#278 Julia Peyre: How Schneider Electric is Pioneering Enterprise AI at Scale

0:00We develop products, connected products, software and services for everything related to energy management and industrial automation. And so it's always a bit hard to explain what Schneider Electric is doing because there is a lot of end markets. So lots of different buildings, residential, commercial, critical buildings like a hospital, industry as well, where we have all of this automated solution, as well as infrastructure or the transportation of energy, for example. We are a manufacturer, so we have a global supply chain that we are trying to bring to a higher standard. So we also use our products in our global supply chain, but also we sell to all our industrial automation clients.

0:48Build the future of multi-agent software with agency. That's A-G-N-T-C-Y. The agency is an open source collective building the internet of agents. It's a collaborative layer where agents can discover, connect, and work across frameworks. For developers, this means standardized agent discovery tools, seamless protocols for interagent communication, and modular components to compose and scale multi-agent workflows. Join Crew AI, Langchain, Llama Index, Browserbase, Cisco, and dozens more. The agency is dropping code, specs, and services. No strings attached. Build with other engineers who care about high-quality multi-agent software.

1:47Visit agency.org and add your support. That's A-G-M-T-C-Y dot O-R-G. Thank you, Craig. Thank you for inviting me to your podcast. I'm very glad to be here. So I'm Julia Per. I'm leading the AI strategy innovation team at Schneider Electric in the AI Hub, which is our center of excellence in AI. So a bit of background about myself. I have a technical background in AI. So I started my professional career in AI around 10 years ago as a PhD in computer vision at INRIA in France. So INRIA is a leading French institute. And I've joined Schneider Electric around three years ago, shortly after the creation of the AI.

2:41And so since then, I have participated to the development of different AI solutions, in particular, taking speed at the release of ChatGPT to understand what this type of technology could mean for us at Schneider Electric, and also more recently, evaluating innovative startups and partnering with them. Yeah. I'm sorry, just tell us about Schneider Electric. Yes, I'm coming to this. Schneider Electric, it's a leader in efficiency and sustainability. And so basically, we developed products, connected products, software and services for everything related to energy management and industrial automation.

3:40And so it's always a bit hard to explain what Schneider Electric is doing because there is a lot of end markets, so lots of different buildings, residential, commercial, critical buildings like a hospital, industry as well, where we have all of this automated solution, as well as infrastructure or the transportation of energy, for example. And Schneider Electric, a bit more than three years ago now, created an AI Hub, so a center of excellence in AI, to accelerate on the adoption of AI solutions for our external offer, all of our clients, but also for internal efficiency, so for our day-to-day activities of for our employees.

4:34And the products are software, is it a SaaS service or are there hardware products? So we have everything. In fact, so we manufacture products. It can be maybe if you see in your electrical panel in your at your home, you will see, for example, circuit breaker that we manufacture, okay? But it goes up to software. So in fact, what we are selling is we have an IoT platform called EcoStruxure. And so that connects products to edge controllers up to softwares that help our clients control all the different electrical appliances and also bring services to our clients. yeah because as i recall uh with covariant i think they were working in with um i'm gonna forget i'll cut this out but with a uh abb i think is that right uh on on robotic solutions in your warehouses am i wrong on that so um we have you're right it's an industrial automation parts okay so we are in energy management domain but also in industrial automation domain and we are also building for example robotic arms that can pick products and place them in different locations we have also some conveyor belts for example and and software to pilot these belts are those products that you're selling or those are in your warehouses it's both in fact It's a good question.

6:25We are a manufacturer, so we have a global supply chain that we are trying to bring to a higher standard. So we also use our products in our global supply chain, but also we sell to all our industrial automation clients. yeah i mean that what what i find fascinating is that you're really applying ai end to end right not only in business processes but in warehouse management and then selling those services and so yeah maybe take us through uh how you well what are you working on right now let's start with that Okay, so it's a good question. So I am leading the AI strategy innovation team in the AI Hub.

7:15So the AI Hub is the center of excellence in AI. And my team is kind of the R &D of the AI Hub, if you will. So what I'm trying to do is anticipate the needs for tomorrow. Okay. And tomorrow is maybe 12 months, 18 months, 24 months, maybe, but not too far neither. And so I'm trying to understand what are the key strategic topic we need to solve problems for our business. And I'm trying to anticipate what are the new solution, innovative solution that we could use. And so for this, I need to identify right partners and also test this solution concretely on our problems. Okay. Yeah. Yeah, and so which solutions are you looking at right now?

