Industrial agents have emerged as a pivotal focus within the realm of industrial intelligence.

Recently, the popularity of AI + industry has soared, with industrial agents taking center stage as a recognized means to elevate industrial intelligence.

“Our large model business has witnessed at least a fivefold increase this year compared to last, with the majority of these gains stemming from industrial agents,” Zhi Zhen, Chairman of Zhonggong Interconnection, a service provider specializing in the deployment of large models and agents in the industrial sector, told Digital Frontline. Following the market enlightenment and scientific popularization initiated by DeepSeek earlier this year, coupled with the continuous policy push in later stages, the market window has officially opened.

Market actions have swiftly followed suit. At the World Artificial Intelligence Conference, numerous domestic manufacturers showcased their industrial agent products and implementation cases.

On August 26, the “Opinions of the State Council on In-depth Implementation of the ‘AI +’ Action” explicitly listed the industry as a standalone item among the key actions for developing the ‘AI +’ industry, emphasizing the promotion of intelligent linkage among all industrial factors and accelerating AI application across design, pilot testing, production, service, and operation.

Has the industrial agent boom truly arrived? Have agents ventured into the depths of the industrial field?

01

Industrial Agents: Genuine or Artificial Heat?

What precisely are industrial agents?

Many industry insiders acknowledge that there is currently no definitive answer to this question. Broadly speaking, industrial agents represent a new generation of cognitive intelligent systems that integrate large models, knowledge graphs, industrial mechanisms, and other technologies. They possess “autonomy” and “synergy,” capable of perceiving the industrial environment, understanding natural language instructions, making decisions, and controlling physical equipment to complete industrial tasks, thereby achieving intelligent control and optimization of production equipment, process flow, logistics management, and other aspects.

Their value is universally acknowledged within the industry. “Customers spanning over 40 major categories and hundreds of subcategories within the industry are now actively pursuing agent implementation, and the market is accelerating,” Zhang Ming, General Manager of Baidu Intelligent Cloud’s Smart Industry Solutions, told Digital Frontline. This year has seen heightened enthusiasm for industrial agents across all facets of enterprise research, production, supply, marketing, and service.

Song Tao, General Manager of Lenovo China’s manufacturing industry business group, emphasized that enthusiasm is particularly high for “industrial enterprise operation agents.” Agents are still in their nascent stages of development, with most leading enterprises cautiously experimenting with enterprise operation and management. Across various links such as research, production, supply, marketing, service, human resources, finance, taxation, laws, and regulations, “there will be numerous intelligent agents focused on enhancing operational management efficiency.”

“The surge in popularity of industrial agents is most evident in the shift from customers being ‘wait-and-see’ to ‘actively demanding,’ especially among leading industrial enterprises that are no longer content with conventional ‘automation + informatization’ construction but clearly aim to build ‘AI-driven smart factories,’” Zhi Zhen, Chairman of Zhonggong Interconnection, revealed to Digital Frontline. In the past, they were primarily the ones actively promoting and explaining the benefits, but now customers are “actively pulling” and “coming to visit,” with many enterprise decision-making teams personally leading teams to investigate agent platform capabilities, and even incorporating “establishing an agent system” into the group’s future development strategy.

IDC data aligns with this trend. By 2025, the proportion of industrial enterprises exploring agents will see a significant year-on-year increase, with the proportion of those applying large models and agents surging from 9.6% in 2024 to 47.5% in 2025.

Du Yanze, Senior Research Manager at IDC China, revealed that the primary reason for this increase is the heightened curiosity of industrial enterprise CXOs towards industrial large models and the lowering of thresholds, which has attracted more industrial enterprises and IT service providers to join, marking the transition of the industrial large model and agent market from initial emergence to a stage of extensive exploration.

Among these enterprises, the proportion that have already implemented applications across multiple links has surged from 1.7% to 35%. Among enterprises that have applied large models and agents, over 73.7% possess application scenarios ranging from 10 to several dozen.

02

Policy Efforts Intensify

When did this industrial agent boom commence?

