Today's episode is sponsored by Configit. To learn more, visit them at configit dot com. Hello, everyone, and welcome to Advanced Manufacturing Now, the podcast for manufacturing professionals. If you haven't already, like and subscribe to get new episodes delivered to your inbox each week. I'm your host, Rachel Thomas, editor at SME Media. I'm joined today by Laura Beckwith, Director of Product Management at Configit, a global provider of configuration, lifecycle management solutions, and business critical software for the configuration of complex products. Welcome to the show. Thank you very much. Happy to have you here today. So today we're diving a bit into generative AI and its relation to manufacturing and solving configuration challenges. But before we get started, Laura, can you tell me a little bit more about yourself, Configit, and what exactly configurable products are? Yeah, so about myself, just a few words. I've been at Configit actually a really long time, about fifteen, sixteen years. In that time, my background is in computer science and in usability. So a lot of UX work that I've done. That's how I joined Configit and I moved through the years doing product management and then into the role that I'm in now of director of product management. And in those years, I've worked with a lot of different manufacturers. I've gotten to work very closely with the number of products that we develop. And yeah, that's really cool. And some of the stuff that, yeah, just to introduce a little bit about what we do. In the one hand, I think people, configurable products, we can kind of imagine what that is. And especially I know that you're in Detroit, a really highly configurable product are automobiles, right? So you can imagine going to a website and there's this build now where configure my car And the stuff that Configit does is both, you know, we can be the engine behind that. If you're going through, you wanna order an electric vehicle, you wanna be cost conscious. And so you're making some selections that won't increase the price too much, but maybe you have a boat that you want to tow to a lake. And so then you, at some point you make a selection to say, I want to tow, I want to hitch on the back of my car. And it then tells you, actually you need to upgrade your battery because we need to make sure. It can be configured technology that's behind that. We both help people define those rules or those constraints, but we can also, make sure that those constraints are handled at the point when somebody is actually doing configuration. And the place where Configured is really special is that we have this underlying technology called VT technology, virtual tabulation technology. And what that does is it allows us to look at what all the available combinations are before, at the point we look at all the rules, and then we can compile that into this VT package, as we call it. And then that makes the user experience on the front end much faster and always accurate. This is really the specialty. So before we get into AI, I think it's kind of nice to understand the problem space that we're working with, with these configurable products. Yeah, absolutely. Thank you for just kind of giving that background into even, you know, like you said, what configurable products are. So yeah, kind of diving into that next question I have for you. Can you explain what it means when you say that generative AI is inherently probabilistic while manufacturing demands determinism? Yeah, so I guess the easiest way to do this, to start, GenAI, many of us have probably used a chat at the very least and maybe done a bit more with it. And, you know, today preparing for this, I thought, what happens if I go in and just ask it a super basic question? So I went in and I created a couple of prompts or a couple of different chats at the same time, and I said, Tell me about the most famous poets, or What are the most famous poets? And I tried to give exactly the same question two times in a row. And the answers I got were similar, but they're not exactly the same. And if you now, you can, so this is the probabilistic nature, but the problem is for manufacturing, imagine that we go in to configure that electric vehicle. And one time I tell it, I want a hitch and it doesn't tell me I need to upgrade my battery. And the other time I tell it I need a hitch and it tells me I have to upgrade my battery. This is a problem for manufacturers. Maybe I can still produce the vehicle, can manufacture it, but my customer's going be pretty upset when halfway to the lake, they run out of battery when they thought they would be able to make it to the lake. And so this is the determinism, this idea that every time I ask a question that I have exactly, I know I can trust, it's going to answer correctly every time and I can trust that answer. And that is this, the difference when we talk about probabilistic models and deterministic models. Yeah, there's definitely a difference between, you know, you need a battery for your car and you don't need a battery. That'll definitely give different outcomes. Yes. Yes. Yeah. So why is that tension such a critical issue when configuring complex products? I think one of the things is it might seem very simplistic when we think about a battery, a battery size versus the hitch on a car. That might seem very simple and like can't AI figure that out? And it can when the models are really small. But as soon, I mean, the number of rules in these systems, in the number of rules in a configurable product can be huge. So you can think about this, that all the rules that determine all the different engineering things, right? What kind of seats do you want seat heating? How does that relate to the mirrors? Do you have memory? You know, then you have an entertainment system. You have some back seats. Do you have speakers