The hidden value of physical AI

0:02You know, if I reach for a cup and discover it is heavier than what I thought it needs to be, 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. And if we start imbuing these capabilities into what a machine can do, it can start manipulating irregular objects. That's McKinsey partner Ani Kalkar. He's talking about the growing sophistication of physical AI. Ani joins me and McKinsey's senior partner, Mark Patel, to discuss the ways advances in physical AI are generating new value and reshaping the future of work.

0:43This is the McKinsey Podcast, where we help you make sense out of the world's toughest business challenges. I'm your host for today, Lucia Raheli.

0:57Let's start with some context. What exactly do we mean when we talk about physical AI and why we're generating so much buzz?

1:04Mark Patel: We're obviously living in a moment where generative AI is transforming many aspects of how we experience the digital world. When we talk about physical AI, we're moving from digital outputs, text, code, images, to the physical, the real assets, real spaces, real places, real operations. And if you want, you can think of physical AI as in the same way that generative AI a couple of years ago started with us using LLMs and enabling chatbots. That's really evolved, right? We don't experience generative AI just as a chatbot now. We experience it in many different forms. It's in the form of code generation, in the form of image recognition and generation, in the form of engineering design, the processes that are getting accelerated, drug discovery, all of these applications that we see the large language models being applied to.

1:59Mark Patel: But just as we've seen that, the generative model, we've seen with physical AI, it's going to look like autonomous systems in the form of autonomous vehicles, industrial equipment, industrial operations that will be more autonomous, warehouses, drones, medical systems, all kinds of infrastructure. Let's turn to you to talk about the economic picture. We certainly spend a lot of time thinking about and talking about the digital economy, but remind us how much economic activity remains tethered to physical work? I think the thing that people miss is the physical economy is enormous. 44 % of paid work hours globally across sectors are tied to physical work.

2:45And the physical world is complex. We move things, we make things, we inspect things, we transport goods, smart set, we operate factories, buildings, vehicles, and infrastructure. So it's a broad set of capabilities that we've deployed for many, many years and optimized. Roughly 40 % of working hours are being spent on physical activity. And using our McKinsey Global Institute estimates, We think about 13 % of all of those work hours are automatable today already. And this doesn't mean that those jobs are going to go away. In fact, those jobs will likely get reinvented. It's just giving you a scale of how massive the physical economy is.

3:27And that's really what makes physical AI interesting. Yeah, that's amazing. So, Mark, let's turn back to you. Give us a sense of the shape this new physical economy will take. In other words, we've said that physical AI is bigger than just robots. What do you expect the form factors and the applications to look like over the next, say, decade or so?

3:52Mark Patel: Even within a form factor, we should expect that there will be different applications. I think that's not necessarily how we've experienced, for example, robotics in the past. Robots in the past tended to be defined and implemented for one specific task. but we may have form factors that are actually capable of multiple tasks. A mobile robot could move inventory in a warehouse, but it could also be inspecting that same location for the purposes of safety, security. You can have a building system that is designed to automate, for example, the heating and ventilation, but is also controlling lighting in the building.

4:31Mark Patel: You know, the use cases that actually will come relatively quickly, things like in the context of moving things around in a commercial context, forks, lifts, tractors, industrial inspection equipment, robot arms that are enabled for security and safety purposes. These may not be visible to most of us, but they're going to be addressing some of those categories of productivity that Annie referred to, which are the baseload of a lot of the economic impact. And in the near term, this is likely to be a lot about mobility and a lot about specialized equipment, because that's where the technology maturity and the commercial use case intersect.

5:09Mark Patel: And you've got a near term opportunity. And just returning to the visible manifestations of physical AI again, I'm just thinking most of us probably watched video clips of the humanoid robot games, which were just jaw dropping in a range of ways. It's starting to feel really like we might be at an inflection point where the reality of robots playing a bigger role in our lives, maybe it's not actually that far off. So Ani, tell us what has changed and what makes robotics feel different now? I think the robotics or Olympics or games are a key way to inspire the community. You have incredible feats, but also some amusing bloopers coming out of it.

5:56I think behind the scenes, the fundamental shift is moving from pre-programmed automation to more adaptable intelligent systems. So 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 the part will arrive and what orientation and what exact motion the robot needs to execute, we've been automating that for decades. What is different is when the machine starts sensing what is happening around it, it interprets that information, figures out what needs to be done, what path it needs to follow, and then is actually able to act in the physical world and increasingly sort of almost anticipate what will happen next.

