Artificial intelligence has been around for many years, but the face of AI is continually changing. Today, by applying well-engineered AI to manufacturing toolsets, we can get closed-loop control and automation that has more humanlike characteristics than ever before. For this episode of Manufacturing Matters, TECH B2B Marketing’s Winn Hardin and Aaron Hand caught up with Kence Anderson, founder and CEO of Composabl, at Automate 2025. Anderson puts Composabl’s multi-agent AI systems into perspective, explaining how applying skill-based, expertise-based, practice-based methodology can turn AI into an expert operator in a tiny fraction of the time it takes humans to learn these jobs.

Winn Hardin: [00:00:06] Hey, everybody, and welcome to “Manufacturing Matters,” where we talk about the technology and trends that are shaping the global manufacturing industry. I’m your host, Winn Hardin. Today I’m with my good co-host Aaron Hand from Tech B2B. How are you doing, Aaron?

Aaron Hand: [00:00:18] I’m doing good.

Winn Hardin: [00:00:19] Beautiful. And we’re lucky enough to be graced with Kence Anderson, CEO and founder of Composabl AI. Is it still appropriate to call you an AI startup at this point or . . .

Kence Anderson: [00:00:29] I think so. I mean, I more would think of it as a more broader controls and automation platform for industrial. But we are an AI startup.

Winn Hardin: [00:00:40] Yeah. Okay. And we’re going to actually . . . and you’re more unique than that. We’re going to delve into that a whole lot more on the intelligent agent side, so I look forward to that conversation. But Aaron, you want to kick us off?

Aaron Hand: [00:00:49] Sure. So, I want to get into what your company is doing specifically, but I want to kind of couch that in what maybe our viewers would understand as artificial intelligence. So, I think about, you know, years back within the automation industry, we’ve been using AI for things like preventative maintenance, kind of the low-hanging fruit area. And then everybody got all excited about generative AI, and suddenly you’ve got it seemed like kind of from the top down, giving orders. Let’s figure out how to use this. And not always so successfully. So, getting into autonomous AI agents in the realm where you . . .  Can you kind of explain where that takes us and how it’s different from what’s been going on?

Kence Anderson: [00:01:33] Super interesting question. The phrase that I’ve been exchanging back and forth with executives is, is finally AI that can actually do something. And that’s a little bit tongue in cheek, but not really. There is this kind of wave after wave after wave of data and machine learning and AI technologies that, you know, first it was analytics. It’s going to change everything about your business. If you can just see this dashboard. I mean, I guess, maybe. Then it was if you can predict something, then that will change everything. But the fact of the matter is, certainly in industrial, I’d say an enterprise in general, all analytics, all predictions serve to make a decision in manufacturing, logistics. That’s usually some high-value decision. You know, it could be anything from inventory management, controlling a specialized machine, process control and automation. And so we’re talking about using AI to actually make the decision in the closed loop. In engineering we call it closed-loop control or optimization decisions, which is a little bit full circle because a lot of folks in industrial will say, well, we’ve been doing automation for a very long time. And that’s true. The PID controller was invented in 1912. What we’re bringing to the table is if you apply well-engineered artificial intelligence, then you can get closed-loop control and automation that has more humanlike characteristics than you were ever able to before. That’s what we do.

Aaron Hand: [00:03:00] Okay, great.

Winn Hardin: [00:03:02] So, how did your background in Microsoft and then that work that led you to form Composabl and then ultimately intelligent agents. How does that fit into what we were just talking about with bringing that intelligence to manufacturing?

Kence Anderson: [00:03:14] Well, I’ll take it even a little bit farther back. I haven’t been in Detroit in a long time, and I was walking over here from the hotel and I realized, I think the last time I was in Detroit was when I was an intern for Ford Motor Company. That was actually my introduction to technology and innovation. And it wasn’t in software. It was, you know, it was manufacturing, air-conditioning compressors. And I got to work with a lot of really great people and see how things were actually made in plants. And the plant was in Connersville, Indiana. And so, after college, I went to Silicon Valley,  worked for IBM and others, and I got into software. I always had that kind of in my history, you know, there’s places where people make things.

Winn Hardin: [00:03:54] The manufacturing DNA component.