A widely accepted answer within the industry is – early 2025. The landmark event was the launch of the domestic open-source large model DeepSeek and the agent platform Manus, which swiftly accomplished market enlightenment and allowed various industries to recognize the potential value of agents.

Especially for industrial manufacturing enterprises that have generally faced challenges such as uncertain external environments, high production and operational pressures, and compressed product profit margins over the past few years, this may be the most promising technology in the past two years, with the potential to deliver tangible value.

Therefore, it is evident that over the past six months, both industrial enterprises and service providers have significantly increased their enthusiasm for exploring agents, with the number of implementation cases continuously rising. The emergence of benchmark cases and the accumulation of engineering experience are further igniting the enthusiasm of more enterprises.

Apart from the endogenous demand of enterprises to enhance their competitiveness, the policy thrust continues to intensify.

Currently, global competition in artificial intelligence is in full swing, with China and the United States emerging as the core competitors. In high-value tertiary industries such as the Internet, law, finance, healthcare, and design, domestic and foreign countries are focusing on implementation. However, in the industrial field, the world has yet to explore a clear path.

Against this backdrop, exploring a differentiated development path leveraging China’s comprehensive industrial system and abundant scenario resources to promote industrial upgrading and create new productivity has become a strategic imperative. Some senior industry insiders believe that this process is akin to nurturing the new energy industry in the past, requiring a “policy-first, industry-traction” approach.

At the industrial level, among the 189 globally recognized “Lighthouse Factories,” China boasts a total of 85, accounting for 45%, ranking first worldwide in total count, with a robust foundation for smart manufacturing and digitization in the manufacturing sector.

At the policy level, last year’s government work report clearly proposed to “carry out the ‘AI +’ action.” This year, more policies focused on the industrial field are continuing to intensify.

In March, the government work report reiterated the “AI +” action. Unlike last year’s focus on technological research and development and industrial cluster construction, this year’s emphasis is more on technology landing and application.

On June 6, the Leading Group Meeting on Integration of Informatization and Industrialization of the Ministry of Industry and Information Technology deliberated the “Key Points of Work in 2025,” clearly proposing to deepen the application of artificial intelligence in industry, with industrial agents serving as the starting point, and drive the innovative iteration of industrial datasets and industrial large models. This is regarded as an important signal by the industry.

“The policy intensity has exceeded expectations,” Zhi Zhen predicted. A series of policies will be issued in various regions next. For instance, shortly after the tone of the Ministry of Industry and Information Technology meeting was set, Beijing held an industrial seminar on industrial agents under the guidance of the Beijing Municipal Commission of Economy and Information Technology in mid-to-late July, with numerous leading enterprises in AI and industrial software fields participating, engaging in in-depth exchanges and sharing cases.

On July 28, the 2025 National Symposium of Responsible Comrades from the Industrial and Information Technology Departments was held, which explicitly proposed to promote the “AI + manufacturing” action to go deeper and more solidly when deploying the key work for the second half of the year, strengthen research on the foundation and application in key scenarios, deepen the application of the Industrial Internet by classification and level, and cultivate a number of industrial agents.

On August 19, Shanghai issued the “Implementation Plan for Accelerating the Development of ‘AI + Manufacturing’ in Shanghai,” proposing multiple deployments such as enhancing the basic capabilities of industrial models and breaking through cutting-edge technologies in industrial intelligence.

In the “Opinions of the State Council on In-depth Implementation of the ‘AI +’ Action” issued on August 26, it was emphasized to promote the intelligent development of all factors in industry, deepen the integrated application of artificial intelligence and the Industrial Internet, and enhance the intelligent perception and decision-making execution capabilities of industrial systems.

“The policy has shifted from encouraging technological exploration to systematically promoting the coordinated iteration of industrial datasets, large models, and agents, adding momentum to the initial market-driven demand at the beginning of the year,” Du Yanze, Senior Research Manager at IDC China, told Digital Frontline. With the further refinement of policies in various regions in the follow-up, the demand side will be consolidated, especially the demand of central state-owned enterprises, driving more extensive industrial enterprises to join the exploration while promoting the development of the supply side, with more vendors entering the market.

