Industrial AI for the Physical World: Siemens’s Peter Koerte

In this episode of the Я, я и искусственный интеллект podcast, host Sam Ransbotham talks with Peter Koerte, a member of the managing board and chief strategy and technology officer of Siemens, about how industrial AI is quietly transforming the infrastructure that powers everyday life. While consumer AI grabs headlines, Peter explains how artificial intelligence is improving factories, transportation systems, energy grids, and buildings behind the scenes. The conversation explores what makes industrial AI different — from the need for near-perfect accuracy to the challenge of working with proprietary, domain-specific data.

Peter shares examples like predicting train door failures days in advance, optimizing building energy use, and accelerating complex engineering simulations. Peter and Sam also discuss the importance of domain expertise, the value of data-sharing partnerships across companies, and why transformation is as much about people and workflows as it is about technology.

Подписаться на Я, я и искусственный интеллект на Apple Podcasts или Spotify.

Транскрипт

Элисон Райдер: Consumer AI makes headlines daily, but industrial AI increasingly enhances and enables nearly everything we do. Learn how one multinational company approaches data management and deployments at scale on today’s episode.

Peter Koerte: I’m Peter Koerte from Siemens, and you’re listening to Я, я и искусственный интеллект.

Сэм Рэнсботэм: Добро пожаловать в Я, я и искусственный интеллектподкаст от MIT Sloan Management Review исследуя будущее искусственного интеллекта. Я Сэм Рэнсботэм, профессор аналитики в Бостонском колледже. Я занимаюсь исследованием данных, аналитики и искусственного интеллекта в MIT SMR since 2014, with research articles, annual industry reports, case studies, and now 13 seasons of podcast episodes. In each episode, corporate leaders, cutting-edge researchers, and AI policy makers join us to break down what separates AI hype from AI success.

Today we’re talking with Peter Koerte, chief technology officer at Siemens. Siemens is a German multinational technology company focused on industrial automation, smart infrastructure, and mobility systems, all increasingly important topics. We’ll discuss industrial AI, what it means for the workforce, and what the implications are for data sharing across industry. Peter, welcome.

Peter Koerte: Thank you, Sam, for having me.

Сэм Рэнсботэм: Let’s start at a high level. Some of our listeners may not be familiar with Siemens. Can you give us a brief overview?

Peter Koerte: Sure. Siemens [has been] out there [for almost] 180 years. What we say is, “We transform the every day of everyone.” What that means is if you think about the chair right now that you’re sitting on, the clothes that you’re wearing, the water that you’re drinking, the electricity that you’re using, the transportation systems such as trains that you’re using every day, all of that was enabled by Siemens. When it [comes] down to the way we design these things, we produce them, how we actually make sure electricity is safe and distributed, how transportation runs smoothly and safely, all of that is coming through Siemens.

As a consumer, usually you don’t see us, but in the industrial world, Siemens is a very, very big brand name, and we are well recognized for high quality but also for the great solutions we bring and the simplicity to our customers.

Сэм Рэнсботэм: I think that’s a great example. Because so much of the world we rely on, we just don’t pay attention to. We don’t notice it unless it isn’t working for some reason. You talked about industrial AI. What exactly is the difference between industrial AI and consumer AI that most people would be familiar with?

Peter Koerte: The big difference is today, of course, consumer AI is making the headlines, while we think industrial AI is quietly but profoundly changing the physical infrastructure, the physical world that we know of.

So think about, for example, the building that you’re sitting in right now. That building has, of course, some climate control. About 30% to 40% of all the electricity that we’re using today goes into buildings. What we’re saying is, “what if we actually can take all the sensors that we have in these buildings, then develop an AI that automatically learns every minute — or 15 minutes, in that case — and then automatically adjusts all the temperature settings, all the lighting settings, and everything in order to cut costs and energy?” That’s exactly what we’re doing.

We just launched an application that saves 30% of your energy bill and therefore reduces greenhouse gases by 30% just by doing that. It runs autonomously in the background. This is what we do for grids. We do this for factories. We do this for machines. We do this for, of course, buildings, and we do this for trains. So everything in the real world, we are making it more efficient, simply by what we say: “Connecting the real world and the digital world.” We try to optimize and make things better.

Сэм Рэнсботэм: That makes a lot of sense. I mean, I’m sitting here on a university campus. It’s spring break. I guess we are probably heating this place about the same as we would be if it was full of people. I don’t even want to ask. I don’t want to know.

Peter Koerte: That’s it.