Kence Anderson: [00:03:56] Yes. Where people make physical things. And when I was at Microsoft, I had the privilege of traveling all over the world to every kind of steel mill, mine, he largest nickel mine in the world. The largest platinum mine in the world. Oil refinery, chemical plant, different factories, talking to executives, plant managers, process engineers, all the way down to operators to try to understand what the pain points were and what they needed as far as process improvement. And what I found was it was kind of very different than what the tech narrative was, which is, well, if you just gather all the data and if you measure everything, then you’ll be able to analyze things and then predict things. That’s not what I found. What I found was very, very high value, specialized skills that took people ten, 20, 30 years to master and no way to automate them whatsoever. Very little opportunity to capture and codify them and then use them to help people go along their way in a quicker way. I went to a chemical plant for a major chemical company, and they said each operator apprentices for seven years before they’re even allowed to touch the equipment in the control room and make the decisions. And I said, okay, that’s that’s where we need AI. We need AI to help with that. These kind of high-value skills.

Winn Hardin: [00:05:26] Is that relatively unique, the time you spend in the field? I mean, normally when we think of IT leadership guys out there, you know, cutting-edge technology, I don’t think we think about them in boots busting around plants, factories, processing.

Kence Anderson: [00:05:41] I think it was. I didn’t think it was because I had walked, you know, plants and done these things. But I think it was pretty rare. And there are some forward-thinking tech companies, big and small tech, that are realizing they need to hire engineers and people from industry, but most don’t. I mean, in fact, when I was at Microsoft, there were different people I talked to that kind of would shrug like, what are those guys even doing over there? But it takes a tremendous curiosity, you know, desire to understand these industries and acknowledgement that there’s physical realities there. You can’t just do anything you want in a steel mill. You can’t just let your algorithm do whatever you want with a piece of mining equipment that’s, you know, ten stories tall. And I think it’s that combination of engineering and physical realities, plus AI and machine learning that actually makes this really interesting.

Winn Hardin: [00:06:37] Yeah. That really, I want to emphasize this a little bit, that really does put Composabl in a unique position because you’ll have a lot of companies who are offering AI tools these days. They were in manufacturing, they were in machine vision. There were in robotics. They did an acquisition, right? Integrated it in. So, they had some manufacturing background before they started trying to apply AI. But in Composabl, I mean, really the manufacturing component and the DNA and the AI component and deep knowledge of the context, of the environment was and the ground floor.

Kence Anderson: [00:07:09] That was part of . . .

Winn Hardin: [00:07:10] The platform was very much developed. And I look forward to hearing more about how intelligent agents can do more than what most people think of when they think of AI.

Aaron Hand: [00:07:19] Yeah, and actually, just to comment, having covered the manufacturing side of the business for so long,there’s still a great deal of animosity between from the guys on the factory floor about IT, for example, that they don’t even understand. You know, your network goes down, maybe your email doesn’t go through, but if their network goes down, they just dropped a whole bunch of product, you know. So, we had listened to you speak at A3’s conference in the fall on AI and smart manufacturing. So, just some really interesting case studies that you had. I’m thinking about one in particular. I don’t know if we can name the company, but they were doing extrusion, which is basically just pushing, just for our viewers, pushing food through kind of like a pasta maker.

Kence Anderson: [00:08:07] Yeah, it’s like that Play-Doh machine where you just put the Play-Doh in and push it down. And extrusion is used to make all kinds of things, you know, films and display screens and soaps and, and yeah, a lot of types of food.

Aaron Hand: [00:08:18] So, I know you got into this a little bit already about just how long that operator has to work to become competent with that equipment. Tell us about what you were able to do with your agent.

Kence Anderson: [00:08:31] Yes. So, and this is a very common story. It’s a very common outcome to this kind of work that we’re doing. Almost every use case I’ve ever worked on, there’s at least $1 million of ROI on the line for a single-digit percentage improvement in the KPI. It’s most commonly 10 million. It can be up to 100 million, depending on the industry.

Winn Hardin: [00:08:55] That’s annualized basis.

Kence Anderson: [00:08:57] Basically annualized basis. Absolutely. And it usually takes the operator almost exactly ten years to be good at that. So, these things require practice. You’re never going to just look at data. It’s almost silly when you think about it, it’s like, who would think of going into a factory and just looking at a bunch of data and that’s going to teach you how to do something? It’s like, oh, I’ll just read a bunch of data about football, and I’m sure I’ll be a great coach. You know, read a bunch of data about chess, i’m sure it’ll work out well for you. And that’s how it works in factories, too. You know, this particular extruder, this snack foods company. I remember taking the call with them and and thinking. I wonder what’s hard about making, you know, this snack food. And . . .

