Bryan DeBois is the Director of Industrial AI at RoviSys. They are working to enable automation for manufacturers throughout the country through digital transformations. He is a fountain of knowledge surrounding the topic of AI. This conversation revolves around the differences between generative and autonomous AI and the best applications for each. He talks about the inner workings of these different AI models, how his company is using it to fight back against manufacturing brain drain, and the strategies that can help businesses begin to implement the infrastructure needed to fully take advantage of this technology.
Watch the full interview on YouTube!
Interview Chapters
- 00:00 Introduction to AI in Industrial Automation
- 02:52 The Evolution of AI in Manufacturing
- 05:45 Understanding Different Types of AI
- 08:27 Generative AI vs. Autonomous AI\
- 11:13 Real-World Applications of Autonomous AI
- 17:10 Challenges and Considerations in AI Implementation
- 19:54 Identifying Use Cases for AI in Manufacturing
- 24:01 The Future of AI in Manufacturing
- 28:58 Building Optimal Schedules with AI
- 30:12 Expert Insights and Legacy Knowledge
- 31:40 Adapting to Equipment Changes
- 33:31 Aging Infrastructure in Manufacturing
- 36:06 Competing with Global Manufacturing
- 38:48 The Need for Automation
- 41:18 The AI Bubble and Its Implications
- 43:48 Taking Action Towards AI Readiness
Nate Wheeler is the owner of weCreate, a nationally recognized marketing agency that helps manufacturers grow, save money, and become more efficient.
Nate Wheeler (00:01)
Welcome to Manufacturing Insiders. Today I have Brian Du Bois with me. He is the Director of Industrial AI at Rovisys. And they are a global player in the industrial automation game. Brian is just an absolute wealth of information on automation, AI. And I really want to pick his brain today about how has AI come into the picture in industrial automation. I know over the years, robotics, traditional machine learning has been a big part of lot of manufacturing operations, but how is AI changing the game, if at all? And we’re also gonna get into what I’m calling the AI bubble, and that’s kind of the basic idea that maybe our hopes and dreams of AI are not coming to pass quite as quickly as we might have expected. And how is that going to play out over the next couple of years. So welcome to Manufacturing Insiders, Brian.
Bryan DeBois (01:03)
Thank you, thanks for having me.
Nate Wheeler (01:05)
So, and just let me make a quick note before I get too far into it. If you’re a manufacturing leader, president, CEO, you hold a leadership position in manufacturing facility and you have insights that you believe other manufacturers would benefit from, what we’re trying to do here is make U.S. manufacturing more successful. So if you have some tips and tricks and ideas that other people would benefit from, go to manufacturinginsiders.com. We’d love to have a conversation with you, fill out the form there. And then if you’re watching this video, please give us a like, subscribe and share. So Brian, give us a little bit of your history at Rovisys and kind of where your main areas of expertise in the business are.
Bryan DeBois (01:45)
Yeah, for sure. So I came to Rovisys as a co-op. I was a computer science major at the university of Akron and really came in knowing nothing about manufacturing, right? So my background, I just wanted to write programs. Like I wanted to code. That was kind of my thing. And since then, know, 25 years later, I’ve still, you know, I’m still at Rovisys and really have fallen in love with manufacturing. I have now been in hundreds of plants. I’ve gotten to see how all kinds of different products are made. And I really believe strongly, I’ve now become a big advocate for manufacturing in general. I really believe strongly that manufacturing is that lifeblood that propels economies forward. I believe strongly in that. So,
Started out at Rovisys. I was writing a lot of custom software for manufacturers, right? So this was back in the day when There wasn’t as many off-the-shelf software products that were targeted at the industrial space that you could just go out and buy So
Nate Wheeler (02:51)
Right.
Bryan DeBois (02:52)
we were writing a lot of that custom and then just kind of moved up through the ranks and then in 2019 Rovisys had our 30 year anniversary and we kind of looked around and said well, what’s next? What’s what’s on the horizon? What do we need to be focused on? And so we started three initiatives after that. One was industrial networking. So this is looking at modernizing the networking infrastructure that runs on the plant floor specifically, cybersecurity locking down those plant floor assets. The second one was MES or manufacturing execution systems. So if some of your listeners are familiar with that, those are the big software systems on the plant floor that talk to the control system on the bottom side and talk to the ERP on the top side. And typically, everything runs through those MES systems. So we got a whole division that just does that. And then the third division, the one that I started, was industrial AI. And so specifically, our focus is on trying to operationalize AI on the plant floor.
Nate Wheeler (03:54)
Right.
Bryan DeBois (03:55)
And what we found in talking to customers was that you know, there were a lot of customers even in 2019. Now this is before ChatGPT, right? This is three years before ChatGPT comes out. But even at that point, there were customers who were looking at leveraging AI on the plant floor. And a lot of them had hired the traditional data scientist, which is great. The problem is, and those folks can build models. I mean, they’re successful at building models. The problem is, that when you go to operationalize on the plant floor, there’s a lot more that goes into that.
And what we’re hearing from clients is we don’t have anyone who can do that part of it. And so we really sat down and became experts on the AI, became experts on the data science, hired the right people so that we can now, we can do the whole thing. We can go from data, build the models, and then of course operationalize it on the planet floor. And there’s a certain amount of organizational change management that’s required too to be able to actually make those projects successful on the planet.
Nate Wheeler (04:53)
Right, right. Yeah, and what I hear a lot of is in order for that kind digital transformation to occur, it has to be a top-down effort. You have to have leadership bought into it. Otherwise, it’s just never going to happen.
Bryan DeBois (05:05)
Yeah, yeah, for sure.
Nate Wheeler (05:07)
So, you know, and I’ve had conversations with so many people about the actual implementation and use case for AI in the industrial space. You know, I was even talking with some of the big guys over there at FANUC when we were over at Fabtech. And I said, what, you know, what are you using AI for? And he’s like, oh, nothing, really. Like, so how are you using? What is the application for AI manufacturing?