Сэм Рэнсботэм: Well, I think we’re all familiar with consumer applications, and I think the failures of AI in consumer applications get a lot of attention, you know, with the hallucinations and these sorts of things. Somehow that seems very different if you’re connecting this to the physical world. It’s not just a funny anecdote that goes across the internet when AI screws up. It could have some real-world consequences when you make that connection. How is Siemens thinking about that?

Peter Koerte: You’re absolutely right. Sam, thank you for saying that. When we compare consumer AI to industrial AI, there [are] three things at the very least that are profoundly different, and the first one you already alluded to is the level of precision and accuracy of those models. Obviously, you don’t need any hallucination when you make recommendations for an engineer to design the next part for, let’s say, your smartphone. Or you certainly don’t need an AI mistake when you think about how to optimize an electricity grid, because that’s critical infrastructure.

So what we need to ensure is the highest level of quality of those models, which, as you can imagine, that’s where we get into 99, 99.9, and so on [for the] percentage of accuracy of the models. And a lot of work goes into that to make sure that these are reliable, safe, and trustworthy. That’s the first part.

The second part is, actually, how do you train these models? Because all of us, we are very familiar with what we call large language models. Now in industry, we’re not necessarily talking about large language models. We’re usually talking about specific data when it comes to — going back to the building example — temperature settings. So we have a lot of time series data. We have construction data. We have engineering data. We have simulation data. This is very different. These are geometries, pictures, vectors, what have you. We have to make these models available in a very, very different way.

The third difference is how do we get that data? Because when we build these models for the physical world, we cannot go on the internet and just download a bunch of data from sensors for your buildings or CAD data or whatever. This is very often even very proprietary data. Customers are only willing to share that data if we are able to express an incremental benefit of when they use our model, then they in return [will] share the data with us. So, of course, in your case, [there will be] better energy savings in the building, but, also, for designers, [they’ll experience a] faster time to market because we can get them designed faster and so on. That way of how you actually get to the data is very different. So the language you’re speaking, the accuracy that we need, how we get the data, this is in the industrial world quite different than, of course, what we use in consumer AI every day.

Сэм Рэнсботэм: That’s pretty fascinating. My naive reaction when you first started talking was, “Oh, what you’re describing is much more structured data,” so I was pretty excited when you were [saying] a lot of this data is temperature data or structured data, but the idiosyncratic nature, and how it applies only to your building or only to your machine and only to your setting, seems very difficult. Tell us a little bit about how you’re getting people to give that data to train machines and how that transfer works between organizations.

Peter Koerte: It’s a very good question. And you’re absolutely right. Because if you think about it, if I say, “It’s a great day” or “The day is great,” the LLM does understand the meaning that actually it’s a great day. In engineering terms, it’s very different, so therefore, we need to adjust and cater for that. The way this works in the industrial settings is you go, of course, after the industries, step by step and say, “OK, what is the semantics in there?” I alluded to buildings, and in buildings there [are] certain standards, and there [are] certain data formats and what we call ontologies. It’s the semantics.

There we try to get that understanding about what is it, that data that is there? It is more structured, as you say, but as you can imagine, right now you’re sitting in a room with Fahrenheit. I’m sitting in a room with Celsius. Therefore, if you then say, “Well, even this is a temperature setting,” actually it is, but it’s quite different if I’m talking 20 [degrees] and you’re talking 20, right? For me it’s warm, and then for you it’s actually really freezing cold. And that’s something to adjust for.

So it’s not a slam dunk, but understanding these use cases industry by industry is really key. In buildings it’s all about energy consumption. But as I said in engineering, very often it’s time to market. It’s in production. It’s usually quality and throughput. Understanding the data and the key variables that drive that is important, which brings us to a keyword that I want to mention, and that is called domain know-how. Because you can argue, “Well, any data scientist can do that.” It’s true. However, you really need to understand the domain that you’re operating in and the key parameters.

I’ll give you just one very simple example, but I find it fascinating. I’m not sure when you last used a train, but maybe the next time you use a train and I ask you, “What is the most critical component of a train?” probably you would say, “Well, probably the brakes.” That’s true; it’s safety critical. But it turns out it’s the doors.

And why is that? Because if you think of the job to be done of a train, [it] is to move people from A to B. That means it stops. It gets people on and off. You go from station to station to station. So the whole day, indeed, yes, the doors of the train open and shut, and thereby they break down. So the most critical part in that regard for the operations is the door. This is the main knowledge; you need to understand that part.

Once you understand that, then it’s fascinating, because then what you can do is you can say, “Give me the voltage reading of that motor that drives the door. Look at, of course, the profile of how that motor operates.” Meanwhile, today our models can predict any door failure 10 days prior [to] its [failure], so therefore we can get it into the depot, and you can fix it, which means higher uptime, higher reliability, all of it, and better passenger comfort. So these are the examples where we have to combine the domain know-how together with the technical know-how, meaning AI, and that’s how you create customer value, industry by industry.