Winn Hardin: [00:09:42] That requires ten years to learn.

Kence Anderson: [00:09:44] Yeah. That’s it.

Winn Hardin: [00:09:44] When the average person probably stays in one position for a shorter amount of time than that.

Kence Anderson: [00:09:48] Absolutely, absolutely. And these folks are retiring by the minute, practically. And they told me a couple of things. They said, one, there’s a lot of variables. So, you have to keep track of . . . When you turn these knobs, imagine the machine has 25 knobs on it. They’re digital knobs. But imagine you turn each of the knobs and then different things happen at different rates and the product changes in different ways. It takes a long time to understand all those relationships and combinations, to kind of build that into your intuition as you practice. But also, they told me you can’t measure everything. That’s the other myth is, you know, if you just go in and measure everything with sensors and IoT, then everything . . . And they said, listen, this is a corn-based product. They said, you can’t tell how much corn is in a corn kernel, how much moisture is in a corn kernel. You can’t tell how much moisture is in there. It depends on where it was grown and how it was stored and where it was stored. And there’s no IoT sensor for that. And so, the operators would learn to judge based on their experience. They would learn that when corn gets introduced into the extruder, where it’s essentially cooking, the product would turn out differently depending on how much moisture was in that corn. And they said, we need, if you’re going to help us with AI, we need an AI that can do that.

Kence Anderson: [00:11:09] And I said, well, okay, you’re going to need an AI that can learn by practicing. And so really, you know, anything that gets good at an industrial scale is going to require practice, whether it’s a machine or whether it’s a person. And so, we did start with data. But you use that data to create a digital simulation of the equipment. Think of it like Microsoft Flight Simulator for whatever operation you’re talking about. And that gives you an infinite amount of practice potential without memorizing the specific things that happened at any one particular time. And the AI actually learns by practicing just like the operator did. You can actually give it a lot of practice on the cloud. It’s kind of like playing a bunch of games of chess at the same time. You ever seen a person play like six games of chess at once, and then they go, move, move, move, move, move. Then you know, the AI is getting six times as much practice or 100 times or 1000 times as much practice. But where this really, really comes to life, and you talked about this earlier, is we have to sit down with the expert. And the expert has to tell us what are the different skills and strategies you need to succeed. So, again, football analogy, basketball, chess analogy. The operator will tell us when this happens, you use this strategy, just like running plays and passing plays.

Kence Anderson: [00:12:34] In football, when the offense sets up like this, or the defense sets up like this, you want to run this kind of play or this kind of audible and that’s the same thing in industrial. I think it’s the same thing for any kind of complex skill. There’s different skills and strategies, and we use those strategies to build different agents that learn each of those skills and strategies. Imagine an agent that learns this particular running play, this particular passing play. And in that extrusion example, there were six specific strategies that the operator said, I promise you, if you do not learn these strategies, you are not going to get good at this right here. And the operator even told us how to arrange those strategies, like which ones are in hierarchy. Some were in sequences. Some use this strategy until this happens, use this strategy until this happens. Or if you’re in this kind of condition, just use one of these strategies. And so, you end up with this kind of playbook or this knowledge graph of these skills and strategies and a team of agents that work together to accomplish the task. We tried other AI techniques and got marginal results. But when we applied this kind of skill-based, expertise-based, practice-based methodology, certified as an expert operator.

Aaron Hand: [00:13:52] Wow.

Winn Hardin: [00:13:53] How would you differentiate an AI intelligent agent from standard AI deployments? What makes it an agent? An intelligent agent?

Kence Anderson: [00:14:01] That is a great question. One of my favorite distinctions I like to make is the difference between perception and action. Most of what people are talking about in AI these days is actually perception, which isn’t a bad thing. You need perception. It’s kind of like the rods and the cones in your eyes. So, your eyes are light sensors, but it’s the rods and cones in your eyes that help your light sensors detect what’s happening and identify objects. You know, predictive maintenance, computer vision. All those things are perceiving that something is happening. That is a part of what needs to happen. But the real moneymaker is deciding what to do. And I would argue that any perception goes with deciding what to do. The only reason you have perception is so you can decide what to do.

