An IDC survey of Chinese industrial enterprises in 2025 shows that the proportion of industrial enterprises applying large models and AI agents was 9.6% in 2024, but surged to 47.5% in the first half of 2025. Now in 2026, an industry expert told us that the entire sector has basically begun to test the waters of AI (Artificial Intelligence) this year. This indicates that the proportion of AI applications in the industrial sector is increasing substantially.
However, rising popularity does not equate to true implementation. Currently, many industrial intelligence projects have chosen to integrate large models and connect AI agents, but the vast majority of these agents have become mere decorations within three months of going live. The root cause is the mismatch between general-purpose AI and industrial scenarios.
Industrial scenarios contain a massive amount of implicit expert experience, private knowledge, or personalized data that AI has yet to understand and invoke. This mismatch at the knowledge level prevents the full utilization of AI's analytical and predictive capabilities. Furthermore, industrial production systems are characterized by high constraints, high accountability, and high complexity. Certain processes require AI to have zero hallucinations and zero errors, making enterprises reluctant to let AI agents enter core production processes. This is significantly different from the high tolerance for AI hallucinations in Q&A scenarios.
Therefore, facing the gap between business scenarios and technical capabilities, how industrial AI can break free from demonstrations and achieve real, effective implementation has become a widespread pain point in the intelligent transformation of the industrial sector.
To this end, we conducted an in-depth interview with experts from Deepexi. As one of the representative practitioners of enterprise-level AI and industrial AI, Deepexi has used its own practices to restore the real scenarios and challenges in the deep-water zone of industrial AI implementation.
In the eyes of today's factory managers, introducing AI is no longer the issue; the real question is whether the introduced AI is actually useful. Because to make AI effective in industrial scenarios, providing only single-point model capabilities is far from enough.
Taking the automotive industry as an example, Deepexi told us that complete vehicle manufacturing is a typical process manufacturing industry. For instance, one of our domestic complete vehicle manufacturer clients has an uninterrupted production process, quality control running through the entire process, a long chain, numerous standards, and complex links. Before the transformation, a single component test report could be hundreds of pages long, and the review relied entirely on the engineers' experience and judgment. The manual review pressure was high, and missed or incorrect inspections could occur in the rush.
Introducing single-point AI in just one scenario brings little value to that scenario. This is because knowledge within enterprises is scattered. A large amount of standard specifications, test reports, review checklists, defect records, and expert experience are dispersed across different systems, documents, and offline processes, with some only passed down orally from quality engineers and experts. Without a complete knowledge system, AI cannot understand business associations and review rules. To change this situation, it requires a complete technical system formed by the collaboration of models, knowledge, data, agents, and processes.
Facing such complex quality control scenarios, Deepexi's technical system is broad enough, and its AI capabilities are rich enough, playing a key role.
First, the model layer. The Deepexi enterprise large model carries ontological professional knowledge, understands business frameworks, business processes, and business logic. It serves as the brain hub of the entire system, collaborating with large language models, multimodal models, vectorization models, etc. They possess capabilities such as page layout parsing, OCR recognition, table recognition, and formula recognition, supporting complex document parsing, understanding, and verification.
Second, the data and knowledge foundation and Skills set. Relying on Deepology enterprise ontological data and the Skills set, Deepexi conducts ontological modeling for automotive quality scenarios, precipitating entity relationships, review rules, business logic, and expert experience into a dynamic knowledge graph. Through the Skills set, it encapsulates capabilities such as review, verification, early warning, and rectification, supporting the intelligent closed-loop of the quality control process.
Third, the agent application layer. Relying on the DeepWorks enterprise agent platform, specialized AI digital employees are built for quality system pre-review, AI review of test reports, software pre-review, and AI verification of review checklists. Subsequently, it will expand to advance pre-review, fault diagnosis, operational statistical analysis, and frontline self-service review.
This AI system, which connects the entire process from the underlying model to the upper-level application, transforms AI from a Q&A tool into a productive force that can truly take up its post. The actual effect is excellent: after the AI digital employees are implemented, they can automatically execute tasks. Users submit SVDC test report PDFs or Words for parsing, segmentation, and task decomposition, and combine the ontological graph and knowledge slices to retrieve relevant standard specifications, outputting the quality pre-review results of the test reports. Currently, it only takes a few minutes for an AI digital employee to review a report, achieving intelligent review. Originally, it was manual spot checking; now, it is full inspection via AI. The efficiency is improved by more than 100 times compared to the original manual method, and the rule coverage rate reaches 100%.
