For decades, we've been told that the factory of the future is just around the corner. Machines would understand when something was going wrong. Production lines would adjust before small problems became expensive ones. Engineers would spend less time hunting through data and more time improving the way things run. And operators would have the information they need, when they need it, without having to ask three people or open five different systems.
It's an appealing picture. In most factories, it isn't normal yet. Walk into a typical plant, and you will find remarkable engineering. Equipment runs at high speeds, tolerances are tight, and teams solve complicated problems every shift. Yet when a line begins losing output, the explanation may still depend on which supervisor is working, who remembers a similar issue from last month, or whether an engineer has time to pull data from several systems. A factory can be highly automated and still depend on human detective work.
Outside manufacturing, AI has become an everyday tool. Inside factories, use cases discussed for 20 years are still being tested rather than widely deployed. Why has it taken so long? And more importantly, is anything actually changing now?
The answer, I would argue, has less to do with whether AI can find useful patterns and more to do with whether the industry has found a practical way to put it to work. For years, deployment required too much custom engineering and too much effort to repeat at the next site. The real opportunity now is not simply a smarter model. It's a different way of delivering intelligence to factories, one that a normal manufacturing business can afford, maintain, and expand.
The Promise Was Real. So Was the Gap
Manufacturing didn't discover AI when chatbots arrived. Predictive maintenance, automated inspection, process optimization, and production forecasting have been sold to plants for years. Some have delivered significant results. The problem has been getting from a successful example to something that works reliably on dozens of lines, across different plants, without rebuilding the project each time.
The numbers are revealing. Deloitte's 2026 manufacturing study found that 84% of surveyed manufacturers were seeing measurable value from AI, but only about 20% of the use cases had been scaled consistently beyond their initial deployment. A separate HiveMQ survey of industrial professionals found that 68% were still in some combination of research, pilot, or limited deployment, while just 7% reported AI embedded in core operational processes. The surveys aren't directly comparable, but both show that demonstrating value is easier than making AI part of everyday operations.
There's a tendency to call that an AI problem. Sometimes the models or data aren't good enough. But RSM's 2026 research on manufacturers points to a wider set of obstacles: security, data quality, older equipment and systems, and a shortage of people with the right skills. These are not problems that disappear just because the next generation of AI is more impressive.
That's why a convincing demonstration can feel so distant from shop-floor reality.
The Pilot Works. Then What?
Consider a familiar scenario. A manufacturer selects one production line for a predictive-maintenance project. Engineers connect the equipment, clean up the data, learn what normal behavior looks like, and build a model that identifies a developing problem. The pilot goes well. Maintenance gets useful warnings, and the plant can point to avoided downtime.
Then leadership asks the obvious question: can we roll this out to the other lines, and perhaps the other plants?
That's when the real project begins. The next line has a different controller, sensor names, and operating patterns. The original team has to investigate it all over again. If the pilot took months of engineering work, the second deployment may require a surprisingly similar amount. Multiply that across a company with several sites and the business case starts to look very different.
HiveMQ's 2026 survey found that 48% of respondents named legacy-system integration and data silos as barriers to industrial AI. Its analysis describes a problem practitioners will recognize: the pipeline that works beautifully at one plant often cannot simply be carried to another. Oxmaint has written about the same pattern in predictive-maintenance projects. The software may have proved the concept, but the knowledge needed to build it is still tied to one installation.
Pilots also benefit from unusually motivated teams and leadership attention. A scaled product has to survive ordinary life: shift changes, maintenance windows, staff turnover, changing production schedules, and limited engineering time. It needs a clear owner and a credible financial case.
A successful pilot, in other words, is evidence that something can work. It is not yet evidence that a company can operate it economically at scale.
Every Factory Speaks a Different Language
One of manufacturing's strengths is also one of its headaches. Plants are built to make things, not to follow a universal software template. Equipment comes from different suppliers and different decades. An older machine may have been modified several times. Even two nominally identical lines can behave differently after years of use.
The information is usually there somewhere. A controller knows when a motor starts. A sensor measures temperature or pressure. A quality system records rejected parts. A maintenance system has a history of failures. But knowing that a value exists is not the same as understanding what it means. Someone still has to establish which signal belongs to which machine, whether a measurement is trustworthy, how it relates to the production process, and what changed when the line stopped. A number on a screen might be useful data or a leftover signal from equipment removed years ago. A person familiar with the plant may know immediately. New software won't.