Winn Hardin: [00:14:46] Absolutely.

Kence Anderson: [00:14:47] And so, agents are the “deciding what to do” part.

Winn Hardin: [00:14:51] Okay.

Kence Anderson: [00:14:51] A team of intelligent agents might have perception in it. Like in that extruder-use case, there was a computer vision system that was able to look at the food items and perceive exactly what the shape and size and all the quality metrics were just by looking at it, instead of an operator taking manual measurements. That’s important, but that doesn’t make the product. What makes the product is controlling that piece of machinery. And so, the team of agents, the multi-agent system, as we now call it, had the perception AI in it. And it had this team of agents that actually take action. And so it was able to do what an operator does–take perception, translate it into action, and produce results towards the goal.

Winn Hardin: [00:15:37] In that corn extrusion application, did you find . . . Because I would assume that if it takes ten years to acquire a skill set, they’re probably, they’re using all their senses as well as the data that’s coming from the machine. Do you find that the process is providing enough sufficient data for an intelligent agent to take action? Or do we say, okay, what we’re seeing is the humans are using two or three other things, you know, whatever the humidity is in the air, the contextual environmental conditions.

Kence Anderson: [00:16:03] The answer is is both. So I’d say maybe 60%, 70%, 75% of the applications I work on . . . Almost everything you need to be an expert is already being measured.

Winn Hardin: [00:16:15] Okay.

Kence Anderson: [00:16:16] But this and this is where the advanced IoT sensors and things come in. So many times an expert operator will tell me, no, hold on a minute. Unless you can see this or you can hear this, you’re not going to be an expert. I’ll give you two examples. National Oilwell Varco expert machinist. His name was David Lawrence, and we were in their machine shop and foundry. And he was explaining to me, he said, do you know why I’m better at this than all that software and all these folks over here and there? Because I hear the sound the tools make as they’re cutting the metal. And he said, there’s 16 different types, and 16 different sounds.

Winn Hardin: [00:16:53] And acoustical signatures.

Kence Anderson: [00:16:55] You got it.

Kence Anderson: [00:16:55] And the AI that I designed, it identified nine different ones. So, there was a perception part that would take a microphone signal. And to be fair, that microphone sensor wasn’t there before. So, in order to do this, they had to install a microphone that wasn’t there. And combine that with the data. They did have speed data, flow of the lubricant data, and all that kind of thing, but they didn’t have microphone data. And the perception part of the AI would identify which of the nine sounds were being made. And then the control part, the AI, would make the decision on how to control the speeds and things. I’ll give you another example. I was in a steel mill not too, too far from here in Indiana. And they were coating steel. So, you coat a sheet of steel by dipping it in a pot of molten zinc, and it comes out, and the operator kept looking out the window. It’s called a pulpit, this kind of reinforced control room. And, I’m like what are you looking at? He’s like, I’m looking at the patterns. And it would have different patterns in the light. Like there’s one called a fish pattern or a rainbow pattern. And he had all these beautiful diagrams and real-time sensor measurements. But he needed to combine it with those patterns that he saw in order to control correctly, to get the thickness of the coating.

Winn Hardin: [00:18:10] And that’s like the drip pattern of the coating as it’s coming out of the extrusion.

Kence Anderson: [00:18:13] Exactly.

Winn Hardin: [00:18:14] It’s funny how much manufacturing, at least advanced manufacturing, has to incorporate both art and technology. I mean, that’s kind of what we’re talking about here, is that creative component that the human brain brings, brings together. And I don’t think, I think it’s, it’s not like you were saying in the case of the corn extrusion. Well, we had to tell them they had to put a hyperspectral system in there to measure water content, you know, as it’s coming into the system. You may need a little bit more IoT sensors, but it doesn’t have to be a super expensive or complex fix. It just, if you want an agent that’s going to work . . . Again, I want to come back to that ground  truthing. You know, that understanding of process and manufacturing and this awareness that there’s more involved than a machine or a robotic arm with a screwdriver on the end doing whatever it’s doing.

