What is physical AI?
Lucia Rahilly: Let’s start with some context. What exactly do we mean when we talk about physical AI, and why is it generating so much buzz?
Mark Patel: Physical AI is the progression from systems that are programmed around a fixed task to systems that sense, decide, and act based on the physical world around them. Generative AI [gen AI] is transforming how we experience the digital world. When we talk about physical AI, we mean moving from digital outputs—text, code, images—to real assets, real spaces, real places, and real operations.
Think of the way the recent wave of gen AI started, with users utilizing large language models [LLMs] and enabling chatbots. We now experience gen AI in the form of code generation, image recognition, engineering design, and drug discovery—all use cases that LLMs are being applied to. We’ve taken a general-purpose capability and created very specific use cases.
Physical AI is a version of that, but in the physical world. When the media talks about physical AI, it often starts with a humanoid robot. But just as we’ve seen with gen AI, physical AI will also look like autonomous systems: autonomous vehicles, industrial equipment, warehouses, drones, medical systems—all kinds of infrastructure.
Why the physical economy still matters—significantly
Lucia Rahilly: We spend a lot of time talking about the digital economy, but how much economic activity remains tethered to physical work?
Ani Kelkar: Many people miss how enormous the physical economy is. Forty-four percent of paid work hours globally, across sectors, are tied to physical work.
The physical world is complex. We move things, make things, inspect things, and transport goods. We operate factories, buildings, vehicles, and infrastructure. It’s a broad set of capabilities. Using McKinsey Global Institute estimates, we think about 13 percent of those work hours are automatable today. That doesn’t mean those jobs will go away; they’ll likely get reinvented. It just gives you a sense of how massive the physical economy is, which is what makes physical AI interesting.
New use cases for physical AI
Lucia Rahilly: You’ve said physical AI is bigger than just robots. What do you expect the form factors and applications to look like in the coming years?
Mark Patel: We think of this as an underlying capability, with applications built on top of it—and those applications may be very specialized, depending on the environment. Even within a form factor, we should expect different applications.
That’s not how we’ve experienced robotics in the past; robots tended to be implemented for one specific task. Now we may have form factors capable of multiple tasks: A mobile robot could move inventory in a warehouse but also inspect that location for safety and security. A building system could be designed to automate heating and ventilation and also control lighting. Behind all that is the concept of being able to sense, make intelligent interpretation, and act. We expect many visible forms of physical AI, but there’s also a lot that’s invisible while still having a large economic impact.
Use cases that will arrive quickly involve moving things around in a commercial context—forklifts, tractors, industrial inspection equipment, robot arms enabled for security and safety. These may not be visible to most of us, but they’ll address the categories of productivity that Ani referred to, which are the base load of a lot of economic impact. In the near term, this is likely to be a lot about mobility and specialized equipment, because that’s where technological maturity and commercial use cases intersect.
What’s changing—and what it could be worth
Lucia Rahilly: Most of us probably watched clips of the World Humanoid Robot Games, which were jaw-dropping. It feels like we’re at an inflection point where the reality of robots playing a bigger role in our lives isn’t far off. What has changed to make robotics feel different now?
Ani Kelkar: Robot games are key to inspiring community—incredible feats, some amusing bloopers. But behind the scenes, the fundamental shift is moving from preprogrammed automation to more adaptable, intelligent systems.
If you think of traditional robots—and we’ve had them for 50 years—they’re extraordinary at repeating trajectories in controlled environments. If I know exactly where a part will arrive, in what orientation, and which exact motion the robot needs to execute, we’ve been automating that for decades.
What’s different is when the machine senses what’s happening around it, interprets that information, figures out what needs to be done and the path to follow, and acts in the physical world, anticipating what will happen next. That’s the sense-decide-act architecture Mark mentioned, and the shift from instruction-following to goal-fulfilling logic.
This sounds intuitive because humans do it constantly: If I reach for a cup and discover it’s heavier than expected, I change my grip. If someone walks into my path, I change direction. I anticipate how the world around me will respond to my actions, and I prepare for it. If we imbue these capabilities into machines, they can start manipulating irregular objects, navigating complex environments with crowds of people, inspecting construction sites, and figuring out the difference between a real toy and a mock-up—recognizing that it needs to handle each differently.
I don’t think this is simply robots getting better. It’s fundamentally changing the technology—moving the unit of automation from programming movement to understanding what the task needs to be.
Lucia Rahilly: Both of you are describing robots capable of multiple tasks. When leaders hear “general-purpose robotics,” they may think that term describes robots that can do just about anything. Is that the right way to think about this opportunity?
Ani Kelkar: We don’t start by asking if we can build a robot that does everything, like a silver-bullet solution to all physical activity. Instead, we ask which physical tasks can now be performed reliably enough, cheaply enough, and safely enough to create a compelling business case. As long as a robot’s ability isn’t limited to a narrow set of preprogrammed motions, it starts entering the world of general-purpose robotics—applying the same machine across a variety of environments and expecting the right level of performance safely, with economics that work.
In our analysis, we broke physical work down into 2,000 constituent tasks. For each, we looked at the capabilities and sensorimotor skills that humans apply to fulfill the task, then mapped that back to the capabilities a robot would need to perform it reliably, cheaply, and safely. Our analysis suggests this market could be as large as $370 billion by 2040—compared with less than $1 billion today. There’s a massive expansion ahead in what robots can do and where we apply them, and that shift comes from moving toward more generalizability while still managing reliability, uptime, accuracy, and ultimately, safety.