all over the place? I mean, on and on and on the list goes, there's engineering constraints, there's marketing constraints. How do we actually wanna to package these different things? There's legal constraints. Am I selling in the United States versus Canada versus another country? This is, there are legal constraints. The rules pile up. And actually, as soon as you start to get more and more interaction, this becomes an enormously complex problem and chat or generative AI cannot reason about it at that space. So the reason that manufacturers care at the bottom line is that almost right or maybe right is wrong. You can't, you need your systems, you know, you don't want to manufacture a car that becomes illegal to sell or drive. That's wasted money for the manufacturers. So I guess that sort of proves itself a point to some degree that it's really important because manufacturers care that they're getting the right answer when somebody orders a product, that it is a legal, buildable combination. Yeah, absolutely. And I love how you mentioned, you know, thinking of all of the different disclaimers and things, especially with the legal components, like you said, every country, every even state, every region has different, you know, rules and everything and policies. Exactly. There is a I mean, I'm always shocked by the enormous number of rules. And of course, I guess one of the things is right now I might sound very anti AI at the moment. Of course, there's a huge I mean, I talk to customers pretty often about AI and how they want to use it in their processes, and we're looking at some different things, which we can get into in just a moment. But one of the things I want to talk about is, you know, of the great things about AI and this generative AI is the ability to kind of meet a user. So who is it who's going to buy a product? Meet either that potential buyer on his or her own terms. So now I'm, I've been working in the space for a long time. I am not a car fanatic, right? A car to me is something that can get me from point A to point B. And if it's comfortable all the better, if it's not too noisy, love that. But I don't know all these details. So an ideal configuration experience for me will be very different than somebody who's super into cars and super into all those details. And that's the power of GenAI is that now the experience of me configuring a car versus some of my colleagues who are really into this, the Gen AI can kind of meet them where they are and say, well, what step is it? What's interesting to you rather than all of us coming into exactly the same configurator and having to answer questions for me that seem irrelevant or like, is this really important? What will it mean if I make this choice? So I think that that's one of the opportunities that we have with the meeting of this space of deterministic and probabilistic models in the Gen AI. So can you explain the concept of combining generative AI with a deterministic backbone and how that approach helps manufacturers capture the benefits of AI without introducing unacceptable risk. Right. And I guess this is a little bit towards what I was just saying, where you want that experience, that generative AI experience that I might want to, if I am to order an electric vehicle, I might want to go onto a website and just say, you know what, I'm cost conscious, I want a comfortable ride, I want it to be fairly quiet, and you or the manufacturer, the people who are going to build this car want to make sure that whatever it tells me is actually something that we can build or that they can build, they can manufacture it. And when I get it, I'm going to be happy with it. And so the idea here, and I have a few different examples to talk about, but the first one is the idea here is that when I type these things in, that it's not just the AI, it's not just the generative AI that's trying to make guesses about what it does it mean that Laura wants a quiet car or a comfortable car. That it's actually mapping it down in back into this product model. This VT technology is sitting there, making sure that every selection I'm making, the cost consciousness, and maybe I do say, by the way, I need a hitch on it, I have a boat, that it then lets me know, well, you're cost conscious, but by the way, we need to upgrade your battery because of this boat. And that's based in this VT technology, whereas the Gen AI might be the presentation layer and the way that I'm communicating with it and the mapping layer. So that's one example. I have another example, and that's a little bit more mapped to something. So when we're talking about these product models and these rules, right, there's somebody sitting behind in a system inside the manufacturer. Sometimes we call them a modeler, and they are using, if they're with Configut, they are using Configut ACE or another tool and they're entering the rules, right? They're kind of saying these are the restrictions. Here's a legal requirement. You must have whatever it might be. There are legal requirements for safety or for Bluetooth rules or whatever that might be. And they're entering all these rules, all these constraints and requirements. And this process actually is quite time consuming. So we have an AI, we're really focused right now on an AI about how do we make this process a bit easier? How do we allow people to take advantage of natural language or reading, you know, documents that are emails that are passed around an organization and use that information to start building up these models or modifying the models? And like I said earlier, they can be enormously complex. And so you can also do things like ask questions of the product models, Hey, which markets is a hitch actually available in? And the AI could try to reason about this, but we're combining again, the AI with the VT technology to say, oh, it looks