Kence Anderson: [00:18:57] One hundred percent. I like to call that more humanlike decision making. It’s very, very common that an operator is looking at a color change. I mean, in mining, looking at the different colors of the ore represents different hardnesses. That’s something that’s also hard to measure, is the hardness of rock. And so, if you’re mining gold or platinum, the rocks around it, that it’s in, are going to be of radically different hardnesses. And one of the proxies that folks use on these lines, manufacturing lines, is they look at the color, and if it’s this color, it’s probably this. If it’s this color, it’s probably that. People look at colors or chemical changes–the size of bubbles in the zinc, manufacturing process. There’s this one part of the process where you have to look at the foam that’s being generated. So, there is a lot of visual and auditory identification.

Winn Hardin: [00:19:51] That’s awesome.

Aaron Hand: [00:19:53] And one of the things I’d like to get into is kind of going back, you’ve talked about the experts that you have talked with in order to build your agent. You know, you’re talking with these guys who have worked in the industry for decades, i’m assuming. How do they feel about this conversation in which you are trying to learn how to do their job?

Kence Anderson: [00:20:13] Yes. I want to tell you a story about . . . It was actually that chemical plant I was telling you where it took seven years to to train and become an operator. They brought me in, and I was having a great time. I love learning about, you know, these new processes that, I mean, I’m an engineer at heart, these new processes that I haven’t known about. And I was wondering, I was thinking, you know, when is the shoe going to drop? And, you know, I’m going to realize these guys don’t want me here. It’s like, oh, this guy’s coming to take my job. And we sat down for lunch, and there was this gentleman who, he started off as what they call a junior maker that takes two years of apprenticeship and then seven years to become a senior maker. And then he started looking after multiple plants. And now he kind of ran optimization for all the plants. And I said, you know, how do you guys feel about this? And, you know, aren’t they, you know, afraid or paranoid or angry that I’m here? He said, let me show you something. He pulls over this clipboard, and all it had on it was a list of names. It was a two-column table: names, numbers. And he said, you know what these numbers mean? The numbers are how many hours it took to complete the changeover. So, the whole thing that they brought me in for was the control system. This was actually another extruder that makes films that go on your computer display. And the extruder is like as long as like three football fields or something, the entire process. But they said . . . Lost my train of thought.

Winn Hardin: [00:21:46] The film processing the numbers, the list.

Kence Anderson: [00:21:49] Thank you. Yep. Thank you.

Kence Anderson: [00:21:50] The changeover. So, the whole reason we were doing the I was because only a human could dial in the process during the changeover. And until you dialed in the process to the point where the quality was good, you’re literally sending the product to the grinder, literally. And the changeover could take anywhere from 2 hours to 47 hours. And so there were numbers that would say like Joe, 26. You know, Susan, two. And he said, if you can put an AI next to the operators so that they never have the number 46 next to them, then they’re all for it because they get judged after every single shift for how many hours of changeover. And that’s when I realized, oh, this isn’t about replacing people at all. This is about making people better at their jobs. So, only about 10% of these folks at every operation I go to are experts. Everyone else is not. They’re trying to get better, and the experts want to be freed up to do other things. They want a second opinion. The AI can totally give them a second opinion. We use this all the time.

Kence Anderson: [00:23:01] That’s why we use like Google Maps and Waze. Like we’re getting a second opinion. And it also frees them up to do a bunch of things that they really should be doing that only experts can do that they end up getting dragged into these kind of normal operations. And that’s my favorite part about this. When I first got into this, I thought, oh my gosh, what am I doing? And then I realized this is filling a crucial skills gap. You know, not to mention the fact that the younger generations don’t want to have anything to do with being in a factory. And one hypothesis I have, and I really hope this is true. One hypothesis I have is that if you give young engineers and maybe even, you know, folks that would go into a skilled trade, the opportunity to work with cutting-edge AI and say, you’re going to get to work with AI agents and you’re going to get to teach AI agents how to perform high-value skills, that will be a draw for people who aren’t going to come and work in the factory anyway.

Winn Hardin: [00:24:00] Yeah. Yeah, we’ve seen that across the board in terms of technology being a culture builder as opposed to an . . . especially for the newer generations. Yeah, there’s definitely shifting expectations about what a job is and what it should entail. And 100%. Do you find  . . . How how how often is Composabl successful when a project . . . Now I’m sure you’re evaluating the projects early on and saying, you know, I’m sorry, but this just really isn’t appropriate for us to try to. But I mean, like Gartner claims that up to 85% of AI projects fail. I think your success rate is going to be well beyond 15%. And I’d love to hear what tips you would give to the audience.