Сэм Рэнсботэм: I like that because I can get my mind around that example. Some of the things that I was reading about Siemens were complicated to understand, but that makes a lot of sense. I think everyone has some sort of application where they would like to know ahead of time that something is going to break before it breaks. Because when it does, it’s a mess.

Siemens doesn’t necessarily own trains though. So how do you get that data about those voltages into your systems versus your customer who has purchased and bought that train? They have to have some sort of way to send that data. They’ve got to share information with you somehow. Weirdly enough, they would benefit from someone else’s train data for a train they don’t own. How do you manage that infrastructure?

Peter Koerte: It’s a great question. That’s why I said it’s very different [from] the way you collect data in the industrial world. Let’s stay in the train example. Truth be told, those customers, they simply don’t. They say, “Give me the train and I’m fine, and then I’ll build my own model.” So we have operators like that. Usually, however, they are not the ones that are most successful.

Usually, the ones that [do consider this:] If you look at the total cost of ownership across the entire life cycle of a train, which is, let’s say, 30 years — in terms of CapEx, the investment is about 10% of the TCO, 90% is operations. So what if I go to you as the OEM? You know your system best. I share the data with you, and you help me to optimize. So you help me to optimize with regards to reliability. That’s the door example. You help me on the efficiency. This really goes down to, of course, the way you operate the train.

Believe it or not, we have AI that helps you to think about how to accelerate and decelerate or break that train in order to save energy. Energy is one of the biggest operating costs that you have on the train. This is where we then take that data. It’s connected. All of these devices are then connected, of course, reliably and encrypted. And then we have the data, and then we make use out of this data, and we build our own models in that regard. And we do this customer by customer, and very often we do have a data-sharing agreement, so we can use that data. We don’t own the data. That’s important. It’s still our customers’ data, but we can use it and train our models for their purposes.

Then, as you said, we can combine it with other data so everybody gets better in that regard. And that’s exactly what’s happening not just in, let’s say, trains, but you see this in many machines. But it turns out it’s not enough data to build your own models because you need to have much more data across different settings. And this is where Siemens comes into play, because usually we don’t build machines, and we don’t build all the trains. Usually, we build components that go into it. So we work with car manufacturers. We work with aerospace manufacturers. We’ve worked with life sciences companies. We work with food and beverage companies, and so on, in order to help enable them. And so they come to Siemens and naturally say, “You know what, how can you help our specific industry to become better?”

Сэм Рэнсботэм: I hadn’t quite thought about it that way, that if one person has insufficient data to train a model by themselves and another person has insufficient data to train a model, but together they do, then the idea of connecting those people together creates value that neither of them could. We had a guest from Land O’Lakes on a prior episode. They’re sharing information with farmers. Farmers build things, they have a lot of data about their crops but how they share that data — I feel like there’s a lot of that going on where we are recognizing that idiosyncratic data is more valuable when combined with other data. At the same time, I’m not naive. People don’t want to share stuff. How do you encourage people to do this?

Peter Koerte: There’s a simple — not an easy but a simple — answer to this, and that is the value. So if I’m not able to translate that and say, “You know what, share the data with me, and then thereby you’re going to improve your availability of the train to stay there, or I [will] improve the efficiency of your building,” then they will not share the data. It’s as simple as that. But if you do, then that’s great. Then they say that’s fine.

Sometimes it’s built into your solution. It’s built into the contract where they say, “Well, we don’t care. It’s fine; you can just use it.” Others are saying, “Hey, I want to also have a negotiated discount,” which is also possible. But the simple answer is you only share your data if you get some value in return. So that’s a little bit like the model. Depending on the industry, it’s slightly different in terms of the kind of value we’re creating, but still there’s some value in return.

Сэм Рэнсботэм: You’re describing largely a partnership but sort of between customers or with customers, but you’ve also done some recent connections with industry, like your partnership with Nvidia. Can you describe what you’re thinking there? I think the goal there is an industrial operating system. How does that work? What’s the plan there? What’s the thinking?

Peter Koerte: With Nvidia we have a very, very close relationship for many reasons. One, of course, is you lose a lot of GPUs in order to train some of our models. Second, [for] tools that we’re providing today, Siemens is the leader in industrial software. So we [have] about 10 billion euros of digital sales. People forget about that. We’re among the top 20 software companies in the world, so we have a lot of simulation software, where you can simulate cars, trains, rockets in the digital world.

