AI helps manufacturing when it solves specific problems better than simpler alternatives, such as managing complex process control or optimizing numerous interacting variables. The key principle is matching the appropriate technology to the actual problem rather than implementing AI for its own sake.
- AI has been used in manufacturing for over 30 years through neural networks, fuzzy logic, and expert systems—not just recent generative AI tools
- AI becomes valuable when problems are too complex for conventional control, such as managing multi-zone dryers with interacting variables and delayed feedback
- Optimization applications like managing 26 compressors across multiple headers can reduce energy costs by significant amounts and pay for themselves in under six months
- Generative AI excels at helping operators and engineers understand large amounts of historical data, maintenance records, and production information—not replacing human decision-making
- Foundational manufacturing practices like reliable instrumentation, quality data, and operator training remain essential; AI cannot compensate for poor data or inadequate process understanding
(Other articles in the Digital Transformation Series: Part 1, Part 2, Part 3, Part 4)
I’ve been implementing artificial intelligence in manufacturing since the early 1990s. Back then, nobody was talking about ChatGPT, generative AI, or large language models. We were using artificial neural networks (ANNs), fuzzy logic, expert systems, advanced process control, and other techniques to solve specific manufacturing problems.
That is why I find today’s conversation about AI in manufacturing both exciting and, occasionally, a little frustrating.
AI didn’t suddenly arrive on the plant floor—it’s been here for decades.
What’s changed is the accessibility of the technology, the amount of data available, and the emergence of new tools such as generative AI. Those developments are opening tremendous new possibilities. But they haven’t changed the most important question manufacturers should ask: What is the simplest technology that will reliably solve the problem?
Sometimes the answer really is AI.
AI has been here all along
Long before today’s generative AI boom, manufacturers were using advanced computational techniques to solve problems that were difficult to address through conventional control alone. ANNs could model complex relationships among process variables. Fuzzy logic could manage conditions that didn’t fit neatly into simple true-or-false rules. Expert systems captured knowledge and applied it to problems such as troubleshooting and production scheduling. Advanced process control and model predictive control (MPC) addressed interactions, constraints, and process dynamics that conventional control strategies struggled to manage.
Dr. Bryan Griffen is the President of Griffen Executive Solutions LLC. He was previously Senior Director of Industry Services for PMMI: The Association for Packaging and Processing Technologies, and he held a number of roles for Nestlé during his many years there.Griffen Executive Solutions
Matching the technology to the problem also means recognizing when advanced technology isn’t needed. Conventional automation remains remarkably good at what it was designed to do. If a PID loop can reliably maintain a process variable, there’s little reason to replace it with a more complex AI solution. If deterministic PLC logic can reliably sequence a process, adding AI may create complexity without adding value.
AI begins to earn its place when the problem itself becomes difficult enough to justify it.
When complexity justifies AI
One application I worked on involved closed-loop control of a large, multi-zone industrial dryer. The product needed to reach a specified final moisture level without exceeding limits on temperature exposure. That sounds straightforward until you consider what the control system actually had to manage.
Incoming product moisture varied with environmental and growing conditions. Product moisture could only be measured offline and then entered into the system. The dryer was more than 100 ft long—creating considerable dwell time—and each drying zone interacted with the others because they were open to the surrounding atmosphere and to one another. Changes made in one part of the process could take considerable time before their effects became apparent elsewhere.
In other words, we had multiple interacting variables, nonlinear relationships, significant process delays, changing incoming conditions, and an important product measurement that was not continuously available.
This was exactly the kind of problem where an artificial neural network made sense.
The ANN became part of the closed-loop control strategy, allowing the system to account for complex relationships among changing process conditions rather than simply reacting to individual measurements. The result was a stable drying process capable of maintaining consistent operation despite environmental changes and variation in the incoming product.
The important lesson was not that an ANN could control a dryer. It was why we chose one. The manufacturing problem justified the technology.
Optimization where humans cannot chase every variable
Some AI applications become valuable not because a process is impossible to control conventionally, but because there are too many interacting choices to optimize continuously.
Industrial refrigeration provides a good example.
In another application, I worked with a refrigeration system containing 26 ammonia compressors serving four temperature headers. Refrigeration demand changed constantly based on environmental conditions, production requirements, and load from refrigerated storage. The challenge wasn’t simply maintaining temperature, it was meeting that changing demand while operating the compressor array efficiently.