Kence Anderson: [00:24:37] Oh, that’s good. So, success is tricky to measure. The first tip I’ll give is identify clear benchmarks. That’s how you keep from becoming a science experiment. And I’ve, I’ve seen it all, you know, from my work previously in Microsoft and now Composabl. I’ve seen benchmarks getting moved. I mean, I did this project years ago where a CTO of a large industrial company for a business unit or a large industrial company said we have this particular problem. The PID controllers aren’t working very well. We beat the PID controllers by, like, 70% on the benchmark, but they wouldn’t lock in the benchmark. And this is . . .

Winn Hardin: [00:25:22] Mission creep.

[00:25:23] Yes. And there were folks that felt . . . This actually is  . . . What’s more common than operators feeling defensive, is engineers feeling defensive about the technology that is they’re kind of baby.

Winn Hardin: [00:25:37] Sure.

Kence Anderson: [00:25:37] And sure enough, the goalpost got moved and the problem got, statement got simplified, simplified, simplified. So, it looked like by the time we were done that we only beat the PID controller by 7%, even though we beat it by 70% according to the benchmark. Because the complex scenarios, which is why we brought in AI in the first place, were taken out. Now, I remember having a conversation with that executive saying, please don’t do this. This is a waste of both of our time. Because you brought me in to solve this problem. So, if you identify the benchmarks clearly up front, then it’s much easier to compare and contrast. I don’t like to play games like this technology is better than this technology and here’s why. That’s not how engineering works. So, it’s like, are nails better or screws better. Well, it depends on what you’re doing.

Winn Hardin: [00:26:30] Yes.

Kence Anderson: [00:26:30] If you’re building a home, and you need 10,000 fasteners to be inserted very quickly where there’s not a huge amount of strength required . . .

Winn Hardin: [00:26:44] Right. Or thermal expansion or other components.

Kence Anderson: [00:26:47] Then nails are clearly the better choice.

Winn Hardin: [00:26:49] Absolutely.

Kence Anderson: [00:26:49] But if you’re dealing with metal, if you’re dealing with the need for much more strength of adhesion in the joint, then screws are better. So, that’s how engineers think about things. What’s the right tool for the job? But sometimes we can get very defensive about whatever technology that we’re kind of interested in maybe because it’s new, or maybe because it’s exciting, or maybe because we spent 30 years working on it. And what I always try to help people understand is, if you set a benchmark, I have a three-step process. After you set your benchmark, one, you decompose your task into skills. That happens by interviewing the experts. The second thing is then you decide how these agents or skills should work together. Pieces of the process. And then third, and only third, do you choose the right technology for the job. And that keeps you from getting into these kind of tribal back-and-forth battles. About I like MPC control. I like real-time optimization, I like this. I like that. I like it all, actually. I like screws, gears, you know. I like all sorts of engineering components. And we’ll engineer the right thing for the job.

Kence Anderson: [00:27:54] One more thing I’ll say about KPIs. Sometimes KPIs are complex. It’s really difficult to distill down to one number sometimes. You know, how to measure a particular process, but you’ve got to do the best you can. You got to put a stake in the ground so that you can show that this beat a particular benchmark. Now, what I find is, if you go through that process in a genuine way, and if you pull in the experts, both the engineers and the operators in a curious way, then you’ll end up with a system that everyone is happy with expanding on. So, let me give you one example if I could.

Winn Hardin: [00:28:31] Sure.

Kence Anderson: [00:28:32] This is a.

Winn Hardin: [00:28:34] Is that talking about the importance of early collaboration amongst disciplines? Is that part i?

Kence Anderson: [00:28:38] One hundred percent.

Kence Anderson: [00:28:39] We deployed an agent in a factory in Ohio for a glass manufacturer last year. So, I would consider that successful. That stat you’re referring to is really talking about things that get built but never deployed. So, that’s definitely one benchmark, one, you know, key milestone is, is it fit for deployment? Did someone certify that this thing is ready to work in the factory? We did that by taking data, creating a simulation, just like I told you, by interviewing experts. Same thing as the extrusion example. This is a very different process of making glass containers, and it had never been automated before, ever. He told us about the agents and the skills. We put it all together, and at the end, there were six deployment tests. So, we put in the factory . . . We actually connected it to the real equipment. And each time, just like you would, you know, evaluate a new operator, it got an evaluation. A very detailed evaluation, and then we improved it, improved it, improved it. All along the way, the engineers were involved. They were driving the process, and the operator drove the process. And on the sixth deployment test, one, it taught him something that he didn’t even know was possible. These are huge spaces. I mean, you can spend your whole career exploring and still learn new things about how the machines work and operate. And then they certified that it was fit to work. Now, I’ll say one more thing. We can hold AI to a higher standard than we do humans. So, think about what we do with Tesla and self-driving cars. It doesn’t matter if it’s Tesla or another type of self-driving car.

