The most useful way to judge whether AI is real in an economy is to stop asking what the models can do and start asking what changed on a Tuesday afternoon in a factory.
Chinese state broadcaster reporting through 2026 gives an unusual amount of operational detail, because the cases are being used as evidence for the "AI+" adoption programme. Here is what the numbers actually look like.
Key takeaways
- Manufacturing: a "navigator-level" smart factory in Xuzhou replans 30 days of production autonomously and reconfigures lines in about 10 minutes, down from five to six hours.
- Quality inspection: an optical fibre workshop in Wuhan runs at over 3,000 metres per minute drawing fibre 250 microns thick, with an AI inspector catching 0.1 mm defects at 200 metres per minute — 99.99% detection, roughly 20× the limit of the human eye.
- Automotive: FAW's integration of Alibaba's Qwen model across R&D, production and service is reported to have cut R&D cycles by 35%, production cost by 30% and defect rate by 90%.
- Appliances: Midea's humanoid "Meiluo 2" on a water-heater line: cycle time −20%, production loss −50%, daily capacity +33%.
- Healthcare: AI-assisted pathology raised a doctor's daily slide review from about 100 slides to 50,000.
- Rare disease: Xinhua Hospital's DeepRare system reaches 57.18% first-diagnosis accuracy and cut the average confirmation time for rare diseases at county hospitals from five years to about three weeks.
The factory cases
The Xuzhou crane plant is the clearest illustration of what "autonomous production" means in practice. When international orders for nine different crane models arrived simultaneously, the system did not wait for planners. It activated, scheduled thirty days of production across lines, and reconfigured those lines in roughly ten minutes.
The old changeover took five to six hours. That is a direct capacity gain, not a productivity metaphor.
At Midea's water-heater facility, the interesting detail is not that a humanoid robot is present — it is that it is doing quality inspection, the job factories have historically found hardest to automate because defects are varied and rare.
In energy, a palm-sized industrial microphone at a Ningxia wind farm captures high-frequency signals from early cracks and abnormal friction that human ears cannot hear. Combined with acoustic large models and inspection robots, operator estimates put the saving at about 3,000 inspection hours a year and roughly 60% lower labour cost per unattended station.
The hospital cases
Healthcare is where the numbers stop being about efficiency and start being about access.
AI-assisted pathology raising daily slide throughput from 100 to 50,000 is not a marginal improvement. It changes what screening programmes are economically possible — mass early cancer screening only works if reading a slide costs nearly nothing.
The DeepRare result is more striking still. Rare diseases affect hundreds of millions of people worldwide, and the defining experience of a rare-disease patient is the diagnostic odyssey — years of misdiagnosis across multiple hospitals. Cutting the average county-hospital confirmation time from five years to three weeks is a change in what the health system can promise.
United Imaging's uAI MedTuring agent takes a different approach: one CT scan, 73 common conditions detected, structured report generated, diagnostic efficiency up by more than 30% — with the explicit goal of reducing missed findings rather than replacing radiologists.
China's "15th Five-Year Plan" outline explicitly links this to building a "Healthy China" through digital and intelligent means, and at the 79th World Health Assembly, international peers converged on a similar framing: health governance in the AI era shifting from "how do we treat illness" to "how do we keep people healthy."
The limits that keep this honest
Two caveats belong alongside every one of these figures.
First, these are showcase cases reported by state media, selected precisely because they worked. The national average is certainly less impressive than the best plant in Xuzhou.
Second, official Chinese analysis is unusually candid about the gap between pilots and scale: AI in manufacturing is described as moving from local demonstration to scaled deployment, with uneven penetration across sectors, and high-end chips, core algorithms and basic software still described as chokepoints.
Both caveats are consistent with what we know: the technology works in favourable conditions, and the hard part — as always — is everything around it.
Case figures as reported by China Central Television, Xinhua and the Cyberspace Administration of China in 2026.