Bryan DeBois (05:37)
Well, and I think it’s important, and you mentioned it in the intro, I think it’s important to understand AI, it’s this umbrella term and it encompasses a lot. But machine learning, right? That aspect of AI we have been using on the plant floor. You mentioned it, we’ve been using it for a couple of decades on the plant floor, machine learning. So modern AI absolutely includes that. In some ways it includes model predictive control too, which is an older technology, very capable technology, but an older technology that’s been around for 15, 20 years. It just now is expanded to include more capabilities. Also includes computer vision. Computer vision’s been around in our industry for a long time. Now, I will say that computer vision, with all the other advances that have happened in AI, computer vision has advanced significantly. So if you’re a customer who has not looked at computer vision in a couple of years, it’s time to go back and look again, because a lot of that has advanced significantly. But then it…it now expands to include a couple other categories. One is called autonomous AI, right? So this is that AI that I talk about that is built on deep reinforcement learning. So this is that AI that’s able to make human-like decisions, that’s able to build long-term strategy, that’s able to reason itself out of problems and get the line back to performing well, right? So that’s autonomous AI. So we can dig into any aspects of these. And then
Nate Wheeler (07:03)
Yep.
Bryan DeBois (07:03)
the other obviously major category, that industrial AI now includes is this generative AI. Now that’s the one that’s getting all the hype. That’s all your chat GPT and LLMs and your image generation and your video generation. All of that is under that umbrella of generative AI. And
Nate Wheeler (07:19)
Yep.
Bryan DeBois (07:21)
there is a lot of capability there. I think we all are recognizing the capability, the generative AI and the impact that it’s having immediately. But there’s some serious limitations to generative AI that make it in my opinion, inappropriate to use on the plant floor today. The limitations are just too great to be just deploying generative AI willy-nilly on the plant floor. Not to say that you can’t use it at all, but you better be very, very, very careful about implementing generative AI.
Nate Wheeler (07:53)
Right. So how do you differentiate that from the previous category that you stated? Because it sounds like there is some crossover there.
Bryan DeBois (08:00)
It sounds that way, but the reality of it is that’s because we attribute way too much reasoning and intelligence to generative AI.
Nate Wheeler (08:09)
Right.
Bryan DeBois (08:10)
And I think people are finally starting to wake up to that, that these LLMs, these large language models, things like ChatGPT, they’re capable tools, they’re fun tools to use, they’ve got lot of advantages, but they are not a generalized intelligence. In fact, what the reality is, because I know what’s happening under the covers, It really doesn’t have any intelligence at all. Generative AI, all the intelligence, we’re ascribing that intelligence to it. We are projecting that intelligence onto it, which it doesn’t deserve. It is effectively a pattern matching engine. It’s a very capable autocomplete. And that’s
Nate Wheeler (08:46)
Yes.
Bryan DeBois (08:47)
really it. It’s very good at picking the next word. In fact, I get really nervous then with the focus right now around these agentic AI. Is solely focused, it seems like, around generative AI right now. That makes me really nervous because it’s been proven not to be able to reason in even the simplest of senses. So what that generative AI, that agentic play is around generative AI today is effectively what you’re doing is you’re asking the generative AI, it does this behind the scenes, but you’re asking the generative AI to accomplish a task. Well, what it does behind the scene is it makes a script, it makes a recipe,
Nate Wheeler (09:23)
Right.
Bryan DeBois (09:24)
and then it follows that script. Well, with what we’ve seen around its inability to reason, what we’ve seen with the prevalence of hallucinations, even on the newest versions of some of these LLMs, we’re seeing hallucinations pretty regularly. I am. I don’t know if you are, but I think most people are seeing still.
Nate Wheeler (09:40)
Right. Yep. Yep.
Bryan DeBois (09:42)
And honestly, we may determine that hallucinations are just so fundamental to the underlying transformer architecture that maybe we will never be able to get away from hallucinations completely, right? But with that all in mind, the thought of it generating a script and then following it and we’re giving it the ability then to actually run code and we’re giving it the ability to call services on our behalf, I get very, very nervous. So to get back to your original question though, how is that different than autonomous AI? Autonomous AI is a completely different technology under the covers. It’s not transformers. It’s a technology called deep reinforcement. And this is what came out of, you remember DeepMind, it was in like the 2016, 2017, it was that Google spinoff.
Nate Wheeler (10:25)
Yep.
Bryan DeBois (10:27)
That’s what is underlying autonomous AI. It’s a completely different type of AI. It learns differently and it actually has been proven to be able to understand causal connections. So it can actually reason. can, this is the same technology that if you remember, beat our grand champions at the game of Go, beat our best champs players, be our best StarCraft players. Like, so it’s been proven to be able to… reason its way out of difficult situations and even be able to adapt to novel situations that it didn’t see during training.
Nate Wheeler (10:57)
Interesting.
Bryan DeBois (10:58)
So that really, the way I look at it is the types of problems that we face on the plant floor are primarily operational problems. So let’s use an operational AI. Autonomous AI is there to solve operational problems. It’s very, very good at acting like an expert operator on the line. The types of problems that generative AI is good at is solving knowledge problems, right? And we have some of those on the plant floor, particularly around maintenance, but primarily the problems we have on the plant floor are operational. So let’s use the right technology, the right AI for the right approach is kind of, I say.
Nate Wheeler (11:30)
Right, right. So can you maybe give me walk me through an example of, you know, the operational, I keep forgetting the reinforcement learning, right? Right.
Bryan DeBois (11:44)
Yeah, deep reinforcement learning, autonomous AI.
Nate Wheeler (11:46)
Yeah. How would that work in practice? Like, who’s using that?
