Written by | Hao Xin Edited by | Wang Pan
Tracing the timeline, the progress of enterprise AI upgrades is far from slow.
In 2023, the inaugural year of large models, the domestic market kicked off the "Hundred Models War," and enterprises began to ponder what large models could do. In 2024, as large model capabilities iterated rapidly, enterprises started experimenting with integrating AI capabilities, with applications like intelligent Q&A and document assistants being launched internally one after another. In 2025, the concept of AI agents exploded. Starting from September last year, enterprises began exploring how to integrate AI into core production processes.
By 2026, Huo Jia, Vice President of Alibaba Cloud Intelligence Group, predicted, "This year is the true inaugural year for the explosion of Agentic AI applications, and the core task for enterprises is to drive the implementation of AI applications."
However, the reality is that most enterprises are stuck at the last mile of AI implementation: transitioning from "usable" to "integrated into production."
The underlying reason is that implementing AI in enterprises requires confronting their internal systems, permissions, data, equipment, and business rules that have been running for years. This means that for AI to truly enter the production process, it is not just a matter of adding another entry point, but rather integrating into this complex system.
Integrating Intelligence into Business
Over the past two decades, enterprise digitalization has primarily addressed the issues of moving processes online, accumulating data, and enabling system collaboration. However, the situation AI faces today is different. Enterprises are not a blank slate; ERP, MES, DCS, and OA systems are already deeply rooted within them. When AI enters, it cannot simply start from scratch, yet it must find the right entry point, naturally making implementation more challenging.
From digitalization to intelligentization, the most prominent change is the role of systems in business. Digitalization focuses on data integration and business restructuring, where data serves as the basis for analysis, humans are the decision-makers, and it relies on information systems and networks; intelligentization centers on algorithmic reasoning and automated decision-making, where data becomes the raw material for decisions, and the decision-makers shift to machine-assisted or autonomous decision-making, relying on a complete digital foundation and AI algorithms.
However, most existing enterprise digital architectures are designed for process solidification and system control. ERP manages resources, MES manages production, DCS manages control, and OA manages processes; each system operates efficiently within its own domain. Systems are accustomed to translating known rules into standard actions, but they are not necessarily good at handling complex scenarios with ambiguous rules, requiring cross-system reasoning, and relying on expert experience. This is precisely the challenge AI must face when entering the production process.
More critically, these systems themselves do not proactively prompt the AI on which data to retrieve at the business site, which systems to switch, who is responsible for decision-making, or what an anomaly means. None of these issues have much to do with whether the large model is smart or cheap.
Beyond system-level barriers, the construction model itself also constitutes a hindrance. Huo Jia told Photon Planet that traditional AI project-based construction involves planning the process first and then starting the build. However, today's model-related technologies are evolving rapidly. From context engineering at the beginning of the year to Harness engineering and then to self-evolution engineering, four versions have been updated in just nine months. Under the inherent digital construction model, it is difficult to keep up with such rapid technological iteration and scenario implementation requirements.
The superposition of system solidification and construction inertia has jointly caused the current dilemma of enterprise AI implementation. AI cannot enter the existing systems, and the traditional construction model cannot keep up with technological iteration.
What has been accomplished in the digitalization era is bringing business into systems; what needs to be accomplished in the intelligentization era is bringing intelligence into business. What enterprises truly lack is a methodology that enables AI to understand business, invoke existing digital systems, undergo validation, and participate in business processes.
Selecting Scenarios, Building Demos, and Saying No
This methodology cannot be built from scratch, nor can it be supplemented by product design far removed from the site. It requires vendors to enter the business site and connect problems, systems, and engineering implementation.
However, over the past three years, most of the paths enterprises have tried are far removed from the site. Some started with large model APIs to create a batch of Q&A assistants; some started from a specific business pain point to conduct single-point POCs; some set up special AI teams to try to drive it top-down. These attempts are not necessarily wrong in direction, but the common problem is that design is done remotely, validation relies on reports, and delivery marks the end. There are still few cases where AI truly enters the production process and forms a closed loop.
In the process of serving large state-owned enterprises and industry-leading clients, Alibaba Cloud has gradually explored a methodology for enterprise AI implementation. This methodology does not require enterprises to build a new system from scratch; its core is to transform general AI capabilities into business capabilities that can truly enter the production process, built upon the enterprise's existing digital foundation.