like the user wants to know about market and hitch. And I can both look for rules, but I can also look at what it is that actually ends up being available for different markets. So that is another way. I don't know if that's confusing or if I'm a little bit too if it's understandable enough, you know, but that's one of the ways that we're trying to do this. Yeah, speaking to that example that you just shared, you know, of combining, you know, VT technology and AI, I did have a follow-up to that. If there were any kind of, you know, challenges that you all have encountered when trying to formulate all of this? Of course. I think, of course, we have the I mean, so AI is moving incredibly fast, right? As soon as, that's both enormously exciting. It also means things that we couldn't do six months ago, we're doing amazingly right now. So I think most recently, one of the things that was kind of a challenge is if you give it a lot of data, sometimes it can hallucinate, right? And one of the things that we're really focused on, because we work with humans, is how do we make sure that we surface the right information to the users, to these people who understand about product models, understand about these constraints, such that they can analyze what is the AI really telling me? What does the AI, I've told I want to, I don't know, do something with that hitch and the battery size. And now what is it saying that it's, how is it going to change the model? How is it going to change what's available? And we are really working on ways to present that to the users so that they have the information to make it easy to say yes or no. Like is AI, is your suggestion good enough or do I need to refine it? And that's one of the things we're doing as a way of trying to keep check and balance on this hallucination. And also, I mean, the AI is just getting better. We can give it bigger and bigger data sets, but the bigger data set we give it, of course, the more checking a user will have to do. And so how do we, as a company who's trying to make sure that people have enough information, we also have to take into account, we need kind of the human in the loop, but we also need to be cognizant of how much a human can actually process and check during that process. So there's a lot of I think it's a really fun space to be in because it is moving, because there are so many possibilities, and because this human intention in the loop, there's a lot we have to do to account for that. Yeah, absolutely. Another question that just came to mind too while we were talking, I'm just imagining, you know, maybe a smaller mid sized manufacturer or just any kind of manufacturer that is still, you know, gaining awareness of industry four point zero knowledge and really just, you know, trying to grapple and understand AI. So what are some things that Configit would tell them, or just help them reassure them that, you know, there's still a human involved, this is still super helpful for you, even though there's just a lot of terminology kinda swirling around? Yeah. Sorry about that. That is true, there is a lot of terminology. I think one of the things There's a couple of different things here. So first of all, some smaller, medium sized manufacturers, a lot of them are on this journey about, do we go to a configure to order process? Like we're not there today, or we're a little bit there, but gosh, it's going to take us a long time. And I think that is part of the problem that we've been trying to look at to say, you have a lot of what these products are in the different variability of your products. You have them in some brochures. That was actually our starting point for when we developed some of this, the AI tooling that we've built, is to say, we'll start there. Just give us some screenshots, give us some of that documentation and let us suggest some models, like suggest a structure or some variability that we're seeing. And then you, we always present it to the user. Okay. Here are all of the things that I saw in this document. Do you want to make those into the variability in your model? And by the way, I see these connections between these, these different parts of your product. Would you like me to write rules for you? And so that is, it really is kind of a feeding information, the AI trying to make sense of it and presenting it back to the user. How should we structure this? And then the final bit is, do you accept that? Is that a good way? And even if you accept it and it's not the right thing, it might be a good place to iterate from. So none of this, I mean, we have a lot of guardrails around. Like even once it's there, it doesn't mean that an end user is going to start seeing exactly what you're doing right away. There's a whole process about, we call it kind of promoting it to be ready for production. So this is part of an overall process about how you mature your data, how you get things ready to be available such that it could go out to a website at some point. Yeah, there's a lot. I mean, this whole area is quite large. We call it configuration lifecycle management. And it's a journey that can take a long time, and AI is meant to help you speed up parts of that journey or get the value a little bit faster. Absolutely. Yeah. Thank you for giving that breakdown. So, you know, generative AI can be used to capture commercial knowledge from unstructured sources like spreadsheets and product documentation. So, what kinds of insights or efficiencies can manufacturers unlock when that knowledge becomes, you know, like structured and governed? We When we first showed one of our existing customers where we were, and this was a while ago, we showed them where we were with our capabilities. The first thing they said is, We just spent two years manually moving a bunch of our products into the system. Like even if this would have gotten