At the 11th IDC China CIO Summit, a set of figures reported by Fujian Energy and Chemical Group left a deep impression on me: relying on a unified digital foundation, they streamlined their original 18 sets and more than 40,000 accounting subjects into one unified group-wide subject set, completed the go-live of 414 financial entities in two and a half months, pushed the voucher automation rate to 95% and the business-finance integration rate to 98%, and their AI business analysis cockpit improved decision-making efficiency by nearly 17 times. More critically, they cumulatively intercepted more than RMB 120 million in non-compliant payments.
This is not a vision on a PPT slide; it is a hard metric written into the group's performance assessment form. At the same conference, Qingdao Hongxing Chemical Group also delivered results of 50~80% improvement in business processing efficiency, 20~40% reduction in operating costs, and over 60% reduction in compliance risk. Put together, these figures actually answer a question many people have in mind: Has AI in Chinese enterprises really been implemented? The answer is not "whether it has," but "implementation has already been written into the performance assessment form."
From Models to Skills: The Engineering Inflection Point of AI Capabilities
While overseas vendors are still comparing "whose model is smarter," domestic enterprise application vendors have already shifted their focus to "skills engineering." At the recent 2026 Tianjin Digital Intelligence Innovation AI Summit, Yonyou released the refreshed BIP enterprise AI product matrix: achieving full-chain connectivity from the data layer, model layer, and digital intelligence platform layer to the intelligent application and service layer, launching a major enterprise-level Skills builder, and already underpinned by an ontology construction and operations platform and an ontology large model.
The value of this move lies in the two words "Skills". Models solve "whether it can understand"; skills solve "whether it can get things done according to my business rules". By precipitating AI capabilities into reusable, configurable skills with business definitions, enterprises no longer have to start over from prompts every time, but can assemble mature capabilities into their own processes like building blocks. For a system like ERP with strong rules and strong permissions, skill engineering is precisely the prerequisite for an agent to "work legally"—it goes on duty with business definitions and permission boundaries, rather than a bare model running wild in the database.
Xinchuang foundation: independent and controllable is not a slogan, it is an entry threshold
At the same summit, there was another easily overlooked move: Yonyou, together with Phytium Information, Kylin Software, Kingbase, Kingsoft Office, and Tingzhi Co., Ltd., jointly launched the "Tianjin Xinchuang Alliance." Looking at this together with the Skills Builder, the logic becomes clear—for AI to enter core business, the foundation must be independent and controllable.
This is not sentiment; it is a hard constraint. For large state-owned and centrally administered enterprises and key industries, ERP carries funds, inventory, HR, and contracts, and no layer can be controlled by others. Only an IT application innovation foundation with full-stack domestic localization of CPUs, operating systems, databases, and office software, combined with the enterprise's own ontology large model and skills library, forms a closed loop that can be "managed and run." In other words, for leading domestic enterprises, "independent and controllable" has changed from a bonus item to an entry threshold; no matter how smart AI is, if it runs on a non-independent stack, it cannot enter core systems.
This path is in sync with the logic of open-source ERP: a transparent foundation, data ownership, and autonomous governance. The only difference is that Yonyou takes the route of 'commercial products + Xinchuang ecosystem,' while open-source routes like Odoo have written 'you own the source code, 100% own your data, and can see through the cost structure' into their DNA. For growth-stage enterprises and small and medium-sized manufacturers, the latter is often a pragmatic alternative when there are no Xinchuang budget constraints—similarly keeping sovereignty in their own hands, but without having to bear the engineering costs of full-stack domestic substitution.
Digital Divide: The Gap Is Not in the Model, but in the Foundation
The most interesting contrast comes from across the English Channel. UK government research shows that agentic AI is adopted by only 5% of UK businesses (slightly higher in agriculture, mining, manufacturing, and energy, at 12%), 16% of businesses have used at least one AI technology, while 80% have neither adopted it nor plan to. At the same time, ERP Artists has already published “8 Agentic AI Examples for Odoo and ERP in 2026,” applying agents to eight specific processes: customer service, accounts payable, inventory forecasting, manufacturing, sales scoring, dynamic pricing, HR, and marketing.
One side is still watching from the shore, while the other has already produced an operational checklist. The gap shows one thing: competition in agentic ERP is no longer about 'who shouts the slogan first,' but about 'who first gets the hard, unglamorous work in place—master data, permission boundaries, and manual approvals.' What the 80% of enterprises still standing by lack has never been models, but clean master data, clear permissions, and exception handling that can be rolled back—these three foundations.
Fujian Energy & Chemical and Hongxing Chemical are able to fill in hard numbers on the assessment form not because of some magical model, but because they first closed the business-finance loop, embedded the rules into the system, and stopped exceptions before payment. The model is only the final layer of amplification.
A Practical Implementation Checklist for Implementers
Connecting the product engineering of the Tianjin summit, the Xinchuang Alliance’s foundational requirements, and the hard metrics of the IDC awards, the roadmap for Chinese enterprises’ AI implementation is actually already very clear. Here are three practical truths for enterprises currently selecting or implementing solutions:
First, stop asking "Is there AI?" and start asking "Has it been connected into the business-finance closed loop?" AI that can generate weekly reports is everywhere; only AI that can complete tasks against real vouchers, real inventory, and real contracts and leave an audit trail is a capability that can be written into the performance assessment form.
Second, treat sovereignty as a verifiable item. Whether choosing a commercial Xinchuang stack or an open-source foundation, ensure that data can be exported, logs can be retained, and costs can be verified. If the foundation is controlled by others, even the best AI cannot enter the core.
Third, patch the three foundation pieces first before deploying agents: clean master data, clear permission boundaries, and rollbackable exceptions. If these three are not solid, the faster agents are deployed, the bigger the holes will be.
Overseas, people are still agonizing over a 5% adoption rate, while leading domestic enterprises have already proven with 95% voucher automation and 120 million in blocked amounts: the bottleneck for Chinese enterprises in implementing AI has never been whether the model is smart enough, but whether they can connect that intelligence into business and turn it into numbers. Whoever is first to make this path work will be first to get the ticket to the second half.
