Composabl's Kence Anderson on Why Manufacturing AI Agents Must Act, Not Just Perceive
Manufacturing Happy Hour
On an episode of the Manufacturing Happy Hour podcast, host Chris talked with Kence Anderson, CEO of Composabl, about what an AI agent means on a factory floor rather than in a sales or office-software setting. Anderson's position is that the valuable part of industrial AI is making decisions and taking action, not just analyzing or predicting. He argues that the best agents are built by combining machine data with the expertise of experienced operators, and that manufacturers need to capture that expertise before those operators retire.
An agent as a "virtual operator" that practices one task
Chris asked for a plain explanation, the kind you would give someone at a bar or a trade show happy hour. Anderson described an AI agent as "like a virtual operator" that practices accomplishing a specific task. He compared this to how people learn anything through practice, whether a football play, a chess strategy, or a recipe: they try something, get feedback, and try again. Operators of machines such as injection molding equipment do the same thing, and in some cases they spend decades mastering how to respond to different situations.
He stressed that expertise is task-specific. A person can be a great driver and a terrible chess player, or a strong basketball player and a poor musician. For that reason, Anderson said, he is not talking about artificial general intelligence. An agent does not need to do everything in a factory. It takes on one narrow task, gets feedback, shows it can perform well, and pursues a goal for that task. As an example, he said nobody would hire him to work in a steel mill, even though he considers himself reasonably intelligent. What qualifies someone for a job is what they have practiced and mastered.
Why engineers and the general public hear "agents" differently
Chris mentioned a recent presentation to manufacturing executives where fewer hands went up than he expected when he asked who had been hearing about AI agents. He noted that most agent talk concerns tools like Salesforce or online digital tasks, and asked how the idea applies to physical manufacturing.
Anderson offered a theory about why manufacturing people react differently. In the broader conversation, which he said has reached the World Economic Forum and the White House, people are excited for two reasons. The first is that AI can finally do something. He said this was partly tongue-in-cheek, but that earlier waves of AI and machine learning promised transformation through analysis and prediction. In his view, what actually moves the needle in enterprise and industrial settings is decisions, and prediction or analysis only matters when it serves a good decision. The second source of excitement is generative AI and large language models. People can talk to them in natural language and get a picture or a paragraph back, without learning a programming language.
Engineers see it differently. Anderson studied mechanical engineering, and he pointed out that software making real decisions is nothing new to engineers. He cited the proportional controller, which he dated to 1912, as evidence that engineers have automated decisions for a very long time. What interests him is that new enabling technologies may let AI make more humanlike, nuanced decisions that previously only people could make. An engineer's reaction to "AI that can make a decision," he said, is closer to "yes, that's what a control system is."
Perception versus action
Chris asked how an agent differs from the machine learning manufacturers already know, such as camera systems that detect quality defects. Anderson framed it as the difference between perception and action.
He said most factory machine learning, which he called very important work, is perception. Detecting a defect is visual perception, and he has also seen auditory perception. He counts prediction as perception too, and gave the example of a farmer who feels the wind, looks at the shape and color of the clouds and sky, and concludes it's about to rain. The farmer is combining many signals into one judgment. Predictive quality belongs in the same category.
Action is deciding what to do. Anderson described working with someone at a large consumer packaged goods company who had built models to detect defects in products and deployed around a hundred of them across the company. Once defects could be detected, people immediately started asking what to do about them, which Anderson said is exactly the question agents address. He referred to a reference architecture he uses for intelligent agents that includes both perception and action, and said an agent "must be able to take action." Chris summarized: where a machine learning model might report three defects out of a thousand parts, an agent would respond with an action. Anderson agreed and gave examples: how to change set points to eliminate a defect, how to change them to prevent it, and what actions would change the situation.
From a Ford internship to eight startups
Anderson said he somewhat fell into mechanical engineering. As a kid he loved cars, space, rockets, and trains, and people told him that meant he should study it. He values the degree because, in his words, it teaches problem-solving along with its two basic flavors, design and manufacturing.
During summers after his freshman and sophomore years, he interned at Ford's world headquarters in Dearborn, and he called this his first introduction to innovation. A group there had invented a new kind of air-conditioning compressor. Although he had barely taken any classes, he was paired with two MIT graduate students and given freedom to redesign the manufacturing system. He said they were the first at Ford to do a form of rapid manufacturing, and they wrote a paper about it. Test parts had previously been sand cast, which took weeks. The team instead sent CAD-generated machining instructions over the phone line to a machine tool in Connersville, Indiana, and got parts in hours. His lesson from this was that innovation doesn't have to come from a seasoned engineer or any particular place.
He also said mechanical engineering is grounded in physical reality. Parts break, vibrate, and make noise, and stress, strain, and tensile strength govern everything. He said this background later gave him a different perspective on AI.
After college he worked as a software engineer at IBM. He moved from Boston to the Bay Area in 1998, just before the peak of the internet boom, and was captivated by the startup scene, where it seemed someone on the train was inventing something every day. Most of his career has been in startups, and Composabl is the eighth he has worked for.
The common thread: platforms need a methodology
Asked what ties the eight startups together, Anderson said they have all been horizontal B2B platforms. He has never worked on a consumer product or a single-solution product. His hypothesis is that platforms in novel, high-value technology areas require a methodology before they can be widely adopted.
His examples: prompt engineering is a methodology for getting useful output from large language models through natural language. Excel is a methodology for organizing data in what we now call a spreadsheet. Before that, when few people worked on databases, structured query language did the same job. His first five startups were ad and marketing technology platforms, and across all of them he found that broad adoption depended on a methodology. He is also drawn to the idea that a platform in the right hands can produce almost unlimited solutions, far more than any single company could build.
