AI is everywhere today … except the factory. What does AI-native manufacturing look like? Is it possible? Can AI agents help manufacturers produce more product at better quality?
And, maybe also enable onshoring or re-shoring?
In this TechFirst episode, host John Koetsier sits down with Apprentice CEO and founder Angelo Stracquatanio to explore what AI-native manufacturing really means, and why traditional AI models fall short in production environments. Instead of chatbots, this new approach uses event-driven AI agents that respond to real-time manufacturing signals: alarms, equipment data, quality issues, and more. The result? Faster troubleshooting, reduced costs, and entirely new levels of automation.
- Get the deepest insights concisely on the TechFirst Substack newsletter
- Subscribe to the TechFirst YouTube channel to never miss an episode
- Or listen on Spotify or Apple Podcasts
This episode of TechFirst is sponsored by Apprentice
Did you think AI was only for digital work? Nope … AI-native manufacturing is here. This month’s sponsor is Apprentice, which offers the first AI Agent built from the ground up for agentic manufacturing. Connects to all your systems, monitors everything, automates all your processes … but keeps a human in the loop. Check it out at apprentice.io.
Watch our conversation here:
Angelo breaks down how their system combines:
- Specialized AI models trained on real manufacturing data
- Role-specific agents (for operators, quality teams, engineers, and leadership)
- Workflow automation that goes far beyond simple prompts
Angelo and John also dive into:
- Why general-purpose AI struggles in manufacturing
- How to eliminate hallucinations with guardrails and workflows
- Real-world ROI: faster investigations, lower cost of goods, improved throughput
- The future of adaptive factories and personalized production
- Why humans remain critical, even in highly automated environments
If you’re in manufacturing, operations, or industrial innovation, this is a deep look at how AI is actually being deployed …and where it’s headed next.
Full transcript: AI-native manufacturing
John Koetsier: So, we’re all pretty familiar with AI right now. We summarize documents, we create pictures, maybe even we create an agent. But what the heck is AI-native manufacturing? That sounds new.
Angelo Stracquatanio: Yeah. Manufacturing teams for the last year have been trying to figure out how they use AI. They’re getting a lot of pressure from their C-suite, their board, their management, whatever it is, that they must have an AI strategy.
But I think manufacturing teams know that a chatbot can’t run production, and a lot of the key workflows within manufacturing are the things that a chatbot can’t do. And so, to be an AI-native manufacturing team, we had to think very differently about how all this works, because teams already have existing systems of record, and teams already have access to these chatbots.
So if we want to move them into this new world, we had to think of something completely different. And so that’s what we built. We built out what we call A1. It’s our AI agent for manufacturing teams, and the whole intention is that it’s not just chat. It responds to manufacturing events because, in a manufacturing facility throughout a given day, there are many things that occur, whether it’s an alarm, whether it’s a material movement, whether it’s an equipment trend.
And so A1 is designed to be able to respond to events, and then that event can trigger an agent. That agent can be as simple as an operator agent, a quality agent, a process engineering agent, a maintenance agent, or it can be way more advanced. It can be an alarm triage agent. It can be a continuous improvement agent.
And all these agents, combined with the existing manufacturing systems they have today, combined with these automated triggers, can move manufacturing teams into a completely new paradigm where they’re automating workflows that they never thought possible.
John Koetsier: I want to learn all about that because we know how most AI is built. We suck down the internet, all the words. We learn about what all that is, and we create models based on that.
Manufacturing is not like that, and so your source data is very, very different. What’s that look like? How do you build that? How do you build the intelligence that goes into this AI?
Angelo Stracquatanio: So we ended up building our own model that we call Apprentice 4.1.
And the reason why was just like you said: the general models were just that. They were general. And what we found, because we’re a 12-year-old company, we’ve been building manufacturing execution systems for over a decade for, in our case, the most complex and highly regulated industries out there. When we tried using the general models off the shelf, what we found was that they were pretty bad. They were generic. They were giving you pretty broad answers. They lacked specificity, they lacked consistency, which in manufacturing is really important, and they were missing compliance because having compliance answers that are either driven from actual regulations or company policy, it was implicitly in conflict with each other.
And so we ended up doing kind of a crazy thing. We built our own model that was fine-tuned and used reinforcement learning from a state-of-the-art model. But what we did is we took our 12 years of domain knowledge, our 12 years of learning, our on-the-floor experience for over a decade, and we ended up fine-tuning these models to then create Apprentice 4.1, which then, compared to, say, an off-the-shelf general-purpose model, greatly exceeded things on these dimensions that, for manufacturing teams, matter: specificity, relevance, compliance, consistency.
And that’s when we really hit this aha moment because we’re like, wait a second, now we can actually use AI in production to work on production things and not get general answers, but get production answers.
And so that was the core basis and foundation before we even built this AI agent that we call A1. It sits on top of it.
John Koetsier: It’s amazing. I mean, Google makes models, right? OpenAI makes models. Anthropic makes models, right? You’ve got a bunch of Chinese companies making models, and you’re making a model. That’s really interesting because that’s not an insignificant task.
Talk about the guy who’s running a factory and what he’s doing right now, and how this AI will help him or her.
Angelo Stracquatanio: Yeah, so through these last 12 years, we’ve learned that there are a lot of people who work at a site. It’s not just Bob the operator, or Jane the quality analyst, or Steve the site head. There are many different roles and many different levels, and in order to be successful in manufacturing, you have to touch everyone and you have to touch all these roles because the manufacturing plant, while they have the same objective, the objective is to drive throughput, create incredible products for their customers, and do so with high quality. It takes a whole army of people in many different teams and roles to do all of this.
