The “Cambrian explosion” of AI development continues apace with complex and diverse tools, models, and applications springing to life across every facet of business. As it expands, maintaining clarity of function amid the growing phyla of categories is increasingly important to minimize confusion and streamline everything from planning to procurement.
For its part, Hitachi continues to build on its domain expertise in industrial AI, a fast-growing subcategory of physical AI. While physical AI lets systems like cameras and robots perceive, understand, reason, and perform or orchestrate complex actions in the physical world, industrial AI does the same for equipment and processes in mission-critical systems in industrial settings.
The distinction, however subtle, is significant against a backdrop of use cases, be it managing the flow of electricity through a power grid or reducing the maintenance costs of high-speed rail. For industrial AI, a host of additional dynamics from regulation and compliance to safety must be built into every design, tested, and then certified. This is the world of the mission-critical system, into which the introduction of AI is, well, critical.
The series editor, Michael Zimmerman, and I recently assembled a group of AI experts from across Hitachi for their insights into this world. Joining us were Frank Antonysamy, Chief Growth Officer for Hitachi Digital; Jason Hardy, CTO of AI at Hitachi Vantara; and Chetan Gupta, PhD, Head of AI at Hitachi Global Research, and the General Manager of the Advanced AI Center at Hitachi Ltd. The following are excerpts.
Q1: What makes industrial AI different from other AI applications?
Antonysamy: In consumer AI, you can tolerate errors. If a chatbot gives a wrong answer, it’s annoying, but has limited impact. In industrial AI, we’re operating in mission-critical applications like power grids, manufacturing lines, and railway systems. A 90% or 95% accurate system is just not fit-for-purpose and cannot be deployed. We need systems, AI included, that are reliable, predictable, and accurate. Not just some of the time, but all the time. These are also often integrated into real-world systems – not just standalone applications.
Hardy: Right, and from an infrastructure perspective, there’s a level of detail that goes into designing for an environment that literally cannot go offline. The stock market, the rail network, and the U.S. power grid are hypercritical systems. While some environments can schedule downtime or have flexibility, there’s a significant cost incurred in designing for something that cannot go down. That’s why the cloud isn’t always the best place for this work: 99.9% availability isn’t sufficient for systems that truly cannot fail.
Gupta: Beyond system- and algorithm-level reliability, industrial AI brings two additional considerations: multimodality and edge deployment. Industrial data is inherently multimodal—ranging from text in manuals and logs, to video from worksites, time-series sensor data from equipment, and discrete event data from operations. In practice, effective solutions often require models that operate across one or more of these modalities. Building on Jason’s point, deployment constraints matter just as much: Many industrial use cases require not only on-prem solutions, but true edge deployment to meet stringent latency, reliability, and data-sovereignty requirements.
Q2: How does this play out in practice?
Antonysamy: Take our work with Hitachi Rail. When we deploy AI systems on trains, we’re not just installing hardware. We must meet rigorous industry certifications. The equipment and systems must comply with railway-specific safety regulations and should be able to operate in a rugged, physical environment. Medical devices have their own certification standards; utilities have their own. For each industry, we must understand the compliance requirements and ensure 100% adherence. There’s no choice if you want to deploy at scale in these environments.
One way in which we achieve this is through extensive simulation. We simulate millions of real-world scenarios using synthetic data. Only when we’re confident these models will behave predictably across every situation do we put them into production. It’s the opposite of the “release and refine” approach that’s common with consumer AI because in our world, you can’t afford to learn from failure in production.
Hardy: I completely agree. Think about the timeline. We’re building AI for infrastructure designed to last 30, 40, or even 60 years of continuous operation. A transformer in the power grid doesn’t get replaced every five years like a laptop. The AI and the equipment supporting it need to be designed for that level of longevity. It’s a completely different design philosophy than consumer technology.