Winn Hardin: [00:30:25] Sure.

Kence Anderson: [00:30:25] We put in the newspaper every single time there’s an accident. Now, imagine if we put it in the newspaper every single time there was an accident with a human driver. Now, you might say, oh, well, there’s a lot more human drivers, but . . .

Winn Hardin: [00:30:37] It would save the newspaper industry.

Kence Anderson: [00:30:40] We hold it to a different standard. It’s almost a standard of perfection. I actually was working with one of my logistics companies that deployed an AI in one of their warehouses. And I said, you have to look at this as if it was a human operator. If a human . . . and we were looking at the statistics. If a human operator was doing this, would you hire them? They’re like, Yeah. Well, there you go.

Winn Hardin: [00:31:03] There you go. And it’s interesting because all too often we hear, you know, the C-suite has decided has told us we need to use AI somewhere, right? So, the company and the engineering department starts looking for a use case. You know, give it a large enough data set, turn loose the AI deep learning algorithm, and somehow magic will come out on the back end. But what you’re describing is a much more strategic, analytical series process. Fully aware that there are human components have to be taken into consideration every bit as much as the the technology and the data that’s available. And so with iterative development and not just of a model, not just iterative development of a model, but optimization of a process, usually bringing together lots of different technologies.

Kence Anderson: [00:31:44] One hundred percent.

Winn Hardin: [00:31:45] So, I mean, it’s back to what automation and integration has always really been, which is applying so many different disciplines to best effect. As you said, it’s about solving the application. It’s not about turf protection or particular technologies benefit.

Kence Anderson: [00:32:01] It really is the last piece of guidance I’ll give today that I give a lot to folks when they say, well, how do I get started, is look at your highest value skills. What is the skill that if it disappeared because of retirement, your operation would stop. Okay, start there. Then cross-reference that with a higher ROI application. That’s where you want to apply. It doesn’t matter whether it’s generative AI or deep learning or computer vision. It doesn’t matter what kind of AI. It matters what is the high-value skill and how can you reproduce it?

Winn Hardin: [00:32:37] Right. Gotcha. Gotcha.

Aaron Hand: [00:32:39] That does get at a question I was thinking about was, you know, what is that first conversation about whether or not Composabl should even be involved in a process? So, you’ve outlined there what makes the most sense for them I guess operationally. Are there jobs where you would say, no, I don’t think this is really a good fit for AI. We can’t get the data in, I guess.

Kence Anderson: [00:33:07] Yeah, that’s a good question. There are one . . . and this is exactly where we started, where you asked if we were an AI company, and I think of us more as a, you know, a multi-agent system company or a controls company because I really got tired, actually, when I was at Microsoft of having these debates of, is this an AI problem. That’s like, you know, is this a gear problem? Is this a screw problem? Is this a nail problem? Even though there are . . .

Winn Hardin: [00:33:33] Problems in life, rarely have a singular cause. Correct.

Kence Anderson: [00:33:36] And complex engineering systems always have multiple components, which is why I talk about multi-agent systems because you have multiple components together. You could have a piece of traditional automation like PID or MPC controls, combined with a prediction algorithm with deep learning combined in the same system with a deep reinforcement learning AI agent. So, Composabl is really an orchestration platform for how to combine things together. But your question is still relevant. One more time.

Winn Hardin: [00:34:09] Okay.

Aaron Hand: [00:34:11] Are there times where somebody might present a case to you and you say, no.

