I. A Clear Chain of Policy Progression

Looking back over the past year, the policy push behind the "AI Plus Manufacturing" initiative in China has followed a tightly paced chain of progression.

In August 2025, the State Council issued the Opinions on Deepening the Implementation of the "AI Plus" Initiative, writing into its targets that the adoption rate of next-generation intelligent applications such as AI agents should exceed 70% by 2027. In September 2025, the Ministry of Industry and Information Technology (MIIT) published the 2025 list of Excellence-level smart factories, with 274 enterprises selected, covering more than 80% of manufacturing categories. At the end of November 2025, the first batch of Pioneer-level smart factory cultivation candidates was announced, with 15 enterprises including Baosteel, XCMG, and Haier on the list—these 15 cultivation candidates have AI scenario coverage exceeding 70% and have deployed more than 6,000 AI models in total.

What truly brought "AI agents" to center stage was the eight-ministry Implementation Opinions on the "AI Plus Manufacturing" Special Initiative released in January 2026. The document sets quantified targets for 2027: driving the deep application of 3 to 5 general-purpose large models in manufacturing, launching 1,000 high-caliber industrial AI agents, building 100 high-quality industrial datasets, and promoting 500 typical application scenarios. For small and medium-sized enterprises, the document explicitly encourages local governments to issue "computing power vouchers" and "model vouchers" to lower adoption costs.

Reading this chain in full, one conclusion is not hard to draw: at the policy level, the push for AI into factories has moved from encouraging statements into an execution phase of setting quotas, publishing lists, and providing subsidies.

II. The Divergence Behind the Penetration Numbers

The other side of policy is real adoption data, and this data presents a clearly two-sided picture.

The optimistic side: according to IDC research (as reported by 36Kr), the share of Chinese industrial enterprises applying large models and AI agents rose from 9.6% in 2024 to 47.5% in 2025; the share of enterprises applying them across multiple functions such as R&D, manufacturing, and supply chain rose from 1.7% to 35%. IDC also projects that China's industrial enterprise AI spending will approach 90 billion yuan by 2028, at a compound annual growth rate of 38%. Globally, Rockwell's 2025 survey of more than 1,500 manufacturers across 17 countries found that 95% of manufacturers have invested or plan to invest in AI within five years, with quality control the top use case for the second consecutive year.

The sobering side: McKinsey's global survey released in November 2025 (covering 1,993 enterprises across 105 countries) shows that 88% of organizations use AI routinely in at least one function, but only 6% qualify as "high performers" that capture significant enterprise-level value. MIT's August 2025 report is sharper still: roughly 95% of enterprise generative AI pilots have no measurable impact on profit and loss.

The two sets of numbers do not contradict each other. Together they describe the same reality: adoption is fast, but monetization is hard. McKinsey offers the key explanation in the same survey—the single factor with the greatest impact on profit is not model selection but workflow redesign. Inserting AI into existing processes and redesigning processes around AI are two entirely different investments; the former is widespread, the latter rare.

III. What AI Agents in Factories Are Actually Doing

Down on the shop floor, industrial AI agents are currently running most solidly in four types of scenarios.

Quality inspection. Vision large models are replacing traditional rule-based machine vision. Among the 12 lighthouse factories newly added by the World Economic Forum in September 2025, AI and generative AI use cases account for up to 50%, with demonstration factories reporting results including a 41% reduction in product defect rates.

Process parameter optimization. In the Baosteel hot-rolling line case disclosed by Huawei Cloud, the Huawei Pangu large model improved finishing-mill width-spread prediction accuracy by more than 5% and raised yield by 0.5%, which by its own account translates into more than 90 million yuan in additional annual revenue. It should be noted that such figures are all vendor-disclosed and still lack independent third-party verification, but directionally they are consistent with public data from lighthouse factories.

Engineering and maintenance assistance. Siemens Industrial Copilot improves the efficiency of generating programmable logic controller code by roughly 60%, and pilots of its maintenance edition show an average 25% saving in reactive maintenance time; the product won the Hermes Award at Hannover Messe 2025. The factory operations agent Microsoft launched at Hannover Messe 2025 lets employees ask in natural language, "Why did Line 3 stop yesterday?"

Scheduling and order management. This is the entry point closest to AI agents for small and mid-sized factories. An AI tool aimed at small and mid-sized factories has disclosed a set of figures: after adopting AI scheduling and order tracking, factories' average order delay rate fell from 25% to within 5%; one factory in Dongguan cut its order-splitting preparation time by 60%; and an annual fee of about 10,000 yuan corresponded to a return of roughly 1:12. These are again vendor figures, but they touch on an important fact—the value of AI agents for small and mid-sized factories lies first not in sophisticated process optimization, but in freeing up the labor tied down by high-frequency, tedious work such as order tracking and scheduling.

IV. The Real Bottlenecks Are Data and Process, Not Models

Placing the deployment cases alongside the failure rates, the bottlenecks for industrial AI agents are already fairly clear.

First, the data foundation. The prerequisite for large models entering a factory is that equipment data, process data, and quality data are online and usable, yet in a great many factories critical data still sits scattered across paper documents and legacy systems. That the eight-ministry document lists "100 high-quality industrial datasets" as a target shows precisely where the gap lies.

Second, the willingness to re-engineer processes. The 95% of pilots with no P&L impact in the MIT report reflects not so much immature technology as enterprises treating AI as an accessory rather than as part of the process. Worth noting is another figure from the same report: the success rate of purchased solutions is roughly three times that of in-house builds—for the vast majority of manufacturing enterprises, building their own large model is not an option; choosing the right tools and reforming the right processes is.

Third, the "three can'ts" of small and medium-sized enterprises. Industry media summarize them as: don't know how to transform, dare not transform, and cannot afford to transform—lacking knowledge, fearing risk, and short on capital. Policy tools such as "computing power vouchers" and "model vouchers" can ease the third point; the first two depend on whether products emerge that are cheap enough, work out of the box, and are industry-specific and judged by results.

V. Implications for Manufacturing Upstream and Downstream

For enterprises upstream in the supply chain—equipment makers, materials suppliers, and industrial service providers—this wave of AI agent deployment has changed at least two things.

First, customers' procurement judgments are increasingly AI-assisted. When nearly half of industrial enterprises begin using large models across multiple functions, whether your product literature and case data can be read and cited by AI systems will gradually affect your position in customer evaluations.

Second, "who is seriously pursuing intelligent transformation" has itself become a usable sales signal. The lists of 274 Excellence-level and 15 Pioneer-level smart factories are public, and tiered smart-factory cultivation continues across regions—this steadily expanding list is, for upstream suppliers, a roster of target customers with ample budgets and a clear commitment to upgrading.

The Tianxia Gongchang Industry Research Institute will continue tracking the deployment progress of industrial AI agents. If you need to locate manufacturing enterprises pursuing intelligent transformation by industry, region, or scale, you can search conversationally in Tianxia Gongchang AI's database of 4.8 million real factories in active production.

Data Sources

The data in this article comes from public sources: the original policy documents from the State Council and the eight ministries including MIIT, reports from People's Daily Online and Xinhuanet, IDC research (as cited by 36Kr), McKinsey's "The State of AI 2025", MIT's "The GenAI Divide 2025", Rockwell's "2025 State of Smart Manufacturing", World Economic Forum lighthouse factory announcements, and official disclosures from Huawei Cloud, Siemens, and Microsoft. Vendor-reported figures are noted as such in the text.