From demo to dependable autonomy
Lucia Rahilly: Let’s pick up on reliability and accuracy, since one of our audience members asked how to move physical AI from demonstration to dependable autonomy in the real world. As you’ve said, doing a set of tasks in a controlled setting differs from doing those tasks day in and day out in a factory. What will it take to see robots at a meaningful scale in the wild?
Mark Patel: This question gets to the heart of what we’re often asked by our clients: When should I expect physical AI to intersect with my business, either as an opportunity or as something to deploy? Enormous excitement is generated by demos, but our mantra is that businesses don’t buy demos—they buy the outcome that comes from a commercially ready product or offering.
Enormous excitement is generated by demos, but our mantra is that businesses don’t buy demos—they buy the outcome that comes from a commercially ready product or offering.
What’s the difference between a demo and getting to that outcome? The product needs to perform reliably. It needs to adapt and recover from failure. It needs to be safe—a huge factor when large, moving objects are making their own decisions. And it needs to be able to integrate into its environment.
That sets up a journey from demo to a realized, reliable outcome. That physical-world data doesn’t exist in a single repository, with one company or environment. It will come together.
Again, we need to think about physical AI differently than gen AI. Many more players and stakeholders will be providing data, interpreting it, educating and informing the systems, and creating not just deployments but the learning that comes from them. In many ways, that’s exciting—it creates a more diverse environment and a more diverse set of participants, not only in the value chain of technology and training but also in how these systems get deployed.
How to think about hardware
Lucia Rahilly: You spoke vividly about robots operating in an integrated, safe, reliable way in the physical environment. How should leaders be thinking about hardware and the supply chain?
Mark Patel: That’s another major question we’re all trying to understand: Where will these systems be developed and manufactured? What critical components and subsystems will be needed to support them?
Physical AI is unusual because the software intelligence manifests through the physical system—you can’t separate the two. You can’t develop a model and worry about the physical system later. Sensors, actuators, power electronics, motors, batteries—all these constitute the system and, in turn, influence how it operates. And don’t miss the compute, connectivity, and how systems get commissioned and operationalized.
Looking at that picture as a supply chain, it’s clearly new, although it does overlap with some relatively mature supply chains. One obvious example is electric vehicles—both the power and control systems and some of the actuators, magnets, precision bearings, motor controllers, and motors themselves. In some cases, these are direct carryovers; in others, if you squint, you see many parallels. That gives an advantage to players with existing supply chains that use electric and autonomous vehicle systems.
Almost every aspect of the product and supply chain will follow an architecture. Whenever we see new growth around a complex engineered system, we see architectures emerge—either common industry standards or proprietary ones that manufacturers provide or use to bring others into their ecosystem. We expect the same here. Some players are already trying to do it for certain layers of the stack; it will evolve as these supply chains develop. What drives that, as in almost every other industry, is unit cost per unit of performance. Here, performance won’t necessarily be miles per hour; it may be safe tasks per hour. We should be thinking about the path to winning unit economics against those metrics, because that’s likely to define who has the winning supply chain.
How to find hidden value
Lucia Rahilly: Ani, give us a high-level sense of where the bulk of the value will accrue.
Ani Kelkar: With gen AI, attention went to models and chatbots. Similarly, a lot of attention here goes to the physical embodiment of AI. But physical AI includes components and hardware—Mark mentioned sensing systems, actuation systems including motors and converters, batteries, and power electronics. There are also AI models controlling robots, whose software is linked not just to the embodiment but also to its integration into the systems in which it operates: simulation, fleet orchestration, data and learning, maintenance, and services.
With enterprise-wide rollout of gen AI, 80 percent of leaders we surveyed said they’d focused on deployment, but less than half said gen AI had actively contributed to positive EBIT [earnings before interest and taxes]. We expect a similar story with physical AI. Yes, the robot’s capability matters, but value gets created where deployment drives productivity, higher quality, and higher throughput—and therefore is an enterprise outcome that can be monetized. That’s not just a question of hardware, software, or the robot itself; it’s about fundamentally redesigning the workflows physical AI participates in and rethinking design to capture the value.
There’s plenty at stake. We estimate that the software and models layer alone for physical AI could represent $330 billion a decade from now. The question is where we can deploy this safely and reliably at scale, so that redesigning workflows and managing maintenance creates EBIT improvements for the customers adopting it.
Lucia Rahilly: Suppose I’m a CEO. What should I do differently today to take advantage of the opportunity physical AI presents?
Mark Patel: Three things: First, focus on the task, not the robot. Understanding the task is how you’ll create value and best apply the technology. Second, think about the talent who will understand the problem and translate it into the data and infrastructure you need.
Third, figure out where you’re going to play. Depending on your business, you have choices: Will you provide and enable this technology? Will you create value for those applying it? Or will you get the most value by applying it in your own business and doing something special for your customers?
The intense momentum is around productivity gains. But there are also opportunities to create disproportionate value beyond productivity—and that’s where I’d encourage people to use their imaginations and not simply assume this is about replacement value.
The intense momentum is around productivity gains. But there are also opportunities to create disproportionate value beyond productivity.