Behind this also lies a brand-new working model, which Huo Jia summarizes in four steps: selecting the right scenarios, rapid validation, scientific construction, and the effect flywheel.
Selecting the right scenario is the starting point. Determining whether a scenario is worth pursuing with large model technology remains the primary factor for the success or failure of most projects to this day. As AI evolves, scenarios are becoming increasingly numerous. Practice has revealed that many scenarios are not necessarily suitable for large models, and forcing it will lead to failure. What is truly worth prioritizing are the core links that are high-frequency, high-value, and highly dependent on experts. In other words, the closer to the core of production, the more measurable the value of AI and the harder it is to replace.
PetroChina Lanzhou Petrochemical is one of Alibaba Cloud's key clients. During the scenario selection phase, the Alibaba Cloud team used a three-tier funnel-style research approach to refine details layer by layer. The headquarters' strategic research covered six functional departments and four industrial co-creation centers, then focused on the four co-creation centers to screen around exploration and development, water injection and oil recovery, refining and chemical units, and equipment inspection and maintenance. Finally, they went to the site to observe the processes alongside unit management personnel, process experts, and control room operators, ultimately selecting the abnormal alarm diagnosis of the refining atmospheric and vacuum distillation units as the pilot.
Rapid validation determines whether the direction is viable. No PPTs, just look at the DEMO directly. Run it with real data and real processes, let business experts directly give feedback on the results, and push the discussion from "whether to do it" to "how to do it better." In the Lanzhou Petrochemical project, the team completed the design and demonstration plan 23 days after entering the site, then spent 13 days debugging it end-to-end on-site, producing a runnable business Demo in 36 days. When the runnable system was put on the table first, business experts directly evaluated the diagnostic results, evidence chain, and handling suggestions to determine how to proceed to the next step.
This mindset of rapid validation is also reflected in the rhythm of project advancement. Huo Jia has a clear requirement for project advancement: compress the cycle. A central state-owned enterprise project team initially reported that going live would take three months. Huo Jia directly said not to do it that way, changing the release rhythm to once a month, and later to once a week. In addition, he insisted on a prerequisite: building a true "One Team" relationship with the client, rather than the traditional Party A and Party B model.
Scientific construction covers corpus data, model strategies, evaluation systems, and engineering methods. Production-grade systems are fundamentally different from experimental environments and require a complete technical path for support.
Taking the Lanzhou Petrochemical project as an example again, the Alibaba Cloud team first built an ontology knowledge base for the atmospheric and vacuum distillation units based on knowledge graph technology, translating the business language of the physical world into the language of the ontology world; then used large models to clean data from operation manuals and various IT systems into Agentic data and fed it into the ontology knowledge base; finally, they built a global reasoning engine, training a vertical domain time-series large model based on the Qwen Max model and the atmospheric and vacuum distillation units. When an anomaly occurs, this model can simulate human thinking processes, assist agents in capturing anomaly signals, and provide root cause judgments. The system has been running continuously since its launch in May, achieving an overall accuracy of 90% and improving handling efficiency by 80%.
The effect flywheel enables continuous value addition for the first three steps. New data generated after the system goes live, new feedback from business personnel, and new experience accumulated by experts should all flow back into the corpus, evaluation sets, and engineering rules to drive continuous optimization of the model and system. With each cycle, the system's understanding of the business deepens, the output value increases, and the unit cost decreases accordingly. When the output value far exceeds the input cost, intelligent construction enters a self-reinforcing positive cycle.
Computing Power, Models, and Agents are Indispensable
The implementation of this methodology cannot be separated from the support of a complete full-stack AI infrastructure.
Overall, this support system is divided into three layers: the bottom layer is AI Infra, the middle layer is MaaS, and the top layer is Agent. The underlying computing power supports inference efficiency, the MaaS layer provides model scheduling and private deployment, the Agent layer completes agent orchestration and engineering constraints, and the security and observability systems ensure production-grade operation.
Alibaba Cloud is a relatively good observation sample. At the AI Infra layer, Alibaba Cloud has T-Head chips. The currently available PPU1.5 is the first domestic chip to support native FP4, and also the first to achieve 144G HBM3E video memory and 800G video memory bandwidth, ensuring sufficiently high inference efficiency for large-scale MoE models.
