Rockwell Automation is applying generative AI to a practical industrial problem: how to capture decades of maintenance knowledge and make it immediately usable by technicians. At its Singapore manufacturing site, an AI-powered Maintenance Copilot combines machine documentation, manufacturing data and experienced engineers' troubleshooting knowledge to help technicians identify faults and follow structured corrective actions.
From Manual-Based Troubleshooting to Knowledge-Assisted Maintenance
Traditional machine troubleshooting often depends on a technician finding the correct alarm code in a manual, interpreting the possible causes and then deciding which corrective action to try first. When documentation is insufficient, the next step is usually to ask an experienced technician who has encountered the same failure before.
This process creates two limitations. First, troubleshooting can consume significant time while equipment remains unavailable. Second, much of the most useful information may exist only as informal "tribal knowledge" held by experienced personnel.
Rockwell's Maintenance Copilot changes the interface to this information. Instead of manually searching hundreds of pages, technicians can describe the machine symptom in natural language and receive information organized around the symptom, possible cause and recommended reaction.
This is an important distinction in industrial environments: the value of AI is not simply generating text. Its practical value comes from making existing engineering knowledge easier to retrieve at the moment a maintenance decision is required.
Capturing Decades of Engineering Experience
The Singapore system incorporates troubleshooting knowledge contributed by experienced Rockwell engineers, including personnel with 20 to 30 years of shop-floor experience.
That knowledge can include details that are difficult to capture in conventional manuals: which symptoms tend to appear together, which component failures are common, what a particular alarm usually indicates and how similar problems were resolved on other production lines.
For maintenance organizations, this creates a potential knowledge-retention mechanism.
Experienced technicians eventually retire or move to other roles. Without a structured method of capturing their experience, the organization can lose practical troubleshooting knowledge even when formal documentation remains available.
An AI system cannot replace that experience, but it can provide a searchable interface to knowledge that has already been documented and structured.
A More Structured Troubleshooting Workflow
One of the technically interesting aspects of the system is its organization of maintenance information into symptom, cause and reaction.
This structure resembles the logic already used by experienced maintenance engineers:
- Identify the observable machine symptom.
- Determine the most plausible causes.
- Check the relevant equipment or signal.
- Apply the appropriate corrective action.
- Verify whether the machine has returned to the expected operating condition.
The AI assistant can also expose comments and experiences from other technicians and connect troubleshooting information with existing machine manuals.
From an engineering perspective, this approach is more useful than treating a large language model as an independent diagnostic authority. The model becomes an interface to controlled technical information rather than the sole source of truth.
Reducing Downtime and Maintenance Cost
According to Rockwell's internal tracking cited by Microsoft, the Singapore operation reduced machine downtime by 33% after implementing the maintenance approach. Maintenance and spare-parts costs were reported to have decreased by approximately 25%.
These figures should be understood as Rockwell's reported internal results rather than universal performance benchmarks for AI-based maintenance.
Nevertheless, the mechanism behind the improvement is technically understandable. If technicians can identify probable causes more quickly, unnecessary troubleshooting cycles can be reduced. Faster diagnosis can also reduce the amount of time equipment remains unavailable and help technicians avoid replacing components without sufficient evidence.
The strongest benefit therefore may not come from the AI model itself, but from shortening the path between machine symptom -> relevant knowledge -> engineering action.
AI Can Also Reduce Technician Learning Time
Another application is workforce onboarding.
Rockwell reports that technicians who previously needed approximately nine months to become proficient at troubleshooting hundreds of machines can reach that level in around three months with AI assistance.
This addresses a common challenge in modern manufacturing: equipment complexity is increasing while experienced maintenance personnel are not always available in sufficient numbers.
For a new technician, an AI assistant can provide contextual access to machine documentation and previously recorded troubleshooting experience. This does not eliminate the need for hands-on training, electrical safety knowledge or equipment-specific competency. Instead, it can reduce the amount of time spent simply searching for information.
The distinction is important. AI can accelerate information access; it does not replace physical diagnostic skills.
Connecting AI With the Industrial Data Layer
Rockwell's implementation uses Microsoft Azure technologies, including Azure OpenAI Service, Azure AI Search and Microsoft's security, identity and governance capabilities.
This architecture illustrates an important direction for industrial AI: integrating language models with enterprise and operational information rather than operating them as isolated chat applications.
