I would like to welcome everybody to our next C-suite session here at Microsoft booth. And for this one, I have two amazing leaders, Roland Busch, CEO of Siemens, and Uli Homann, CVP of Engineering at Microsoft. Gentlemen, thank you for coming and being here with me.
I was really looking forward to this conversation. This is the highlight of my day because we're going to talk about physical AI. And this is where the future of AI, in my mind, really resides. I want to start with a question just to ground us in this subject. Roland, I'm going to ask this to you. Just tell us, what is it? How should we think about physical AI?
Well, if there's physical AI, what's the other part? Call it maybe consumer AI. Just to differentiate to start with. What's the difference? I'll give you one example, which you can also see when you visit our booth, and then a little bit of context. Think about what we showcase on our booth: how to manufacture a personalized shoe. So it's a sneaker. You stand there on a pad, it measures your feet, then translates it into a sole, into a shoe, and then it's manufactured according to your specification — a little bit stronger on the left side or right side, depending on what you want. The whole thing goes from the design of a product all the way to manufacturing. In order to make that work, you need industrial AI. It starts with translating the footprint of your foot into a structure which you manufacture, and how to manufacture that. This is really when AI hits the shop floor. Once the shoe is in the store and a customer comes in, what they normally demand, what color — this is another kind of AI, which is about personal preferences. When you're a frequent buyer, you get that. If they say, 'Oh, you like yellow shoes,' and you say, 'No, no, today I'm more into green' — no problem. However, on the manufacturing side, if I print a shoe and it just doesn't fit you, maybe two sizes too small, that does matter. So in physical AI, you cannot allow hallucination. Number one, it really affects the real world. It changes the real world. And it does something at a completely different level of quality expectation. Now, a little bit of context. Most models are trained on language — that's why they're called large language models — maybe on numbers. These models haven't seen everything that physical AI needs. They haven't seen time series data, design data, visual inspection data, any kind of data which is physics-related. And these data have different context. For texts, the language models got it. If you have a sentence, 'What's the next word?' — somehow they got it because they have the context. The context of the industrial world is not there yet. How does a part of a machine relate to another one in a very complex machine? No idea. So you need to train the model on industrial data, and then you can bring it to the context to be really super precise. And that makes a big difference because it really changes how we manufacture, how we design, how we operate. Because the context changes. It's no longer a language context. It's the context of the physical reality that we are dealing with. That's a different way to think about it.
Uli, do you want to comment on this one?
I would add to this: at the end of the day, what we are looking at is the fusion of intelligence with process and design. You start at the design side, and then you end up in the process side, and we put intelligence in there. That's fed by data, by process understanding, and by dynamic decision making, which we have done. But if you write code to make dynamic decisions, you have to foresee every possibility. And with physical AI — AI in the physical world — we're now starting to see that the system will understand the scenario, look at the real world physically, and then make a decision. That's not something that normal software does. That's really the big difference between AI-driven process manufacturing and standard process manufacturing. I think that's going to be where we see a lot of innovation, because right now we're relying on people to provide the flexibility angle. But if you have no people in the factories anymore, or people who don't really know what they're doing — we're working with one customer whose operator turnover is six months. So they have a person there for six months, then they're off, which means they have no idea what they're doing. Therefore, you have to have intelligence in the system that can make up for the experience that's not there anymore. I think that's the other half of where process manufacturing and physical AI come together. AI is already bringing a lot of change into how we think and operate. Physical AI takes it to the next level. You need a new model, a new way of thinking about the world. You need to include processes into it. That's a lot of change for everybody involved.
When I see this much change, I'm thinking, what's the first thing? What's really most important in dealing with it? And then I think about you, gentlemen, because I think it's leadership. Roland, how do you change your leadership? How do you lead your team and help them deal with this transformation? You go first.
I knew he would do that to me.
I think Roland actually does it well, and Satya does it well too. You lead from the top, meaning you show that you actually live AI. Roland, I heard you built your first agent yourself. That's something I think everybody — I would not release it to the shop floor yet. But anyhow, you did it, which means for me you show leadership. You embrace the capability and you show that it can be done, because a lot of people will look at you or me or Adam and say, if you don't live it, why should I do this? So that's where it starts. The second thing is, especially with generative AI — Roland already mentioned it — it's not deterministic. Therefore, we have to have a very strong governance model in place to make sure that the intelligence we want to apply stays within the guardrails we set. And a lot of enterprises don't take this serious enough from my perspective. They talk a lot about it, but then there are a lot of POCs and use cases being built without it. I personally think that's a problem that I would love us to avoid, because if we put the right foundations in, we can apply a lot of AI without the risk really going up. That's the second element. The third thing is we have to give people the opportunity to use these capabilities. I see a lot of companies where technologies like Microsoft CoPilot are only given to 10 percent of the organization because they need it, and the other people don't. If you really want an AI-based transformation, you have to give the tools to everybody so they can advance, learn, and apply them to their jobs. And most likely you'll be surprised, because people are smart. Once they figure out what these things can do, they will achieve amazing results. If you combine leadership, governance, and access to the capabilities, you'll see a lot of advancement in your organization and the transformation that AI ultimately represents.