Of course, all these simulations take an awful long time when you think about computational fluid dynamics, which is very complex. But [it] turns out you really can accelerate them. So what we’re doing together with Nvidia is to say, “What if instead of waiting eight hours for a complex computational fluid dynamic simulation, let’s say, of the air drag on a car — we can reduce that to minutes?” And that’s exactly what we’re looking at.

So it’s accelerating simulation, accelerating design, when it comes to chip design, which is really interesting as we get to lower nanometers — two nanometers and less — the complexity of verifying those chip designs is enormous. It exponentially really rises. So instead of having human engineers going through every circuit and really testing it to every gate array, actually you can start to have an AI go through this and do this over and over and over again. So the chip design verification is one.

Then, lastly, the design transfer to manufacturing is a key issue because these really hold you up in how fast you can get these chips out there. There again, as you are the designer, we can have the AI in the background verify whether what you’ve designed is correct and whether it can be manufactured.

These are examples that we have announced also at [the Consumer Electronics Show] earlier this year with Nvidia. We are really excited about [them] because we think we can further accelerate — and this is always the keyword: acceleration of design, acceleration of manufacturing, acceleration of operations. That’s why we are so excited about it.

Сэм Рэнсботэм: I get the appeal of switching eight months to eight minutes. It doesn’t take much quantification; we can do that in Fahrenheit or in Celsius, either way that works. But the other thing it makes me think about is that you probably have a lot of processes designed around the idea that it was going to take eight months to do that. And when it takes eight minutes, it feels like, sure, it compresses it, but it also might change the types of things you do, the order that you do them in. It seems like it could just have this ripple of upheaval. How do you manage that? Or maybe am I extrapolating too much? It feels like it could be a mess.

Peter Koerte: That is very true. That’s why I tend to say, always, AI is about 20% technology and 80% is actually transformation. What that means is, we talked a lot about data, that’s one thing, but then it is really changing the processes of how you do things. And, usually, what the AI is now doing is it really changes workflows. So instead of thinking sequentially, where I do one task, let’s say I do the design. The next one is doing the verification. Then the next one is looking at how do I design to transfer, and transfer it to manufacturing. It’s very sequential.

Now what if you could do this all in one step because the AI is doing it? Obviously you’re disrupting a very well-established workflow process. The first question that comes is, who is doing this? Is that the designer from the very end [or] from the beginning? Is it somebody completely else? Who’s the persona that you’re actually talking to? Some very interesting questions.

Second, how is that process then going to go? And who is verifying that whatever the AI is doing is really correct? Then a third question is, where do I do this? Where is the AI sitting? Is that a new application? Is that embedded into an existing application? Is it talking to all applications? All of these interesting questions arise, and they are not usually all technical. Very often, we find this is very much about the people [who] use it every day and involve them, and then start to think — rethink — what wasn’t possible before and thereby addressing also some anxieties, because many would then argue, “The AI is going to take my job away.” So then you have a lot of resistance. Then all of a sudden a technology conversation becomes a cultural-change transformation conversation. We find this time and again.

Сэм Рэнсботэм: Now, the natural follow-up is free for me to ask about workflow and these types of issues. They’re all important, and I don’t want to discount those or whatever, but you’re pretty fired up about smart glasses and workers wearing smart glasses. What’s next for them? How do you see them in the industrial world?

Peter Koerte: I’m very excited about smart glasses. If you think about, in particular, U.S. manufacturing: I just spoke to a major new electrical vehicle car manufacturer, and they told me in their manufacturing, their churn rate — so the attrition of their blue-collar workers — is 35%. What that means is you constantly have to retrain your employees. And it’s not just retraining them, but also the other question is, “How do you capture that knowledge?” What if you can take your glasses, you have that camera, and, let’s say, you are a specialist in operations and you are a maintenance engineer for a specific machine.

That camera and that AI is [looking over] your shoulders, literally, and really checking off what you’re doing. Maybe you’re narrating it even. You record this. You do this over and over again, thereby, you’re democratizing that knowledge, actually. You can capture this for future people coming in. But even better than for the new worker, working the night shift, 2 a.m. in the morning, a machine breaks down, usually people are just tinkering around having no idea. But what if you had those glasses on now, and those glasses are saying, “This is a CNC machine. Usually the failure code of E345 means actually it is a Jam 2. Check that lid and open this one, two, three, four, five,” and off you are. How amazing is that?