Moving from a large automotive country to a strong automotive country, improving quality and efficiency for domestic automobiles is imperative, and the intelligence of quality control will be an important component of competitiveness. For industrial AI service providers, the breadth of the technical system is the prerequisite for supporting full-process AI applications in complex industrial scenarios. Relying on its rich, full-chain AI capabilities, Deepexi has seized the initiative in fields such as automotive manufacturing.
Depth of Integration: Close Collaboration to Embed AI into the Industrial Field
Currently, many manufacturing enterprises are embracing AI and undergoing intelligent transformation, but widespread cognitive blind spots exist. They neither understand their own digital endowments nor the capability boundaries of AI. They are highly familiar with processes and production, but when facing AI, they either have excessively high expectations, believing large models can do everything, or harbor doubts, thinking AI cannot handle production scenarios.
For example, when Deepexi first contacted a domestic laser cutting equipment solution provider, the other party did not understand AI and only raised a broad demand: to make the equipment smarter and easier to use. The reason is that the enterprise's customers are distributed globally with varying skill levels. If the equipment could feature foolproof one-click operation, it would greatly reduce the service pressure on after-sales engineers.
Facing the initial demand, Deepexi did not rush to provide a solution or build a system, but first sorted things out together with the client.
First, Deepexi abandoned extensive project delivery and first sorted out the data foundation, knowledge assets, and personnel capabilities together with the client. During this process, it was found that the client had over 40,000 pieces of original existing data, but after cleaning, only over 600 pieces were truly effective and usable, indicating a severe shortage of high-quality data. Next was a small-scale POC (Proof of Concept) verification, which ultimately achieved acceptable results despite the insufficient data volume and low data quality. The team objectively issued a diagnostic report, analyzed the current data quality situation, and proposed optimization directions. They first guided the client to supplement the data to build a solid foundation for subsequent AI implementation, ultimately dispelling the client's doubts about implementation.
Based on the POC verification results and business priorities, a targeted solution was proposed, deciding to first implement AI in four core scenarios, and then proceed in an orderly manner to steadily embed AI into the entire production chain. High-frequency rigid demand scenarios are prioritized, such as intelligent knowledge Q&A for business modules like operation and use instructions for equipment and software, software fault alarms, hardware fault maintenance, and process parameter optimization guidance, before moving on to intelligent programming, equipment fault diagnosis, and process parameter optimization.
In the knowledge Q&A scenario, AI is relied upon to handle 80% of routine after-sales issues, such as software alarms and hardware maintenance, alleviating the pressure on after-sales engineers and allowing them to focus their time and energy on higher-value tasks. The remaining 20% of difficult and complicated issues are resolved through equipment fault diagnosis.
At the same time, AI capabilities are introduced into the intelligent programming scenario. After the operator submits processing requirements, the system can understand graphics, dimensions, and process requirements, and generate corresponding NC (Numerical Control) codes and processing graphics, helping novices complete processing preparations faster.
Furthermore, in the high-value process parameter optimization scenario, AI has also begun to be implemented step by step. Through AI standard process parameter recommendation, if a user encounters a plate they have never cut before, they can input the plate type, plate thickness, gas type, and rated power to automatically recommend parameters such as cutting focus, gas pressure, and cutting speed. It can also automatically optimize process parameters. Traditional laser cutting relies on the experience of veteran workers. After this AI capability is implemented, it can be cut well within three passes.
If AI is compared to a hammer and the factory's pain points are the nails that need to be hammered down, then the reason for the failure of many industrial AI projects is that industrial practitioners do not understand AI and cannot figure out what this technological hammer can do, while AI practitioners do not understand the industry and cannot see where the real nails are. Technology companies and industrial experts cannot just talk past each other; instead, they need to jointly hold this AI hammer in their hands to solve business pain points.
During this process, Deepexi did not simply graft a general solution, nor did it just deliver a Q&A tool. Through sufficient preliminary communication, they accurately found the nails from the broad demands, enabling AI to truly fit the business pain points, hitting the mark with one strike, and embedding itself into the production site.
The depth of integration with the industry determines the depth at which technology solves problems, and it also serves as the starting point of trust for Deepexi to win long-term cooperation with industrial clients.