Compressors don’t operate at the same efficiency under every loading condition. The objective was to keep operating compressors in efficient ranges and bring additional swing compressors online as refrigeration demand required them. With 26 compressors, four headers, and constantly changing system demand, the optimization problem was highly dynamic.
An ANN-based control strategy helped manage those changing conditions. The system stabilized the temperature headers while significantly reducing energy consumption.
The project paid for itself in less than six months.
That’s the kind of AI application that matters on the plant floor. It wasn’t interesting because an ANN was involved, it was interesting because the technology solved a difficult operational problem and produced a measurable financial return.
A different kind of AI
Today’s generative AI and large language models bring a different capability to manufacturing.
The ANN controlling an industrial process and an LLM analyzing production information may both fall under the broad umbrella of artificial intelligence, but they aren’t doing the same job.
LLMs are particularly powerful at helping people work with large amounts of information. A modern manufacturing operation may contain years of historian data, maintenance records, quality information, production schedules, operator notes, alarm histories, MES records, and other operational knowledge. Finding useful relationships across all that information can require significant time and expertise.
This is where generative AI becomes interesting.
Imagine an operator or process engineer asking why a particular batch is taking longer than normal. Rather than manually searching trends, maintenance history, previous batches, and production records, an AI-enabled analytics system could help identify what is different about the current run, highlight unusual relationships, and point the user toward conditions worth investigating.
The AI is not necessarily controlling the process. It is helping someone understand the process.
That can be enormously valuable. An experienced operator may recognize relationships almost instinctively after years of working with a process. A newer operator may not. Giving people better access to operational knowledge and helping them interpret what they are seeing can shorten the path from data to understanding.
In this role, AI helps steer the operator rather than replace the operator.
Do not use AI when you do not need it
The current enthusiasm surrounding AI creates a predictable temptation: Once an organization decides AI is strategically important, teams begin looking for places to use it.
That’s backward. Instead, start with a manufacturing problem, and then choose the appropriate technology.
Sometimes that will be an ANN, machine-learning model, optimization algorithm, or generative AI application. Other times it’ll be a PID loop, a better sensor, improved PLC logic, or simply presenting existing information more clearly to an operator.
More sophistication doesn’t automatically mean better.
The same caution applies when considering how much authority to give newer AI technologies. An LLM may be extremely useful for analyzing operating information, identifying possible relationships, and recommending areas for investigation. That doesn’t mean it should directly manipulate a critical process variable. Closed-loop process control requires predictable, bounded, and appropriately validated behavior. Established control technologies remain better suited to many of those applications.
AI should be given the job it’s good at, not every job we can find for it.
The foundation still matters
None of these technologies eliminate the fundamentals of good manufacturing.
An advanced model cannot compensate for unreliable instrumentation, poorly calibrated sensors, inconsistent data structures, or missing process context. Nor can sophisticated analytics compensate for a workforce that doesn’t understand the underlying process well enough to recognize when a recommendation doesn’t make sense.
AI doesn’t fix bad data, it learns from it.
That makes the foundational work discussed throughout this series even more important. Connectivity, contextualized information, well-designed operator interfaces, reliable measurements, and disciplined operating practices create the environment in which advanced technologies can deliver value.
The human element remains equally important. Operators, engineers, maintenance technicians, and supervisors need to understand what the system is telling them, when to trust it, and when to question it.
The goal isn’t to remove people from manufacturing decisions, it’s to give them better tools for making those decisions.
When AI actually helps
Artificial intelligence will undoubtedly play a growing role in manufacturing. Better computing, richer data, improved algorithms, and generative AI are creating opportunities that would’ve been difficult to imagine when I began applying these technologies more than 30 years ago.
But the fundamental principle has not changed.
AI helps when it solves a problem better than the alternatives. It helps when it recognizes patterns we would otherwise miss, manages complexity conventional approaches struggle with, optimizes interactions too numerous to evaluate continuously, or helps people understand information quickly enough to make a better decision.
It doesn’t help simply because it’s AI.
The plant floor doesn’t need more artificial intelligence for the sake of having it. It needs useful intelligence applied to problems that matter. Sometimes that intelligence will come from the newest AI technology available—and sometimes the smartest solution will still be a well-tuned PID loop.