03

Enterprise Demands and Actions Undergo Transformation

Amidst the boom, industrial enterprises have begun to advance from the pure “trial-and-error” stage to a new phase of pursuing deep integration and value output.

Zhang Ming from Baidu Cloud told Digital Frontline that last year, many enterprises had implemented some agent applications. At that time, they primarily required the ability to mount enterprise knowledge bases or even achieve knowledge Q&A and retrieval solely based on the capabilities of general large models to meet preliminary usage requirements. However, this year, enterprises are no longer satisfied with such basic applications and are more eager for agents to integrate into business scenarios, solve specific problems, and generate tangible value.

Take the R&D scenario of chemical enterprises as an example. It used to rely on the deep knowledge and experience of industry experts to understand molecular characteristics and conduct repeated experiments. Nowadays, they expect agents to learn various molecular characteristics, automatically recommend molecular synthesis routes to reduce the number of experiments, and even predict the properties of new substances, such as their wear resistance, flame retardancy, etc.

Another example is the energy-saving scenario of industrial air compressors, which previously relied on fixed control strategies, adopting different optimization modes during peak and off-peak seasons, requiring manual on-site adjustments, and the technology was difficult to link to produce significant value. Now, it is anticipated that agents can integrate various technologies to achieve fully intelligent dynamic optimization, which not only allows remote control of the cabinet to wirelessly transmit data and issue precise control instructions from the cloud but also utilizes AI models to predict gas consumption and formulate optimal scheduling control strategies based on real-time collected data. It can also send real-time alerts and push them to management personnel when issues such as low energy efficiency arise.

“An integrated product that fuses agents, AI algorithms, cloud control, and other technologies, if it can be polished, produce efficiency, and be replicable, will have immense promotional value,” Zhi Zhen said. Now that large models have passed the initial frenzy stage, those who can closely integrate business and create superior agents in the next year or two will undoubtedly hold value. This mirrors the development trajectory of the Internet, gradually transitioning from early general networks and portals to vertical fields such as e-commerce and social networking.

As demands evolve, corporate actions have also become more systematic.

Over the past two years, many enterprises have gradually become more rational and professional in their understanding of technology through trial and error. For instance, they once believed that building a knowledge base was straightforward, but in reality, to guide production, it requires domain-specific knowledge enhancement, agent support, and closed-loop business correction, far from achievable through “zero-cost trial play.”

Therefore, as a new round of booms unfolds, more industrial enterprises are shifting their focus to specific scenarios with more precise demands. Simultaneously, they have begun to pay attention to platform capabilities, the depth of data integration, the continuity of model training, and cross-system integration capabilities, no longer merely停留在 the surface of “AI display,” and in this process, they have also increased the threshold for budget investment.

“The entire market is transitioning from ‘heat’ to ‘substance,’ a change we hold in high regard,” Zhi Zhen said.

Notably, leading enterprises are promoting the implementation of industrial agents in a more systematic manner.

Using Lenovo as an example, in May of this year, at the Lenovo Innovation Technology Conference, the company officially launched the Super Agent, which encompasses all scenarios for individuals, enterprises, and cities. Among these, the enterprise-focused Super Agent, Lenovo Enjoy, serves as Lenovo's next-generation official entry point for customers, business partners, investors, media, and the public. It can access chain data information across devices and ecosystems within the enterprise to autonomously complete work tasks. For instance, it can proactively anticipate user needs and initiate interactions, assisting users in placing orders independently.

It forms an "agent army" with agents from various fields such as R&D, supply chain, marketing, sales, service, and operations, embarking on a journey to help Lenovo reduce costs and enhance efficiency. Currently, Lenovo is applying Enjoy to its own operations. After automatically identifying customer intentions, the Super Agent disassembles tasks and collaborates with the aforementioned domain agents to process them jointly, ultimately delivering the results.