6:41We anticipate if I reach for a cup and discover it is heavier than what I thought it needs to be, 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. And if we start imbuing these capabilities into what a machine can do, it can start manipulating irregular objects. It can navigate in complex environments with crowds of people. can start inspecting construction sites every day. It can figure out the difference between a sort of a real toy versus a mock-up of a toy and recognize that it needs to handle all of that differently.

7:23So I don't think this is simply robots are getting better. I think this is fundamentally changing what the technology is and moving the unit of automation, if you will, from programming a motion to understanding what the task needs to be. Both you and Mark are talking about robots that are capable of multiple tasks at this point. Clarify for us, when leaders hear the term general purpose robotics, they may think, oh, okay, well, that's about robots that could do a whole panoply of tasks, could do just about anything. Is that the right way to think about this opportunity? I don't think we start by asking the question of, can we build a robot that can do everything, An art analysis, and we actually broke down physical work into 2 ,000 constituent tasks.

8:16And for each of those tasks, we really got an appreciation for what is the capability, what is the sensory motor skill that humans are applying to fulfill the task? And then map that back to what capabilities does the robot need to have to be able to perform the task in a reliable, cheap, and safe way. Our analysis suggests this could be as large as a$370 billion market by 2040. And compared to the market that we think of today in 2025, that's less than a billion dollars. So there's a massive scale up and expansion in what robots can do and where we apply them that's ahead of us. And that shift really comes from the shift towards more generalizability while still managing things like reliability, uptime, accuracy, and ultimately safety as well.

9:10Ani, I want to pick up on something you said there related to reliability, accuracy, and so forth. We've all seen as discussed a growing variety of robot demonstrations, but doing a set of tasks once in a controlled setting seems very different from doing those tasks day in and day out in a factory or in a warehouse. And in fact, one of the questions we receive from our audience is how to move physical AI from demonstration to dependable autonomy in the real world. Mark, why don't you help us with that? What will it take for us to see robots in the wild that meaningful scale?

9:50Mark Patel: This is a question that Ani and I debate quite a bit, actually, because it goes to the heart of the question that we're often being asked by clients, which is, when should I expect this to intersect the business, either because I'm seeing it as a business opportunity or I'm seeing it as an opportunity to deploy. And there is an enormous excitement that often gets generated by demos. The sort of mantra that we've got into is that businesses don't buy demos generally, they buy the outcome that comes from a commercially ready product or offering. So what's the difference between a demo and getting to that?

10:25Mark Patel: Well, it needs to perform, as you said, reliably enough. It needs to be able to adapt and recover from failure or have a robust way of doing that. It needs to obviously be safe. That's a huge factor when we talk about large weighted moving objects that are making their own decisions. And it needs to be able to integrate into the environment. This sort of sets us up to need a way to think about this, this journey, if you like, from demo into really realized and reliable. That physical world data doesn't exist in some single repository. There's no one company, no one environment that's going to give you it.

10:59Mark Patel: It's going to be something that comes together. There's going to be many more players and stakeholders who are providing that data, who are interpreting the data, who are educating and informing the systems, and then creating not just the deployments, but also the learning that comes from the deployments. And in many ways, that's quite exciting, right? It's creating a much more diverse environment and potentially a much more diverse set of participants in the value chain of technology and the training, but also in the way that we're going to see these technologies deployed. Mark, quick follow there.

11:31You spoke very vividly about robots operating in an integrated, safe, secure way, reliable way in this physical environment. How should leaders be thinking about hardware and the supply chain picture?

11:49Mark Patel: I mean, that is another one of the major questions that we're, I think, all looking at and trying to understand when we look at where will these systems be developed? Where will they be manufactured? What are the critical component parts and subsystems that need to support it? This physical AI is unusual, right? And the software intelligence has to manifest. It has to be applied through the physical system. So you can't just separate two. You can't say, I'm going to develop a model and then I'm going to worry later about the physical system. No, you have sensors, actuators, you have power electronics, motors, batteries, all of these things that constitute the system and in turn will influence how the system operates.

