At this year's WAIC (World Artificial Intelligence Conference), industrial AI agents have become one of the main protagonists. Almost all enterprises involved in industrial AI are showcasing agent-related products. However, compared to the imagination of AI agents in the consumer sector, industrial enterprises are more concerned with one question: Can AI really solve long-standing problems on the production floor and bring substantial returns? After all, industrial customers face complex equipment, complex processes, and complex workflows. Any new technology entering the production system must be validated by its return on investment.
On the eve of WAIC, Gongkong.com had in-depth exchanges with Siemens, Schneider Electric, Supcon, and Texas Instruments (TI). Starting from the actual implementation effects, we calculated the financials of AI agent implementation together and explored the true industry anxieties in their eyes.
Finding the Entry Point First: Exploring the Implementation Focus Determined by Corporate DNA Through Products
At this year's WAIC, the display of industrial AI agents has shown clear differentiation, with companies leveraging their own experience and technical accumulation to find the most familiar and advantageous industrial segments.
International manufacturers rely more on existing industrial software and engineering systems to transform years of accumulated industrial knowledge into agent capabilities.
▶ Siemens Eigen Engineering Agent is among the world's first officially commercialized AI systems capable of autonomously planning and executing industrial automation engineering tasks. It can understand project requirements, write PLC (Programmable Logic Controller) code, and complete hardware network configuration, addressing the efficiency of automation engineering itself. In discussions with Gongkong.com, Siemens noted that industrial AI agents should prioritize entering high-value, replicable scenarios rather than seeking applications based solely on technical capabilities. It focuses on the entire industrial value chain—from product design, production planning, and engineering implementation to production operations and services. Every link presents opportunities for intelligence, provided the scenario can generate clear commercial value and meets the conditions for scaled promotion.
▶ Schneider Electric focuses more on the implementation of specific industrial application scenarios, such as engineering development, technical support, and operational optimization. In its view, current industrial AI agents are more suited to first solve highly repetitive, knowledge-intensive tasks, such as assisting engineers with configuration and commissioning, and participating in operational links like equipment management. As data integration progresses, agents will gradually engage in more complex operational decision-making. At this WAIC, Schneider Electric started from the equipment and energy ends. Its native power management agent connects the entire chain from equipment access to intelligent diagnosis, emphasizing a shift for enterprises from passive response to proactive maintenance. The difference in paths between the two international giants lies in the fact that Siemens anchors on engineering efficiency, while Schneider anchors on data standardization, scenarios, and the closed loop of data application.
Domestic manufacturers have more diverse approaches, which also reveal a relatively obvious commonality behind them: a stronger emphasis on rapid implementation combined with specific industrial scenarios.
▶ Supcon believes that the application center of industrial AI agents is anchored in core production links such as prediction and early warning of production operation parameters, equipment health status monitoring, intelligent control, operation optimization, and carbon and energy optimization. Rooted in the process industry for many years, Supcon takes industrial time-series data as the cornerstone and has taken the lead in launching the world's first trusted large model deeply integrating first principles—the Time-Series Foundation Model TPT. Based on the MoE (Mixture of Experts) architecture and directionally fine-tuned with actual plant data, this model combines reliability, trustworthiness, and engineering applicability. Building on this foundation, Supcon has constructed an industrial AI agent generation platform—the Supcon Time-Series Foundation Model Platform TPT. Relying on the five core capabilities of SCOPE, this platform supports engineers in building plant-specific agents with zero code and a low threshold, achieving a full-process closed loop of "identification-evaluation-decision-execution". By applying their own production data, enterprises can quickly adapt to the plant, achieving low-cost generation of agents and coordinated deployment with industrial control systems, driving intelligent operation and autonomous optimization across all scenarios in the process industry.
COSMOPlat chooses to cut in from industrial world models and platform capabilities, creating the Tianyu Industrial Flexible Operation Agent Development Platform. Combined with "digital humans", it lays out an agent product matrix covering task decomposition, code generation, simulation verification, and production execution, with the goal of building intelligent capabilities for the entire manufacturing process. Black Lake Industrial Agent adopts a more vertical path, focusing on the high-frequency business scenario of order management, covering processes such as order splitting, quotation, scheduling, quality inspection, and fulfillment, helping enterprises cope with production management challenges brought by multi-variety, small-batch, and short-delivery orders. It can be seen that domestic manufacturers focus on finding AI value landing points in industrial data, business processes, and specific scenarios.
On the other hand, the entry logic of Baosight Software's AIC integrated software lies in directly connecting the equipment control layer, focusing on incubating AI chief operator agents for each process. These agents can take over the equipment chief operator work originally undertaken by manual labor in the routine production of complex processes. Currently, they have achieved autonomous operation in multiple key processes of steel cold rolling and have expanded to the chemical and mining fields. This is the logic of sinking from the management layer to the execution layer, which is different from making platform tools or strengthening business systems.