At the MaaS layer, Alibaba Cloud provides the Qwen series of models, including a large model with 2.4T parameters, as well as the 27B small model of Qwen3.8. It also provides five types of exclusive closed-source models, ranging from large language to multimodal, to speech recognition, speech generation, video generation, and image generation and editing, all of which can be privately deployed.
At the Agent layer, Alibaba Cloud provides the AgentScope agent framework, Qoder AI-native programming tools, Harness engineering platform, industrial ontology platform, and more.
The Lanzhou Petrochemical project verified the supporting capabilities of this full-stack system. The ontology knowledge base enables agents to understand the units, the Harness layer makes every step of reasoning auditable and trustworthy, 13 professional AI assistants are embedded into the daily work of teams, and industrial software is uniformly scheduled by large models via the MCP protocol.
This full-stack capability also builds a positive-cycle flywheel. Cloud vendors use AI-native products to solve problems at the client site, and the clients' usage generates real data and feedback. This feedback flows back to the product team, driving product feature iteration and architecture evolution. After product capabilities are enhanced, scenario value and work efficiency are further improved.
Huo Jia told us that in the process of serving clients, Alibaba Cloud has expanded step by step from initially only using basic model and cloud products to more AI-native products such as the Bailian platform, Qoder, and Qwen Workspace. The requirements obtained in real scenarios can serve as a feedback source for future product iterations, continuously maintaining the product's advanced nature.
For cloud vendors, once the usage and product flywheel starts spinning, it can drive the underlying model and data flywheels, thereby enhancing overall commercialization. For clients, it solves the three pain points of finding no scenarios, validating technical feasibility and ROI, and production-grade launch all at once. Once a positive ROI is generated, enterprise AI implementation naturally yields value returns.
Technology is Not the Barrier; Transforming Technology into Enterprise-Exclusive Intelligence Is
As enterprise AI implementation gradually moves into deep water, more fundamental issues are emerging. General large models are becoming increasingly powerful, but an enterprise's core competitiveness has never been about possessing an exceptionally smart model.
General large models are the technological foundation, callable by any enterprise, with capabilities tending toward homogenization. What truly creates the gap is the enterprise's own data, knowledge, and processes, as well as the process parameters, expert experience, institutional rules, and business logic accumulated in the production process.
Huo Jia repeatedly emphasizes a judgment: no matter what the technology is, for enterprises, the value ultimately boils down to the eight words: "increase revenue, improve quality, reduce costs, and enhance efficiency." The essence of enterprise operation is to create value and earn profits. Technology itself does not constitute a barrier; transforming technology into exclusive capabilities does.
Exclusive intelligence is not bought; it grows bit by bit within the production process. There are no shortcuts in this process. Which data is usable, which knowledge is trustworthy, which rules must be strictly constrained, and which judgments can be handed over to the model all need to be confirmed one by one through real validation. Once accumulated, these exclusive assets constitute the most difficult-to-replicate barrier in the enterprise's intelligent transformation.
After model capabilities become homogenized, what can truly command a premium is not the model, but the ability to use exclusive intelligence to solve specific business problems. The stronger the model, the higher the ceiling for exclusive intelligence; but the depth of exclusive intelligence depends on how much the enterprise itself has accumulated in business scenarios.
For exclusive intelligence to be implemented, specific product carriers are also needed. Taking Qwen Workspace as an example, in addition to large-scale personal office scenarios, a privatized version for enterprises has also been launched, which can be deployed in the enterprise's own environment and adapted for permissions, data isolation, and auditing around security and trustworthiness requirements.
What Qwen Workspace brings into the enterprise is not just an AI assistant, but digital employees who understand their roles, can collaborate, and are traceable, undertaking specific tasks in enterprise services, operational management, and industrial collaboration. When digital employees truly take their posts, the depth of exclusive intelligence has a measurable foothold and room for continuous thickening.
General intelligence is the starting point; exclusive intelligence is the endpoint. When enterprises transform general capabilities into autonomous and controllable exclusive intelligence, AI truly moves from "usable" to "integrated into production," transforming from a technological tool into productivity itself.
WeChat ID | TMTweb Official Account | Photon Planet