Bonsai, Microsoft, and the idea of teaching AI
Anderson said the move into industrial AI was full circle for him and that he has been doing this work for about seven years. Composabl is two years old, but the work started at Bonsai, a startup he participated in but did not found, which Microsoft acquired. He continued the work at Microsoft. Across those three companies, he said, his obsessions were three ideas.
The first is AI that can do something, meaning it can act in the physical world. He recalled that in 2017 everyone seemed focused on making predictions. The second is AI-based control systems making real decisions in physical settings. The third is that if algorithms can learn, "you better teach it something."
He said the teaching idea was not well received at first, because some people thought teaching would restrict the AI and keep it from doing something amazing. He rejected that view: "nothing intelligent has ever been restricted by a good teacher ever," whether human or machine. He cited DeepSeek. He acknowledged the controversy around it but said that, reading the paper, what the developers were doing amounted to teaching, putting boundaries around specific areas rather than letting the model read the internet and learn whatever it would.
Lots of machine data, scarce expertise
Chris noted that manufacturing is often called an AI sweet spot because of how much data it generates. Anderson said there are two sides to this. He recalled a wave of IoT and digital transformation starting about ten years earlier that sometimes sounded like "get out of the stone age by measuring everything in your factory." When he visited factories, people pushed back and pointed to their historians, time-series databases that track everything happening on the machines. Adding sensors or cameras may be legitimate, he said, but factories have been measuring things for a long time.
The data that is hard to get is the expertise that lives in people. His favorite illustration comes from chess. He described a 2016 agent, which he called "Alpha Chess," and its companion AlphaGo, which learned by playing against themselves. The chess agent beat many chess masters, and AlphaGo beat the top Go player of the time. What struck him was that the chess agent discovered and used the 12 most common opening strategies, at least one of which is a thousand years old. Millions of dollars of compute went into rediscovering what people had known for centuries.
He said factories are similar. An AI that only studies your data might eventually discover things that "Susan and Joe over there could have told you in 15 minutes." That is why he argues for capturing and codifying high-value expertise before it disappears. Some of it is proprietary and central to the company's operations. It has to be captured alongside the machine data: you can't just read the data the way ChatGPT reads the internet. You have to study the data, interview the experts, and match the two together.
Chris called capturing this expertise one of the most underused AI strategies, given how often the industry talks about retiring workers. He corrected himself from "data" to "expertise," and Anderson agreed, adding that the expertise would be amplified and explained by the machine data.
The Cheetos extruder case
Chris asked about an example Anderson has often used: an extruder making Cheetos. Anderson said consumers rarely appreciate how hard it is to make everyday products. Around 2018 he got a call about the problem and first learned that PepsiCo makes Cheetos along with Doritos, Fritos, Funyuns, Lay's, and other snacks.
He explained an extruder with the Play-Doh toy: dough goes in one side, you pull a lever, and it comes out shaped like a star or circle. Extruders are used for soap, pet food, snack foods, films, sheets, plastics, and more. For Cheetos, cornmeal goes into a long metal tube with a screw inside, where it is heated and cooked, then forced through a small slit. The pressure difference across the slit makes it puff up. The company told him it takes about ten years to become good at running it.
A big reason is variation in the input. Corn differs depending on whether it grew in Iowa or Illinois, whether the season was rainy, and on the soil, and those differences change its moisture and fat content. Anderson said these properties can't be measured in the process. To people who say "just measure everything," he answered that some things can never be measured, and that operating through them takes practiced expertise. The extruder had more than 25 control variables and more than 50 sensor variables. The product also had to meet quality criteria for length, diameter, curl, internal bubbles, and bulk density. Operators needed a long time to learn all the situations, including how to compensate for missing information like corn moisture.
Simulation, practice, and then the operator's insight
Anderson described the project in order. The team first collected a lot of machine data and built a simulation. He explained why simulation matters: you don't want to practice on real equipment and, in a steel mill for example, make bad steel or endanger people. Practicing on the real thing would also take too long. A simulation can be built from data or from engineering equations, and the agent practices on it. In the cloud, the agent can practice on many copies at once, the way a chess player plays seven boards simultaneously, so it could practice on ten extruders at the same time. That means more practice in less time. He also said you don't want to train only on historical data, because the agent might memorize what to do in your specific past situations. You want an effectively unlimited supply of representative data so scenarios can be varied.
The first agent learned only by practicing, with little teaching, and Anderson said the result was "just okay." It reached expert level after the team interviewed an operator, who told them the job required five distinct skills. Anderson compared this to a soccer team. The goalie's whole job is to stop the other side from scoring, and the striker's job is to score. They are on the same team with opposite roles, and different skills apply at different times. He used to call these "skills" but now calls them "skill agents": individual agents that together form a team controlling the extruder. When the team built the system that way, he said, the result was expert control.
Chris summarized that the 25-plus knobs are in some sense the easy part, and the hard part is what can't be measured, which an experienced operator judges by looking at the output and inferring what the batch of raw material is like. Anderson said every process has this, including mining. He described a conversation that morning about cheese making, where a consulting company noted that milk varies widely in fat and sugar depending on the cow and its diet, and that milk prices go up and down. In his view, variation in inputs drives much of the optimization manufacturers are looking for. Chris encouraged listeners to identify the variables in their own processes that they can't measure today.
