Inside Kunshan’s 14-Person Office That Brokers AI on Factory Floors

At eight on the morning of 12 October, several members of Kunshan's scene cultivation and application coordination group picked up their notebooks and walked out of the office. "Let's go look at some factories," one of them said. Science and Technology Daily, which reported the scene, used it to make a point: this is a government body whose work happens on factory floors, not at desks.

The group is known to everyone as the "scene office" (场景办). It was set up in early 2026 and is described as the first county-level body of its kind in China to exist as a real staffed team rather than a coordinating memo. Its stated job runs in two directions — find applications for products, and find technology for scenes. It does not approve permits, issue documents or hand out subsidies.

Kunshan is a county-level city in Jiangsu with roughly 35,000 industrial enterprises and two clusters that matter far beyond its borders: electronics worth about 800 billion yuan and equipment manufacturing worth about 300 billion. A base that dense ought to be an obvious test bed for AI. It wasn't, because the two sides couldn't see each other: factories with inspection problems had no route to vendors, and vendors with working technology had nowhere to put it. The office exists to do the walking between them.

Two lists, one platform, fourteen people

The headcount is small and the reach is not. Each of the 14 staff is assigned to a district or town and linked to liaison officers in districts, departments and state-owned firms, forming a network of more than 150 "specialists plus liaisons."

The working method rests on two lists and one platform. Through "scene call" notices and site visits the office collected more than 400 application scenarios, had them screened by industry experts, company representatives and department staff, and published the survivors as a demand-side "scene opportunity list" and a supply-side "scene capability list." In May 2026 the city launched the "Kunshan Scene Window" online platform, where a firm types keywords and gets matched technology or needs back.

Projects then run a closed loop: solicitation and reserve, expert evaluation, publication and matching, implementation, tracking and supervision, review and replication. Each project sits in a ledger, and batches are promoted as they mature. As of the October 2026 reporting, 78 scenarios had been built and 32 were pending across 19 key fields; 207 had been published, 69 as opportunities and 138 as capabilities; close to 20 manufacturers had been matched with AI partners.

What the matches look like

The clearest case is Qingtao (Kunshan) Energy, which makes solid-state batteries and inspects the appearance of every cell before it leaves the line. Production director Jing Shaodong put the old arithmetic plainly: one inspection shift took 10 people, two shifts more than 20, and each person judged appearance by slightly different criteria, which produced a steady trickle of small customer complaints. The company wanted AI visual inspection and could not find a suitable supplier. The scene office went to its pool of local machine-vision firms, screened them, and brokered a point-to-point, custom-built system.

What came out of it is reported only qualitatively. Jing said AI learning would bring labour costs down substantially and improve consistency; the report says consistency rose markedly and complaints were basically resolved. No defect rate, no headcount figure, no payback period has been published. That is worth holding onto, because it is the difference between a promising pilot and a proven one.

The second case is a straight logistics swap. Luxshare Precision's Kunshan plant needed to move high-value goods inside the building; Suzhou Youying Drone Technology had the technology and no site to fly it. After the introduction, drones work above the plant. The report cites transport efficiency up more than 50% and overall cost down about 60% — figures attributed to the report and participants, not independently verified.

The third is a supplier looking for markets rather than a buyer looking for tools. Shi Hang Intelligent Technology's marine general-purpose working robot, described as the only full-degree-of-freedom model of its kind, is now used for ship cleaning, port dredging, marine photovoltaics, offshore wind, marine ranching and deep-sea survey. R&D centre director Gao Penghui said the office helps connect the company to users' actual pain points so a scenario can move "from 0 to 1, even from 1 to 10." A batch was being tested in a simulated marine environment and was due to ship to Shandong that month for port ship cleaning and marine ranching.

The office also builds scenes, not just matches them. When eVTOL maker Peak Flight Aviation (Kunshan) reported that a shortage of test-flight sites was slowing iteration, the city's transport bureau and the scene office worked across departments to put up a civil drone test base — 24 square kilometres of airspace on 64 mu of land — within a few months.

The risk problem is the real problem

The obstacles are stated openly, which is unusual. Officials and public bodies hold the resources but fear trial and error; "dare not open" and "do not know how to open" are described as common attitudes. AI carries high upfront and maintenance costs. And a new scenario has no precedent, so technology admission, safety assessment and performance certification lack clear rules — review is slow and rollout expensive.

Kunshan's response is procedural: turn an uncertain innovation into a controllable workflow. Expert review before anything is published. Ledger management to shift from "afraid to open" to "opening in order." Cross-department consultation with an explicit tolerance for error, so that a stuck step has an owner instead of a scapegoat. Whether that reassures the official who has to sign is not something a workflow chart can settle.

The policy context explains the timing. After the 2024 Fair Competition Review Regulations curtailed tax-break and subsidy-based investment attraction, the State Council General Office issued an opinion in November 2025 on accelerating scenario cultivation and opening, framing a path of "technology breakthrough — scenario validation — industrial application — system upgrade." Kunshan's 2026 business-environment action plan makes building 500 benchmark application scenarios a government task — a shift, as the report puts it, from giving policies and subsidies to giving scenarios and opportunities.

Jiangsu is pushing the same direction at scale. High-tech manufacturing grew 14.8% in the first half of 2026. In June the provincial industry and IT department issued an "AI plus manufacturing" plan offering diagnostic services across 33 scenarios in R&D, production management and product service, with 422 diagnostic providers in the resource pool and a goal of covering essentially all above-scale industrial firms by 2027. Since March, the provincial government has held "AI plus" matchmaking roadshows every two months; embodied-AI firm Jiangsu Dinghui's CTO said a May roadshow brought 200 robot orders within a month — a company statement, not an audited number.

Set against 500 benchmark scenarios, 78 built is a baseline, not a triumph. What Kunshan has actually demonstrated is narrower and more useful: that fourteen people who leave the office, two lists that get published, and a ledger that keeps every project moving can turn a county's information problem into a queue.