us seventy, eighty percent of the way there, it just would have been a huge time saving. And when they in particular had a lot of PDF documents, I think that's my understanding, a lot of PDF documents and somebody was manually having to go in and it wasn't something that they could automate at that point in time. But now basically with AI, it's like, okay, give me that PDF and it will try to extract some data. And then, you know, also one of the things we're looking into is how do I make sure it has instructions for company, you know, car manufacturer X, who is trying to, these are the rules that we use. Like, this is how we think about building up our products. Here's how we code our different things. You can kind of provide that so that you don't only give it the PDF, but you also give it the company rules as a way of making sure it's following all that. So I think the really exciting thing about this is you don't have to change everything at once. It's both efficiency gains, but you know, the AI is moving very fast. So we try to keep up with that to make sure that maybe you can expand the type of people who are giving this information, right? I think one of the, maybe just, this is perhaps a little aside, I was just reading something that basically said, if your job is to take piece of paper A and transport it into function B, that's when you really need to start to worry. I think a modeler is much more than that. However, maybe they could take that transformation A to B and just do the part that they're really good at, which is the making connecting dots and kind of coming up to the big view and say, how is this going to perform? And use that set of skills instead. I'm wondering if I moved very far from your original question, but I think again, the idea is really to take advantage of all of that that you already have in order to accelerate and improve the process such that your underlying data foundation is the quality you need in order to be able to use maybe GenAI and the other projects, right, on your website, that you can trust what it's doing because it can talk to a reliable data foundation. Yeah, and you kind of gave an example during this answer you just gave us, but do you have any other examples of like customer success stories with this? So we have a customer who also, one of them who, they moved onto our technology a bit ago, and now they're in the process of really appreciating the speed that they've gotten in the kind of front end, the kind of configuration speed that they've gotten. But they have a number of other products and they're starting to figure out, okay, how can we best use this to speed up that, again, that transition for those products that aren't yet in your system, but we want to get them there. We have another customer. Lots of the customers that we're talking with are saying we bring on new people very often. And one of the hardest things about bringing on new people is that these models, these products are incredibly complex and it takes them a long time to get onboarded. And one of the other things that the AI is able to do is you can start asking it questions. Hey, I have this hitch. I want to understand if we want to make a change to the hitch, what other areas of the product do I need to consider? Is that going to have an impact on a particular market? Is it going to have an impact on the battery? Is it going to have like, what exactly is it connected to? And maybe not just one degree of connection, but a kind of wider degree of connection. These, this type of analysis, while it it's, it can be quite complicated, and also getting a sense of the model, just being able to ask questions before you do it, that's a huge gain for some of these customers to onboard new people in the organization. And then there's something about, can we optimize our model in some way or another? So that's an area that we're actively also looking at. How do we make sure that the performance of that model? So we have this VT technology. It has also, like anything, it has its limits. Its limits are quite broad, but our customers are also quite huge. Some of them have very big product models. And how can we make sure that they're able to optimize some of the rules that they've written to make sure that they behave in most performant way? Because that really can make a difference as well. So there's lots of different ways our customers are exploring using it and finding some various successes in the areas as well. Yeah. And I'd love to bring back the, I know we've kind of woven VT technology within this conversation, but is there anything else that you'd like to share about VT technology that kind of really drive home, like what it means and how it relates to Gen AI and everything? So I think in order to explain that, I'm going to kind of explain a little bit about what isn't VT technology. So very often, if we think about all these rules, and especially if you have a little bit of a programming background, then you think about a rule. You choose a hitch, then I need to have a bigger battery. It's something that we execute like at the point that we choose a hitch, then we have to get this other battery. And by the way, if we had this bigger battery, maybe we need to also have a different motor size or something else. I don't know. But that's not how VT technology works. Like we don't execute it at the point of doing a configuration. We actually look at all of these rules and we compile them, which basically means we look at all available combinations and we build that up not in a real physical table, but in a virtual table. And what this allows is that it allows us to ask questions about what is actually available. This is quite huge. This capability, we call this the solution space. It allows us to answer questions about, hey, you want to understand the relationship between that