Bryan DeBois (11:50)
Yeah, so and I’ll give you a use case here. So we’ve got a glass bottle manufacturer, right? And so these are world leaders. It’s name that we would all know, but they are world leaders in making glass bottles. And so they had on their glass bottle line, on one of their lines, they had a problem. And the issue was that the process was very finicky. It was very drifty. And so they would get dialed in, but then it had a tendency to drift where they’re not making unspec bottles. And part of the challenge was to move all the knobs and change the parameters to bring that process back to making good on-spec bottles. And
Nate Wheeler (12:29)
Right.
Bryan DeBois (12:31)
they really only had two expert operators who were the best in the world at being able to get back to making on-spec bottles. it took, depending on who the operator was, would take them anywhere from seven minutes to 20 minutes to get back to making on-spec bottles if the process drifted. And it took a light touch.
Nate Wheeler (12:46)
Okay.
Bryan DeBois (12:47)
So even if you got back to good, if you weren’t real careful about how you adjusted settings and you really stayed on, if you went to lunch for an hour and you came back, it’d be making bad bottles. It really took a constant effort. And so that’s a problem when you’re trying to run 24-7. So we came in and we built for them an autonomous AI agent. So our agent, and the way we built it, you asked about how we do this approach. So we took their historical data, okay, so they had a process historian, they had years of data, so we took that data, and with that we built a simulation of the real world, it’s called a data-driven model. It’s a simulation of the real world built on real world data. And with
Nate Wheeler (13:34)
Okay.
Bryan DeBois (13:34)
that data-driven model, that’s ultimately what the autonomous AI agent needs to learn from. So it has to learn in some kind of a simulation. So this was a real world simulation we built with their real world data. And then we also give it through a process called machine teaching. That’s where we sit down with those two experts and we say, hey, how did you get so good at this? And they’ll say, well, when it does this, then we do this. Or when we see this particular pattern, then we tend to do this.
Nate Wheeler (14:01)
Yep.
Bryan DeBois (14:01)
And so we captured all that through a process called machine teaching. And we fed that into the Autonomous AI agent as well. So now it’s got a simulated environment and it’s got kind of a starting workflow playbook on how to accomplish its tasks.
Nate Wheeler (14:16)
Yep.
Bryan DeBois (14:16)
And then we give it what are called reward and penalty functions. So reward it when it’s advancing towards a goal, penalize it when it’s moving away from the goal. And that’s really all it needs.
Nate Wheeler (14:23)
How do you do that?
Bryan DeBois (14:25)
It’s not as hard as it sounds. Like there are obvious metrics for, hey, you’re moving towards making good bottles versus you’re moving away from making good bottles. so building those functions is actually pretty straightforward. it’s a little bit of an art in science, but yeah, we’re pretty good at building those reward and penalty functions.
Nate Wheeler (14:45)
Yep.
Bryan DeBois (14:45)
So that’s really all it needs. And then you let it loose. This thing trains in the cloud, takes a decent amount of resources to train, not like LLM level resources, but it trains for about a day in the cloud. By the time it’s done training though, it’s like that autonomous AI agent has worked that part of the process for 20 to 30 human years.
Nate Wheeler (15:05)
Wow. Wow.
Bryan DeBois (15:07)
So by the time we were done, our autonomous AI agent, I told you, it takes their experts anywhere from seven to 20 minutes to get back to making unspec bottles. Our agent has never taken longer than five minutes and on average takes about two minutes to get back to making one spec bottles. So it’s about a 50 % improvement over, I mean that’s 50 % less time that you’re making bad bottles, right? That’s a huge drawback for the.
Nate Wheeler (15:29)
Right, right, yeah, that is huge. Yeah, that’s a really, really good example. then this AI model is then telling a less experienced operator perhaps that turn this knob this much, do this, do that,
Bryan DeBois (15:44)
Exactly.
Nate Wheeler (15:45)
OK.
Bryan DeBois (15:46)
Yeah, and here’s the thing. How can it get, so how is it so, because again, it’s not controlling it directly. Just like you said, there’s a human in the loop who has to actually take those recommendations, right?
Nate Wheeler (15:56)
Right.
Bryan DeBois (15:56)
So how is it that it’s able to get back to making unspoken bottles so fast? Well, here’s why. So when a human is faced with a problem like that, what do they do? They tend to move one knob and then they kind of guess and check and then they move another knob and it’s making bottles every two seconds. This is very fast process. So they guess and check, and then they move a knob, and then they see if that improves it, if, that didn’t work, they put it back, and then they move it to the knob.
Nate Wheeler (16:18)
Right.
Bryan DeBois (16:18)
That’s how humans try to fix problems like that. The autonomous AI doesn’t work anything like that. The autonomous AI just moves all the knobs at once, because it’s already mapped out the trajectory to get back to making on-spec bottles.
Nate Wheeler (16:31)
Yep.
Bryan DeBois (16:32)
So in a matter of just a handful of moves, it’s able to get everything back to making on-spec bottles. So that’s…That’s the difference between how a human approaches a problem like that and how the Autonomous AI is able to reason out a path and map out a path to get back to good.
Nate Wheeler (16:48)
Yeah, absolutely. And I think it’s such a great example, because I think it’s very easily transferred to a million different scenarios, you know, like you’re machining apart and you’re getting some deflection occurring and it’s creating some sort of distortion in the machining and, you know, and so then this model could understand that and then adjust the machine or tell you how to adjust the machine right off the bat instead of having to have 30 years of experience.
Bryan DeBois (17:19)
100%. And let’s say, so for that specific example, minus AI, how are you going to solve that? Well, you’re probably going to go to that 20 or 30 year guy or gal, right?
Nate Wheeler (17:27)
Right.
Bryan DeBois (17:28)
You’re going to say, Hey, I’m getting deflection constantly on these parts. Like, what am I doing wrong? Do I have a setting wrong? Can you come take a look at it? Right. Well, by the way, those people are leaving the industry. So we have a giant brain drain right now, where primarily baby boomers are leaving our industry in droves. We are losing massive amounts of expertise.