Augmenting operators, not replacing them
Anderson said he has built more than 200 of these agents over seven years, mostly for companies he described as roughly Fortune 500 scale, and that "hardly anyone wants to replace anybody." According to him, customers want agents to train new operators, help novice operators perform better, and give experts either a second opinion or time back from routine work so they can do more valuable things.
Chris asked where an agent physically lives, so listeners wouldn't picture "some orb" above the extruder. Anderson said it is a small piece of software deployed on a commodity Linux server, often under a desk in the control room, inside the plant's edge infrastructure on the OT network. It connects to machines through the standard protocols controllers already use, such as MQTT and OPC UA. It can be wired into the HMI so operators see its recommendation on the same screen as the sensor variables, or into the DCS for closed-loop control.
How the agent teaches
Chris asked how an agent can teach less experienced operators. Anderson described two ways.
The first works like Google Maps, Apple Maps, or Waze. Even in the Bay Area where he lives, following navigation has taught him roads and timing based on traffic patterns he didn't know. He compared this to learning Texas hold'em during the 2000s poker craze from a Doyle Brunson book. The book listed the top ten hands and said to play them and fold everything else. Anderson played that way for about six weeks at a weekly game he attended for about a year. He said the rule was never meant to be permanent. It was a lesson-based starting point that led to decent results while he built his own judgment, noticed exceptions, and asked others. Following an agent's recommendations can work the same way for a new operator.
The second way is through the skills themselves. He said the chess agent was described as having an "alien" playing style because it did not follow the strategic structures humans use, which made it not very useful for learning chess. What helps learners is a system whose agents represent recognizable skills. He noted that a line supervisor can't monitor thousands of variables. What they watch is whether an operator is using the right strategy at the right time. If a multi-agent system has skill agents with distinct strategies, like strikers and goalies, or in American football terms defensive plays, running plays, passing plays, and reverses, experts can use it to supervise and train newer people, and novices can learn by watching which strategy the AI chooses. Chris summarized this as building agents that follow learnable best practices rather than alien strategies. Anderson added that the agent should know those strategies, use them, and make them transparent.
Where manufacturers should start
Asked for first steps, Anderson gave a sequence. First, identify the highest-value skills most at risk of "extinction." If 50 operators in their twenties work a given process, that is not the place to start. Start with the valuable skills held by the fewest people who are closest to retirement.
Second, decide who owns innovation. Larger companies may have innovation departments or manufacturing R&D that can take on codifying skills and building AI systems. In operations without a dedicated group, someone still needs the authority and the time. He said he has worked with very large companies whose engineers had no time to innovate, so they ended up doing "stupid POCs" rather than anything meaningful.
Third, give those people tools. Anderson recommended a platform they can use to build agents themselves rather than buying point solutions, because there are too many use cases. One manufacturer told him it had a thousand use cases for intelligent agents, and nobody buys a thousand separate solutions. A platform lets a company build agents for high-value skills in priority order. Chris noted that Composabl is such a platform, and Anderson acknowledged it.
Closing: empower the engineers who already have the expertise
For his final remarks, Anderson told manufacturers that they already have the expertise: operators hold it and engineers know the process. He urged companies to empower their engineers to build intelligent systems, since engineers designed and built everything else in their plants. He questioned the tech-industry assumption that industries need to be disrupted ("well, maybe not"), and said he believes there are 100 million engineers in industry who, given the right tools, will engineer intelligent systems the same way they have engineered everything else.
Kence, good to have you here at Manufacturing Happy Hour, and I imagine your ears may have been burning lately because I've been mentioning your name a lot in a lot of conversations around AI agents. So, good to have you here to hear it from the source in today's episode, if you will.
Yeah, thanks for having me.
Well, my first question, in the spirit of Manufacturing Happy Hour, as if we're having this conversation over a beverage. Let's say you're at a bar out there in the Bay Area or you're at the trade show happy hour and someone asks you, "Kence, what in the world is an AI agent?" How do you describe that to them in the context of manufacturing, in the context of an over-a-drink scenario? Candid and to the point.
Yeah, absolutely. An AI agent is a... it's like a virtual operator that practices accomplishing some specific tasks. So, you know, you're an operator of an injection molding machine. It's not really that much different than any other task that we practice, you know, whether it's practicing a football play, a chess strategy, you know, practicing, you know, almost cooking, you know, some sort of recipe. There's this aspect of trying, getting feedback, and practicing. That's what operators do. They spend decades in some cases mastering what happens in different situations as you, you know, play the game, and that's true for a lot of things.
And so an agent is something that is going to master a task, a very specific task. Now remember, expertise is task specific. So Chris, you could be a great driver but a terrible chess player. And I could be, you know, a great basketball player but a terrible musician, because expertise is task specific. So we're not talking about what people call, you know, artificial general intelligence. You don't have to do everything in a factory. In fact, you know, no one's going to hire me to work in a steel mill, even though I'm pretty intelligent. So are you. It's actually what we practice and what we've mastered. So, an agent is taking on a specific task, getting feedback and proving that they can perform well, pursuing a goal for that specific task.
And we're going to take a deeper dive into this about halfway through the conversation. I want to ask you about your background a little bit, but I think the first question I need to ask before we get there is: I hear about AI agents a lot these days, and I feel like a lot of folks are, but I could be wrong. We were chatting before the interview. I was giving a presentation recently to a room full of manufacturing executives. And maybe it was the fact that we were ordering food at the time and things like that and people were a little bit distracted. But when I was asking, "Hey, who's been hearing a lot about AI agents lately?" I didn't see as many hands go up as I thought. So my question is, I hear about AI agents a lot in the context of, like, Salesforce, for example, or the context of someone that's going to help you with a digital task online, if you will. Can you frame an AI agent real quick in the context of manufacturing, like physical manufacturing?