8:12Or maybe, I'm sorry, I'm not doing this very logically, but I'll edit it. Maybe take us, first of all, within Schneider Electric,

8:27the range of AI applications that you guys are using, and then how what what are the ai applications that you're selling and then we'll talk about what you see on the horizon okay so um okay so if we take a setback so for uh if we take the application we have you see we have two kind of application internal efficiency and then external for external offers that we sell. Internal efficiency, maybe 80 % of our use case today. So it's to help day-to-day activities. And with Generative AI today, what we empower is everything related to knowledge management, information retrieval. Okay? So that's the type of topics that we are working on.

9:23An external offer for our clients, there is a wide diversity of tasks that we are trying to solve. One of the examples I can take to make it more concrete. We have an offer called Microgrid EcoStruxure Microgrid Advisor. So it's basically for residential building to help homeowners who have a modern home with maybe an EV charger, a battery, a solar panel. some heating system as well, to control this automatically, to answer the demand, the demand which might be in 24 hours the car needs to be charged, and the temperature in the orb needs to be set at 19 degrees when the hormone air comes back.

10:11And so, to respect the demand, but also the constraints, so minimal costs of energy, and trying to use as much green energy as possible from the renewables. So this is a problem that makes, in fact, forecasting of the demand and supply of energy, as well as optimization of this complex problem under constraints. So that's the type of AI application that we are selling today to our customer. we have a first version out, but we are still working, for example, on making the model more robust, improving these forecasts, modeling the uncertainty. And so for this, for example, we are looking at how generative AI for time series in particular can help us to model all of these different scenarios.

11:11And then internally, when you're saying,

11:18can you walk us through some of the processes that you apply AI to internally? So from maybe filing expense reports or managing corporate finances to controlling robotic arms, picking products and placing them in bins in the warehouse. Okay, yes. So here maybe I can give an example. So the example I can give here is for customer care center, in fact. Okay. So we need to satisfy our customer and to satisfy the customer, we need to understand exactly what are the problems that he met. So one thing that we are doing is leveraging our historical base of FAQ, question-answer of problems that occurred in the past, as well as all our documentation, product manuals, documentation that can be maybe a bit less official as well.

12:32So leveraging all this diverse set of documents to help customer care experts to answer more efficiently and so with better accuracy the clients. um and uh and then on i'm interested in the robotic side can you talk about how you're using ai and robotics or am i mistaken on that um i think we don't have uh i'm not sure craig on this i'm not sure what i can talk about actually okay okay that's fine So, I mean, just what interests me is, you know, I talk to people a lot about enterprise adoption of generative AI. And while there are a lot of pilot programs, it's been extremely slow. It's been extremely slow in other flavors of AI, supervised learning even.

13:43and schneider just seems way out ahead or one of the companies that's way out ahead on adoption why is that it's a good question craig thanks for asking this i think maybe there's two components i would say there's an organizational component so we have made the choice three years ago to have a central AI hub to support business and function. And so we have put in place this AI hub and SPOC model where we help business units with our expertise. We help business units to structure their problem, formulate it, make sure that all the constraints, the environment is well described. And we go up to through different stage from exploration to industrialization, leveraging the 350 experts we have internally, but also leveraging the different partnerships that we develop with big companies like Microsoft, AWS, Databricks as well.

14:55and also smaller companies, startup, innovative startup as well. So that's the organizational part. I think that works quite well today. And the second part is also that we invest a lot of efforts in AI awareness, education for everyone at the company at different levels, of course, awareness for leadership, awareness for users people we will use the solution internally because it's key for the adoption and also keeping our ai experts at the forefront of the technology and so the lab works with all the various business units do you have sort of representatives from the lab embedded in each business unit?

15:56So what you call the lab is the AI Hub, right? So the AI Hub, yes. So it's still, so, you know, we were still a central organization and we have a reference in each of the business units that talk to one person in the AI Hub. Okay. So currently we don't have a mini lab in each business unit. It's still very centralized. Yeah. I'm sorry. I said lab, I meant hub. Right. And then is this coming from, obviously it's coming from senior leadership, but do you have a CEO who's particularly focused on AI adoption. I mean, it is an electronics company, so I presume that people are paying attention to what's happening in technology.

16:56But I just find it unusual the speed at which Schneider has adopted this stuff. So you know the AI lab, before even becoming a hub, it existed before we we always had an r d department looking at ai and so senior leadership indeed decided to scale it to to 350 people today uh um sorry i i am so can i redo it yeah sure okay so but basically my answer is that um this comes from senior leadership this come also from signals that now ai is a key solution if we want to bring more efficiency and more sustainability in our products and for a company of 150 ,000 people, we needed at least a few hundred people to support correctly our business unit and function to scale.

18:25So this really comes from the fact that we cannot play and be a leader in our domain without AI. Yeah, yeah.

18:40I mean, I'm not sure that you would be aware, but are there other companies who are as deep into AI application internally as you guys are that you're familiar with? I think in our domain, we are quite well positioned.

19:06Yeah, I cannot, yeah. And so let's talk about what you see on this 12 to 18 month horizon. I mean, obviously, multi-agent systems are going to be a major part of it. What kinds of things are you working on, both primarily for Schneider internally and then for products? So it's a good question. I like it. So one thing that is very important for us at Schneider Electric, so we manufacture physical products, okay? And so these physical products, they live in the world, and so they are submitted to the law of physics, if you will, okay? So it's really how, it's not just AI data-driven, but it's also how we embed this physical law in our system.

20:13Okay, so that's what we call hybrid AI, and we are not the only one to look at this. So, lots of people, like a lot of companies like NVIDIA, for example, they work on building this simulator that can create a more realistic representation of the world in which we can train AI model. Okay. So that's one first thing of hybrid AI that we think is important. Everything also for us related to expert knowledge, how to embed beyond embedded knowledge from documentation, already formalized in documentation in our system, how to include domain knowledge in our system. And so, of course, we talk about large language model fine-tuning.

21:12So that's one thing, of course, that we are exploring. But we are also looking at how we can have systems that are more interactive with the user to kind of embed their knowledge that is maybe not so well formulated in our system. On the first, the simulation, are you talking about building a digital twin of Schneider Electric and then using that to explore different scenarios for decision-making? Yeah, I think that's the idea. And again, we are not the only one to do this. That's extremely important to build this kind of simulated environment or digital twin because data acquisition, for example, is extremely costly.

22:11Okay. So if you can build with the knowledge you have on a problem, physical equation, you know, on how the system works. If you can build this digital twin and using to simulate different scenario to then, you know, maybe train control system or these robotic arms that we were talking about. And that can also adapt to new environment, new scenarios that's extremely valuable to us, of course. yeah this digital twin idea is something that's kind of fascinated me and i've never really understood uh how you do that i mean is it in a single system so you have one model that uh, that, that simulates the entire organization.

23:07So you can, um, ask it about, uh, you know, your supply chain and it would, if you have a, if you're considering a new vendor, then it could look at the supply chain and, and, you know, explore all of the implications of switching vendors so is it an overall digital twin or do you build simulations of specific business units or or transactions it's a good question i think it's like for generative ai it starts maybe by a single application for for a business unit or even an entity in a business unit and then trying to make maybe this digital twin communicate with each other to unlock even more value because for example you might you in certain application you might want to have information about the design of a product to to better maintain it maybe to to know also how to operate it.

24:25So from design, operation, maintenance, you need to have all of this database communicating, so hopefully through a digital twin. I think it's a complex problem. So just like what we see in generative AI, probably it will start with more silos, application but hopefully breaking the silos in the future yeah and and to build that simulation is that purely with generative AI or using other technologies no it's with different technology so I cited Nvidia for example you might be familiar with Nvidia Omniverse toolkit that allow to build this kind of simulator. And generative AI is one component of it.

25:22For example, if you want to generate from this environment different videos, for example, or images to train your AI model.

25:38I mentioned multi-agent systems. Are you working on that? We are starting to look at this, of course, and it's kind of the same thing. So we start by making applications that work for well-defined and more silos application and starting to see how we can make all of these communicate with multi-agent system. and for multi-agent systems are you guys uh do you build that would you build that yourself using the tools available or would you adopt software as a service i'm thinking i mean it's not widely available yet but Everyone talks about MANUS, the Chinese multi-agent system. How do you approach that?

26:40Is that something that you guys do internally, or do you look to an external vendor that has a product that you can adapt? It's a good question. It's a general question. What is our make versus buy versus partner approach, in fact? Schneider Electric, you understood. we have this team of AI experts, but we are not building generic AI products. So AI products of the Sheffield AI products that we can adopt and use directly without customization. We don't hesitate. We buy these products. We can also partner with providers to use their solution and customize our needs on top of this. So for AgentsEKi, we are at the moment looking at the ecosystem with a different partner.

