Staff Reporter |
Editor | Wen Shuqi
The competitive focus of industrial large models is now stretching beyond raw model capability into the nitty-gritty of engineering deployment.
In the early days, companies mostly cared about whether a large model could answer questions or handle a single, isolated task. But as AI gradually works its way into equipment management, energy management, production analytics, and other operational areas, the questions companies ask have started to shift. AInnovation told us that some of its top-tier manufacturing clients are moving agent development from one-off pilots to replication across core business processes. Their concern is no longer just “does this feature exist?” — it’s about build cost, operational reliability, and the ability to scale.
On August 27, AInnovation held an industrial agent technology launch event in Beijing to upgrade its industrial ontology intelligence platform. The update brings three new capabilities: automated ontology construction, automated workflow construction, and automated evaluation and optimization. The company also rolled out three new agent suites: AEAM v2.0 for equipment management, AEMS v2.0 for energy management, and AMI v3.0 for manufacturing data analytics.
AInnovation is an AI company focused on “AI + Manufacturing.” In the first half of 2026, the company posted revenue of RMB 829 million, up 18.6% year over year, with its “AI + Manufacturing” business accounting for 81.5% of total revenue. The company currently focuses on 8 manufacturing sub-sectors and has served more than 1,800 enterprise customers to date.
The industrial agents highlighted at this launch can be understood as AI systems that combine enterprise data, business rules, and software tools to carry out analysis, decision-making, or hands-on execution tasks.
Take equipment failure as an example. When a production machine goes down, the industrial agent needs to pull up real-time equipment data, maintenance records, operating procedures, and spare-parts inventory all at once, then reach a conclusion based on the company’s own repair rules. Sometimes, the system also has to generate a repair plan, create a work order, and even trigger a spare-parts request.
This means that for a large model to truly enter the production environment, it can’t just have general-purpose language understanding and generation capabilities — it also needs to grasp the business logic unique to each enterprise.
To tackle these challenges, AInnovation has built an “Ontology + Harness Engineering + AI Coding” development approach into its industrial agent platform.
Ontology is mainly about solving the “helping agents understand the business” problem.
Enterprise data is usually scattered across different systems, and there are tons of business relationships between data points that only the company itself truly knows. For example, which production line a piece of equipment belongs to, what production task it handles, and which orders, process requirements, and maintenance records are tied to it. These relationships may be everyday common sense for employees, but they don’t naturally show up in a general model’s training data.
Li Fan, CTO of AInnovation, told us, “An ontology should be understood as an ‘asset’ rather than a runnable system — it’s an explicitly expressed asset.”
An ontology can structurally represent objects like equipment, personnel, orders, materials, energy consumption, and processes, as well as the connections and business rules between them. This lets agents not only know what data a company “has,” but also understand how that data relates to each other in real business scenarios.
Understanding the business is just step one. To scale up industrial agent adoption, you also need to bring development costs down.
In the past, building an industrial agent often required endless back-and-forth between business staff and developers, followed by developers manually configuring nodes, data, tool calls, and processing rules. The more complex the business, the longer the workflow that had to be assembled. A key focus of this AInnovation upgrade is getting AI more involved in developing the agents themselves.
Previously, attributes and relationships in an ontology were mostly sorted out manually. AInnovation, however, has adopted an “AI + FDE (Forward Deployed Engineer)” model, where AI handles data analysis, relationship reasoning, and first-draft generation, while FDEs work with business staff to confirm business boundaries and final relationships.
The second change is in workflow development.
In a live equipment-repair demo at the launch, a business user typed in a natural-language request with equipment codes and fault descriptions, asking the system to look up relevant data, run diagnostics, and generate a report. The platform first turned that sentence into a more complete requirement spec, and then produced a working equipment fault-diagnosis workflow.
Once a workflow is generated, you still need to judge whether it actually works.
The third new capability added to the platform is automated evaluation and optimization. The system can run workflows on test data, generate an evaluation report based on the execution process and results, then produce optimization suggestions. After human confirmation, the workflow is adjusted and re-validated.
This creates a closed loop for agent development: build, evaluate, and optimize. It also clarifies two other concepts. Ontology handles whether the agent truly understands the business; AI Coding handles whether development and iteration costs can be reduced; and Harness Engineering handles whether the system can run stably once it enters the production environment. These are the engineering challenges that industrial agents must overcome on the path from development to production systems.
The shift in how industrial agents are developed also reflects a shift in what enterprises actually need.
At the 2026 Global Digital Economy Conference’s Industrial Agent Development Forum held in July, the organizers released a list of enterprise demand scenarios for industrial agents. Fourteen requirements from 11 industry-leading companies covered product R&D, smart manufacturing, equipment operations and maintenance, warehouse and logistics, and production safety — and all of them required quantifiable quality-and-efficiency improvement metrics.
Policy requirements for industrial agents are also getting more specific. In May this year, Beijing released the “Beijing Implementation Plan for AI-Empowered High-Quality Development of Industrial Internet (2026–2028),” which calls for building industrial agents with high reliability, real-time responsiveness, and strong security for industrial scenarios, while making their development, deployment, runtime monitoring, behavior auditing, and tuning iterations visual, controllable, and traceable.
These demands echo the product evolution directions of other enterprise software vendors. For example, Siemens, the industrial technology giant, launched its industrial automation engineering agent, Eigen Engineering Agent, in April 2026. The agent can plan and execute automation engineering tasks based on specific project environments and verify output results against the customer’s established standards.
The competitive dimensions of industrial AI are changing. Models still define the ceiling for what an agent can accomplish, but in a production environment, companies also need to deal with questions like: Is the data accurate? Are permissions controlled? Can operations be traced? Can results be verified? And what happens when the system hits an anomaly?
At the same time, industrial agents are moving deeper into the enterprise’s core business processes.
The three industrial agent suites AInnovation released this time target equipment management, energy management, and manufacturing data analytics scenarios. Take equipment management: AEAM v2.0 includes nine types of agents covering fault prediction, fault troubleshooting, SOP generation, repair assistance diagnostics, spare-parts identification, and decision analysis — spanning risk identification, fault diagnosis, repair handling, and spare-parts assurance.
This means industrial agents are extending from single-point functions into more granular, deeper business processes. In the past, a company might just need an agent that could look up equipment info or answer repair questions. Now, agents need to understand the relationships between equipment, production lines, orders, processes, and spare parts — and collaborate across different stages to get things done.
In its 2026 manufacturing trend forecast, International Data Corporation (IDC) predicts that AI in China’s manufacturing sector is moving from single-point automation to system-level autonomy. IDC projects that by 2028, 65% of China’s top manufacturers will integrate AI agents into their design and simulation tools; by 2030, 70% of China’s leading manufacturers will use AI agents to build data models and manage hybrid cloud workloads.
As industrial agents move into production environments, data governance, engineering capability, and delivery efficiency are becoming the next set of variables that determine how far they can scale.
