
By Ahmed Farahat, Director, Applied AI R&D, Hitachi America Ltd., R&D
Artificial intelligence tends to make headlines for what it can do in the enterprise, whether it’s generating text, detecting anomalies in datasets, or predicting customer behavior. But the more interesting story, at least at Hitachi, is what happens when AI trades the office for the shop floor.
That’s really where the stakes are higher. Mistakes can result in lost time, investment, and safety. Manuals aren’t always available or up to date. Veteran expertise is stretched thin. And yet, the margin for error is almost zero. U.S. manufacturers collectively lose more than 100 billion dollars annually from delayed or defective production caused by maintenance issues.1
Over the past decade, we’ve focused on bringing AI into these environments, not as a flashy new system, but as a constantly evolving, trusted partner for the people who keep the world’s infrastructure running. Our goal has always been to build tools that respect a worker’s knowledge, respond to the realities of their job, and make critical tasks faster, safer, and more accurate.
We saw this need clearly in our work with Penske, one of the largest fleet operators in the U.S. Amidst an industrywide shortage of technicians, AI offered an extra layer of support for both experienced and developing technicians. For developing technicians, it bridged the skills gap and enabled them to become more effective more quickly.
A common challenge developing technicians often contended with was finding the root cause of a problem. While a fix itself might take only an hour, identifying the root cause could take much longer, and this time was often non-billable. With hundreds of thousands of vehicles on the road, the cost of this inefficiency added up quickly.
So we co-developed the Guided Repair solution, a tool that would offer likely diagnoses based on repair history, vehicle data, structured logic, and driver feedback. But crucially, the system didn’t dictate a course of action. It offered options. Technicians remained in control, free to accept, ignore, or bring the issue to someone’s attention. Providing feedback made the system stronger over time.
That’s where the magic happened. Each technician’s decision fed back into the system, making it smarter. Over time, the AI started surfacing better suggestions faster. It stopped feeling like software and started feeling like a second set of hands, or a quiet partner who’d seen this problem before.
Repeat repairs dropped. Diagnosis time shrank. But perhaps more important, trust grew. Technicians began turning to the system not because they were told to, but because it consistently helped them do their jobs better.
The lessons from Penske became a template for applying AI in other industries, where workers are often dispatched to remote or resource-constrained locations, support is hard to access and the margin for error is small.
A worker might have limited connectivity, outdated documentation, and no immediate access to expert help. Historically, they would rely on experience, intuition, and even luck to identify issues. And if the problem was unfamiliar, they would often have to escalate, taking hours, sometimes days.
So we built a natural-language digital assistant trained on hundreds of equipment manuals. The worker could describe the issue in their own words—“It’s making a grinding noise and won’t start”—and the system would identify the likely causes, suggest a diagnostic path, and even link to the relevant sections of documentation.
Navigating a fault tree used to take hours, sometimes longer, when one was even available. Now, it could be done in minutes. But more importantly, the system restored autonomy. Workers no longer had to stop what they were doing, track down a supervisor, or wait for a callback. They could move forward with confidence, guided by intelligence tailored to the equipment and situation at hand.
Another area where industrial teams needed help was visual inspection, one of the most routine yet risk-sensitive parts of industrial operations. Miss a crack, overlook wear, or skim over a warning sign, and the consequences can be serious. Especially for newer employees, the question wasn’t just “what” to look for, but “how” to look.
We designed an augmented reality (AR) inspection assistant that combines real-time guidance with computer vision. Workers use a tablet or headset to follow step-by-step prompts. When the device is correctly positioned, it automatically captures an image, then scans it for signs of damage, wear, or other issues using onboard AI.
By removing the guesswork around both inspection steps and what constitutes a “correct” view, the system helps bring consistency to a process that traditionally depends heavily on personal judgment. Pilot users told us they felt more confident, not just because the system caught things they might have missed, but because it confirmed what they already knew.
It was less about replacing the human eye, and more about giving it trusted backup.
Some jobs, such as inspections of remote pipelines or visual checks in hazardous zones, aren’t safe or practical for people. Robotics has long promised a solution, but in reality, traditional robots need explicit programming. Every action, every contingency has to be scripted. Which is why many industrial bots feel more like expensive fixtures than adaptive partners.
We’re changing that with language-based planning for robots. Instead of writing scripts, a technician can issue a natural command: “Go to the compressor room and check the filter.” The system translates that intent into a sequence of actions — navigation, image capture, analysis — and sends a mobile robot or drone to carry them out.
We’ve tested this in our labs, and early results are promising. The system isn’t just responding. It’s improvising. It’s adjusting to real-world variables like a human would. It’s early days, but we believe these general-purpose assistants could become indispensable in unpredictable or high-risk environments.
Across all these applications — repair, diagnostics, inspection, robotics — one principle ties everything together: trust is earned, not assumed.
That’s why our systems make suggestions, not decisions. The worker stays in control. The AI learns from the worker, not the other way around. And we never position our tools as surveillance or control mechanisms. They’re there to assist, not oversee.
Even the way we deploy these tools matters. We spend time onsite, working with field teams and leadership to integrate AI into existing workflows. We explain what the system sees, and what it doesn’t. We take feedback seriously, and we make room for business rules that allow organizations to tailor AI recommendations to their environment.
It’s this trusted partnership, between people, processes, and technology, that makes the tools stick. And it’s why adoption doesn’t feel like disruption. It feels like progress.
When we started this journey, AI was still mostly a curiosity in frontline settings. It was the stuff of PowerPoints, not power tools. But as the technology matured, and as generative AI unlocked new possibilities in language understanding and context awareness, we saw something shift. Combined with physical AI embedded in devices, sensors, and wearables, intelligence was no longer confined to the cloud. It moved closer to the work itself, enhancing how people see, decide, and act in real time.
Today, expectations are higher. Workers want tools that understand them, not just the equipment. They want flexibility, responsiveness, and respect for the realities of their work. And they deserve all of it.
That’s why we continue to build with the worker in mind, not just the workflow. We don’t chase the flashiest new models. Sometimes the answer is a smarter interface. Sometimes it’s better guidance. Sometimes it’s just listening.
Because real intelligence isn’t about replacing people. It’s about helping them do their jobs better—with clarity, with confidence, and with trusted tools that finally speak their language.
1 Economics of Manufacturing Machinery Maintenance , National Institute of Standards and Technology, U.S. Department of Commerce, https://nvlpubs.nist.gov/nistpubs/ams/NIST.AMS.100-34.pdf