are today’s routes
Own code: an insurer pays per use through the AI's own price item or code. Patients or others: per-use payment by patients or third parties. Budget: a hospital or government buys it, including central purchase. In device price: the AI comes inside the price of the device.
Source: Cleared to Paid payment cases from pricing rules, payment data and filingsHealthcare AI Insights
So what?
- Product managers: for help with paid work, expect no lasting billing code. Sell to whoever saves the time, and measure the saving after the clinician's check.
- Startups: for a result nobody could get before, design the first study for coverage (does the result change care?), then a code, and expect years on a temporary one (FFR-CT took six).
- Outside America: in China, plan on hospital purchase or provincial tenders; in Germany, insurers pay for apps patients use (DiGA), not clinicians' reading tools.
Which AI product do you know that insurers pay for per use even though it only helps a clinician, and will that payment last?
Join the discussion on LinkedIn: the post
Explore in the full report
Sources
Cleared to Paid payment cases (42: America 20, China 16, Britain 4, Germany 2; we found few result-type cases in Germany or Britain; the 5 treatment cases pay for the product, rarely its AI); CMS Medicare physician data 2024 (traditional Medicare only, none from commercial insurers; two retinal-photo codes that people read are left out); CMS CY2027 outpatient rate-setting file (preliminary claims, Jan-Sep 2025: 18 of the 24 AI codes with hospital claims were used fewer than 500 times) and proposed Addendum B (payment status; the new status is for software services); CMS CY2026 outpatient rates (Addendum B status indicators); IPPS final rules FY2021-FY2027 (new-technology add-on payments, which end by design after two or three years); CY2017 physician fee schedule final rule (81 FR 80170; the CAD payment folded into the exam's price); American Medical Association, CPT Appendix S AI taxonomy (ama-assn.org, described in our own words; CPT is the coding system American insurers use); NHSA radiology pricing guide and its explanation of 2024-11-20 (AI cannot yet replace a doctor's diagnosis or deliver a service of its own), pathology guide (2025-12-19) and digestive-system guide (医保价采函〔2025〕132号, usage note 18), as Shanghai implements it (沪医保价采发〔2025〕52号); NHSA cardiology guide (2025) and Shanghai FFR price notice (沪医保价采发〔2025〕31号; one price item for FFR within an angiogram, from a pressure wire or software; FFR from a CT has no national price item); Rao et al., JACR 2010 (CAD in Medicare mammography, 2004-2008: on 74% of Medicare screening exams by 2008); Lehman et al., JAMA Internal Medicine 2015 (625,625 screening mammograms); RadNet SEC filing, Form 425, April 2025 (reprinting STAT on its $40 mammography AI add-on) and Radiology Business on patient-paid AI mammogram reads; Japan MHLW 2024 fee schedule, K721 lesion-detection support add-on (60 points), as quoted by Fujifilm (2024-06-03); the 'Who keeps the gain' section is our analysis, not a finding (surgical robots get no box because no payer prices a robot's AI, so the theory makes no prediction for them); 'So what?' is advice; CMS Clinical Laboratory Fee Schedule CY2026 Q4 national limits; Medicare coverage NCDs 90.2 and 210.3 and MolDX LCDs (Decipher, Afirma, Envisia, DecisionDx-Melanoma); machine-learning statements (we count a test only when its maker, the FDA or Medicare says a trained or machine-learning model produces the result) in Veracyte's (Decipher, Afirma, Envisia) and Caris's (GPSai, an AI add-on inside its paid tests that earns nothing extra: the second question again) 2025 10-Ks, FDA summary P230009 (Guardant's Shield), Castle's validation study and Medicare's Envisia LCD; NHSA lab-test pricing guide (检验类医疗服务价格项目立项指南（试行）, 2026-08-14), AI-assisted diagnosis in the price of 659 of 662 items; LabMedica, 2026-09-02 (Epignostix brain-tumor classifier and the NHS Genomic Medicine Service); Chongqing 渝医保发〔2026〕1号 (computer-aided cytology fee charged at the main item's price); Japan fee schedule item D296-3 (nodoca, an AI flu test that reads throat images, paid per test at 305 points, about ¥3,050; covered since 2022-12); 知见生命 (2026-01-13) on Chongqing's category-B insurance cytology item; Dian Diagnostics investor-relations record (2026-01, via 体外诊断观察) on provincial computer-aided cytology fees (tens of yuan to 200 yuan a case; in Chongqing through insurance; folded in as provinces adopt the national pathology guide); Shanghai and Qinghai adoption of the pathology guide (June 2026); Burning Rock Biotech Form 20-F for FY2025 (OverC: machine learning, sold through a few pilot hospitals, the company's lab tests paid out of pocket); 早筛网 2026 H1 screening review (NHSA reply on screening and insurance); CMS CLIA FAQ on lab-developed tests (CLIA does not address clinical validity; most of the six covered machine-learning tests are lab-developed); ACLA v. FDA, E.D. Tex., 2025-03-31 (FDA's 2024 lab-developed-test rule vacated); FDA final rule 90 FR 45134 restoring the prior regulation (effective 2025-09-19); Hinge Health Form 10-K for 2025 (employer-paid programs); GKV-Spitzenverband DiGA report (2026-04); NCD 210.3 (blood-based colorectal screening, 2021) and the multi-cancer early detection law (2026-02-03); UnitedHealthcare Medicare Advantage in-home vendor programs (diabetic eye exam gap); Singapore SELENA+ (public-sector AI eye grading since 2021)