A bank teller that never sleeps. A delivery van with no driver. An office assistant that files itself. China's AI stopped performing for the cameras this autumn and started doing the jobs.
At the 2026 China International Fair for Trade in Services (服贸会) in mid-September, the most-quoted line from exhibitors was not "look what it can do" but "look what it already runs." That shift — from demonstration to deployment — is the single most useful signal for anyone tracking whether Chinese AI is landing or just loud.
The platform play: Tencent opens WorkBuddy
On 2 September, Tencent switched on the open platform for WorkBuddy, its desktop agent (智能体) tool. The launch mattered less for the product than for the shape of the ecosystem it revealed: Tencent said it brought in more than 100 ecosystem partners, unveiled 9 co-branded smart-hardware devices (including Rokid, Insta360, iFlytek and Anker), and opened the underlying agent capabilities to hardware makers, industry apps and developers at once.
That makes WorkBuddy the first Chinese agent platform to explicitly bridge three layers — hardware, applications and developers — in a single move. Rivals are moving the same direction: Kingsoft Office shipped two AI office agents in July, while Alibaba's "Qianwen Office" (千问办公) and ByteDance's "Doubao Work" (豆包工作) both landed in August. The race is no longer about who has the best large model (大模型). It is about who controls the layer where the model actually touches work.
Across the city: finance and logistics at scale
Deployment, not demos, is where the numbers get interesting — and where they stop being hypothetical.
- China Construction Bank put digital-human tellers "Long Zhiwei" and "Long Zhiyuan" onto smart teller machines in some branches. Staff at the fair said the avatars handle a 98% customer problem-resolution rate on those machines.
- ZTO Express is running more than 3,500 autonomous delivery vehicles across 260-plus cities, moving over 8.7 million parcels a day, according to figures the State Post Bureau cited in a June briefing and the company restated through the autumn. At one ZTO node reported by Economic Daily, unmanned vehicles cut transport cost by roughly 50% and about ¥0.1 (US$0.014) per parcel.
- The same logistics story is repeating at scale: JD Logistics and SF Express run their own autonomous-vehicle fleets, and the open road-rights that more cities granted in 2026 are what turned pilots into daily routes.
None of these are lab results. They are operating fleets and live branches.
In the clinic and the classroom
Healthcare and education remain the two scenarios where AI shows the steadiest grip, because the payoff is measured in time saved by professionals, not novelty.
At the fair, Guang'anmen Hospital (China Academy of Chinese Medical Sciences) demonstrated its "AI + doctor" system built on the Guangyi·Qizhi large model (大模型) 2.0, covering patient service, diagnosis, ward management, quality control, multidisciplinary consultation and teaching, with 7×24-hour TCM service across the whole loop from consultation to prescription pickup. Separately, at the Guangdong AI application conference on 22 September, iFlytek told attendees that at Guangya Middle School its terminal lifts homework-grading load and pushed subject-mastery rates from roughly 50% to 90%, while its "Zhiyi Assistant" (智医助理) completed more than 40 million auxiliary diagnoses in just over a year. Those are vendor-reported metrics, not audited outcomes — but they show where companies are pointing their roadmaps.
What the government is betting
Shenzhen published its 2026–2028 AI action plan in early September with concrete targets: by 2028, the city's AI core industry scale will exceed RMB 300 billion (about US$41 billion / HK$320 billion), agent (智能体) and smart-terminal penetration will pass 90%, and the city will host more than 3,000 AI firms. The plan also names humanoid (人形机器人) terminals, unmanned logistics vehicles and the "Shenzhen–Hong Kong R&D + Shenzhen conversion" model as priorities, and sets aside industrial-fund and compute support to pull prototypes into production.
A local government putting a three-year, multi-billion-dollar number on AI adoption is itself a deployment signal: the buyers are now also the planners. When the stated goal is 90% penetration of smart terminals and agents, the question stops being "will AI arrive" and becomes "who supplies it."
Honest limitations
- The ZTO vehicle and parcel figures are company- and State Post Bureau–stated, not independently audited, and were anchored to a May–June 2026 snapshot; daily volumes swing with seasonality and holidays.
- WorkBuddy's 100+ partners and 9 devices are launch-day claims. Ecosystem traction is unproven until renewal and usage data appear, and "agent (智能体) as operating system" is still a thesis without a public per-user productivity study.
- CCB's 98% resolution rate applies to specific smart-teller machines at some branches, not the whole bank, and is a vendor-stated metric.
- The education and health numbers (50%→90%, 40M+ diagnoses) are presented by the vendors at conferences; none are peer-reviewed or independently verified, and "mastery rate" definitions vary by school.
- Government targets like Shenzhen's RMB 300 billion (US$41 billion) are aspirations written into plans, not contracted revenue; missed deadlines are common in this sector.
What readers can do now
- Treat "AI落地 (deployment)" headlines as signals, not proof. Before acting on any number, pull the original source — a government plan, a company release, or the fair's own materials — and check the date it was measured.
- If you run operations in China, pilot one narrow agent (智能体) workflow — invoicing, dispatch, first-line support — and measure cost and hours saved over a quarter, not a demo afternoon.
- Use Shenzhen's 2028 targets as a barometer. When the city reports progress against the RMB 300 billion (US$41 billion) goal and the 90% penetration mark, you will see whether policy ambition is turning into purchased capacity or just more slides.