Song Tao told Digital Frontline that, apart from central state-owned enterprises, the AI implementation of leading private enterprises is largely similar to Lenovo's, involving strategic resource investments to explore business transformations, with large enterprises typically establishing their own teams. For example, Lenovo has a dedicated DTIT (Digital Transformation, Information Technology) department, encompassing business application delivery and AI R&D teams, responsible for designing exclusive hybrid AI bases to support the group's business intelligence transformation.

Particularly this year, when promoting AI implementation, many enterprises have shifted from early extensive exploration to requiring various departments to "detail implementation with KPIs".

Song Tao explained that when DeepSeek gained attention previously, many enterprises quickly deployed full-fledged versions for various departments to test after the Spring Festival. However, due to the surge in visits, the development and operation teams struggled to handle it, leading them to gradually shift to various departments independently promoting agent-related work.

At the same time, "there have also been numerous demands from small enterprises this year, but most of them have not purchased commercial platforms, instead using open-source tools, models, or even small computing power cards to run related scenarios," Zhang Ming of Baidu Cloud told Digital Frontline. These enterprises have not yet formed a large-scale system spanning from computing power, model training and tuning tools to business scenarios. Instead, they start with specific business scenarios that can help them save money or increase revenue, building agents themselves to understand their capabilities. Once the effectiveness is confirmed and their own capability boundaries are realized, such as discovering that the self-developed tuning effect is not as good as professional-level, they will consider commercial payments.

04

Where Have Industrial Agents Gone?

The current market is evolving extremely rapidly. On the application side, industrial agents are continuously expanding the breadth and depth of their application. On the technical side, iterations such as collaboration between large and small models, tool chain updates, and context engineering optimization are frequent, with new changes occurring every one or two weeks. Some scenarios that did not perform well in the previous two years have achieved a leap in value this year with the improvement of engineering capabilities and the application of agents.

Taking the scenario of querying numbers as an example, which requires extremely high accuracy, a single data error may pose significant risks to the business decision-making of enterprises, necessitating 100% accuracy. However, large language models are prone to hallucinations when processing data. "The approach in the 1.0 stage was to use large language models to write SQL and fixed questions, and then query SQL after understanding the specific problems corresponding to the agent through intention, but the accuracy could only reach 90%, making it difficult to meet the demand," Zhang Ming of Baidu Intelligent Cloud explained. In the 2.0 stage, it was recognized that the hallucination problem could not be solved by relying solely on the large model itself, necessitating the construction of an agent team. First, conduct data domain governance analysis, fix common problems, and then build a workflow to decompose tasks and match processes. After writing SQL, comparing it with indicators in the indicator domain can greatly solve the hallucination problem in the scenario of querying numbers.

"There has always been a high demand for the scenario of querying numbers. Due to immature technology last year, many projects failed to be promoted after verification due to poor results. With technological progress this year, we are re-communicating with customers," Zhang Ming said. "This may still require a certain project initiation cycle, and it is expected that more landing projects will be seen next year."

There have also been breakthroughs in the scenario of graphic and text generation. In the past, large models often generated garbled text images when generating images with text. Nowadays, when generating promotional posters, agents can first let the model generate textless images and then superimpose text layers to solve the problem.

Currently, the two common general agents of knowledge base Q&A and intelligent question-asking have become basic capabilities and are no longer the primary factors when enterprises select service providers. "Similar to basic assisted driving after the equalization of intelligent driving, it is only used for basic judgment," Du Yanze of IDC told Digital Frontline. Everyone has turned their focus more towards subsequent general agent solutions, such as smart office, contract review, document generation, etc., as well as agent and application solutions in sub-segments of various fields such as R&D, production, supply, marketing, and service.

For example, the field of operation and management is currently the field with the most general scenarios, the highest application maturity, and the highest reproducibility. Du Yanze said that agents in this field are gradually deepening from early general scenarios such as human resources, finance, and customer service, expanding in more refined directions such as compliance review, production and purchase document review, and supply chain optimization.

Especially for agents related to supply chains, the demand is booming. Currently, a large number of enterprises have the need to go global, and supply chain management has become a core scenario that top managers of enterprises focus on.