Nate Wheeler (17:48)
Yep.
Bryan DeBois (17:49)
So those people, a lot of times, are becoming harder and harder to find. it will solve, the Autonomous AI will solve that problem in the same way that that expert person will.
Nate Wheeler (17:58)
Yep.
Bryan DeBois (17:59)
So again, they have, because it’s run effectively for so long in the simulation, it’s seen everything. And it goes, I know exactly. Yeah, here’s what you’re doing wrong. You’re this particular material with this heat and this parameter and yada, yada, yada. And if you change these things then to compensate for that, then boom. Now you’re back to making good product. It’s the same way that a human solves those types of problems. so back to that expertise loss, that’s the primary driver, I would say, right now, behind why my customers are looking to adopt AI.
Nate Wheeler (18:31)
Yep.
Bryan DeBois (18:32)
Because they’re like, look, we don’t have these people anymore. They’re like, we’ve got high turnover in these positions now. We’ve got six month, one year people where before we had a 15 year person running that asset. It’s a real problem. And so, know, AI is not the whole story, but it’s a big part of the story about how we’re going to become more competitive in the future, particularly in the US world.
Nate Wheeler (18:56)
Yeah, absolutely. mean, it’s a critical issue right now. So if you’re a, you know, if you are a midsize manufacturer, how do you look at the applications for this? So do you start with, you sit down with your experienced workers and say, you know, what do you really do on a day-to-day basis and try to figure it out from there? Or do you look at, you know, where do you start to define the areas of application for this type of
Bryan DeBois (19:26)
Yeah. So it’s a real common question that I get. the easiest answer I would say is we always start with use cases. So even before we talk to the expert, we’re back a step and we’re looking at what are the broad use cases where we can apply this and have an immediate impact. So when I started my career, there were still lots of customers that were like, well, you we’re looking for a three year payback on this investment or, you know, that kind of thing. Those days are gone. Every single one of my customers wants a 12 month or less payback on their investment, right? Now we can argue whether or not that’s realistic or not, but that’s just the reality. that’s, that, you know, I don’t know what the economic drivers are there, but everyone wants a near immediate payback on their investment. So AI is really there.
Nate Wheeler (20:11)
I thought you were going to say that they aren’t worried about that anymore because that’s not the main issue. The main issue is they don’t have the people to do the work.
Bryan DeBois (20:20)
Well, one would think, one would think, Hey, I’m trying to solve this real big problem that you have.
Nate Wheeler (20:25)
Right.
Bryan DeBois (20:26)
It’s still the same story. Purchasing departments and being counters it nowadays and every single one of my customers, need, tell me how we’re going to get to a 12 month or less payback on this AI investment. Right. So, right. Right. We could go on and on about that, but it doesn’t matter. Like that’s, that’s the reality I live in.
Nate Wheeler (20:38)
Well, you aren’t going to be producing parts if you don’t do it, so… Right. Yeah.
Bryan DeBois (20:48)
Right. So, okay. So in some ways that does kind of distort the use cases that we have to pick, right? Because we’ve got to pick ones that we know are going to have an immediate impact, that are going to be able to get to that one-year payback. Now, the goal is, the hope is, is that we knock it out of the park with this first one, and then we can maybe get to some of those later ones that maybe have a longer, you have to have a little longer vision to see the payback on that. But either way. So we’re looking at the use cases that that customer has, and we’re down sampling it to, here’s the ones that we think are going to be the right combination of the lowest initial starting cost and the highest ROI. That’s the combination that we’re looking for. And then we pick one or two of those. And then if we need to de-risk it, we’ll do a paid assessment where we go in and we look at all the systems, and we talk to the people, and we look at the process changes, and we look at the KPIs. To be honest, though, we’ve been doing this for so long, or six years, there’s a lot of companies, OT system integrators that are just now starting down this AI journey. So we’ve been at it long enough that we kind of have a pretty good feel of, so a lot of times we’ll just skip all that, we’ll go right to a proposal. Here’s what it’s gonna cost to do this project. I’m gonna do it. And so then, those projects then, it could be more of a traditional AI project, right? So maybe it’s a computer vision or it’s a predictive model or it’s just more of an ML model like the old, like, models that have been around for a long time. Maybe it is an autonomous AI project and we’re just jumping right in because those are bigger, more complex, longer projects, those autonomous AI projects. But the ROI, the ceiling for ROI is way higher with the autonomous AI projects. And
Nate Wheeler (22:27)
Right. Right.
Bryan DeBois (22:29)
then we are getting some interest in doing generative AI projects. I we’ve got customers that are asking us about it. And so we’re not, again, we’re not completely saying we won’t do a generative AI project. We’re just saying we need to be really, really careful about leveraging generative AI.
Nate Wheeler (22:45)
Right. Yep, exactly. So can you give me an example of a successful generative AI project? Yeah. Yep.
Bryan DeBois (22:54)
I mean, none of them are done yet. This is all very new for our industry, right? So I’ve got a customer that, so we’re working with this vendor and they basically capture all of that tribal knowledge, but specifically the vendor primarily targets OEMs, so equipment manufacturers.
Nate Wheeler (23:20)
Okay.
Bryan DeBois (23:21)
Not a terrible way to do it because nobody knows more about that piece of equipment than the machine builder who built it, right? So effectively, the idea is that every one of these machines that that OEM builds comes with the generative AI model. The chatbot is included with it, ships with it to solve problems.
Nate Wheeler (23:39)
I see.
