Looking back at the timeline, the progress of enterprise AI upgrades has not been slow.

In 2023, the first year of large models, domestic companies initiated a hundred-model battle, and enterprises began to think about what large models could do. By 2024, the capabilities of large models were rapidly iterating, and companies started to integrate AI capabilities, with applications like intelligent Q&A and document assistants gradually going live internally. In 2025, the concept of intelligent agents exploded, and starting from September of the previous year, enterprises began to explore integrating AI into core production processes.

By 2026, Huo Jia, Vice President of Alibaba Cloud Intelligence Group, judged that 'this year marks the true explosion of Agentic AI applications, and the core task for enterprises is to drive AI application implementation.'

But the reality is that most enterprises are stuck in the last mile of AI implementation: from 'usable' to 'entering production.'

The reason behind this is that AI implementation in enterprises must confront their internally established systems, permissions, data, devices, and business rules that have been running for years. This means that for AI to truly enter production processes, it is not just a matter of adding another entry point but rather integrating into this complex system.

Integrating Intelligence into Business

The digitalization of enterprises over the past two decades has primarily addressed the issues of process onlineization, data accumulation, and system collaboration. However, the situation AI faces today is different. Enterprises are not blank slates; ERP, MES, DCS, and OA systems are already deeply rooted internally. When AI enters, it can neither start from scratch nor must find suitable entry points, making implementation naturally more challenging.

From digitalization to intelligence, the most prominent change is the role of systems in business. Digitalization focuses on data connectivity and business reconstruction, with data serving as the basis for analysis and humans as the decision-making subjects, relying on information systems and networks. Intelligence focuses on algorithmic reasoning and autonomous decision-making, with data becoming the raw material for decisions and machines becoming the decision-making subjects, either as aids or autonomous decision-makers, requiring 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, with each system operating efficiently in its own domain. Systems are accustomed to converting known rules into standard actions but are not necessarily adept 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 production processes.

More critically, these systems do not proactively indicate to AI which data should be retrieved, which systems should be switched, who should be responsible for decision-making, or what an anomaly signifies. These issues have little to do with whether large models are smart or cheap.

Beyond system-level obstacles, the construction mode itself also poses resistance. Huo Jia told Photon Planet that traditional AI project-based construction involves planning processes before starting construction. However, today, model-related technologies are rapidly evolving, with four versions updated in just nine months, from context engineering at the beginning of the year to harness engineering and then self-evolving engineering. Under the inherent (inherent) digital construction mode, it is difficult to keep up with such rapid technological iterations and scenario implementation requirements.

The combination of system solidification and construction inertia has jointly created the current dilemma of enterprise AI implementation. AI cannot integrate into existing systems, and traditional construction modes cannot keep up with technological iterations.

What has been accomplished in the digital age is enabling business to enter systems; what needs to be accomplished in the intelligent age is enabling intelligence to enter business. What enterprises truly lack is a methodology that allows AI to understand business, invoke existing digital systems, undergo validation, and participate in business processes.

Selecting Scenarios, Creating Demos, and Saying No

This methodology cannot be built from scratch or supplemented by product designs far removed from the field. It requires vendors to enter the business site and connect problems, systems, and engineering implementations.

However, over the past three years, most enterprise attempts have been far removed from the field. Some have started with large model APIs to create Q&A assistants; others have focused on specific business pain points for single-point POCs; still, others have established AI task forces to drive change top-down. These attempts may not be wrong in direction, but the common issue is that design occurs remotely, validation relies on reports, and delivery marks the end. Cases where AI truly enters production processes and forms closed loops are still rare.

Alibaba Cloud has gradually developed a methodology for enterprise AI implementation through serving large state-owned enterprises and industry-leading clients. This methodology does not require enterprises to build a new system from scratch but rather focuses on transforming general AI capabilities into business capabilities that can truly enter production processes on top of their existing digital foundations.

This also represents a new working mode, which Huo Jia summarizes in four steps: selecting the right scenarios, rapid validation, scientific construction, and the virtuous cycle of effectiveness.