hitch and a seat and a market, which you think may not have any relation, but because there is something, you can ask that and we can show you all available, like in this market here, you can get the hitch, you don't have to have the hitch. And by the way, here it's only available with two types of seats, but in Canada, it's available with seven types of seats. We're able to get that information within seconds. It's incredibly fast because we've already kind of compiled it. So that's the virtual, this VT technology is really what underlies this ability to know assuming you haven't made mistakes near rules, it's going to answer very precisely and it will give a very good also configuration experience. Every time I make a selection, I will never get into what's called a dead end. I will never end up at a place where, oh, you really want the red paint, but now you have to, you know, now you have to, you can't get that because of something else. It will say, yes, you can get it, but you might have to remove something else, But it will never say kind of, Oh, you're an incomplete configuration. You're done. Right? Go start over. So that's another side effect of what the VT technology can do. Yeah, no, I think that's really helpful to kind of get a breakdown of what it isn't to also understand like what it is. It's very complex. I feel like even after having been here sixteen years, like the inner workings of it, that is not something I can talk so much to, but I think the high level impact towards the users is a much better configuration experience. Yeah, absolutely. So, you know, looking ahead, how do you see the balance between AI acceleration, these deterministic configuration models, and just human oversight evolving as manufacturers adopt more advanced digital tools? I think that there's sort of a nice rule that I have in my head, or that we use each tool where it's best, right? So we've talked a lot about the determinism, this idea of the VT being the fundamental of that, the core of that, and that will continue, in our view, that continues to be the case. You need that. And the layer above that is this human reasoning, right? We as humans, we need to check that any suggestions that the AI makes that we either have knowledge that that looks correct to our best knowledge, or that we, you know, we kind of approve it and we second, not second guess it, but make sure it's right. And then, but we really also need to give space to the AI that it can take all of this data in such a way, in such a fast way and such a complete way and handheld with a human, it's incredibly powerful. So to me, there's sort of this, make sure that you use each space for what it's best at. And that's what we're at Configit really trying to take advantage of as well. And I think the other thing is I have in the back of my head that we really want to be the trusted partners in this AI journey for these configurable products, right? We want to make sure that you feel sort of held by Configit technology as you start to embrace it in these configurable Yeah, I really think at the end of the day that trust is really important, especially getting into some, you know, new unchartered territories for, you know, some of these manufacturers, so. Agreed. Yeah. Is there anything else that you'd like to add to our conversation? Anything you didn't touch on? I think it's one of the things that's really fascinating and a question that our sales team has been getting, is whether or not, like, do you really need, you really need a ConfigUt? And of course, I've kind of tried to make the argument today that without this determinism underneath that it really doesn't work. There's interesting that, you know, through with our technical team that I've been in touch with them. And one of them sent me a paper last week. And it basically was trying to make this argument that actually when you get complicated models, like a lot of connections, a lot of variability in your products, I mean, it simply can't do it. The size, the amount of information that you need, that the Gen AI, that the LLMs need to hold onto is so large that it is like theoretically impossible. Of course they can try to break down the problem in other ways, but in the end it's, it is the level of complexity. The problem space that we're in is one that even as AI improves, it will never be able to, the generative AI will never be able to solve these types of problems with the type of skill set. You really do need this underlying technology that tells it right from wrong, so to speak. Yeah, well, thank you so much, Laura, for joining us today and sharing all of these amazing insights, you know, about generative AI, Configit, these complex products, all of that. So, it's really been informative, and I'm sure our audience will have gained lots of insights from our conversation. Thank you, Rachel. Yeah. I do have one last question for you. If listeners want to learn more about Configit, where should they go to get that information? Yeah, so they can go to configit dot com, which is spelled C O N F I G I T dot com. And I would highly encourage people, we have actually something called the CLM Summit coming up May sixth and seventh, and I think that is a great opportunity. You can just register for it, and there's a lot of our customers who are giving presentations, and there's some analysts, and then there's those of us from Configut who will be talking about these topics. And I think that it's a great opportunity to just tune in and learn a little bit about what it is that Configut does. Yeah, so that's a virtual opportunity for people to register for? It is. Yes, exactly. Awesome. Thank you so much. And for our audience to keep up with the latest manufacturing trends and hear more podcasts, visit us at advancedmanufacturing dot org. Thank you for tuning in, and we'll see you all again soon.