"Nowadays, manufacturing enterprises generally pay attention to how to use AI to ensure the security and controllability of the supply chain, avoid risks, reduce costs, and improve the efficiency of idle materials, etc. Many top key customers have already communicated with us on this topic," Song Tao told Digital Frontline. Lenovo has combined the latest technology to upgrade its Supply Chain Intelligence Control Tower (SCI), which was launched during its intelligent transformation, to a more powerful "iChain Lenovo Supply Chain Agent".

Currently, this agent has been implemented within Lenovo, supporting the multi-agent collaboration of its global supply chain, from demand forecasting to component delivery, from manufacturing to logistics delivery, improving decision-making efficiency by 30% and shortening workflow cycles by 50%.

In the field of R&D and design, the application of agents is expanding from initial text-to-3D modeling and AI simulation to more stages and deeper scenarios such as design changes, design reviews, and process document generation.

Taking simulation calculations in the equipment industry as an example, previously, simulations such as high-speed rail aerodynamics and motor temperature control relied on foreign industrial software. A single high-speed rail wind tunnel experiment simulation required 128 HPC machines running for 1-2 days, which was costly and time-consuming. Now, through AI simulation and prediction, results can be obtained in just a few seconds with one GPU. "We have collaborated deeply with a large central equipment enterprise, implementing it in multiple factories, and some factories are replicating this achievement," said Zhang Ming.

In the field of manufacturing, IDC observes that the application of agents has initially achieved replicated applications in equipment, safety supervision, processes, and other fields, extending from providing single-point auxiliary information to the entire business process of data query, cause analysis, and report generation. On the other hand, it is also exploring new directions such as factory-level production performance analysis, industrial control programming, and robot intelligent control. For example, China Industrial Internet helped a large leading water enterprise create a factory director assistant agent that can retrieve factory operation data and predict anomalies in real-time.

Predictive maintenance of equipment is a popular scenario. Industrial manufacturing processes have a large number of critical equipment, and once a failure occurs, it may lead to downtime of the entire production line. The emergence of large models and agents has provided a new solution, and many enterprises and manufacturers are focusing on this area. China Industrial Internet helped an electrical equipment manufacturing enterprise create a predictive maintenance agent that can achieve an average equipment fault warning accuracy rate of 92% and reduce maintenance frequency by 40%.

In addition, quality inspection agents, intelligent scheduling agents, and others are also widely explored and have projects implemented. In these scenarios, the collaboration between large and small models is the key path. For example, in quality inspection scenarios, the front-end relies on a small model of machine vision to complete quality inspection, while the subsequent analysis and summary of quality inspection data and logical reasoning are handled by a large model. "Industry definitely uses both large and small models, which we call hybrid AI," said Song Tao.

IDC also stated that these scenarios can be abstracted into two paths: one is the integration of large models with small models, mechanism models, and time series models to provide business optimization through agents; the other is to rely on large models to precipitate the experience of senior technicians and initially verify them in the directions of fault diagnosis, energy conservation, and yield analysis. It is predicted that the penetration rate of AI in industry will reach 25% in five years, of which small models will still account for 70%.

It is worth mentioning that although the depth and breadth of industrial agent applications are expanding, the industry generally believes that there are not many that have good implementation effects and can be replicated in batches. Most are still being explored in the form of projects and have not entered the core deep-water area of manufacturing.

05

What Challenges Remain to Be Solved as the Market Accelerates in the Second Half of the Year?

The industry generally agrees that industrial agents are no longer as simple as "installing software" in the past but represent a dual revolution of system engineering and organizational change. In this process, there are still many challenges to be addressed.

First, there is a shortage of high-quality data, and enterprises need to make up for lost time in advance.

The deeper agents delve into the production environment, the more professional data they require. However, data is currently the biggest challenge in industry implementation. High-quality datasets not only determine the effectiveness of the model but also whether the agent can truly generate value in the business. In industrial scenarios, due to the numerous "dialects" of data, inconsistent formats and quality, as well as issues such as data silos, missing data, noise interference, and fragmented systems, contradictions will only become more prominent.