Selecting the right scenarios is the starting point. Judging whether a scenario is worth pursuing with large model technology remains the primary reason for the success or failure of most projects. As AI develops, more scenarios emerge, but many may not be suitable for large models, and forcing implementation can lead to failure. The core areas truly worth prioritizing are those that are high-frequency, high-value, and strongly expert-dependent—the closer to the core of production, the more measurable AI's value and the harder it is to replace.

PetroChina Lanzhou Petrochemical is one of Alibaba Cloud's key clients. In selecting scenarios, the Alibaba Cloud team used a three-level funnel-style survey to refine options layer by layer. Headquarters strategic surveys covered six functional departments and four industrial innovation centers, then focused on four innovation centers to screen around exploration and development, water injection and oil production, refining devices, and equipment inspection and maintenance. Finally, they went to the field to work with device managers, process experts, and internal operators to refine processes, ultimately selecting abnormal alarm diagnosis for refining atmospheric and vacuum devices as the pilot.

Rapid validation determines whether the direction is valid. Forget PPTs; focus on DEMOs. Run through with real data and real processes, let business experts provide direct feedback on the results, and advance discussions from 'whether to do it' to 'how to do it better.' In the Lanzhou Petrochemical project, the team completed the design demonstration plan 23 days after entering the site, then spent another 13 days achieving end-to-end connectivity on-site, producing a functional business Demo in 36 days. When the functional system was presented first, business experts directly evaluated the diagnostic results, evidence chains, and disposal recommendations to determine the next steps.

This rapid validation approach also runs through the project's rhythm. Huo Jia has a clear requirement for project progress: shorten the cycle. A central enterprise project team initially reported a three-month timeline for going live, but Huo Jia said not to proceed and instead changed the release rhythm to monthly, then weekly. Additionally, he insisted on a premise: building a true 'One Team' relationship with clients rather than a traditional vendor-client mode.

Scientific construction covers corpus data, model strategies, evaluation systems, and engineering methods. Production-grade systems differ fundamentally from experimental environments and require a complete technical path to support.

Taking the Lanzhou Petrochemical project as an example, the Alibaba Cloud team first built a knowledge base for atmospheric and vacuum devices using knowledge graph technology, translating business language in the physical world into language in the ontological world. Then, they used large models to clean operation manuals and data from various IT systems into Agentic data, which was fed into the ontological knowledge base. Finally, they built a global reasoning engine based on the Qwen Max model and trained a vertical-domain time-series large model for atmospheric and vacuum devices. When abnormalities occur, the model can simulate human thinking processes to assist intelligent agents in capturing abnormal signals and providing root cause judgments. Since going live in May, the system has continuously operated with an overall accuracy rate of 90% and an 80% improvement in disposal efficiency.

The virtuous cycle of effectiveness ensures continuous value addition from the first three steps. New data generated after system launch, 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 model and system optimization. With each cycle, the system gains a deeper understanding of the business, outputs higher value, and reduces unit costs. When the output value far exceeds the input cost, intelligent construction enters a self-reinforcing positive cycle.

Compute Power, Models, and Intelligent Agents: All Are Indispensable

The implementation of this methodology relies on a complete full-stack AI infrastructure for support.

Overall, this support system is divided into three layers: AI Infra at the bottom, MaaS in the middle, and Agent at the top. Bottom-layer compute power supports reasoning efficiency, the MaaS layer provides model scheduling and private deployment, and the Agent layer completes intelligent agent orchestration and engineering constraints, with security and observability systems ensuring production-grade operation.

Alibaba Cloud serves as a good observation sample. At the AI Infra layer, Alibaba Cloud has T-Head chips. The currently available PPU1.5 is the first chip in China to support native FP4 and the first to achieve 144G HBM3E memory and 800G memory bandwidth, ensuring high reasoning efficiency for MOE large-scale models.