For maintenance applications, useful data can come from several sources:
- Machine instruction manuals
- Maintenance records
- Alarm and fault information
- Technician troubleshooting notes
- Manufacturing software
- Equipment history
- Production-line observations
The engineering challenge is therefore not simply selecting an AI model. It is establishing reliable data sources, access controls, information retrieval and governance around the model.
AI Should Support Engineers, Not Bypass Engineering Controls
There is a practical limitation that should not be overlooked.
A maintenance recommendation generated by an AI system should not automatically become a machine-control command. Industrial equipment can involve electrical hazards, mechanical energy, process hazards and safety-instrumented functions.
For this reason, AI-assisted troubleshooting is most appropriate as a decision-support layer. The technician remains responsible for validating the physical condition of the equipment and applying established maintenance, lockout/tagout and safety procedures.
For critical control and safety systems, the separation between AI-generated information and deterministic control logic is particularly important.
The AI can help answer:
"What should I investigate?"
It should not independently decide:
"Execute this safety-critical action."
That boundary will be important as manufacturers move from AI-assisted maintenance toward increasingly autonomous operations.
Beyond Maintenance: Quality and Predictive Operations
The Singapore factory is using AI for more than maintenance.
Rockwell also describes a multi-agent quality system designed to detect and address product defects during assembly, together with predictive maintenance capabilities intended to identify conditions that could lead to unplanned equipment downtime.
The combination creates a broader industrial architecture:
Operational data -> AI analysis -> engineering insight -> human or automated action -> feedback
This is more significant than deploying an isolated chatbot. It represents the gradual integration of AI into multiple stages of the manufacturing lifecycle.
The World Economic Forum's June 2026 designation of the Singapore facility as a Global Lighthouse site also cited improvements including a 43% increase in units per person-hour, a 35% reduction in defects and a 67% reduction in time-to-competency.
These results encompass multiple technologies and operational changes, so they should not be attributed solely to the maintenance copilot.
My Engineering View: Data Quality Is the Real Foundation
The most important lesson from this project is not the choice of GPT model.
For industrial AI, the underlying engineering data is usually more important than the conversational interface. If maintenance records are incomplete, machine documentation is outdated or troubleshooting knowledge is poorly structured, a more sophisticated model will not automatically produce better maintenance decisions.
A successful implementation therefore requires several foundations:
- Accurate equipment and asset information
- Consistent fault and maintenance records
- Structured engineering knowledge
- Clear information ownership
- Appropriate cybersecurity controls
- Role-based access
- Human validation of recommendations
- Feedback mechanisms for correcting incorrect information
In other words, AI exposes the quality of an organization's industrial knowledge base. Companies that have spent years maintaining clean equipment data and disciplined maintenance records are likely to have a stronger foundation for industrial AI adoption.
From Individual Expertise to Organizational Knowledge
The deeper value of Rockwell's approach is its attempt to turn individual technician experience into organizational knowledge.
A highly experienced engineer may recognize a failure pattern within minutes because they have encountered it many times. A newer technician may need hours to reach the same conclusion.
If that experience can be documented, indexed and presented at the point of need, the organization becomes less dependent on a small number of individuals.
This does not make experienced engineers less important. In fact, their knowledge becomes more valuable because it can be captured and reused across a much larger technician population.
Scaling From One Factory to Global Operations
Rockwell plans to extend the maintenance copilot to manufacturing locations in Twinsburg, Ohio, as well as Poland and Mexico.
Scaling an industrial AI application from one plant to multiple facilities will introduce additional engineering challenges. Machines, procedures, languages, maintenance practices and historical datasets can differ between sites.
A globally scalable system therefore needs a common architecture while still allowing site-specific knowledge.
This is where cloud-based data, identity, security and AI services can provide an advantage, provided that the underlying operational data remains properly governed.
The Next Stage of Industrial Maintenance
The Singapore deployment illustrates a broader shift in industrial automation.
Earlier generations of digitalization focused heavily on collecting machine data. The next stage is increasingly focused on making that data useful to people who must diagnose, maintain and operate physical equipment.
The most practical industrial AI systems are unlikely to succeed simply because they can generate convincing answers. They will succeed when they can connect real machine conditions with trusted engineering information and an appropriate human workflow.
For maintenance teams, that could mean less time searching manuals, faster access to proven troubleshooting procedures and a more consistent approach to recurring faults.
The longer-term opportunity is even broader: preserving engineering knowledge, accelerating workforce development and using AI as an additional information layer around existing PLC, DCS, SCADA, MES and maintenance systems.
