Artificial intelligence is becoming increasingly visible on the manufacturing floor, from predictive maintenance and machine vision to production analytics and AI-assisted troubleshooting. Yet the real question for plant managers is no longer whether AI can be used in manufacturing. The more practical question is whether it can solve a clearly defined operational problem and deliver measurable improvement without disrupting existing control systems.
AI Is Expanding the Role of Automation
Traditional industrial automation is designed to execute predefined logic consistently. PLCs, DCS platforms, SCADA systems, drives, safety systems, and instrumentation remain responsible for controlling physical processes according to established rules.
AI introduces another layer. Instead of simply executing a programmed sequence, an AI system can analyze large amounts of historical and real-time information and identify relationships that may be difficult for a person to detect manually.
For example, production data can be combined with vibration measurements, motor current, temperature, maintenance records, alarm history, and operating conditions to identify patterns associated with equipment degradation.
From an engineering perspective, this distinction matters. AI should not be viewed as a replacement for the control layer. In most plants, its immediate value is more likely to come from supporting decisions around the control system rather than replacing deterministic control logic.
The Knowledge Gap May Be More Important Than the Labor Gap
Many manufacturing organizations have accumulated decades of practical knowledge through experienced operators, technicians, maintenance personnel, and control engineers. Much of that knowledge, however, exists outside formal databases.
It may be contained in maintenance notes, handwritten records, troubleshooting procedures, old drawings, operator habits, engineering documents, or simply in the memory of a technician who has worked on the same machine for 25 years.
As experienced personnel retire, this creates a significant knowledge-transfer problem.
Industrial AI can potentially connect machine data with manuals, standard operating procedures, maintenance history, alarm records, and other documentation. A less-experienced technician can then receive more context when investigating an abnormal condition.
This does not eliminate the need for experienced engineers. Instead, it can make their knowledge more accessible across shifts and production areas.
In my view, this is one of the strongest near-term applications of AI in manufacturing: preserving and distributing operational knowledge rather than attempting to automate every human decision.
The Best AI Projects Start With a Plant Problem
A common mistake is beginning with the technology.
A company selects an AI platform, develops a pilot, connects some data sources, and only afterward asks what business result the system is supposed to deliver.
The engineering approach should be reversed.
Start with a measurable problem:
- Which assets cause the most unplanned downtime?
- Which process deviations generate the highest scrap cost?
- Where do technicians spend excessive time troubleshooting?
- Which alarms repeatedly require senior engineering support?
- Which manual inspection activities consume significant labor?
- Where is energy consumption higher than expected?
- Which production parameters are difficult to optimize using conventional methods?
Once the problem is defined, AI becomes one possible engineering solution rather than the objective itself.
A useful project should ultimately be connected to plant metrics such as downtime, throughput, yield, scrap, maintenance cost, quality, labor utilization, or energy consumption.
Pilot Projects Are Not the Same as Production Value
Manufacturers have conducted numerous AI experiments, but a successful demonstration does not automatically translate into a successful production system.
A pilot may correctly identify a machine anomaly while operating on a limited dataset under controlled conditions. The production environment is different. Sensors may fail, instruments may be recalibrated, PLC logic may change, production recipes may be modified, network connections may become unavailable, and operating personnel may respond differently from the assumptions used during development.
This is where many AI initiatives encounter difficulty.
The real test is not whether an AI model works in a demonstration. The real test is whether it continues to provide useful information after being integrated into the plant's normal operating procedures.
Existing Automation Infrastructure Should Not Be Ignored
Manufacturing plants rarely have the luxury of starting from a clean architecture.
A typical facility may contain multiple generations of PLCs, DCS controllers, HMIs, SCADA systems, historians, remote I/O, variable frequency drives, condition-monitoring equipment, industrial Ethernet networks, and legacy serial devices.
Replacing this infrastructure simply to introduce AI is usually impractical.
A retrofit strategy is therefore more realistic. Data can be collected from existing historians, OPC interfaces, PLCs, condition-monitoring systems, databases, and other available sources while keeping the primary control architecture intact.
This approach also reduces the risk of introducing unnecessary changes into validated production systems.