Kence Anderson: [00:34:17] There’s what I call six, six superpowers. Five or six superpowers that humans bring to bear in decision making that you’re not going to find anywhere else except these kind of AI systems. One is perception, we talked about that. Two is strategy. Strategy is the ability to apply a completely different approach or run a different play in a different scenario. Third is adaptability. If you, and that usually requires learning, learning and practice. Adaptability means there’s a lot of different ways that this particular thing could go down. The fourth is forward planning, when you have to plan multiple steps ahead in order to succeed. Like if you do the right thing for what it looks like now, that might not be the right thing because you have to look ahead. That happens a lot in energy consumption, right? I know that the sun’s coming up and the temperature is going to change. So, it’s actually best to turn on the chiller now even though that might bring the temperature down a little bit more than you would think you would want to but because you know you’ll save energy later when the sun is beating down and energy costs more. That’s where forward planning or anticipation helps. And deduction is the fifth. That’s this ability to intuit, if you will, things that can’t be measured like that moisture in the cornmeal.

Kence Anderson: [00:35:37] And then the sixth is actually language. And that’s something that generative AI is amazing at. You know, the the ChatGPT and those large language models, they’re amazing at translating language into . . . Our language into something else, which might be action, which might be perception. And so, if those things are needed to make a good decision, you’ve got to have this. If none of them are needed, then you might have a situation where traditional automation and only traditional automation, you don’t even need this kind of AI-orchestrated system. When there’s simple, kind of more repetitive tasks, that is usually where you don’t need this kind of . . . And I like to make the distinction between tasks and skills. Skills are something you learn that’s like a very complex task. It requires practicing. A task is something smaller that can get checked off and got done. You know this these files got moved from here to here. These machines got turned on, these machines got turned off. But sometimes even inside tasks, there’s greater intelligence. Like, why do you turn these machines on and versus these other machines and when? Well, now you’ve got yourself into a skill. But if you’re talking about simple, repetitive tasks, that is where I would draw the line.

Aaron Hand: [00:36:59] Okay. Makes sense.

Winn Hardin: [00:37:00] Now, I think I’ve heard you talk about the six superpowers during some of the certified artificial intelligence courses that you’ve taught for A3 recently. Are you the only instructor to date on the . . . Is there an official name for that program yet? I wanted to call it the AI, I don’t know, what are you guys calling it?

Kence Anderson: [00:37:15] We’ll have to look it up on the agenda, but we are teaching two of those courses at Automate this week. One is Introduction to Industrial AI, and the other is Designing Autonomous AI Systems. I will have my head of learning, Sarah Cohen, with me teaching that first class with me. So that’s great.  But I’ll be teaching the second class, so.

Winn Hardin: [00:37:41] And I’m not sure, but I guess the ultimate goal is to get these classes online so it’ll be available to a larger audience, if I’m not mistaken.

Kence Anderson: [00:37:47] That’s right. A3 is working on that as we speak.

Winn Hardin: [00:37:48] So, I would very much encourage everyone to give A3 chance. Make sure you drop by A3.org Am I getting that URL correct?

Aaron Hand: [00:37:56] Automate.org.

[00:37:58] Okay. And look for the standards and resources available. And you can see perhaps some of Kence’s work. If not, keep an eye out for it. If you’re not here in Detroit in this in this particular week, you may not be able to see it in person, but it’s going to blow your mind when he talks about how he develops intelligent agents, and it’s going to make artificial intelligence become real. I think in a way, for me, that it hadn’t up until this point. It was still I mean, obviously I don’t have the technical acumen that you guys have, but for me, it was still a lot of smoke and mirrors. We’re not exactly sure . . . There’s a model . . . it gets trained where . . . it’s a mystery, but the way you break it down for me is . . . How you solve very specific industrial complex processes is awesome, and it’s very easy and intuitive to grasp with, with a little bit of technical background. So, I hope everyone will have time for that. And Kence, thank you so much for taking time today. I know we’re kicking off the show. It’s always a pleasure to see you, my friend.

Kence Anderson: [00:38:48] Absolutely.

Kence Anderson: [00:38:49] Have a safe trip back home.

Winn Hardin: [00:38:51] And enjoy the rest of Automate. And until that time, folks, thanks for joining us on “Manufacturing Matters.” If you’ve got any questions for Kence or for us, bleep-bloop those guys below, and we’ll make sure it gets to the right person. You can also see past episodes on manufacturing-matters.com or on your favorite podcast episode platform. So, for now, do us a favor. Subscribe, hit the likes, and be sure to tune in for our next episode when we’ll talk more about artificial intelligence, robotics, machine vision, and so many other important topics. Thanks.