12:28Mark Patel: And not to miss the compute, the connectivity, the way that the systems become commissioned and are operational. So when we look at that picture and say, okay, what does that look like in terms of a supply chain? It's clearly a new supply chain. It's not one that obviously exists today, but it does have overlap with some relatively mature supply chains. One of the obvious ones is around electric vehicles and both the power systems and the control systems of electric vehicles, but also some of the actuators, the magnets, the precision bearings, the motor controllers and the motors themselves. These are systems that in some cases are direct carryovers.

13:08Mark Patel: In other cases, if you squint, you see many of the parallels. So we think that gives advantage to the players that have existing supply chains around electric vehicles and autonomous electric vehicle systems. The other thing that I'll mention is that almost every aspect of the product and the supply chain is going to follow an architecture. Typically, we see whenever there's new growth around a complex engineered system, you start to see architectures that become either common industry architectures or at least proprietary architectures that manufacturers provide and then use as a source of propelling and bringing in others into their ecosystem.

13:47Mark Patel: And I think we should expect to see the same here. There are obviously some players who are already trying to do that for certain layers of the stack, but by no means is this mature yet. It's going to evolve as we see more maturity develop around these supply chains. And what drives that? As with almost every other industry, unit cost per unit of performance, right? Now, performance is not necessarily going to be miles per hour, but it may be tasks per hour, or it may be safe tasks per hour. And so we should be thinking about what is the path to winning unit economics against those metrics, because that's probably going to define who has the winning supply chain.

14:24Ani, talk to us very quickly about where we see the bulk of the value and the opportunity accruing. Give us a high-level sense. I think there's a lot of parallels with what we saw in generative AI. I think the attention in generative AI went to the model, the chatbots. I think similarly, a lot of the attention here is going to the physical embodiment itself. But when you think about physical AI more broadly, you have components and hardware. Mark just talked about the sensing systems, the actuation systems include motors and converters. You have batteries and power electronics, but then you have the AI models that are controlling the robot.

15:02You have software that is linked to not only what happens on the embodiment, but integration into the systems that the robot operates in. You have simulation, you have fleet orchestration, you have data and learning, you have maintenance and services. And so intuitively, as we saw with enterprise AI and the rollout of generative AI across companies, while 80 % of the respondents that we surveyed in that space have said that they have focused on deployment, less than half say that it has actively contributed to positive EBIT. So when we think about physical AI, it's going to be a similar story where, yes, the capability of the robot matters, but value will be created where the deployment of physical AI creates productivity, creates higher quality, creates higher throughput, and therefore is an enterprise outcome that can be monetized.

15:56And that's not just a question of the hardware or the software or the robot itself. It's a question of fundamentally redesigning the workflows that physical AI participates in and then thinking through how that redesigned workflow creates value. And there's plenty of value at stake here, right? Our sense is just the software layer and the models layer for physical AI could represent 330 billion in a decade from now. So the question is thinking about where we can deploy it safely, reliably at scale, so that redesigning workflows, managing maintenance, all of that creates EBIT improvements for their customers who are adopting.

16:36I suppose I'm a CEO. I'm sold on the idea that this is coming. What should I do differently today? to make sure I'm taking advantage of the opportunity that physical AI is presenting?

16:47Mark Patel: Okay, three things. First of all, don't focus on the robot, focus on the task. Your understanding of the task is where you'll create value and how you'll understand best how to apply the technology. Secondly, think about the organization and the talent that's going to understand where the problem is and translate it into that data and that infrastructure that you need to deliver it. and then thirdly figure out where are you going to play depending on what your business is you've got some choices are you going to be part of providing and enabling this technology are you part of servicing and creating value for those who are applying it or are you going to get the most value by applying it into your business and then doing something special for your customers The intense momentum is around where is the productivity gain?

17:37Mark Patel: There are also opportunities to create disproportionate value beyond just productivity gains. And that's where I would encourage folks to say, let's use our imaginations, not let's just assume that this is replacement value. Ani and Mark, thanks again for joining us today. Thank you, Lucia. Thank you, Lucia. Thanks so much for listening to the McKinsey Podcast. I'm Lucia Raheli. And I'm Roberta Fasaro. Find us on McKinsey.com. We'll have a transcript of this episode up shortly. And if you have any questions based on this episode, please email them to us at themckinseypodcast at mckinsey.com. If you enjoyed the show, we'd love for you to leave a rating and a review.

18:19We'll see you next time.

18:31Thank you.