Yeah, that's actually a really good question, and I have an idea about that I'll test out with you about why people in manufacturing are thinking about agents differently than people at large. Right? So right now in the general conversation, you know, in society, almost everyone's talking about agents. Like, you know, they talked about agents at the World Economic Forum, they're talking about agents at the White House, you know, these pieces of software that can do something. And from that perspective, people are really excited for two reasons. One is people are thinking, "Wow, there's AI that can finally do something." And I say that kind of tongue in cheek, but not really, because previous waves of AI and ML were about, okay, your life is going to change if you can analyze data. Well, kind of. I mean, you know, if you can predict this, it's going to change everything. Well, not really. I mean, in certainly every enterprise and industrial situation, the thing that moves the needle is decisions. Great decisions. And so if you're predicting something, analyzing something, it's only to serve the purpose of making a great decision. So some people are very excited because they're like, "Oh man, there's AI that can actually make decisions or actually do things."
The second thing that captured the general imagination is, you know, the generative AI, the large language model, this idea that there's this AI that I can talk to. Like, I can ask it a question in natural language and it will draw a picture for me, write a paragraph. And so it really sparked people's imaginations, like it's AI that I can actually interact with. I don't have to learn some programming language or something to make it do something, and then it can do something.
Now here's the other side of the coin. If you're in engineering, like I studied mechanical engineering, if you're in engineering, the thought that there's a piece of software that can make a real decision is not novel. I mean, the PID controller was invented in 1912. So we've been automating things as engineers for a very, very, very long time. So what's interesting to me about that is, okay, now there's certain enabling technologies that allow AI to maybe make more humanlike decisions or more nuanced decisions or more decisions in manufacturing that previously only human beings could make. But no engineer is going to, you know, jump out of their chair and go, "Oh, there's an AI that could actually make a decision." They're like, "Yeah, that's what a control system is." So there's kind of two ways to think about it.
I like the way you phrase that. I also like your example of going back to a PID controller, where we've had instrumentation and controls that have helped manufacturers make decisions for a long time. The way I want to maybe frame this up in the context of an application is if you could describe how an AI agent is different from what I think manufacturers have recently thought of when it comes to AI in the context of machine learning, right? Looking for quality defects, if you will. You know, I think that's an example manufacturers are familiar with. Hey, there's a camera. It's looking for defects. It's been given some parameters to help pick up those defects and learn on the way. Machine learning. How does an AI agent take this to a different level relative to what we've thought of as kind of a default AI machine learning example for the past few years?
Beautiful. Chris, it's the difference between perception and action. So most of what people are doing with machine learning in factories, and it's very important work, is perception. Can I identify a defect? That's visual perception. You know, I've seen auditory perception. Can I hear... or even prediction. You say, "Is prediction perception?" Yeah, I think so. When a farmer, you know, kind of does this and feels the wind direction and looks at the shape of the clouds and the color of the clouds and the color of the sky and says it's about to rain, that's a prediction, but it's a perception. They're combining all this information into a perception about whether it's going to rain. So predictive quality, all those kinds of things, that's perception. Action is what to do.
And so, for example, I worked with someone who is at a large CPG company, and he had developed some perceptive models, some that could detect defects in these consumer product goods. And he actually deployed many, many of these, like a hundred of these, within their company, and then the very next thing people started saying was, well, if I can detect that the defect is happening, I need to do something about it. So the very next thing people start asking about is what agents do. So what am I going to do about it? And so if you look at a reference architecture I have, which we may talk about later in the podcast, for intelligent agents, there should be perception, but there also has to be action. An agent must be able to take action.
So action is what you're saying is the big differentiator here. Am I hearing that correctly? Where in the past, you know, you would be able to get all that information from a machine learning algorithm. You'd be able to say, "Okay, here were the three defects out of the thousand parts that were just produced." Now, with an AI agent, it is going to give you some sort of action to take in response to that.
Yes. How do I change the set points to eliminate that defect? How do I change the set points to prevent that defect? You know, what actions do I need to take to change this situation?
I love it. I had a feeling this might be a little bit of an extended intro today. So this is good. We're going to look under the hood here in a second, but you have a very long, fascinating career in the manufacturing and startup world and beyond. So not everything you've done has been in manufacturing, but you got your start in the manufacturing world, you were telling me. Give us where Kence's career got started and how things have led to today. And I'll probably ask some questions along the way.
Sure. I studied mechanical engineering in college. I kind of fell into it. I've always loved... I love cars. You know, as a kid, I was super into space and rockets and trains, and people said, "Well, if you're into that kind of stuff, you should study mechanical engineering." Sure. I don't really know what that is, but sounds great. I'm glad I studied mechanical engineering. Mechanical engineering teaches you how to solve problems. It teaches you design and manufacturing. Those are the two kind of basic flavors. And there's a lot of branches of mechanical engineering, like aerospace engineering is a branch of mechanical engineering, and, you know, there's quite a few of them.
