Factories Aren't Run by Chatbots — The Four-Layer Sequence Behind China's Manufacturing AX

Factories aren't run by chatbots. CEO Park Ji-min's column on China's manufacturing AI transformation (AX) defines manufacturing AX not as a project to install a single large language model, but as work that ties together equipment, processes, industrial software, and on-site knowledge into a single operational flow. Everything must be connected — the state of equipment, what sensor readings mean, which process a defect originated from, and whether the AI's decisions stay within safety rules.

The column divides shop-floor transformation into four non-overlapping layers without redundancy. The first is the physical layer — sensors, robots, controllers. The second is the data layer, which aligns equipment, process, quality, and energy data on a common timeline and semantic framework. The third is the operational software layer, connecting MES, ERP, PLM, warehousing, and maintenance systems. The fourth is the governance layer, which defines safety limits, human approval, emergency stops, and accountability. AI predicts, plans, and coordinates execution on top of these four layers — and the column's diagnosis is that if even one layer is missing, a pilot may succeed but scaling across multiple plants becomes difficult.

China's manufacturing reality is likewise not a single, uniform picture. While leading companies recognized by the World Economic Forum as lighthouse factories are expanding adaptive production and AI-based quality control, a considerable number of small and medium-sized manufacturers remain at the stage of connecting ERP, MES, and PLM systems. Leading factories are scaling AX on top of existing DX foundations, while the industry average is undergoing a compressed transformation that pursues automation, digitalization, and intelligence simultaneously. Policy goals, too, are designed as bundles rather than single model deployments. China has set a target of deeply applying three to five general-purpose models in manufacturing by 2027, along with building 1,000 high-level industrial agents, 100 high-quality industrial datasets, and 500 representative application scenarios — while simultaneously emphasizing manufacturing computing infrastructure, a cloud-edge-device model architecture, and collaboration between large and small models.

The corporate cases cited in the column all follow the same sequence. Sany Heavy Industry first standardized and digitized its R&D, procurement, production, and service processes, then built an industrial internet platform and a data middle layer, replacing welding paths and conditions once adjusted by skilled workers with laser scanning and process models whenever products changed. In other words, before AI could learn, the company first converted human tacit knowledge into a form machines could read. CNPC (China National Petroleum Corporation), after long-term development of distributed control systems, equipment monitoring, geological and drilling data, and an integrated production platform, unveiled its Kunlun industrial large model with 300 billion parameters in 2025. Haier Group extended its own factory DX experience into the COSMOPlat industrial internet platform, and by 2026 is positioning a collaborative structure in which multiple agents divide up production, energy, quality, and supply chain tasks.

However, a platform does not automatically fix aging equipment or incomplete data. Since processes and safety rules differ by industry, integration costs arise each time a standard solution is repeatedly applied on-site, and the responsibility for equipment connectivity, data ownership, and process redesign remains with the client company. For small and medium-sized manufacturers, the burden of connecting outdated equipment, cleaning up data, and on-site engineering costs may outweigh model usage fees. This is why the column argues that the likely winner of AX will not be the company with the largest model, but the one that can replicate the same solution across multiple factories at the lowest cost.

Implications for Korean Companies' AX

For companies considering adoption, the sensible order is to first identify which of the four layers is missing before selecting a model. The maturity scale presented in the column can be used directly for self-diagnosis. Visual inspection and equipment failure prediction correspond to Stage 1, assistive AI. When AI integrates MES, ERP, warehousing, and quality systems to propose production plans and work instructions, that is Stage 2, task integration. When multiple agents coordinate orders, materials, production, and quality — escalating only exceptions to humans — that is Stage 3. When a system detects process conditions, adjusts plans or control values accordingly, and relearns from the results, that approaches Stage 4, limited autonomous operation. The column also notes that if data remains scattered across PLCs, MES, ERP, spreadsheets, paper records, and the experience of skilled workers, a project will stall at the detection-and-assistance stage.

Performance metrics also need to be reset. The column's standard is that industrial AI performance should be measured not by app user counts but by yield, downtime, energy intensity per unit, frequency of plan changes, and safety incidents. Since efficiency improvement figures disclosed by vendors and companies are often self-reported aggregates from a specific line and period, the real criteria for commercialization should be whether the results replicate across multiple lines and factories, and whether the investment pays back once maintenance and data-cleanup costs are included. This is precisely why, even while discussing the application case at Baoshan Iron & Steel/Baosteel under China Baowu Steel Group, the column adds the caveat that since the figures and effects are company-disclosed, independent verification is needed.

When safety-critical processes are involved, the questions themselves change. It is rare for a general-purpose large model to directly operate valves and robots; a more realistic structure has the large model plan the work while verified industrial software, smaller models, and control systems handle actual execution. The column proposes that the maturity standard for safety-critical industries should be based not on "how much the AI decides" but on "under what conditions it stops and hands control back to humans." When planning to raise the level of automation, it is safer to also document the stop conditions and handover rules together.

What to check this week can be narrowed down to one thing. Pick the single defect that caused the largest loss last month, and check whether you can trace back, using only the records that remain, exactly which process it originated from. If you cannot trace it back, what is lacking right now is not the model — it is the data layer.

Source: Factories Aren't Run by Chatbots... The Real Sequence Behind China's Manufacturing AX [Park Ji...