Bryan DeBois (23:39)
I’m OK with that. That’s not terrible. That, to me, seems like it’s fairly de-risked. But this is General idea that we’re gonna slap generative AI down on the plant floor across all kinds of different equipment Some of this equipment like the manufacturer doesn’t even exist anymore They’ve gone out of business or whatever and like and somehow the generative AI is gonna know how to fix it all and do it in a way that’s completely safe 100 % safe zero failures where because that’s the problem on the plant floor right is is that if you give the wrong information to the wrong person at the wrong time on the plant floor you could kill somebody like the stakes are so much higher. And that’s the part that I don’t feel like a lot of these AI vendors came from the IT world, the IT space, and they don’t understand that the stakes on the plant floor are so much higher. This isn’t about generative AI generating a funny marketing blurb and everyone kind of laughs about it and then they go in and fix it or a funny email to your boss and you’re like, well, I’ve never sent that. And then you go in and fix it. This is about telling somebody on the plant floor how to do something in the heat of the moment. And again, like you could, you could, we’re talking about the loss of life, and property, and that’s a real big deal. So that’s the stakes that we play at on the plant floor. And so that’s all I’m trying to say is, I’m not saying we got to slam on the brakes, but let’s tap on the brakes and make sure that we’re approaching the right problems on the plant floor. Regenerative AI can actually make an impact. That’s the other thing is, if you’re going to spend all this money, let’s make sure it’s actually going to be able to solve the problem that you want it to solve down there. But then let’s make sure that it’s doing it in a way that’s safe and grounded.
Nate Wheeler (25:17)
Right. Yep. So is it fair to say then that the biggest application for manufacturing for AI and manufacturing is in these judgment based scenarios?
Bryan DeBois (25:29)
Yeah, yeah. So what we do, we actually have kind of a cheat sheet of like, here’s where you want to look for what in trying to determine what are good use cases, you know, what are the parameters to look for? And one of those is exactly that, right? So where it’s as much that part of the process, it’s as much of an art as it is a science, you know, where the operators kind of has to go like this to be able to make good determinations.
Nate Wheeler (25:54)
Yep.
Bryan DeBois (25:54)
Great place for a great place for autonomous AI. great place.
Nate Wheeler (25:59)
Very interesting.
Bryan DeBois (26:01)
Where we also say kind of along the same vein, where like, the difference between an expert operator and a novice operator has a material impact on the final quality, on the final throughput, on, you know, effectively, you know, if you’re saying, look, Bill’s in today, we’re gonna have a good day on the line versus Nathan’s in today, could get a little hairy, right? So
Nate Wheeler (26:21)
Right.
Bryan DeBois (26:22)
like, those are great situations. You know, we want to then empower that novice operator to be as good as Bill, right? So let’s put a system in that looks over their shoulder. And what we call it is a decision support system. Let’s put a system in that looks at it’s like then you bills there 24 seven looking over the shoulder of
Nate Wheeler (26:41)
Yep.
Bryan DeBois (26:42)
that novice operator and saying, Hey, do this, do that. Hey, make this one change. you know, because here’s the other thing. It’s not just about, you know, like, like we were talking about, like just hitting the basic bare minimum of making on spec product, right? Like, yes,
Nate Wheeler (26:58)
Right.
Bryan DeBois (26:58)
they need to do that, but that’s kind of table stakes. If you’re not making good product, then you don’t have a company, right? So you got to be able to do that. But the difference between novice operators and expert operators is the expert operators then also have kind of secondary and tertiary goals, right? An expert operator is going to say, look, yeah, it’s on spec, but it’s on the heavy side of spec. So let’s make some minor, very minor adjustments to get it down to the lighter side of spec, still on spec. So we’re not giving away product, right?
Nate Wheeler (27:24)
Yep. Right. Right.
Bryan DeBois (27:27)
So that decreases the per unit cost. Let’s look at, and then, or they may say, look, yeah, we could do this and we could keep running this way and we’ll make on spec product, but boy, we’re wasting energy. If you make these three changes, now we can make the same on spec product, but we’ve reduced our energy consumption by 50 % or 10% or something like that, right?
Nate Wheeler (27:42)
Right, right.
Bryan DeBois (27:43)
That’s the type of things that expert operators have learned over decades.
Nate Wheeler (27:48)
Yeah.
Bryan DeBois (27:49)
Novice operators aren’t doing any of that.
Nate Wheeler (27:51)
Yeah.
Bryan DeBois (27:52)
They’re happy that you’re just finally, they got it dialed in and they’re making on-spec products. So that’s where this decision support can have those extra parameters and conditions that it’s trying to, constraints that it’s trying to work.
Nate Wheeler (28:03)
Yeah, it’s very interesting. And the thing I like, I think the most about the decision support model is it serves a class of manufacturers that have been traditionally unserved by the automation space. And so I think that this is a great way for them to start moving towards automation without having a robot that’s physically doing something.
Bryan DeBois (28:26)
Right, right. Because, mean, like, look, everyone wants the robot and we probably need to get more robots and more of that. But I mean, that’s huge investments. Like that’s
Nate Wheeler (28:36)
Right.
Bryan DeBois (28:37)
significant investments. This is something where, you know, the goal is to be able to take someone in off the street and with a little bit of training and this system, DecisionSport system looking over the shoulder can now make product like an expert operator, at least approaching an expert operator.
Nate Wheeler (28:52)
Right.
Bryan DeBois (28:53)
Let me give you another example in a completely kind of different part of the process. So we talked more about, we’ve been talking about using autonomous AI more on process control. Same technology, but we can use it to build production schedules, right? Exact same technology. We sit down with those. So we did a project for a paint manufacturer we’ve all heard of, and they were trying to, they had two human schedulers in a plant that runs 24 seven. And so if an asset went down or they had some issue, their options were, continue to run the plant suboptimally until one of their human schedulers came in in the morning, or wake them up in the middle of the night. Like these guys, they get sick, they go on vacation, right? There’s two people.
Nate Wheeler (29:36)
Yeah.
Bryan DeBois (29:36)
And so, you know, their goal really then was just to try to build a production scheduling system that would work as well. And again, that a novice person could at least build decent production scheduling, reasonable production schedules based on this system. And so that’s what we were able to do. Again, we sat down, we built a simulation of their process. We sat down with those two human experts. We said, give us the tips, tricks, the things that you would typically sit down when you’re teaching a new scheduler how to do things. Teach us.