This year, the National Data Administration Bureau has issued multiple policies to promote the construction of high-quality datasets. Major service providers are also addressing this pain point by providing related tools or collaborative construction models. However, enterprises still need to make their own efforts to make up for shortcomings.

Some enterprises have already started to make up for lost time. The ICONIQ Capital "State of AI 2025" report shows that data has become a major component of enterprise AI budgets, with expenditures on data storage and processing potentially exceeding those on inference and training.

Song Tao from Lenovo also told Digital Frontline that currently, many enterprises that wish to transform with AI but still have doubts are prioritizing data governance and making up for shortcomings in digital transformation before using high-quality data to conduct spot explorations.

Large enterprises are investing even more heavily. Many enterprises have specifically established teams to support data governance and knowledge management within the enterprise, including the updating and correction of knowledge, to ensure the accuracy of agents. Song Tao introduced that Lenovo has established a knowledge management system.

Second, there is an urgent shortage of talents who understand both business and AI and can implement them in engineering.

Although service providers are striving to get closer to the industry, according to industry practice, it may take five to ten years for algorithm experts to thoroughly understand industry details, while industry experts can begin to see results in understanding AI in one to two years. Therefore, whether from the perspective of building competitiveness or ease of friction, enterprises need to build their own talent teams.

"Factories do not need people to train production 'machines' (agents), but they must cultivate people who know how to operate 'machines' (agents)," Zhi Zhen told Digital Frontline. These talents do not necessarily need to fully understand artificial intelligence but must understand how to use artificial intelligence in their own scenarios.

Many industrial enterprises have embarked on actively constructing internal talent pipelines. For instance, in the ninth iteration of Baidu AICA's "Chief AI Architect Training Program," which recently commenced, among the 96 participants from various enterprises, the representation from the energy and heavy industry sectors surpassed half for the first time, demonstrating a keen interest in engaging with the real economy.

Thirdly, the challenge of quantifying Return on Investment (ROI) stands as a pivotal pain point for enterprises investing in AI.

From 2023 to the present, ROI has consistently been a central concern for businesses. As AI deployment scenarios grow more intricate, so do the demands on model capabilities and the costs associated with hardware, software, and deployment. Nevertheless, the benefits derived from these investments are often elusive to quantify. For example, when a task that previously took 10 minutes is now accomplished in 5, how does one measure the value of the additional 5 minutes saved?

Concurrently, AI extends beyond mere IT budgets, displacing roles that involve simple, repetitive tasks, thereby intersecting with HR budgets. Consequently, some enterprises are contemplating that ROI assessments must transcend departmental silos for a holistic perspective. Furthermore, the role of "silicon-based employees" merits reevaluation during top-level strategic planning.

Moreover, issues such as model hallucinations, the inability of current algorithms to meet the real-time and reliability demands of complex industrial settings, the fragmentation of industrial scenarios, difficulties in replication and scaling, the absence of uniform standards, and the nascent state of the ecosystem are all constraints impeding the widespread adoption of industrial agents.

Despite these unresolved challenges, the industry consensus is that the industrial agent market will undergo an accelerated growth spurt in the latter half of the year.

"The pace of implementation this year will undoubtedly surpass expectations," remarked Zhi Zhen. Following the national policy's explicit focus on "industrial agents," an increasing number of enterprises have accelerated their investment timelines. He forecasts that the second half of the year will mark a pivotal shift towards "agent platformization," with enterprises transitioning from experimenting with individual agents to establishing their own "agent ecosystems." They are also poised to advance the "platform as ecosystem" strategy, integrating the closed loop from scenario co-construction, data accumulation, to model services, and expediting the "grounding" of platform capabilities.

Song Tao added that from the second half of this year onward, there will be a collective resolve among manufacturing enterprises to embrace AI for exploration. Through persistent trials and breakthroughs, they will continue to deepen their journey from AI models to AI engineering, ultimately transforming them into revenue-generating "silicon-based productivity."