--- title: "AI Is Already in Chinese Hospitals. The Bill Is the Experiment" date: 2026-10-19 category: China AI Decoded site: NeuroAI canonical: https://neuroai.site/a/na-decode-ai-hospitals-reimbursement language: en --- # AI Is Already in Chinese Hospitals. The Bill Is the Experiment > China has approved more than 110 clinical AI products, but payment is the missing piece and hospitals are absorbing the cost of finding out. In the small hours the reading room is lit like an aquarium. A resident works through a stack of chest scans, moving between a grey image and a report template, while the queue on the wall monitor stays stubbornly long. The software approved to help her is not the software she is using. It sits in a different queue — one that runs through procurement, price lists and budget, not through the imaging console. That gap between a certificate and a purchase order is the part of Chinese medical technology that rarely makes the trade press. The approvals are real and increasingly routine. Who pays for them is the experiment. ## Key takeaways - NMPA had approved more than 110 deep-learning Class III standalone software products by October 2025, across nine data types. - 2026 additions include cervical pathology, gastric endoscopy, an orthopaedic surgical robot and a first certificate built on a large model (大模型). - United Imaging Intelligence's chest agent claims 97.8% accuracy, about 100 abnormalities, 30 hospitals and >2.5 million cases. - Approval does not create a payment code; most adoption runs through hospital budgets and vendor-funded pilots. - The brain implant approved in March 2026 won a payment pathway almost at once — the exception, not the pattern. ## The approval pipeline works The regulator's software list has become a production line. By October 2025 the National Medical Products Administration had cleared more than 110 deep-learning Class III standalone products across nine data types (NMPA). Read that as a regulatory achievement rather than a market: each certificate is a safety and efficacy judgement about a defined use, not a purchase. What changed in 2026 is breadth. The additions run from cervical pathology and gastric endoscopy to an orthopaedic surgical robot, ROPA6, cleared in March, and nuclear-medicine triage, which gave United Imaging Intelligence its twentieth certificate (NMPA). In May came AI AutoVision's chromosome karyotyping product, described as the first Class III certificate built on a large model rather than a task-specific network. Breadth matters because it tells you where the regulator is comfortable. Imaging was the easy case: a still picture, a radiologist's ground truth, a clear workflow. Pathology, endoscopy, surgery and genetics are harder — moving tissue, live video, physical risk, rarer events. Each approval in those areas implies a body of clinical evidence that did not exist a few years ago. ## Four gates between a certificate and a patient A cleared product still has to travel. The gates are sequential, and each one can stop it: - **Approval**, which the regulator supplies and which is now the least scarce of the four. - **Procurement**, where a hospital decides whether the software is worth a line in its capital budget. - **Payment**, where a price is set and someone other than the vendor agrees to bear it. - **Adoption**, where a clinician changes a habit under time pressure and accepts liability for an output she did not produce. The first gate produces a public record. The other three do not. That asymmetry is why the approval count is quoted so often and the installed base so rarely. ## The agent wave arrives through the side door The most ambitious product in the current wave is not a single-disease detector. United Imaging Intelligence's chest agent reads one scan for roughly 100 abnormalities at a claimed 97.8% accuracy, is deployed in about 30 hospitals and has processed more than 2.5 million cases (company disclosure). It has entered NMPA's innovation review channel as the first multi-disease product built on a large model. That framing is revealing. The innovation channel is a faster, more consultative route for products the regulator considers novel; entering it is not approval. But the design points at the payment problem rather than around it. A tool that reads one scan for a hundred conditions does not map cleanly onto a payment system built around billing for individual findings. If one scan substitutes for several billable reads, the hospital gains throughput and loses revenue lines at the same time. That is the quiet argument inside the agent wave: the technology is arriving faster than the accounting. ## What would unlock scale Three things, in order of difficulty. Payment codes that pay for an outcome rather than a procedure. Evidence that a patient did better, not that a model matched a radiologist on a test set. And a liability arrangement a department head can defend after a missed finding. The brain-computer interface (脑机接口) case shows how much difference the first item makes. The implant approved in March 2026 treats a small, expensive, visibly disabled population, and it received a payment pathway almost immediately — perhaps because a surgically implanted device fits an existing category of hospital payment in a way that a software subscription does not. Software has no such anchor. Scale, if it comes, will probably arrive through hospital group purchasing rather than consumer demand. A single large hospital can absorb a pilot; a group with a standardised workflow can amortise the integration cost. That is an operational story, not a technology one. ## Honest limitations The approval count is cumulative and covers standalone Class III software only, so it says nothing about daily use. The 97.8% accuracy, 30 hospitals and 2.5 million cases are company disclosures, and entry into the innovation review channel is not an approval. No public data shows what hospitals actually pay, how many pilots convert to paid contracts, or how many approved products are commercially dormant. The payment gap described here is inferred from the structure of the system and from the contrast with the implant, not from a published survey of hospital budgets. ## What readers can do now - **Operators:** ask vendors for paid installations and named departments, not certificate numbers, and put liability and integration cost in the contract before the pilot. - **Investors:** track reimbursement events as the real milestone; an approval is a permission, not revenue. - **Researchers:** study the gap between the approval register and procurement records, which is where the sector's actual adoption curve is hidden. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-decode-ai-hospitals-reimbursement Free to quote with attribution and a link to the original.