Gupta: It means moving fast without breaking anything. Domain experts must be first-class stakeholders at design time; deployment must be backed by rigorous testing and clear acceptance criteria; and rollout must be handled with care so that frontline workers trust the technology and meaningfully integrate it into their everyday workflows.
Q3: What about trust? When you apply AI to critical environments, how do you assuage any concerns people might have?
Hardy: We need to exceed the existing standards. With AI, there’s an immediate fear factor, whether it’s the “Terminator,” or whatever gets into everyone’s head when they hear “the machine is now autonomous.” So there needs to be an expectation that if a human can do X, the AI needs to do X plus Y because regulators, operators, and the public all demand demonstrably higher assurance. That’s how we prove this technology is trustworthy.
Antonysamy: Exactly, that’s a critical point. We don’t say that AI improves safety because that would imply the existing systems aren’t safe. Industrial systems without AI already go through rigorous safety-critical verification and validation. What AI does is augment these already-safe systems, helping them operate more efficiently, with better yields, while maintaining that safety standard. It’s about enhancement, not replacement.
Hardy: We’re already seeing this level of precision and recall in other domains. In cancer detection, for instance, AI now picks up things that would have been missed just a few years ago. That’s the standard we’re bringing to industrial systems.
Gupta: Agreed. Building trust starts with transparency. In critical environments, that means being explicit with stakeholders about both the strengths and the limitations of the technology. It means designing “human in the loop” system where appropriate, building solutions with domain experts as integral members of the team, and deploying them with frontline workers as partners—so trust is earned through clarity, collaboration, and demonstrated reliability.
Q4: You’ve mentioned rail, power grids, and manufacturing. How does Hitachi’s experience working across those industries shape your AI work?
Antonysamy: The key elements of making these kinds of systems include access to high-quality data, domain expertise in these industries, and deep expertise in data science and AI. Hitachi has a heritage in industrials that goes back to our founding 116 years ago. Mission-critical systems are in our DNA. We’ve been doing OT/IT integration for decades, and we have deep expertise in data and AI. We’re uniquely positioned in this corner of the AI market because we understand both the operational technology side and the AI technology side. This isn’t new territory for us. It’s an evolution of what we’ve always done.
Hardy: What’s exciting is that we’re actually putting ‘One Hitachi’ [a Hitachi Group philosophy that encourages intra-company collaboration] into practice. We’re taking the best of our rail, energy, and industrial expertise across all our business units to solve challenges with the most complicated industrial problems on the planet.
Q5: Where is this space heading? What will industrial systems look like in a few years?
Antonysamy: We’ve already seen real impact in productivity improvements, energy consumption reduction, and throughput increases for customers. We’re on a continuous path to offer better productivity, yield, and quality, while lowering energy consumption. That trajectory will continue. We’re focused on making sure we have the right sensors and designing our systems to adapt to the changes. We are going towards autonomous infrastructure – self-balancing grids, manufacturing lines optimizing for higher yield and quality, human frontline workers increasingly using agentic systems to augment their productivity, and machines that can self-diagnose.
Hardy: I think we’ll see a symbiotic relationship develop where AI identifies inefficiencies and automates improvements, while humans make the bigger decisions. That’s the future: Understanding and processing tremendous amounts of information at the machine level, interpreting and forecasting to head problems off before they start to occur, and bringing that high level of precision into very complex industrial systems.
Gupta: Beyond the gains Frank and Jason mentioned, industrial AI is increasingly evolving and moving beyond prediction and recommendation toward direct actuation. We are already seeing robots of different form factors—drones, quadrupeds, and potentially humanoids—being deployed to address workforce shortages in industrial settings. Even more transformative will be the shift toward designing industrial systems with an AI-first mindset. In mining, for example, we are seeing a move to smaller haul trucks as autonomy removes the dependence on human drivers. We saw a similar structural redesign with the introduction of electricity in manufacturing in the early twentieth century, and AI-driven design changes may prove even more consequential.
For more on Hitachi’s work in Industrial AI, visit these sites:
