Agentic AI and the Future of Predictive Maintenance - with James Loach
Show notes
Welcome to the Trend Detection podcast, brought to you by Senseye Predictive Maintenance – which gives you visibility and insights into all your assets, from single machines to full plants to help you reduce downtime, increase knowledge sharing and accelerate digital transformation across your organization.
What Agentic AI Really Means: Discover how Agentic AI differs from traditional and generative AI, and why it's best understood as a digital colleague that can take action, not just provide answers.
The Role of Agentic AI in Predictive Maintenance: Learn how technologies like Agentic AI can enhance Senseye’s capabilities—making predictive maintenance systems smarter, more autonomous, and more helpful for maintenance teams.
Real-World Maintenance Use Cases: Hear practical examples of how AI agents can support maintenance tasks—from spotting problems and synthesizing information, to optimizing strategies and even configuring systems.
The Future of AI Agents Working Together: Get a glimpse into what’s ahead as AI agents from different systems begin to interact, share data, and collaborate across industrial environments.
Trust, Teamwork, and the Human Factor: Understand how AI in maintenance isn't about replacing jobs, but augmenting human expertise—helping teams work more efficiently while still relying on their critical on-the-ground knowledge.
You can find out more about how Senseye Predictive Maintenance can reduce unplanned downtime and contribute towards improved sustainability within your manufacturing plants, by visiting: www.siemens.com/senseye-predictive-maintenance
Transcript
Yeah, I, I think all of the same things apply, right? And, you know, there's a transition, I think people make in how they think about the interactions with the computer when they're interacting with, um, You know, a generative AI based system, whether that's a traditional one or a, uh, a agentic one or what have you. Um, and I think that is thinking it is a bit more like, um, again, like a colleague, right? It's just some system with all these kind of powers and abilities and knowledge, right? But it's not a fool foolproof thing. Yeah. um, and yeah, just like, um, your, your human colleagues or what have you, you have some issue that you, you want to solve. Welcome to the trend detection podcast powered by Siemens, and now your host Nel Sullivan. Hello and welcome to the latest episode of the Trend detection podcast. And today I'm joined by my colleague James Loach, who's the head of research for Sensii Predicting Maintenance. Um, and James has been on a few times before and we've we've of course talked about some really interesting but also technical subjects, but this one is really, um, really exciting. And potentially a really exciting development in the world of predictive maintenance as well. Um, I'm sure it's a term you've heard around Agentic AI and we're going to talk about how it relates to, to predictive maintenance. So I guess before we go into it, for those who don't know James, maybe just a quick line from you in terms of the work you and the team do within the Sensai team. It'd be great. Yeah, hi, Niall. um, yeah, very, very happy to be here. And yeah, in terms of a, a, a brief introduction, so I'm head of research at Sensi and have been since the, the early days of, of the company. And, and during that time, myself and my group have been, um, basically responsible for, for, for the algorithms that, that underpin the product, um, and then algorithm, you know, um, related functionality in the, in the user interface and, and, and so forth. OK. Fantastic, fantastic. So yeah, I mean, let's, let's dive straight in, and I guess we, we should set the scene first of all for and, and start with the basics. So, um, a lot of people won't be aware of this term agentic AI or they might have heard about it, but don't know how it relates to them. So maybe you could just start by giving us sort of a, a brief explanation what agentic AI is and how does it differ from traditional AI in an industrial context. Yeah. OK. I mean, I, I, I think the best way to think about it is in terms of the sort of the, the user experience of, of this. So an Eergenic AI system, um, relates to you some way, um, in, in the, in the way that a colleague or a, a worker would, OK? So it's, um, it's an entity, um, that, you know, has some general knowledge of the world, has some knowledge of your particular use case or your particular setup. And importantly, has the ability to, um, to do actions of different types. Everything from looking up information to perhaps, you know, interacting with the, with the physical world, um, and the ability to go off and, and do sort of delegated tasks for you. So this is a sense in which it's like a worker. It's a system that you can specify the roughly what you want it to do. You can expect it to go away and approach that task with, with common sense, um, you know, doing, um, a series of actions. Figuring out what we wanted it to do and, and coming back to you when it's done or it, or it needs your help. So it's, it's that sort, sort of experience. And, and you asked how it differs from traditional AI. So, I mean, uh, traditional AI is, I guess is sort of a, an interesting term in itself, right? What, what we should take that, that to mean, but perhaps the best thing to do is to contrast it with, um, a simple chatbot, um, you know, which obviously people have plenty of experience with now. So, um, you know, a simple chatbot can be ergentic, you know, in a very sort of simple way, right? So, I mean, typically when you, you talk to a chatbot, it's just a thing that knows some stuff and you, you ask it some questions, you, you get some answers, something like that. And it may have the ability inside it to do some simple things like do some information lookup for you, right? Or maybe delegate you to another chatbot or, move you to some bit of the app, something like that. OK? So I, we, we might call that traditional, like a sort of an example of traditional AI, you know, in an interactive way. Now, the agentic version of that is just something much deeper, right? So it's not something that you're just doing a little Q&A with to get some info. It's something, you know, you're, you're specifying a task to it and you're, you're letting it get on with it, right? You know, maybe it come back, come back, it comes back to you. In an hour with its reports or, or, or what have you, right? Yeah, so it's a similar kind of thing, I guess, but just, just deeper, right, and more sophisticated. Yeah, and so, and when we, so when we think about sort of like generative AI as well, what, what is sort of the genetic AI in relation to that? Is it sort of the, the next sort of step along, you know, in terms of technology breakthrough? Is that the best way to describe it, or are they sort of sat alongside each other? Yeah, I mean, generative AI to me is, is, is like, um, uh, underlying technology, you know. So in, in the, the, the context of the things we're just talking about, so, you know, the, The most important generative AI technology that would be, would be a language model, OK, which is something that really just can can continue text for you, OK? It has some model of all the text on the internet and so forth and given some prompt, you know, it can continue that, right? So that's like sort of an underlying thing and not just continuation of text is something that can be turned into a chatbot, right? Because instead of continuing text, you know, just like in paragraphs or what have you. Um, the thing can be formed into something that's completing like a movie script, you know, and you, the user is playing one part and it's, you know, completing by filling in the next part as a helpful assistant, you know. So you have this text model that can be used to create this sort of Q&A, um, back and forth. Right? Um, and then you sort of build on top of that, right? So instead of just doing this Q&A prompted, um, you can allow the model to, to use tools. So instead of sending a message to you, it can send a message to some computer system to retrieve some information or, or do something like that. OK? So the simple back and forth is a simple chatbot. You give it a little lookup tool that becomes a slightly more sophisticated chatbot. And yeah, you just make it more sophisticated, right? So you, you, uh, fine tune these models, um, to do extended tasks, right? And maintain coherence over extended periods of time, right? You have them reason in the background about things. You give them dozens of different tools to use, you give them a sort of memory that they can manage, and it, it all sort of builds up, right? So generous VIA to me is a sort of, it's either like um an umbrella term or it's a description of the fundamental technology and then these systems that we're talking about today are just things that you sort of build with that build with that technology, if that makes sense. Yeah, no, no, it does, it does. And so if we're taking that, you know, that wonderful description you just gave and apply it to predicting maintenance, so how, how does agentic AI enhance sort of predictive maintenance solutions like Senseii, for example? Um, I, I think the, the best way to think about this actually sort of harks back to the, um, to Sensai before the, the generative AI days, right? Because it's sort of, um, an image that we've had of the product in a way that we've structured it. That, that still applies and is sort of made a lot, a lot more real and powerful by agentic AI, OK? And that thing is basically thinking about the Sensai application, um, it's something like, uh, like a colleague, OK? It's, um, It's a system that sits there and has this ability to look at all the data at once that's coming in from all your assets. That makes it quite an unusual colleague, right? And then in its, in its traditional days, it's got some ability to reason over that using traditional machine learning, so forth and tell you when it thinks there might be a problem and explain what, what's, what, what, what's, what's, what's happening, right? So, you know, it's a system that's got access to all this data. It can look at it. It can tell you when there's something up and, you know, help, help you decide what to do. It's sort of a decision support system. So Sensi is always. Um, always sort of been that, right? Um, what generative AI in general and this agenic AI, um, specifically does is it sort of, in some way changes the, the characteristics of that colleague that, that Sensi is, right? You know, and that it makes it, um. A lot smarter than it was. It brings, you know, general knowledge and common sense to bear. And then via the tools that we were just talking about, you know, um, the, the ability to pull in information from different places, um, to, to synthesize information and really just to give you, um, to do this tasks much better, right? So to be more accurate and more precise in telling you when that's something that might be wrong. Um, giving you much richer and more detailed and again, more precise explanations of what that, that could be, right? Using information from the. The, the, the sensor already, already has, right? OK, and coupling that in with documentation and so forth. Um, and then in the interactive part of the application when you're trying to dig into something to figure it out and, and work out what to do, you just have a much more powerful system, um, at, at your disposal, right? Um, so yeah, again, something that can browse across the data, across the documentation, you know, ultimately something that can do web searches and what have you. And just bring all that stuff into one place, OK? So Site is like this kind of strange maintenance colleague that you sort of work with as part, part of your team. All of these systems just make it a, a much, you know, better colleague, like a much more useful one, a much smarter one, you know, yeah. It's, it, it, you know, it's interesting, James, what you say about, you know, sort of before generative AI. If there was a time, it feels like it's been around forever so much talk around it, but it is interesting that, cause it's a really good point to make. Sensi was, you know, even before Gen AI, before Gene AI was already demonstrating, you know, big success and stuff. So I guess this is about taking it a step further in terms of those like new efficiencies and. Make, make people work more or helping people to work more, more efficient. I guess it's just taking that a step further with new technology, right? Um, yeah, and it's, it's a very strange thing. I mean, I, I've not really gone on, gone into it on this call, but, you know, discuss a lot about what, what agents actually are and the technical definitions of these things. Um, but, you know, there there is a way of looking at the sensor application before generative AI long ago and saying it is an aergentic application. Because it's a system that you, you hand off a task to the monitoring stuff and getting back to you when it's, it's got some, some, some problem, right? And it can do that task and in a sense, act as an agent. Um, you know, without any generative AI in it, right? Um, and, and this is, this is a nice thing. It's just the way that we sort of formed the product and, you know, it's, in some sense, it's, it's by accident that these technologies have just come along, right? But they just slot in very, very nicely, right? You know, you've got the system doing a task in a traditional, some, somewhat traditional machine learning way, and you just have the ability very naturally to inject this kind of, um, yeah, intelligence and fantastic power into it. And just have it just do its job better, right? With the same general form, right? Exactly, exactly, and I guess, er, cos you've mentioned already a few, um, I guess outside of industrial environment, you know, examples of how agents might, you know, help to make the sort of task more efficient, but if we talk about in a maintenance environment specifically, can we, could you give an example of sort of where, where this might apply? Where you could apply the um AI agent. Yeah, so, um, you, you can think of it in, in diff different levels, right? So to begin with, you might think about it doing more, um, sophisticated, um, information lookup and, um, synthesis of information, right? You know, to help with standard application tasks like spotting problems, explaining what might be wrong, um. And, and, and so forth, right? Um, you can, OK, so, so you can think of in terms of the sort of things the product has, has already done, um, like that. Um, in terms of exploratory things, which AgenicIA really does excel at, I mean, we have, we have this basic task of trying to figure out what, what's going on, right? So, so maybe the system is. Raised some alerts, giving you some possible explanations and things and back that up, right? And then, you know, yeah, maybe, you know, you, you want to ask some questions, the documentation or something to decide whether to go and do, um. You know, um, an, an investigation or what have you. When you go and do an investigation, you take some photos or something like this. You can imagine uploading those into the system. The system then takes into account what it's seen, right? Um, you know, goes off and, and just generally does that in a more sophisticated way. OK? So it's sort of like debugging problems, um, kind of thing. And, but, but you, you can also imagine a little bit, um, more broadly. Um, you, you can give these systems, um, yeah, very general and sort of user, um, you know, very, very general tasks coming from, from the users, um, needs or interests that we've, we've, you know, not, not built into the app in terms of structures, right? So, so one of those examples would be, is like, um, optimizing your maintenance strategy, right? You know, maybe you're monitoring 1000 robots, you have all these maintenance events in the sense I app, all this data, all these things happening. There's just a lot of stuff, right? You know, but you, you can have, uh, an agentic system, you can give it that general task, you can give it access to all of that information. Yeah, and you can just have it, browse around, think about it, reason, you know, come up with some hypotheses about how you, you might improve things, go and follow up on those hypotheses, da, da, da, da, da, and in the background, you know, can compile you some report, right? So that, that would be an example of things where that's in some sense, not the core purpose of the app, but the application has all this information there. And if you, you think, oh, you know, I wonder how I could schedule my oil checks on my robots a little bit better, is there anything you could help me with that? Kick off for an hour using all our data and come back with some, you know, very, very detailed, sophisticated analysis, you know, ultimately of. How, how to address that, right? And, further down the line, um, there are, there are things that you can do. So, for example, you can imagine using this kind of systems to help, um, configure the application, right? You know, so, you know, you just got some robots, um, with some sensor on it and you just want to set it up and get some good results. And, you know, maybe rather than talk to somebody at Sensa. You know, you, you can have the system that, again, go off and, and dig into this, right? You can understand the analytics, have the ability to configure settings and things, look at historical data, look at your maintenance records and say, OK, I think this is a good setup, right? What do you think of this? And then actually apply those things, um, in the app. Yeah. And also down the line, you know, once people have, you know, more trust in these systems, yeah, you can imagine them doing more like actual kind of like actions to do with the physical world, right? So, um. You know, um, raising, um, maintenance orders, right? Filling them in with the suggestions, and you know, sending them into an external system, um, you know, that, that, that kind of thing, right? Yeah, but I, I guess trust is, um, maybe, yeah, even, you know, putting agentic AI to the side, you know, even with a predictive maintenance trust is at the forefront. Do you think, do you think it'll be harder, harder to build trust in sort of agentic AI agents, or does it just become part of the natural workflow like generative AI, uh, is now, as it were, in all forms of life? Yeah, I, I think all of the same things apply, right? And, you know, there's a transition, I think people make in how they think about the interactions with the computer when they're interacting with, um. You know, a generative AI based system whether that's a traditional one or a, uh, a, a genetic one or what have you. Um, and I think that is thinking it is a bit more like, um, again, like a colleague, right? It's just some system with all these kind of powers and abilities and knowledge, right? But it's not a fool foolproof thing. Yeah. Um, and yeah, just like, um. Your, your human colleagues or what have you, you have some issue that you, you want to solve, you might go to someone who you think knows about that. You listen to what they say and you put some value on it, right? But you don't treat it as, as gospel, right? You know, even experts in any area make mistakes and what have you, right? But, you know, you, you, you go to somebody who, who has a, has a relevant knowledge, you listen, you understand, and you, you choose how to act based, based on that. And I think, you know, that's a sensible way to, to interact with any generative AI system, you know, um, chat GPT or all the way to the things that we have in our, on our app, right? It's fundamentally an advisory system, um, but one with great, great power and, you know, intelligence. And basically, like at the moment, you know, you, you make some suggestions, you very clearly say this is just to support you or to, to help you, and then you give reference information to support that and you let the user understand to make that decision. And I think you would just follow a curve over time, um, when these systems just become naturally better, right? You have better models and implemented in better ways, people learn more about them, and more comfortable with them and just offload more and more onto them. OK? Um, and it's, it's just a curve that I think would just be followed every time all over the place. Um, but it's a smooth curve, right? There's no point where these systems in a sense, become, you know, completely reliable, right? So. You know what I mean, you should go from ignoring it to doing exactly what it says, right? It's just a, it's just a smooth transition. No, no, exactly, exactly, and there's back and forths and there's successes and things that don't rub below, so nothing's perfect, right? Um, and it's interesting what you say. It's something we emphasize that sense, isn't it, but, you know, about, you know, decision support, you know, it's about, you know, pointing people to a problem that, that the human still needs to investigate. Um, but it's all, yeah, it's interesting because we've, we've talked about this, you know, um, since I was like a, I've heard it before, you know, like a re reliability, a virtual reliability engineer on your team. So, um, and so we're certainly talk about today, this is to help augment a team because obviously that's that, well. …