The objective should be to add intelligence around proven automation infrastructure, not rebuild the entire plant because a new technology has appeared.
Data Quality Determines the Quality of AI
AI cannot compensate for poor industrial data.
If instrument tags are incorrectly identified, timestamps are inconsistent, sensor values contain gaps, equipment states are undocumented, or historical data lacks sufficient context, the resulting model may produce technically impressive but operationally misleading conclusions.
This problem is often underestimated because manufacturing data is generated continuously. Large quantities of data do not necessarily mean high-quality data.
Before implementing an AI application, engineers should understand:
- Where the data originates
- How sensors and instruments are calibrated
- Whether timestamps are synchronized
- How equipment states are identified
- How missing values are handled
- Whether historical maintenance events are documented accurately
- Which personnel or systems own the data
- How changes to PLC and DCS configurations affect data meaning
A simple, well-documented dataset can be more valuable than millions of poorly contextualized records.
Safety Must Remain Separate From AI Experimentation
The consequences of an incorrect AI recommendation depend heavily on the application.
Using AI to summarize maintenance records is fundamentally different from allowing an AI-generated decision to directly influence a safety-related shutdown.
In process industries, chemical plants, refineries, power facilities, and other high-consequence environments, deterministic safety functions must continue to operate according to established engineering principles and applicable safety requirements.
AI can assist with monitoring, diagnostics, anomaly detection, and decision support, but that does not mean it should automatically be given authority over safety functions.
A practical architecture should clearly define what AI is allowed to observe, what it is allowed to recommend, and what actions must remain under conventional control or safety logic.
Plant Managers Do Not Need to Become Data Scientists
Another misconception is that plant managers must suddenly become experts in machine learning, large language models, and AI agents.
They do need to understand what these technologies can and cannot do, but their most valuable expertise remains operational.
A plant manager understands production constraints, maintenance priorities, workforce capabilities, quality requirements, safety expectations, and the financial consequences of downtime.
That knowledge is necessary to determine whether an AI application is solving a meaningful problem.
The strongest implementation teams therefore combine several disciplines: operations, maintenance, controls engineering, IT/OT networking, data engineering, cybersecurity, and AI specialists.
AI projects are much more likely to succeed when the people building the system understand how the plant actually operates.
ROI Should Be Measured Before Scaling
An AI project should have a baseline before deployment.
For example, if predictive maintenance is intended to reduce downtime, the organization should establish the historical frequency, duration, and cost of relevant failures. If AI is being used for quality inspection, the existing defect rate and inspection workload should be documented.
After implementation, the same metrics can be compared against the baseline.
This makes the investment decision much clearer.
A project that reduces troubleshooting time by 30 minutes per event may be valuable in one plant and insignificant in another. Likewise, reducing a small percentage of scrap may represent a major financial improvement on a high-volume production line.
The important point is that AI ROI is plant-specific and should be calculated from operational outcomes, not the sophistication of the technology.
Human Oversight Will Remain Important
Manufacturing processes operate under physical constraints that cannot always be represented perfectly in historical data.
An experienced engineer may recognize that an unusual sensor value is caused by an instrument problem rather than a process problem. A technician may know that a particular machine behaves differently during startup. An operator may understand a production condition that is not properly documented in the database.
AI can identify patterns, but humans provide operational context.
For this reason, I expect the most effective manufacturing systems to follow a human-in-the-loop model. AI identifies anomalies, organizes information, recommends possible causes, or prioritizes actions, while qualified personnel remain responsible for decisions with operational consequences.
The Next Stage Is Integration, Not Just Automation
The manufacturing AI discussion is gradually moving beyond individual pilot projects.
The next challenge is integration: connecting AI capabilities with existing automation, maintenance workflows, engineering documentation, historians, asset-management systems, and operational procedures.
That integration will require more than purchasing an AI platform. Plants will need clearly defined data ownership, standardized information, cybersecurity controls, documented workflows, and people who understand both industrial systems and modern software.
The manufacturers that benefit most from AI are unlikely to be those that deploy the largest number of models. They will be the organizations that identify specific operational problems, connect reliable industrial data to those problems, validate results with experienced personnel, and scale only after measurable value has been demonstrated.