27:54And we are at the same time looking at all our requirements, all the different types of agents that we need to build, how we will need to also integrate this with our current software, sometimes also hardware. and we will make the decision with all these different criteria of business, scalability, technology that will answer our needs. Yeah. And again, looking out 12 to 18 months, and I mean, I'm sort of obsessed with agentic workflows, but you mentioned simulations. And so what is the thing that you think is going to have the biggest impact on Schneider Electric, but enterprises generally in the next 12 to 18 months?

28:50Build the future of multi-agent software with agency. That's A-G-N-T-C-Y. The agency is an open source collective building the internet of agents. It's a collaborative layer where agents can discover, connect, and work across frameworks. For developers, this means standardized agent discovery tools, seamless protocols for interagent communication, and modular components to compose and scale multi-agent workflows. Join Crew AI, Langchain, Llama Index, Browserbase, Cisco, and dozens more. The agency is dropping code, specs, and services. No strings attached. Build with other engineers who care about high-quality multi-agent software.

29:49Visit agency.org and add your support. That's A-G-M-T-C-Y dot O-R-G. So for me, the biggest impact now, I think it's for internal application, so internal process efficiency, I think it will come if we manage to break these data silos. And this is not just a question of AI model. This is also, do we have the right data connector for this? And can we also standardize this data to understand what relates to what? So this is where I think is the biggest value coming. Also in everything related to knowledge management, I think we have the solution now available. but we still have to adopt them to take benefit of this solution of knowledge management.

30:57And then for external products, for Schneider Electric,

31:06a big part will still, I believe, not necessarily relate on generative AI, because in the end, we need a solution that we can also embed in our product in specific hardware. And so for me, the biggest value will come more if we can integrate, as we say, this more physical law, this expert knowledge in this system, but also respecting the hardware constraints, the latency constraints, and being able to, for sure, to maintain this system and integrate them well in the current software stack. And on that end of it, on the external side of it, can you give us a use case that you're looking at? We are looking...

32:05So I mentioned hybrid AI, so how to integrate physical equation in the system. So all of this, you can think of it in the context of predictive analytics. So we are developing a solution for predictive maintenance for our clients to understand when there is a failure, a risk of failure in our equipment, but also how to optimize the maintenance of this equipment. And so we need to understand what is, for example, the remaining useful life of our electric appliance. So this is an area where there's a large body of literature explaining how to measure remaining useful life for these different types of equipment.

33:01And I think the challenge is to see how we can integrate all of this work, all of this knowledge into more modern AI system. Yeah. When you said hybrid, you're talking about traditional control theory, hard-coded equations, marrying that with generative AI. Is that what you mean? Yeah, I'm really talking about, so you have data-driven approach that can benefit from all the lab data, operational data that you are gathering. And on the other end, you have all expert rule, physical equation. And so how can you create a hybrid system to take benefit of both? Yeah, it is.

33:59Don't worry, I edit. This idea of breaking data silos and simulating specific applications and then eventually, hopefully simulating the entire enterprise. Do you see a day when Schneider Electric or enterprises in general are going to be largely managed by AI and the senior management or the technical workforce will be... tweaking and improving and watching the AI operate. Do you understand what I'm saying? Yeah, I understand. Currently, I'm more thinking AI is augmenting our workforce. So AI will probably automate things that are very much standardized. But what I see everywhere is lots of exceptions.

35:18also environments that are changing. You see also Schneider Electric, there's a wide diversity of clients, of applications. So I don't see this as miraculously, we'll have a multi-agent system that will handle everything, but more really helping and empowering our employees in their day-to-day job to automate completely things that you know that are no added value that are very standard and the rest of it to to to to have the expert re concentrate and the very concentrate on spending more time with the clients or preparing or so the future yeah And how far out, I mean, obviously this is an evolving process, both in the adoption and in the technology that you can apply.

36:24how far out do you see an enterprise that really has taken care of all of the standardized work or non-value-added work with AI? When will it occur? That's your question. Yeah, I mean, is that something that you imagine, and I mean on a widespread basis, both at Schneider and globally, do you see that something happening in a year, in five years, 10 years? For me, it's a never-ending process. We are already automating some of the things. The technology keeps on changing very fast. So this will change, you know, hand in hand. And in 10 years, we will still have more to augment, you know. And our job is just that our job, of course, will be different in 10 years from now.