Traditionally, making that data usable has meant sending engineers into the plant to map equipment and interpret signals manually. That's necessary work, but it becomes expensive when the process starts from scratch every time. The industry has been very good at connecting equipment one project at a time. It has been less successful at making that work reusable.
The conversation should move beyond whether factories have enough data. Most have plenty. The harder question is how to turn messy, local, sometimes poorly documented information into something useful without treating every machine as a brand-new consulting assignment.
We Built a Delivery Model Around Scarce People
There's a human side to this problem, too. Industrial AI has often depended on an unusual combination of people: someone who understands manufacturing processes, someone who understands plant control systems, someone who can prepare data, and someone who can build and maintain AI models. Occasionally that's one extraordinary engineer. Usually it's a team.
Those teams can do excellent work. But they don't scale easily, and manufacturers are already competing for limited engineering resources. Smaller plants rarely have spare specialists for lengthy experiments. Research on AI adoption among manufacturing SMEs identifies financial resources, infrastructure, and skills as recurring barriers. Industry discussions of forward-deployed engineering point to the same fundamental cost issue: expert labor becomes part of every deployment.
This creates an awkward business model. To grow, a technology provider hires more specialists to repeat the same kind of integration work at more plants. The customer pays for it. And the resulting knowledge often stays with the individuals who did it, rather than becoming something that makes the next deployment dramatically easier.
The problem isn't the engineers; it's how much they must do manually. The future of industrial AI cannot depend on finding a small army of people who understand every machine, every software system, and every manufacturing process in America.
A Model That Understands One Machine Isn't Enough
The same pattern has shaped the AI itself. Much of the earlier generation of industrial machine learning was purpose-built for individual problems: a model for one machine's vibration, another for one line's quality defects, and another for predicting one type of downtime. Each model might be valuable. But moving it elsewhere often meant gathering more data, revisiting assumptions, and doing substantial work again.
Industrial processes do differ, so some adaptation is unavoidable. Two machines can behave differently even when they share a model number. But having to rebuild an entire analytical approach every time a sensor changes or a line is reconfigured is an expensive way to operate.
This is one reason recent work on industrial foundation models matters. Researchers, including the authors of An Asset Foundation Model for Industrial Asset Performance Management, are exploring models that learn more general patterns from equipment data and can then be adapted to particular assets. The idea is similar in spirit to what has happened elsewhere in AI: start with a broader base of learned knowledge rather than a blank page for every task.
It's promising research, not a guarantee that a pretrained model will understand a new factory on day one. Local validation and safeguards still matter. But if the amount of new work falls with each deployment, the economics could change substantially.
Factories Need Confidence That Their Data Stays Under Control
Another reason industrial AI has moved slowly is that manufacturing isn't a typical software environment. A plant may run equipment that cannot tolerate unexpected interruptions. Production data can reveal sensitive operational details. Sending it outside the facility may be restricted or trigger a lengthy security review.
For a long time, many AI projects assumed that operational data would be sent to the cloud, analyzed there, and then returned to users. That works for some applications. But it introduces questions that plant teams understandably ask: What leaves the network? Who can access it? What happens when the internet connection fails? How will software updates be managed? And what happens if the vendor disappears?
The emerging alternative is to run more of the AI inside the plant. Instead of moving every measurement to an outside service, the factory can keep its data and processing on local hardware, including systems that work without a continuous internet connection. ARC Advisory Group has described this appliance approach through work by Edgescale AI and Red Hat. It's part of a shift toward bringing computing power closer to equipment.
Local AI isn't automatically secure, and installing a computer in a plant doesn't eliminate the need for careful access controls, monitoring, or maintenance. HiveMQ's research on edge AI makes that clear. The difference is that security can become part of a familiar plant-controlled setup, rather than a new dependency on an external data environment. For companies that have held back on AI because of where the data goes, that is a meaningful change.
So What's Different This Time?
There are several changes happening together, and they matter more in combination than they do individually.
First, AI systems are getting better at working with unfamiliar data and recognizing relationships that previously required substantial hand-built preparation. That doesn't mean software can safely guess what every signal means. It does mean that more of the initial investigation and mapping may become assisted or automated, leaving engineers to verify the result rather than construct everything by hand.