My freshman and sophomore year, during the summer, I interned at Ford. So I was at world headquarters in Dearborn, and I had an amazing opportunity. This was actually my first introduction to innovation also. I went out and there was a group there that had invented a new kind of air conditioning compressor, and here I am, this freshman. I mean, I haven't even hardly taken any classes yet. And they paired me up with two MIT grad students, and they let us loose on redesigning the manufacturing system, and we were the first ones at Ford, and we wrote a paper about it, that was able to do rapid manufacturing. So to get a new part to test, they used to have to sand cast, which would take weeks. And we took a CAD program where the machine tool would cut the part, and we'd send the instructions over the phone line to the machine in Connersville, Indiana, and get the part made in hours. And I was able to do that as a freshman, you know, in college. And that taught me that innovation doesn't have to come from a specific place. Like, you'd think that it would be some, you know, very well-seasoned engineer. You'd think that for a variety of different reasons I wasn't the person who should be involved in doing that. And it taught me a lot about innovation. I really appreciated that opportunity.
But also, mechanical engineering is really grounded in physical realities, like the part's going to work or not. There's tensile strength and there's, you know, stress and strain. There's so many things that, like, the thing will break. The thing will vibrate and make noise. The physical world affects everything you do. And so you'll see later when I talk about how that affected my career in AI, it gives you a different perspective when you come from that kind of physical world.
But I spent almost my entire career in startups, though. So I graduated from college, ended up getting an opportunity to work as a software engineer at IBM. And when I moved to the Bay Area, I was fascinated by the technology. I didn't know anything. I'm from Boston. I didn't know anything about San Francisco. It was like Bryce and the Golden Gate Bridge. That's about all I knew. And the Silicon Valley startup and technology scene... I got here right before, you know, kind of the height of the internet boom in 1998, and I was fascinated. I was like, people are inventing things, you know, every day. You know, every day you're riding on the train with someone who's inventing something amazing. And I really got into this kind of entrepreneurial thing. And so I've mostly had a career in startups. Composabl is the eighth startup that I've worked for. But it all started with engineering.
I have to ask, then, what has been a common thread, or the common thread, across these eight startups?
That's a great... Oh yeah, it's a great question, but it's easy. They're all platforms. So I've actually never worked on a consumer product, like a software product that a consumer uses. They've all been B2B platforms that were very horizontal. I've never actually worked for a kind of solution platform either. It's always been horizontal, a platform that enables many users to do many different things in a kind of category or class of problems. Which to me, that's fascinating, because my hypothesis is that platforms in novel, high-value technology areas require a methodology. So what do I mean by that? Well, a great example of that is prompt engineering. Prompt engineering is a methodology for getting generative large language models to, you know, produce something useful for you through natural language. That's a methodology. In some ways, like, you know, Excel is a methodology. It's like a methodology for organizing data in what, you know, we now call a spreadsheet. But that was very, very novel. And so when databases were early, very, very, very few people worked on databases at the beginning. But then in order to get massive widespread adoption, you have to have a platform that implements a methodology, and that's kind of what Excel is, or even before Excel, structured query language is.
And so I became really fascinated with... regardless of the platform — my first five startups I worked for were platforms in ad and marketing technology — so regardless of the type of platform, widespread adoption requires some sort of methodology. I became fascinated with that, and I was fascinated with the fact that a platform can produce almost an infinite amount of solutions. So if you give a platform to the right audience, they'll be able to produce, you know, exponentially more than any one company could ever produce in that area. That's kind of the theme of all the startups I've worked on.
I'm curious if this platforms-requiring-a-methodology has something to do with this next question I'm going to ask. What led you then, in this most recent venture, to focus on artificial intelligence in the manufacturing industry specifically?
It's a bit full circle for me, and I've been doing this for the last seven years. Composabl was only two years old, but it started at the last startup I worked at, which was actually not my startup, but that I participated in. It was a startup called Bonsai, and it was acquired by Microsoft, and then so at Microsoft. So really at three different companies... it was full circle for me, this idea that, okay, now there's AI that can do something. That means you could do something in the physical world. And I was like, now that's interesting, because at the time, I don't know if you remember this, but in 2017 everyone was obsessed with predictions. Like, just make a prediction, just make a prediction, it'll change everything, just make a prediction. I'm like, oh gosh. And then when I realized, oh, there's AI that can actually, you know, make these real decisions in the physical world, you know, control systems, if you will, I
was fascinated. The third thing I was fascinated by was, well, if AI can learn, you should probably teach it something, right? Like, if there's algorithms that can learn, well, you better teach it something. And that was honestly — that idea was not super well-received when I first started working on it. There was this thought that you're gonna restrict the AI, you're going to keep the AI from doing something amazing by teaching it. No. No, nothing intelligent has ever been restricted by a good teacher ever. That's a human being or a machine.
And even if you look at things like very, very recent developments like DeepSeek, and there's a lot of controversy about DeepSeek, but if you read the paper and you see what they were doing, what they're doing is teaching. They're going, "Hey, let's draw boundary boxes around, you know, different areas instead of just letting it read the internet and kind of learn whatever it's going to learn. Let's teach it some specific things." And that's actually a lot of what I've been doing over the past few years. But it's the idea of an AI-based control system controlling things in the physical world, AI that can do something, and, you know, teaching AI something — those have been kind of my obsessions for the last seven years.
Yeah, I like your comment that nothing intelligent has ever been restricted by a good teacher. You know, just from my research and my preparation looking at AI, listening to some of your past podcasts and presentations, another thing that sticks out that I'm not sure every manufacturer grasps yet is one of the things that makes a particular industry, a particular vertical so ripe for artificial intelligence is when there's just a ton of data in that space as well. And that's one of the reasons that I'm hearing not just yourself, but other people are doubling down around artificial intelligence in manufacturing, is because of just the vast swaths of data that are there. Can you give us a little more context on the mass amounts of data that are just available to manufacturers right now and why it's an AI sweet spot?