Nate Wheeler (30:04)
Yep. Yep.
Bryan DeBois (30:05)
We put that into the AI and then it runs against that simulation by the time. And we didn’t go easy on it, right? We sent things like, okay, this asset went down. Now build me an optimal schedule. Lowe’s just came in with a big hot order for paint. now building an optimal schedule, right?
Nate Wheeler (30:23)
Right.
Bryan DeBois (30:24)
So how do you handle all these crazy scenarios and things?
Nate Wheeler (30:27)
Yeah.
Bryan DeBois (30:28)
And by the time it was done, it was able to build schedules as well as their human schedulers were. So completely different problem, same technology to solve.
Nate Wheeler (30:32)
Wow. So how did the experienced schedulers respond to this process?
Bryan DeBois (30:44)
It’s also one of the most common questions I get. How do these experts receive this and do they see it as a threat? That’s one of the most common questions I get. Do they see it as a threat? How do they respond to it? Honestly, it’s the exact opposite of what you would think. They were excited. They were thrilled to work on this. I did a project with a vinyl extrusion manufacturer up in Canada. both of these guys, his two… So I sat down with the plant manager. And he said, I’ve got two guys who are the best at running this vinyl extrusion line, and they are both less than five years from retirement. So sitting down with those two experts, they got one foot out the door.
Nate Wheeler (31:23)
Yeah.
Bryan DeBois (31:23)
For them, they saw this as their legacy. They were so excited to get into all the details and the minutia of, so when it’s running hot and you’re running this particular product, you’ve got to make sure you turn down the RPMs. That’s the kind of stuff they get all jazzed about. Why? Because they spent their entire career learning how to be the expert at this.
Nate Wheeler (31:41)
Right.
Bryan DeBois (31:41)
I was at a street show once and I got asked that question. said, look, when we sit down with these guys, who are they going to talk to about this stuff? They’re, they’re buddies at the bar after they get off work. They don’t want hear anything about vinyl extrusion. Said, their wives definitely don’t want to hear anything more about vinyl extrusion. Right? And here we come, we sit down with humility and we say, Hey, teach us everything you know, how’d you get so good at this? Right.
Nate Wheeler (32:02)
Yeah.
Bryan DeBois (32:03)
And they’re so excited to teach us. And then they, again, they see it as their legacy because otherwise all that knowledge leaves with.
Nate Wheeler (32:09)
Right.
Bryan DeBois (32:10)
And most of these folks, grew up at that organization, or at least they spent a long time there. And so they don’t want to see it go down the tank. They want to make sure that that gets captured. And then we do things like we’ll name the system after them or something like that just to…
Nate Wheeler (32:23)
That is a cool thing. I have heard other people that have different ideas and ways of transferring knowledge from these experienced workers. One of them includes video snippets. Searchable AI database that if somebody is having trouble with a process. But then this guy is in all the videos. Then these people start to feel like they know him even though he is not there anymore.
Bryan DeBois (32:47)
Yeah. Yeah.
Nate Wheeler (32:49)
He has a legacy like that.
Bryan DeBois (32:51)
Yeah, I know.
Nate Wheeler (32:52)
That’s right.
Bryan DeBois (32:54)
I’ve never had a situation where one of those experts wasn’t excited to work with us on this.
Nate Wheeler (32:59)
Right.
Bryan DeBois (33:00)
You got to kind get over the typical credibility gap, right? So you got to make sure that they feel like you know what you’re doing. But once you get over that, and that’s every project, when you’re dealing with people who know that process inside and out, you got to establish some credibility. But once you get past that, they’re excited.
Nate Wheeler (33:15)
Right. Yeah, so it also sounds like a lot of these nuanced art sort of processes in some ways may be related to having some antiquated machinery that you really have to stroke and be nice to and all that kind of stuff. So, you know, let’s say this company decides to update that piece of equipment. How do you revise the model based on that? big of a project is that?
Bryan DeBois (33:45)
It’s not a huge deal. It will need revised. If you’re making significant equipment changes, can handle the typical like, you know, equipment just changes its properties over time as it ages and stuff. It can handle that kind of drift. Typically we may have to do a quick retrain and redeploy, you know, to handle drift,
Nate Wheeler (34:04)
Yep.
Bryan DeBois (34:05)
but it can handle that. But if you’re making a significant change to how the equipment operates, we will have to kind of rebuild that model. We don’t have to start from complete scratch by any means,
Nate Wheeler (34:14)
Yeah.
Bryan DeBois (34:15)
but it’s definitely something that you have to keep in mind that the model has to be updated to reflect, you know the equipment
Nate Wheeler (34:21)
Right.
Bryan DeBois (34:21)
But it’s the same I mean honestly It’s the same thing like if you completely change the equipment you’re gonna have to retrain all of your human operators, too So now
Nate Wheeler (34:28)
Great.
Bryan DeBois (34:29)
you got to retrain one more thing. You got to retrain the model to reflect,
Nate Wheeler (34:32)
Yep.
Bryan DeBois (34:32)
you know this new state I wanted to address one thing though. You mentioned there, you know this situation of aging equipment First off we take that as as a given We never walk into these companies assuming that we’re to be able to upgrade this equipment. It’s almost never an option.
Nate Wheeler (34:50)
Right.
Bryan DeBois (34:50)
So we go into these situations understanding that this brand new whizbang AI model might be running on a piece of equipment that’s 25 plus years old. That’s just kind of the world that we live in.
Nate Wheeler (35:01)
Right.
Bryan DeBois (35:02)
That being said, I will make the broader statement that if we want to continue to onshore and we want to continue to make US manufacturing competitive, we have got to do something about this aging infrastructure. is rough out there. I
Nate Wheeler (35:15)
I agree.