At the MaaS layer, Alibaba Cloud provides the Qwen series models, including a 2.4T-parameter large model and a 27B small model (Qwen3.8), along with five types of proprietary closed-source models for large language, multimodal, speech recognition, speech generation, video generation, image generation, and editing, all available for private deployment.

At the Agent layer, Alibaba Cloud provides the AgentScope intelligent agent framework, Qoder AI-native programming tools, the Harness engineering platform, and the industrial ontology platform.

The Lanzhou Petrochemical project validated the support capabilities of this full-stack system. The ontological knowledge base enabled intelligent agents to understand devices, the Harness layer made each reasoning step auditable and trustworthy, 13 professional AI assistants were embedded in daily team operations, and industrial software was uniformly scheduled by large models through the MCP protocol.

This full-stack capability also builds a positive feedback loop. Cloud vendors solve problems on-site with AI-native products, and client usage generates real data and feedback. This feedback flows back to the product team, driving product feature iterations and architectural evolution. Enhanced product capabilities further elevate scenario value and work efficiency.

Huo Jia told us that while serving clients, Alibaba Cloud has expanded from initially using only basic model and cloud products to incorporating more AI-native products like BaiLian Platform, Qoder, and Qwen Office. Demands identified in real scenarios serve as feedback sources for future product iterations, continuously maintaining product advancement.

For cloud vendors, when the usage and product feedback loop is activated, it can drive the underlying model and data feedback loops, enhancing overall commercialization. For clients, it resolves three pain points at once: finding scenarios, validating technical feasibility and ROI, and achieving production-grade launch. Once a positive ROI is achieved, enterprise AI implementation naturally yields value returns.

Technology Is Not the Barrier; Transforming Technology into Proprietary Intelligence Is

As enterprise AI implementation gradually moves into deeper waters, more fundamental issues are surfacing. General-purpose large models are becoming increasingly powerful, but an enterprise's core competitiveness has never been about owning the smartest model.

General-purpose large models serve as a technological foundation and can be called upon by any enterprise, leading to homogenization of capabilities. What truly makes a difference is an enterprise's own data, knowledge, and processes—the process parameters (process parameters), expert experience, institutional rules, and business logic precipitate ( precipitate : accumulated) in production processes.

Huo Jia repeatedly emphasizes a judgment: Regardless of the technology, for enterprises, value ultimately boils down to eight words: 'increasing revenue, improving quality, reducing costs, and enhancing efficiency.' The essence of business is to create value and earn profits. Technology itself does not constitute a barrier; transforming technology into proprietary capabilities does.

Proprietary intelligence is not purchased; it grows bit by bit within production processes. There are no shortcuts in this process. Which data is usable, which knowledge is trustworthy, which rules must be rigidly enforced, and which judgments can be delegated to models all need to be confirmed one by one through real validation. Once precipitate (accumulated), these proprietary assets form the most difficult-to-replicate barrier in an enterprise's intelligent transformation.

When model capabilities become homogenized, what truly commands a premium is not the model itself but the ability to solve specific business problems with proprietary intelligence. The stronger the model, the higher the ceiling for proprietary intelligence; however, the thickness of proprietary intelligence depends on how much an enterprise has accumulated in business scenarios.

Proprietary intelligence also requires specific product carriers for implementation. Take Qwen Office as an example: beyond large-scale personal office scenarios, it has also launched a private version for enterprises, deployable in their own environments and adapted for security and trustworthiness with permissions, data isolation, and auditing.

Qwen Office brings more than just an AI assistant into enterprises; it introduces digital employees who understand roles, can collaborate, and are traceable, taking on specific tasks in enterprise services, operations management, and industrial collaboration. When digital employees truly take on roles, the thickness of proprietary intelligence gains measurable anchors and room for continuous growth.

General intelligence is the starting point; proprietary intelligence is the destination. Only when enterprises transform general capabilities into autonomous, controllable proprietary intelligence does AI truly move from 'usable' to 'entering production,' transitioning from a technological tool into productivity itself.