I really think in terms of the keyword augmentation. So augmenting the workers, the blue-collar workers, but also white-collar workers on the shop floor and, of course, capturing that knowledge as they are exiting. Isn’t that amazing? I think it’s going to make us all much more productive and much more enjoyable, because you get faster time to results, and thereby you get the factory running, and so on and so on. And you reduce a lot of anxiety and fear, because very often people don’t know what to do. Now all of a sudden they have a companion. They have a copilot, colleague, whatever you want to call it, that helps them, and that is there for them 24-7, as opposed to calling somebody who’s probably somewhere home and sleeping.

Сэм Рэнсботэм: That makes a lot of sense. I want to draw a little contrast though. Earlier we were talking about data, and you were talking about a need for deep expertise and deep domain knowledge. But it sounds like this is maybe a push against, or you’re not needing to know that the E345 error code means this, that, or the other. Is it deeper? Is it more specialized? Those seem in conflict to me in some ways.

Peter Koerte: Obviously, we need both. But, actually, the example is pretty comparable if you think of it. So yes, I can tell you the door is going to break down, and this is now preventative maintenance. The other case was more as a reaction. But in both cases it’s maintenance. So the preventive maintenance means that still a worker has to go out there and replace the motor. Now, on the other hand, in our case here, it’s the same thing. It just gives you the intelligence of what to do. And the doing itself still has to be done by somebody who’s operating that machine. So I think it’s pretty comparable.

The interesting thing about this is because it still requires humans, could we at some point automate that through the whole conversation about robotics and humanoids and everything? This is certainly then also a big push right now that we’re seeing in the market. Whether this is going to come soon or not, we don’t know, but for sure we’re missing at least 2 million people in the workforce in the United States already today … on the shop floor. So the only way to stay productive is by automation. This is where Siemens helps many companies to automate their processes in the factories.

Сэм Рэнсботэм: Maybe I’m reading too much into it, but I read something you’d written about humanoid robots and some skepticism about the actual humanoid shape, and you were kind of hinting at that right there. For one, I’m totally with you. The human shape is not anything magical, and there are a lot better shapes for industrial machinery in particular. Are things going to look like humans, or are they going to look like machines, or different?

Peter Koerte: Well, that’s the big debate. To be honest, it’s too early to tell. I’ve seen both. As a matter of fact, today I just had two conversations of that sort. One of them [was] going in the direction of we need to have humanoids, the other one [was] saying “No, no, no.” I think in the end it comes down to the ROI and the value, again, that we’re creating.

Let’s take a very simple example. Let’s say material handling is a big one in a factory. You have to always make sure that there’s an ample supply of material. Let’s say, in particular, if you’re in a stamping plant, it’s metal sheets, and so it’s heavy. Taking a humanoid is probably not a good idea, although there [are] use cases; I’ve seen them. And there [are] many reasons. One, the payload is very, very, very limited. Number two, humanoids are quite slow if you look at them, at least today. The question is, can you accelerate them? But today they are slow. And then lastly, up to 30% of the energy consumed in a humanoid is just to make sure that you’re standing upright. What if you actually had different form factors that would give you higher payload, faster speed, less energy consumed, and then it becomes an ROI conversation? It depends. It’s very hard to generalize.

In this case, though, I almost would bet probably a different form factor to a humanoid is a better one. But there [are] others where you could argue a humanoid could do a better job, for example, wiring harnesses, clipping them together, where you need to learn dexterity and versatility and all of it. Maybe, but that’s exactly why it’s a fascinating field. I think anybody who claims [to] know it, I think it’s too premature, but it’s a fascinating field.

Сэм Рэнсботэм: Actually, I like that because I think so many things are increasingly “it depends,” because we don’t have these one-size-fits-all models that are going to work. And you know that defeats our ability to make some sort of prognostications here.

Thanks for taking the time to talk with us and sharing your insights about industrial AI, which is probably a different idea for some people, and also data sharing in the future of work. And listeners, thanks for joining us on Я, я и искусственный интеллект.

Peter Koerte: Thank you, Sam. It was great.

Сэм Рэнсботэм: Thanks again for listening today. Next time, Vineet Khosla, CTO at The Washington Post joins us for a conversation about AI innovation in publishing. Please join us then.

Элисон Райдер: Спасибо, что послушали Я, я и искусственный интеллект. Наше шоу может продолжаться, в значительной степени, благодаря поддержке слушателей. Ваши потоки и загрузки имеют большое значение. Если у вас есть минутка, пожалуйста, оставьте нам отзыв в Apple Podcasts или оценку на Spotify. И поделитесь нашим шоу с теми, кто, по вашему мнению, может найти его интересным и полезным.

ME, MYSELF, AND AI^® is a federally registered trademark of Massachusetts Institute of Technology. All rights reserved.