Bryan DeBois (35:16)
It was at a steel manufacturer a couple of weeks ago and I’m talking to some hardcore, these are hardcore operators, these are the ones who have boots on the ground, they’re there every day fixing this stuff. And the one guy said, he’s like, I will not sign my name that this piece of equipment is going to last more than a year. He’s like, and honestly, he was like, I probably should have said that five, 10 years ago. But now, mean, now it’s just ridiculous. Like they’re buying parts off of eBay replacement parts. And he’s like, I will not make a claim that this piece of equipment will last more than a year. Now, here’s the thing. We’ve got investment that is coming in. Right. We’ve got, you know, billion dollar investments in different manufacturing customers coming in and half a million, you know, half billion here. And we got, we’ve got money that’s starting to flow back into US manufacturing. But I am, and I’m an advocate for the plant floor. I’m an advocate for manufacturers, right? And I’m trying to tell those folks on the plant floor, if you don’t speak up, if you don’t advocate for yourself, because we’re not good on the plant floor at doing that, if you don’t advocate for these dollars, they will all get siphoned away into IT. And I like IT, I’ve got a ton of respect for IT, right? But do we need another SAP upgrade? Or do we need that money to go towards the 45 year old piece of equipment that we’re buying parts off eBay and no one will guarantee it’s going to run in a year? Because guess what? If that thing goes down, now you’re done. Now you’re not making any product, you’re not making any money.
Nate Wheeler (36:40)
Yeah. Right.
Bryan DeBois (36:45)
That’s, you know, we don’t do a good job down on the plant floor of making it clear that we’ve got to have these dollars as this investment, as this money now is starting to flow in to on-shoring and US based manufacturing. We’ve got to make sure that it makes its way down to the client floor so that we can get this equipment upgraded.
Nate Wheeler (37:04)
Yeah.
Bryan DeBois (37:05)
It’s a significant concern of mine because if we don’t, then that billion dollars will get sucked up into a move from SAP 3.5 to 4.5 or whatever and all that billion will be gone and 10 years from now that guy’s still going to be trying to patch up that piece of equipment that actually makes the product.
Nate Wheeler (37:23)
Yeah, no, think that’s a great point. Something you don’t hear a lot, but that’s absolutely right. And I think in general, whether it be in the machinery itself or in these types of automation systems, I think we need to figure out a way to be more aggressive in our approach in the US because we are way behind a lot of Asian countries and a lot of people that you hear these manufacturers, yeah, our business has gone to China and India and whatever and it’s not a huge surprise. mean, in a lot of cases, these guys are doing a better job at a cheaper price. And obviously that’s not always the case, you know, the better job part, but in some cases, you know, the China always had this reputation for these cheap chintzy crap products, but they’ve come a long, long way. And a lot of times they are producing at least equally as good product at a much cheaper price. we need to kind of figure out where that investment’s going to come from. You know, in a lot of cases, it just doesn’t seem like it can come from the shop itself simply because they’re already struggling with profit margins. So yeah, I don’t know who’s going to solve that problem.
Bryan DeBois (38:39)
Yeah, well, mean, you know, better, worse, better and different, like with the tariffs, like I think that’s kind of the idea is that some of those, those, those flows start to come back because if they’re selling product, if they’re able to sell product at a competitive price, now they can maybe start to allocate that money. again,
Nate Wheeler (38:57)
Right. Yep.
Bryan DeBois (38:58)
you’re going to have those big IT consultancies that are going to be slobbering at the mouth once they see those revenue flows starting to inch up. And they’re going to try to gobble up as much of that as they can.
Nate Wheeler (39:13)
Yep.
Bryan DeBois (39:14)
part of my role, feel like I’ve grown up in this industry is just to be an advocate for the plant floor. you know, these are guys who are typically guys and gals who are head down and working hard to keep things running. And they just don’t, it’s not in their nature to speak up and say, you know, one of the things I try to tell them is, that, look, use me. AI can be that carrot. Go to, go to corporate and say, look, I want to do this AI project. I know you want to do this AI project, but we cannot get the data off this equipment in the state that it’s in. We have got to do XYZ upgrades to be able to get to this AI future that you want to get to. I’ll make one other point.
Nate Wheeler (39:50)
Okay.
Bryan DeBois (39:51)
You brought up, who are we competing with? And you mentioned some of the countries that we’re competing with, right? One of the advantages they have is in their workforce because they can just throw bodies at these problems. They can just throw armies of people at these problems. And a lot of times working in very unsafe conditions, which I certainly don’t want the US to adopt, but…That’s how they solve these problems, right? We’ve got a very different workforce here.
Nate Wheeler (40:17)
Right.
Bryan DeBois (40:19)
We’ve already talked a lot about that. we have to, instead, we have to be smart about it. We have to adopt automation. We have to adopt AI. We have to make it so that it’s push button and most of the intelligence lives in the machine and you can take almost anyone off of the street. I use this analogy and it’s the best analogy I’ve been able to come up with so far on this. You know, when you look at McDonald’s, they don’t hire a cook like a diner would hire a cook, right? They hire a person and they give them a couple hours of training and they’re like, here, you’re going to load these frozen pucks in and you press a button and the machine’s going to do all the work and then you just unload it, right? I mean, in McDonald’s, better, worse and different, like they have automated pretty much as much as you possibly can, right? They’re using humans basically for the things that it’s so hard to do, you would need a pretty sophisticated robot to do. So they’re using humans for that, but everything else they’ve automated. That’s the level of automation we have to get to to be able to use that same workforce. We have to get to that level of automation and manufacturing. And we are so far from that.
Nate Wheeler (41:22)
Yep, we have to. I know.
Bryan DeBois (41:26)
Like if you walk into the average factory, there’s still so many judgment calls that are happening. There’s so much expertise that’s required on the plant floor to make good product. You’re so far from that level of automation. Implants in that but that’s where we have to get to if you’re going to have high turnover and you’re going have people who going only stay for a couple months and you’re going to like you know and unskilled workers like that’s that’s the level that we have to get to of automation so i’m just trying to help where i can you know and i like
Nate Wheeler (41:50)
Yeah.