It's super interesting you say it that way because there's two sides of it. There is vast amounts of data. It's funny. There was this wave of IoT. I'd say about 10 years ago, it started this wave of IoT and digital transformation that to me sometimes sounded like, hey, get out of the stone age by measuring everything in your factory. But then when I went around to a bunch of factories, people were like, "What are you talking about?" Like, we have this thing called a historian. It's a time series database that tracks everything that happens on these machines. Like, yeah, we might need to add new sensors or add some cameras, which is a totally legitimate point, but we've been measuring things for a really long time here. So, there's a lot of data in what are called historians. They're these databases for factory information.
Juxtapose that, though, with the kind of data that's really hard to come by, sitting in your human beings, and I call that expertise. So my favorite example of this is actually not from manufacturing. It's from chess. So back in 2016 there was an agent called Alpha Chess and there was a companion agent called AlphaGo, and later there was one that played multiple games, but that learned to play chess and Go just by practicing against itself. And it beat many, many, many chess masters, and the AlphaGo version beat the best Go player in the world, at the time at least. There's a whole movie about it. But what was interesting about chess is it discovered and used the 12 most common opening strategies in chess. At least one of those strategies is a thousand years old. And so you go, "Oh my gosh, they spent millions of dollars of compute to have this thing practice and learn some stuff that we've known for a very, very, very long time."
That's kind of how it is in factories. You could have an AI just look at your data, look at your data, look at your data, and discover a bunch of stuff that, you know, Susan and Joe over there could have told you in 15 minutes. And so there's this need to capture and codify high-value subject matter expertise, frankly, before it goes extinct. I mean, before it's gone forever. Some of it's very, very proprietary too and very valuable. It's core to the company's operation. And you have to do that in the context of this vast amount of data. So, you're not just going to be able to look at the data like ChatGPT reads the internet. You're going to have to look at the data and interview the experts and then match those two things up together.
I'm going to give a little spoiler to one of the big lessons from this conversation today, which you've already started hinting at, and I think one of the most underutilized strategies around artificial intelligence right now for manufacturers is building up that database of expertise, as you said, the data that's sitting inside of the humans, before — as we hear about all this time on this show, at trade shows, wherever you go, people are talking about how people are retiring, and we're going to lose a lot of that expertise in the manufacturing space. So, you know, I know we need to start capturing that data and putting it to use before those folks walk out the door. I shouldn't call it data. We need to start capturing that expertise before those folks walk out the door.
Yes. And it will be amplified and explained and fleshed out by the data that you have in the machines.
You know, as we get into this conversation, I want to ask you about an example that I've heard you talk about multiple times over in terms of what is an AI agent and how does it really work, and we're going to tie it into, you know, how to get that database of skills built. You've talked about an example of an extruder for making Cheetos a lot. That was — I think is just a perfect example for an AI agent application because of the variables that go into this piece of equipment to get that perfect poof and that perfect crunch inside of a Cheeto. So can you describe that? I've given folks a preview there. But what is this example and why is it such a sweet spot for an AI agent?
It's funny. I was just on a conversation with a major industrial company and they were highlighting exactly what you're talking about, because if you're not manufacturing things, it's hard for us to understand, when we consume these products, the difficulty in making them and how challenging it is. So, you know, we eat Cheetos and, you know, my son loves Cheetos, and it's like, well, how hard could that be? And it was about 2018 when I got a call and I got explained. First of all, I didn't even know that Pepsi made Cheetos, but they make Cheetos and Doritos and Fritos and Funyuns and Lay's and all these snack foods. Okay, cool.
And so what's so hard about making Cheetos? Well, first of all, Cheetos is made on an extruder. So, what's an extruder? Remember those Play-Doh toys where you would put the dough in one side and then you'd pull a lever and then it would get pushed out in the shape of a star or a circle or a hexagon? That's an extruder. That's literally an extruder. So extruders are used to make soap, pet food, snack foods, films, you know, sheets, plastic, all sorts of things. And so the cornmeal goes into the extruder, and the extruder is basically a long metal tube with a screw inside, and extruders always heat up what's inside. So you have to melt it, or in this case cook it. So the cornmeal actually gets cooked in there and then it gets forced out — always forced out a slit, just like the Play-Doh. And because the slit is small, there's high pressure on one side and low pressure on the other side, it poofs, it puffs up.
And I said, "Well, okay, that sounds difficult." And they said, "Yeah, it takes like 10 years to get good at things changing." Now, one of the things in manufacturing that not a lot of us think about is the composition of the input material is always changing. The corn is different depending on whether it grew in Iowa or Illinois or wherever, whether the season was rainy — I mean, there's lots of different things, the soil — that will make the corn different: moisture different, fat. And you can't measure these things. So to all the digital transformation IoT folks that are like, oh, just measure everything, it'll work out — no, no, no. There's some things you're never going to be able to measure. And so, it's actually expertise of practice, and this particular extruder had more than 25 knobs on it, more than 25 control variables that you change, and over 50 sensor variables that you're reading. That's a lot of information to measure in real time, and then, you know, to track the quality measurements, and every manufacturing operation has quality measurements. In this case, the Cheeto has to be within certain length criteria, certain diameter criteria, has certain curl criteria. Even the bubbles inside of it, the bulk density of it, needs to be within parameters. And so it took operators a very, very, very long time to practice with all those different situations, getting the things right, including the information that was frankly missing, like the moisture level in the corn.