Speed of Evolution: The Knowledge Flywheel, Accelerating Scaled Implementation
After contacting hundreds of enterprises, Deepexi knows well that the business value and ROI (Return on Investment) after implementation are what enterprises currently value most.
As early as the small model era of machine learning, factories had already begun to attempt intelligent transformations such as visual inspection and intelligent production scheduling. However, fragmented scenarios can only rely on customization, making them difficult to scale. The costs are also extremely high, and the price of trial and error is too high, which is a difficult problem hindering the popularization of industrial AI.
Entering the era of large models and agents, enterprises are equally sensitive to the implementation costs of AI. Investing millions in intelligence, what can it ultimately bring to the enterprise? This is a mandatory question for industrial AI.
In practice, Deepexi has also explored a scaled path, which may bring some inspiration to the popularization of industrial AI.
First, through an internal circulation mechanism, Deepexi ensures that project assets are no longer cleared upon delivery. Scenario knowledge ontologies, fault data, process rules, expert experience, and standard processes can all be precipitated to continuously enrich its own knowledge, creating a knowledge flywheel.
The knowledge flywheel consists of several parts, including Deepexi enterprise ontological data, industry ontological knowledge foundation, Skills capability set, and a full-industry project case library. Deepexi's effective experience in over a dozen segmented industrial tracks, through continuous precipitation, has become reusable and combinable industrial digital assets.
Next, when docking with the same industry, same type, new projects, and new scenarios, these knowledge assets rooted in real industrial scenarios can reuse existing assets for pre-sales solution generation, POC verification, agent orchestration, and delivery configuration, thereby shortening the trial-and-error cycle and improving delivery quality and efficiency.
Not only Deepexi itself, but also the domestic complete vehicle manufacturer mentioned earlier is a direct beneficiary of the knowledge flywheel. Deepexi helped the complete vehicle manufacturer precipitate its scattered quality review capabilities into knowledge assets that can be understood and invoked by AI, allowing expert experience to be continuously invoked by more teams, forming a pre-positionable, reusable, and expandable intelligent quality capability.
In the early stage of the quality management project, SVDC components were used as the entry point. After this scenario was run through, it was promoted to other components and full-category quality scenarios, and fault diagnosis and operational statistical analysis capabilities were gradually added. This "build once, reuse across multiple domains, and continuously expand" approach solves the pain points of high customization costs and difficulty in scaling traditional industrial AI, helping car companies transform efficiently.
Currently, Deepexi is continuously upgrading. Its technical solutions and tool systems iterate along with AI transformations, changing every few months, continuously promoting the updating and upgrading of the entire industrial AI solution. This continuously evolving knowledge flywheel breaks away from the traditional model of mere customization and piling up man-days, making the entire technical system more proficient and efficient, continuously empowering the entire process of pre-sales, delivery, and iteration.
Through scaled replication and industrialized implementation, helping industrial enterprises carry out full-domain intelligence constitutes another major advantage and barrier for Deepexi in the industrial field.
Industrial AI in 2026: The Dragon Leaps from the Abyss
The characteristic of the industrial field being one meter wide and a hundred meters deep means that scenarios are severely fragmented. If enterprises only chase concepts, pile up models, and make demonstrations, they are likely to consume their budget, patience, and confidence before even entering the main business line.
Reviewing Deepexi's industrial AI practice, it can be found that the premise of technology implementation is to always take value as the ultimate orientation.
Precipitating in hard and tiring work, building barriers, and the technical concept of pursuing real business value enable it to completely abandon conceptualized and demonstrative actions. All strategies revolve around operational pain points and implementation, which directly determines Deepexi's overall strategy in the industrial field.
Facing enterprise clients, it systematically improves quality and efficiency, moving AI from concepts to the field, from tools to productive forces, and delivering business value. Facing the AI industry, the knowledge flywheel provides an effective popularization path for industrial AI to achieve full-domain scaled implementation and improve ROI.
The fourth industrial revolution centered on AI will gradually penetrate 41 major categories and hundreds of medium categories in the industrial field. Industrial AI has just leaped from the abyss, and the subsequent industrial space is even broader. While building a solid foundation for industrial AI, Deepexi has also built an advantageous barrier in this enterprise-level market with a long slope and thick snow.