Behind these product routes lie their respective distinct judgments on the implementation points of AI agents.
▶ Texas Instruments (TI), based on underlying chip technology, proposes from the perspective of industrial system evolution that intelligence will first land in local optimization scenarios. Highly repetitive applications such as spot welding and material handling primarily pursue high speed, high efficiency, and high repeatability accuracy, which is essentially an enhancement of the capabilities of existing automation systems. In contrast, core production links involve critical functions such as motor drive, power control, and safety interlocks, which have extremely high requirements for system determinism, real-time performance, and functional safety levels, and currently still rely mainly on traditional real-time control architectures.
Behind each of the above routes is a strategic choice determined by corporate DNA. Those making platform tools want to secure their position through ecosystems; those focusing on vertical scenarios bind the industry with data; those working on the control layer tackle the hardest-to-replicate core barriers; and those doing whole-plant solutions cultivate customer stickiness with integrated solutions.
Seeking the Foundation Again: What is Missing for the Universal Era of AI Agents?
For industrial AI agents to truly enter scaled application, the difficulty lies not only in model capabilities. Agents in the consumer sector can spread rapidly relying on broader user demands, but industrial scenarios must face complex equipment, professional knowledge, and stringent reliability requirements. From pilot to universal application, there is still a gap to bridge in terms of making up for industrial knowledge, system architecture, and engineering capabilities.
The judgments of the interviewed enterprises almost revolve around the same core logic: for industrial AI agents to truly enter scaled application, the difficulty lies in whether the data, knowledge, and tool systems accumulated on the industrial floor over the long term can be transformed into capabilities that agents can understand and execute.
▶ Siemens believes that the biggest challenge in industrial scenarios lies in the complexity of industrial data. Unlike the text-dominated data in the Internet sector, the industrial floor simultaneously hosts various types of information such as 3D product models, simulation results, control signals, and sensor time-series data. How to integrate these multimodal data and establish correlations between them is the foundation for agents to understand the industrial world.
This is also why general large models struggle to directly solve industrial problems. Siemens is promoting the construction of industrial foundation models, including proprietary models trained on industrial time-series data. Through years of accumulated industrial data and domain knowledge, the models are equipped with the ability to understand equipment status, process workflows, and engineering logic. At the same time, industrial AI agents also need to be combined with software tools and hardware systems to complete truly complex industrial tasks. Siemens believes that future competition among industrial AI agents will not depend solely on model capabilities, but on the comprehensive capabilities of multimodal models, industrial foundation models, industrial knowledge accumulation, and software and hardware toolchains.
▶ Schneider Electric believes that a problem currently faced by enterprises is data governance. To support AI implementation, the key lies in relying on open architectures to build a data foundation that is high-quality, highly timely, fully covered, highly semantic, and secure and compliant, promoting the coordinated upgrade of standards first, scenario-driven, platform-supported, and organizational guarantee. In specific practice, enterprises need to utilize standardized data platforms and build an AI ontology architecture for industrial scenarios based on data insights, further integrating industry knowledge bases, knowledge graphs, and the capabilities of both large and small models.
▶ Supcon focuses on the deep understanding capability of agents regarding industrial production processes. The process industry has long faced multiple bottlenecks: long process R&D cycles, slow project construction progress, difficulty in leaping production efficiency, high labor input, and high material and energy consumption. If agents are to truly enter the core production links, they must first profoundly understand complex industrial processes, which depends on the capabilities of the underlying foundation models. However, although general AI can understand text and knowledge information, it is difficult to understand the implied working condition evolution such as the temperature curve of a distillation column in the process industry, nor can it truly master the operation logic in DCS (Distributed Control System) systems. Therefore, industrial AI agents must rely on industrial time-series data to train specialized models, enabling AI to possess the core capabilities of perceiving production processes and identifying operation patterns.
The core agent business of Supcon takes the Time-Series Foundation Model Platform TPT as the hub, providing industrial customers with full-lifecycle services for agent generation, deployment, and operation. For example, its production prediction and early warning Agent achieved a prediction accuracy of 99.79% on a petrochemical ethylene cracking unit, real-time monitoring production parameters and equipment status, accurately locating anomalies and generating disposal plans; the intelligent control Agent can link over 15,000 monitoring points on the production line at the Hubei Xingrui Factory of Xingfa Group, autonomously identifying risk points such as raw material composition fluctuations and equipment operation anomalies. The system response speed is 10 times faster than manual operation, and the AI reliability exceeds 98%.