And so the first step in solving that problem was, yes, we did need to take data for sure. We took a bunch of machine data, but then I interviewed — things got really interesting when — so let me, sorry, let me tell you what happened. First thing, we took a bunch of data, we created a simulation. Okay, why do you care about a simulation? Well, let's pretend you're in a steel mill. You don't want to make a bunch of bad steel, potentially hurt people, by practicing on the real thing. And plus it would take forever practicing on the real thing. So you practice on a simulation. So you use data or you use engineering equations to create a simulation, and then the agent gets to practice. Now the agent can practice on the cloud. The agent can practice like playing a bunch of games at once. It's like — I don't know if you've ever seen people play seven chess games at once and they go boop boop boop boop boop boop and they play seven opponents. You can do that in the cloud, right? So the AI can practice on 10 extruders at the same time. The computer doesn't care. So it's more practice and less time. And also you don't want to just train on the data because it could memorize what to do in your specific situation. You want an infinite amount of representative data so that you can modify, you know, the situation and the scenario.
Okay, cool. And we trained an AI that just learned by practicing, and we didn't really teach it much, and it was okay. I mean, it was just okay. Things got really, really interesting, to kind of an expert level, when we interviewed an operator and the operator said, "Oh, did you know there's five different skills that are required?" It's almost like a soccer team — and I think of these things as teams of agents, actually — because on a soccer team the goalie does something that is literally the opposite of what a striker does. Like the opposite. The goalie's 100% job is to keep people from scoring. The striker doesn't care about that at all. The striker's job is to score goals. So, they're on the same team, but they have completely different roles. And that's how these skills work, and you use different skills at different times. So, I used to call them skills, but now I call them skill agents. These things are individual agents that form a team to control this extruder. And when we did it that way, bam, expert control. Expert control.
Excellent example. And I'm going to try to do my quick over-a-beer play-by-play really quickly on some of my main takeaways from this. When we're talking about an extruder, one of the things that you mentioned is, hey, there are like 25 knobs, 25 variables that, to an extent, that's the easy part of this process, but the hard part is the stuff that you really can't measure — in this case, the moisture and the corn composition, which, you know, it just takes an experienced operator to look at that when it's coming out of the extruder to say, "Oh, based on what I'm seeing right here, we're going to adjust this, because now I kind of have a feel for what this batch of corn, or more generally this batch of raw materials, was like." And you're —
Yes. And that's the amazing thing about manufacturing, is every process has that. It doesn't matter — even mining, think about it — there's always a different composition. Like, I was just talking to a manufacturer of cheese this morning, and they were saying — actually the consulting company was saying — in making cheese the main ingredient is milk, and milk composition changes tremendously. It has wildly different amounts of fat depending on, you know, a variety of things about the cow and what the cow is eating and all that, and different amounts of sugar, and different prices. Milk prices go up and down. And so the variation on the input is what really drives a lot of the improvement and optimization that people are looking for in manufacturing.
And I like that you bring up this dairy example as well, because I want the manufacturers out there listening to this to be thinking of this in the context of their process. What are the things there that are variables that you really can't measure today? And I love your goalie-striker example in the context of, hey, you're on a team, but these two folks on the same team are doing things polar opposite, completely different. And that's where these skills that you interview for come into play. So you build these — in this case you said there are five skill agents — and that's what allows an AI agent to do what in the past was only able to be done by your best operator at the facility.
And I should say something too, that I've done this a lot over the last seven years. I mean, over 200 of these agents for Fortune 500-ish companies, and hardly anyone wants to replace anybody. What people want to do is they want to use these agents to train new operators and to help novice operators do better, and to give expert operators either a second opinion or time back from things that were more mundane, so that now they can go do things that are more valuable.
I love — we're absolutely going to dive into this. Before we get there, I have to ask, where does this AI agent live? I just want to make sure manufacturers aren't thinking this is like some orb that's sitting on top of the extruder, that's like beamed in from space and all that kind of stuff.
No, exactly. It is literally a small, you know, piece of software that's deployed on your commodity, you know, Linux server, usually sitting under the desk in the control room, you know, wired into the — it's all inside your edge infrastructure, in your kind of OT network, and wired to the machine through your kind of standard protocols that your controllers already use. Yeah, MQTT, OPC, OPC UA. And it's wired into each of the HMIs so the operator can see the recommendation at the same time as they see the sensor variables, on the same screen that they currently do. Or it could be wired into the DCS so we can actually accomplish closed-loop control.
So, long story short, it is attached — you're saying it lives in this Linux server, but it is attached to be able to get the data and the variables you need from the various systems within your facility. And you can also teach it those different skills as you learn that from your operators. Did I capture that correctly?
Yes, sir. Yeah.
Excellent. Well, you know, the main thing I want to get to then is something very important that you just mentioned, and that, you know, not only does this allow an agent to in some ways have that knowledge that your best operator that
might be retiring very soon has, it can also teach that next generation of operators that don't have that 10, 20 years of experience. Go into that a little bit more. We've talked about how the AI agent learns now. How does the AI agent teach?
Yes. That's really good. So at the most basic level, you can use it the similar way to the way we use Google Maps or Apple Maps or Waze. Even in the Bay Area, you know, I live in the San Francisco Bay Area. I'm learning new ways to navigate, which roads to travel at different times based on different traffic patterns that I didn't actually know. And so you're kind of following along, and this is how we learn and develop expertise.
So when I first learned how to play poker, I don't know if you remember, Chris, there was a big craze in the 2000s about playing Texas hold'em poker, and I read this book by Doyle Brunson, and he basically said, "Here's the top 10 hands of poker, and here's how you practice: if you get one of those top hands, play it, and if you don't, fold it." And I did that. I went to a game every week for about a year. And for the first like six weeks, that's how I played. I played exactly how he told me to play.