Second, broader pretrained models offer the possibility of reusing knowledge across equipment and applications. Manufacturers may be able to adapt a common model and focus on what is unique about each line.
Third, local computing has become a more practical delivery option. AI can increasingly run on hardware at the plant, close to where information is produced and decisions are made. This provides manufacturers with another path when cloud dependence is undesirable, although they still need a credible plan for maintenance, updates, and security.
Fourth, the delivery approach is changing. There is growing value in working with teams that know a particular manufacturing process well, rather than asking generalist software consultants to learn the basics at each installation. An engineer familiar with extrusion, assembly, or machining starts with useful context. Combine that experience with reusable software, and the work can become more predictable.
None of those changes makes the factory effortless to understand. Together, though, they offer a different possibility: that the second deployment could be materially easier than the first, and the tenth easier still. That is the point at which industrial AI starts to look more like a scalable product and less like a series of custom projects.
The Real Test Is the Ordinary Factory
There's a temptation, especially during an AI boom, to judge progress by the most impressive demonstration. A highly automated flagship plant isn't the best test of whether the industry has solved adoption.
A better test might be a medium-sized manufacturer with a mix of older and newer equipment, a small engineering team, and production targets that don't pause for a technology rollout. Can useful AI be installed without months of custom development? Can an operator understand the recommendation? Can a supervisor see whether it prevented downtime, reduced scrap, or improved output? Can the plant expand to another line without repeating the same cost and effort?
These are the questions that determine whether industrial AI becomes normal.
For manufacturers, the definition of value is refreshingly straightforward. A tool should help people run the plant better. It should identify problems worth acting on, explain what changed, and fit into the way the operation works. The measure of success isn't how sophisticated the technology sounds. It's whether the plant wants to keep using it after the demonstration team has gone home.
The Factory of the Future May Finally Be a Practical Question
We have spent years treating the intelligent factory as something that would arrive when the technology became clever enough. That may have been the wrong question. A great deal of the underlying intelligence has been possible for a long time. What has been missing is a reliable, repeatable, affordable way to deliver it across the messy reality of manufacturing.
That doesn't mean every factory will suddenly become autonomous, or that operators and engineers are about to disappear. Quite the opposite. The most useful near-term version of the intelligent factory is one where people can see what's happening, understand why, and respond sooner. A technician gets an early warning with enough context to investigate. A supervisor understands why output is slipping before the shift ends. An engineer spends less time collecting evidence and more time solving the underlying problem. There is plenty left to prove: safe generalization, maintainable local systems, and results beyond carefully managed pilots. Manufacturers should be skeptical of anyone promising that installing a box will make decades of plant complexity vanish overnight.
But there's a reason to be more optimistic now than during previous waves of industrial AI. The conversation is beginning to move from what a model can predict to what it costs to put useful intelligence into the hands of people who run factories, and how easily that can be repeated. If that cost keeps falling, adoption can reach far beyond the handful of plants with the largest technology budgets.
Perhaps the factory of the future will arrive when adding intelligence to an existing line becomes so ordinary and useful that nobody thinks to call it the future at all.
Sources
Oxmaint, “Why Your Predictive Maintenance Pilot Is Failing: The 80% Scaling Trap” (2025).
HiveMQ, “Industrial AI Pilot: Why 68% of Manufacturers Can't Scale Past the POC” (2026).
HiveMQ, “Edge AI in Manufacturing: Why Is It Harder Than Building the Model?” (2026). iFactory, “On-Premise AI Cybersecurity: A Manufacturing Guide.”
iFactory, “Industrial AI Cybersecurity: Protecting Models and Data in OT Environments” (2026).
DeepInspect, “AI in OT Environments: What IEC 62443 and NIS2 Require” (2026). Shetty, Lam, and Abinader, “An Asset Foundation Model for Industrial Asset Performance Management” (OpenReview).
ARC Advisory Group, “The Appliance Approach to Industrial AI: Inside the Edgescale AI and Red Hat Partnership” (2026).
Deloitte, “AI in Manufacturing 2026: From Pilot Value to Scaled Industrial Impact” (2026).
RSM, “How Manufacturers Are Using AI in 2026” (2026).
International Journal of Research and Innovation in Social Science, “AI Adoption Among Manufacturing SMEs” (2026).
Ilmtec, “Forward-Deployed Engineering as a Service: Cost & Model Comparison” (2026).