Bryan DeBois (41:50)
i said ai is part of it it part of the story.
Nate Wheeler (41:53)
Yeah. Yeah, and it’s I mean, it’s not going to get better. I mean, I don’t know if some people are not moving towards automation because they think it’s going to get better, but it’s not going to get better. You know, our biggest generation in the history of the country, the baby boomers is either out or on their way out. And even if you have the same percentage of people that were interested in working in blue collar jobs, you still don’t have the same number of people. And so you have a growing economy, you have a more need for product, and you have less people to work with. So it’s just going to continue to get worse. And so we need to figure out how to supplement with automation and from making what used to be a complex job accessible to a unskilled or minimally skilled worker.
Bryan DeBois (42:47)
Yeah, 100 % agree with that. Yes. Yes.
Nate Wheeler (42:49)
Yeah, wow. So, we talked a little bit about this the other day and I just read a stat that there was some Harvard economists that estimated 92 % of our GDP growth is due to data centers and related technology, which is a pretty scary number because when you look at the incorporation of this data into production and into the real GDP production of things, it’s fairly minimal. So do you think we’re in an AI bubble?
Bryan DeBois (43:24)
I do, I do. And I know that people don’t want to hear that, but I 100 % do. think that we are being, the economy is being bullied right now by the data center work. And we’re very close to it. So Robasis does a huge amount of data center work. Is so we see it, we’re kind of on the front lines. And it’s been great. It’s been a boom for us. And we’re very, very busy on executing those data centers. But I do think that all that investment and all those flows going to AI and building these data centers. I think that unfortunately, it probably is a bubble that we’re inflating. And we’re probably getting a little over our skis here. And at some point, it may pop. They are starting to know, that Gartner cycle around AI. I think we’re starting to get towards that trough of disillusionment where we drop down, we start to realize what the actual limitations. I think that we’re seeing a slowdown in some of the innovation happening around AI. think we’re starting to, you know, a year or so ago, we were just seeing like constant innovation there. I think that’s starting to slow down as they’re starting to realize the limitations of these algorithms and
Nate Wheeler (44:38)
Yep.
Bryan DeBois (44:38)
what’s going to be possible. I’m not excited about anything I’m saying. I mean, this is kind of the reality, I think, of where we’re at. And so, now, that’s not such a bad thing, right? There’s a reason why we have to get, there’s a reason why Gartner created that hype cycle, right? We’ve got to get past all of that. We’ve got to get past the hype curve. We’ve got to get past the trophic disillusionment before we can finally start to have reasonable adoption and understand where it can actually have an impact. And look, I wouldn’t have a job if this AI can’t have a significant impact on the plant floor,
Nate Wheeler (45:11)
Right. Right.
Bryan DeBois (45:11)
when it’s implemented well. So let’s pick the right use case. Again, I feel like we’ve been doing Clume, so let’s end it on a high note. Let’s pick the right use case. Let’s get this AI in there and let’s start. I will say I love my job because I love how excited when we finally deliver this thing and it’s running on the plant floor and people are using it, they get so excited about it and so pumped up. Typically, by the time we finish the first project, they got five more projects lined up for us because they’re so excited about what the possibility is for this technology. So this is the path forward. This is how we’re going to outperform everyone. Again, it’s going back to American ingenuity. This is how we’re going to, this is how we win. And so we just, we’ve got to make sure that we’re taking that first step, right? We can’t get so caught up in the risk, right? Like that’s business. Like there’s always going to be some kind of risk. If you partner with an expert, we can help you de-risk these things. We can help you navigate that risk in a smart way. But let’s go ahead and we need to bite this off. We need to take that first step. Everyone listening, if you’re a manufacturer, look, even if you don’t want to do an AI project, it’s not for us yet, we’re not, then do an AI readiness project, right? Let’s get your house in order. Let’s get your data house in order.
Nate Wheeler (46:21)
Yeah. Great.
Bryan DeBois (46:23)
Let’s get your infrastructure in order.
Nate Wheeler (46:25)
Right.
Bryan DeBois (46:26)
All those kids that are disconnected, let’s get all those networked together. So we can at least have a chance of doing AI at some point in the future.
Nate Wheeler (46:32)
Exactly. Yeah, almost like, you know, before you get your ISO certification, you got to go through the checklist and make sure you have all your documentation and your processes in order and all that kind of stuff. No, I think it’s a great point. And I think it is exciting. And it would be very exciting to 99 % of the manufacturers I know to know that if their experienced worker left tomorrow that they would have at least a reasonable chance of passing that knowledge along to the next guy and keeping their operation running. So I think it’s so vital. And I think the other bright side about the AI bubble is at least I don’t think we’re as close to the robots exterminating the human race as we maybe thought a year ago.
Bryan DeBois (47:19)
We are not. We are not. You know, people always ask me in the role that I’m in, like, how close are we to artificial general intelligence, to Skynet? Do we need to be worried? I’m like, I’m not worried. Like, these AI algorithms are really, really capable in the very narrow scope of things we ask them to do, but they can’t solve their… I mean, outside of that very narrow thing we’ve asked them to do, like, they can’t do any… Like, it’s not that. We’re so far from Skynet. We do not need to worry about that.
Nate Wheeler (47:47)
That’s good. It’s comforting to me anyway.
Bryan DeBois (47:49)
Yeah.
Nate Wheeler (47:51)
Well, thanks a lot for coming today, Brian. Really enjoyed the talk. think you’ve given us some points that I think any manufacturer can take action on, is one of the big goals of this podcast. And really appreciate it. We’ll drop your info in the bio. And I’d encourage anybody that listen and wants to learn more about it to reach out to Brian. So thanks for coming today.
Bryan DeBois (48:16)
Thanks, Nate. Appreciate it.