Now, he didn't ever intend that that's the way I was always going to play poker. He was walking me through a step-by-step process on how I could develop expertise and identify which hands I actually did want to play, and identify exceptions to the rules and which hands were not good to play. He was doing it in kind of a lesson-based approach by saying, here, let's start simple and then practice. And when you get these kind of Google Maps or Waze instructions, then at first you're just doing what you're told because it'll lead to decent results. But soon you start developing your own expertise, and then you start seeing exceptions and you go, oh, that's interesting, and you maybe ask another operator, and then you start developing more expertise.
The second way that it helps teach is it teaches in terms of the skills. So one of the things about Alpha chess, even though it beat some of the best chess players in the world, was they said it had an alien playing style. They said it was like playing against an alien because it didn't obey the kind of structure of strategy that us humans use. So it's not actually useful for learning how to play chess. What's useful for learning how to play chess is a system that has agents in it that actually do know the skills.
Because a supervisor on a manufacturing line isn't monitoring every move that you make. It's not possible to monitor thousands of variables. What they're monitoring is, are you using the right strategy at the right time? And so if your multi-agent system has skill agents like strikers and goalies that have different schools of thought and different methods of playing and different strategies, then you can look at it and you can go, "Oh, I get it. This is when you run this kind of defensive play." And now switching to the American football analogy: this is when you run this kind of running play, this kind of passing play, this is when you run a reverse. And the expert can use that to supervise and train the younger folks, but the novices, the people that are gaining expertise, can also learn by watching the AI and what strategies it uses.
Yeah. So basically you need to build an agent that follows, let's say, the best practices, the strategies that one can really learn from, versus some of these alien strategies that you described in the context of chess.
Yes. That knows, uses, and provides transparency into.
Okay. I follow that. Well, you know, Kence, as we get to the end of this conversation, one thing I need to ask is: what is step one, maybe step two, for manufacturers? Because we've talked about the importance of building a database of skills, and honestly I hope manufacturers are listening to this realizing that artificial intelligence in these scenarios is a no-brainer, because we're literally talking about leveraging it to solve what many manufacturers would describe as their biggest issue, which is retaining the knowledge of their retiring workforce. And in my mind, I feel like building that database of skills right now, while you have those people for two, three, five more years, just makes a ton of sense. But I don't want to answer the question for you. Can you share what your recommended step would be for manufacturers that want to take advantage of this?
It's a very astute question. I say one, I really highly recommend identifying your highest value skills that are most in danger of extinction. Now, if you have 50 operators that are, you know, between 20 and 30 years old working on a specific process, that's not where you start your focus. You start your focus on the highest value skills that have the rarest expertise that is most likely to retire. You go, okay, that's where we start.
Then you have to get your innovation strategy. So some larger companies have innovation departments or manufacturing R&D, and you say, you guys are going to be responsible for codifying these skills and building AI systems. Others, you know, the operation owns, maybe smaller manufacturing operations that are still large but don't have a dedicated innovation group. You have to decide who's going to have the purview and time to do the innovation. I've worked with huge companies that said, our engineers don't have time to innovate, so we do stupid POCs because we don't have time to do anything material. So you have to figure out who's going to innovate.
Then you have to give those folks not just the purview but the tools, and I recommend you look for a platform, look for something that's going to help you build these things yourselves, because if you take a solution approach, there's too many use cases. You know, I worked with one manufacturer that said, "We have a thousand use cases for intelligent agents." So you're not going to buy a thousand different point solutions. So you need to look for a platform as leverage that's going to help you build, in priority order, the agents that learn to execute those high value skills.
So I like the way you describe it. It's not just identifying all the skills; it's the skills that are in danger of extinction, the folks that could potentially walk out the door tomorrow, that are rarest. For example, maybe there's a skill five people on the factory floor have versus a skill that one or two have. Start with the one that's one or two. And then step two is working that in with your innovation strategy. And very much going back to an area you've been experienced in your entire career is, you know, don't do a thousand point solutions. Look for that platform that allows you to scale and build out those AI agents.
100%. Which coincidentally is Composabl, but we don't need to... I figured we'd just say it out loud here at the end.
But you know, as my last question, I have to ask. We covered a lot of ground today. I really appreciate you breaking this down in a way that I think manufacturers can easily understand, and hopefully start thinking about AI beyond the context of machine learning or, you know, what you do in Copilot, for example. So I think this was great. Is there anything else you'd want to leave the audience with, or something else you want to bring up before we wrap today's conversation?
The only last thing I'd say, and this is a really fun conversation, is: you already have the expertise, manufacturers. You have the expertise, your operators have the expertise, and your engineers know the process. Empower your engineers to build intelligent systems. They designed and built everything we have anyway. And you know, there's in technology this thought that industries need to be disrupted. Well, maybe, maybe not. What I actually think is that there's 100 million engineers out there in industry that, if you empower them with the right tools, they'll innovate and engineer intelligent systems the same way as they have everything else.
Kence, I appreciate you taking the time to jump on Manufacturing Happy Hour today. I also appreciate all the things that I've learned through you over the past year. I'll definitely mention this in the intro/outro, but hey, if you're out there and you're looking for someone to follow on artificial intelligence and really pragmatic approaches for manufacturers, Kence is your guy. So, Kence, thanks so much for jumping on, and I will be seeing you around the industry very soon. I know that.
Thanks, Chris. Great conversation. Cheers.
[Music]
Article published