From the perspective of more underlying system architecture, ▶ Texas Instruments believes that the development of industrial intelligence also faces the issue of control architecture evolution. Traditional automation systems rely on centralized control, but with the increase in the number of sensors and actuators, future industrial systems need to gradually evolve towards a more distributed architecture, sinking some computing and decision-making capabilities to the edge. At the same time, a large number of tasks on the industrial floor have extremely high requirements for real-time performance and determinism, and intelligent capabilities must be deeply integrated with real-time control systems, which is also an important challenge distinguishing industrial AI agents from ordinary AI applications.
It can be seen that for industrial AI agents to achieve scale, it is necessary to reconnect industrial data, software tools, control systems, and industry knowledge. This tests not only model capabilities but also the industrial knowledge, software platforms, and field experience accumulated by enterprises over the long term. For industrial enterprises, truly valuable agents ultimately need to become part of the production system, rather than remaining peripheral auxiliary tools.
Revisiting the Business Model: When Will Industrial AI Agents Enter the Harvest Period?
Whether industrial AI agents can scale ultimately comes back to a financial calculation. For industrial customers, whether the efficiency improvement brought by AI is sufficient to cover the input costs is more important than the technical capabilities themselves.
From current industry explorations, the commercialization of agents first depends on whether they can be standardized and replicated.
▶ Siemens judges that the engineering field is easier to form a business model because its task boundaries are relatively clear, and relying on mature industrial software systems such as TIA Portal, experience can be precipitated into standardized capabilities. Therefore, the Eigen Engineering Agent is delivered using a SaaS subscription model, where customers can purchase licenses according to their needs without redeveloping for each project. Siemens revealed that in global customer trials, the agent can achieve up to a 50% efficiency improvement in some scenarios and reach a return on investment balance within months.
However, after entering the production operation links, the commercialization logic becomes more complex. Although scenarios such as equipment management, process optimization, and energy management have higher value, they are difficult to simply replicate due to the involvement of different factories, different processes, and different operating conditions.
▶ Schneider Electric believes that for industrial AI to truly enter the harvest period, three conditions must be met: moving from single-point pilots to replication in similar scenarios, transitioning from auxiliary tools to the combination of business processes and data decision-making, and forming quantifiable return on investment.
Also exploring this "from validation to replication" path is Supcon. In the process industry field,
▶ Supcon's TPT platform is currently open for free trial to industrial users. For formal commercialization, it adopts an annual fee subscription system calculated by the number of plant sets, with contracts mainly for 3 to 5 years and payment before use. In the long run, Supcon prefers the subscription model, shifting revenue from one-time project-based to continuous operational service charges, which is conducive to establishing long-term and stable customer relationships while creating greater value for customers through continuous model optimization and service upgrades. To accelerate this process, Supcon relies on the TPT CAMP co-creation training camp to build an agile transformation path of "trial before subscription"—after customers verify TPT capabilities with real data on machines, user reports and implementation suggestions are directly output, seamlessly connecting to subscription signing, greatly reducing the decision-making threshold, and effectively promoting the scaled replication of industrial AI subscription services among core customer groups.
From these explorations, it can be seen that industrial AI agents are still some distance away from comprehensive popularization, but the direction of commercialization has gradually become clear. In the short term, scenarios with higher standardization such as engineering design and technical support are easier to form a closed loop; whereas high-value scenarios such as production control need to find a balance among safety, reliability, and long-term returns.
Epilogue: Industrial AI Agents Enjoy High Popularity and Have No Need for Anxiety
The popularity of industrial AI agents at this year's WAIC significantly exceeded that of last year. But behind this hustle and bustle, we also feel that the industry is not without anxiety.
The anxiety comes from two aspects. On the one hand, enterprises worry whether the application of large models will bring high computing power costs; on the other hand, facing the complex equipment, processes, and workflows in the manufacturing industry, how can agents avoid falling into one-time project development and truly achieve scaled application.
Regarding Token consumption, the judgments of several enterprises are basically consistent. Industrial AI is still in its early stages, local deployment is the mainstream, and computing infrastructure is mostly existing resources, so there is no issue of burning through Tokens. Siemens put it directly: instead of worrying about Token prices, it is better to think about how to turn agents from customized projects into replicable products. Supcon's approach is to cooperate with Inspur Information to launch an intelligent all-in-one machine, lowering the deployment threshold, and supporting the industrial implementation of TPT with faster inference speeds and lower Token costs.
The real difficulty lies in the fragmentation of industrial scenarios. Each factory has different equipment, different processes, and different workflows. A set of agents might struggle to adapt when moved to a different location. Siemens' approach is to dive into the scenarios from the product definition stage, thoroughly understand them before abstracting, rather than holding onto technical capabilities to find applications.
WAIC has showcased the aspect of industrial AI agents accelerating towards the industry, but beyond the popularity, the industry is still in the exploration period. How to make AI truly understand and serve the industry, and help enterprises calculate the financials first, will still be a question that enterprises need to answer together in the future.