Hitachi research finds that manufacturers are deploying AI in inspection, maintenance and production while providing far less support to frontline teams. The imbalance raises questions about whether industrial AI is being designed around operational needs or the applications that are easiest to deploy.
Factories are increasingly using AI to inspect products and predict equipment failures, but far fewer are putting comparable intelligence into the hands of the people who keep production running. That imbalance is easily overlooked when progress is measured by the number of applications deployed rather than the decisions, they help employees make. Manufacturers may have several successful AI systems operating across their facilities while operators and maintenance teams continue to tackle everyday production problems with little additional support.
That uneven development emerges from Hitachi’s October 2026 research report, Beyond the Pilot: Industry’s Physical AI Progression. The study surveyed 236 senior industrial AI decision-makers across six markets, with manufacturing accounting for more than two-thirds of responses. Most represented large enterprises, making this a view of established industrial organisations rather than every factory. Although 97 percent reported using industrial AI to some degree, only eight percent placed their organisation at the stage of continuous optimisation and innovation. More revealing than that maturity gap is where the technology is being deployed and who can use it.
An uneven presence on the factory floor
Within manufacturing, the most widely deployed applications reflect processes that already have defined measurements, established workflows and relatively clear criteria for success. Computer vision for inspection was reported as being in production by 43 percent of manufacturing respondents, followed by automated quality inspection at 38 percent and predictive maintenance at roughly a third. These systems identify defects or emerging equipment problems without necessarily changing the wider organisation of production. They address tangible losses and can be evaluated against familiar quality and reliability measures.
The adoption picture becomes narrower with applications that require more extensive changes to production. Digital twins and process optimisation were each reported by fewer than a third of manufacturing respondents, while AI-powered robotics and autonomous mobile robots stood at around a quarter. Collaborative robots were in production at fewer than one in five respondents’ organisations, and smart computer numerical control equipment and intelligent personal protective equipment were less common still. The survey cannot establish the reasons for these differences, but adoption is evidently dominated by specific applications rather than comprehensive changes to factory work.
The report adopts a broad definition of industrial AI, encompassing technology applied to physical industrial and manufacturing settings to improve efficiency, safety and decision-making. Its adoption figure therefore does not mean autonomous systems have become commonplace. A vision model inspecting a component and a system coordinating physical actions across equipment both qualify, but demand different levels of validation, supervision and operational control. Grouping them together can obscure changes in working practices.
The people closest to production have less support
The clearest indication of the imbalance appears in Hitachi’s examination of how extensively AI supports people working directly with equipment. Almost half reported limited AI support for frontline workers, meaning it reached no more than a quarter of that workforce. Maintenance and repair workers followed a similar pattern, with more than four in ten respondents placing them in that category. Only around one in six reported significant AI usage for these groups, and fully AI-driven work was almost absent. These are estimates of the extent of support within organisations, not a count of individual workers, but the difference from the enthusiasm surrounding adoption is substantial.
There is an organisational dimension to this uneven reach. Information and operational technology departments were identified as leading industrial AI adoption by more than a third of respondents, whereas manufacturing production and operations led in fewer than one in five cases. Chief information and technology officers featured prominently among decision-makers. Their leadership is necessary for managing infrastructure and security, but programmes directed primarily through these functions may overlook the difficulties of managing production and maintaining equipment.
The question is not simply whether a factory worker is provided with an AI interface. An inspection model can flag an anomaly, but somebody must assess whether the indication is credible, decide what action is appropriate and understand the consequences of stopping or adjusting a process. A maintenance prediction has value only if it fits with access to equipment, planned downtime, available skills and the practical demands of the shift. Tools that ignore those constraints can add another system to consult without making decisions easier.
The research provides another reason to question whether operational needs are shaping implementation sufficiently. Three-quarters said leadership supported and invested in AI, yet only 38 percent felt their organisation had sufficient internal resources. The most frequently cited obstacle was limited AI expertise, while integration with existing systems and uncertainty about business value were also prominent. Executive commitment may secure the programme, but it does not supply the engineers, production knowledge and time required to make a new capability useful during everyday operations.
Operational knowledge cannot be added afterwards
Hitachi’s findings on underperforming deployments suggest that the difficulty goes beyond a shortage of algorithms. Poor data quality was the most identified contributor to failure or disappointing results, followed by constraints imposed by legacy systems. Insufficient data availability, inadequate integration and limited human knowledge of how to advance industrial AI were also frequently mentioned. Each can prevent a technically promising application from operating reliably in the environment for which it was intended, particularly when data and decision-making responsibilities extend across several systems or departments.
Industrial decision-making contains practical knowledge that does not always appear in a dataset. Experienced operators recognise when a sensor reading reflects a process change rather than an instrument problem; maintenance teams understand the consequences of taking a machine offline at a particular point in production. Those judgements are neither infallible nor a reason to resist automation, but they help define what a useful AI recommendation must contain. The people expected to act on it need to participate in deciding which problems are worth addressing, what evidence the system should provide and when human intervention remains necessary.
The report describes industrial environments as demanding reliable, explainable and auditable AI outcomes. Those requirements become more immediate when a recommendation influences physical equipment, worker safety or production continuity. There is a considerable difference between using AI to prioritise an engineering investigation and permitting it to change machine behaviour without further approval. Manufacturers therefore must establish where decision authority sits, how an outcome can be challenged and what happens when the system encounters conditions it has not previously seen.
The measure of value is what changes in production
Respondents already report benefits from industrial AI, most often improved production efficiency and workforce productivity, followed by lower operating costs. Yet those responses describe perceived outcomes rather than independently verified performance improvements, and more than four in ten said current solutions had only partially met expectations. An application may work correctly within its defined scope without delivering the wider improvement imagined when the investment was approved. That possibility becomes more likely when the organisation measures whether a model has been deployed but pays less attention to how decisions, responsibilities and working routines have changed.
The research also suggests that the approach to change remains heavily directed from above. Leadership-led initiatives were more widely reported than peer mentoring or communities of practice, despite the need for frontline experience when shaping applications. Training in AI is valuable, but familiarity with the technology is not the same as confidence that an alert is trustworthy or that an automated recommendation reflects what is happening on the line. Bringing operators and maintenance specialists into design, testing and evaluation would give manufacturers a firmer basis for deciding whether a tool is ready to become part of normal work.
For manufacturers, a more useful test of AI progress would examine whether the technology helps someone identify a developing fault sooner, make a safer intervention or avoid unnecessary disruption, and whether those results persist beyond a demonstration. This requires measurements of operational performance alongside evidence that the people using the system understand its limitations and retain appropriate control. It also means accepting that some of the most valuable applications may be modest improvements to established work rather than highly visible advances in autonomy.
It is perfectly possible for a manufacturer to operate effective AI inspection systems while leaving most routine decisions untouched. The research provides no reason to dismiss those applications, many of which are delivering benefits, but it exposes the limits of treating their presence as evidence of wider transformation. The more demanding question is what changes for the person who must respond when a machine drifts, a defect appears or a production plan unravels.