And so we didn’t approach A1 in a general sense where you just have one agent to rule them all. We created a bunch of these sub-agents, and these sub-agents are designed for every role in the manufacturing facility: an operator agent, a quality agent, a process engineering agent, a site leadership agent, a supervisor agent, a quality systems agent.
And we created all these sub-agents on purpose because they’re not only meant to help the person, but they help the different teams and they help the different cross-functional objectives that happen within a plant.
And then what we did is we created this series of what we call workflows. These workflows, we even have a visual builder that will feel very familiar to manufacturing teams. This workflow builder then has these consistent, repeatable AI steps that, for the operator, let’s just say they’re on the floor, they’re troubleshooting, it’s 2:00 a.m. Instead of having to go through a stack of SOPs to find the relevant troubleshooting step, it will go and look through all those SOPs for that operator and put together a step-by-step troubleshooting guide.
This in and of itself is a bit unique because it’s not just giving an answer where it’s text and it’s like, “Hey Bob, go do this thing, and here’s the four bullet points.” It will put together a visual step-by-step guide that Bob can click through, and then it feels very familiar to Bob, whether it’s his existing MES or his existing work instruction system.
Bob doesn’t have to go through 20 SOPs. It’s the same thing for Jane in quality. Instead of having to manually go through and review every exception that occurred on the shop floor, which for a given day can be hundreds throughout the day, the agent will go and process it itself, put together a list of recommendations for Jane.
Jane then can say, “You know what? This one I want to go investigate further.” And the agent then will go into those subsystems or flat files, because many plants don’t even have systems. They can go check from flat files, and then it’ll come back to Jane and say, “Jane, this is what I saw across these different systems. Here’s the analysis that I saw.”
Because instead of taking Jane weeks to do an investigation, it can do it in a couple seconds. And so it’s a completely different paradigm shift, but the whole goal is to build agents for people, not an agent to rule them all. We want it to be for Jane, we want it to be for Bob, and we want to help them as they’re going through their day-to-day.
John Koetsier: What’s fascinating about that, I was talking with the gentleman who leads AI for AT&T about two months ago, and they had implemented something somewhat similar to what you’re talking about in their world of networks and switches and billions of gigabytes being sent all over the world and everything.
And they had put an agent in there, and they had reduced the time necessary to triage a problem, a network problem, to something like 2% of what it was before, like a 50x improvement. It was off the charts because of exactly what you’re talking about. An agent would look through all the data, all the issues, all the whatever, have all the understanding of the system, and would say, “Hey, check this, check that, your issue’s here,” type of thing. Right? That’s amazing.
Now, if I’m running a factory, do I have to instrument specifically for this kind of agentic help, or do I have that already? I mean, agents work on data, right? They need to know what’s going on in order to help you out with fixing things or improving things. How do I have to instrument my factory, or are most modern factories already instrumented for that?
Angelo Stracquatanio: So throughout the industry, when I say industry, there are many different market sectors when you think about manufacturing. It’s not just one amorphous blob. Each market sector has its own unique challenges, its own unique systems, its own unique process, and within it there are different levels of digital maturity.
It could be everything from everything is still on paper and spreadsheets to the most advanced lights-out manufacturing. There are different ranges along the spectrum.
The way that we approach this is that not every plant is the same. Not every plant is digitally mature, and not every plant wants to go full-blown AI on day one, because that’s a bit ridiculous too for some manufacturing teams.
You have some teams that are very advanced and already understand this technology really well and, like that network example that you just gave, are ready to take an agent connected to a manufacturing event, have the agent trigger without a prompt, trigger autonomously off of it, and run on its own.
But other plants are not there yet. Other plants just very simply want to provide an easier and faster way for their operators, their quality folks, their process engineers, to do the things they do today, but faster.
And that’s where our approach to this is that there are different levels, and we want you to ramp at what you’re most comfortable with. If you want to just simply provide a great tool to your operators, awesome. If you want to fully automate this thing and have it trigger off manufacturing events, awesome.
But what we recognized over these last 12 years is that every plant is different. Every plant has different systems, and we designed the tool, to the best we can, to serve these different ends of the market so that no matter if it’s just Bob the operator or Steve the site head in a fully lights-out manufacturing facility, we can use AI in both of these areas so that all manufacturing teams, all sectors, all maturities, can grow with this technology as we’re all figuring this out at the same time.
John Koetsier: What I love about that strategy is that even if you are this lights-out manufacturing facility that’s instrumented for data flows and it’s amazing, incredible, you’ve got digital twins, whatever, you’re probably going to be a little bit cautious about bringing in new technology, especially AI, especially an agent, because this is a multimillion-dollar facility and downtime here is a big deal. And you’re thinking about AI and you think about ChatGPT, and you think about hallucinations, and it sometimes can’t answer what’s two plus two.
So how is this going to work for me? Talk about hallucinations and what you’ve built into your AI so that Bob the operator doesn’t get some alert that is totally fictitious.
Angelo Stracquatanio: Yeah. So over the last 12 years, the hardest thing that we had to do, because before AI we were building out these manufacturing systems, and we’ve been building software for a long time, software wasn’t the hard part. Trust and credibility was the hard part. Building enough trust and credibility so a manufacturing team trusts my software to run their production, because their production ultimately, at the end of the day, drives revenue. If you don’t produce, you don’t drive revenue. And so the criticality of this is such an extreme.
And where we got started, we were focusing on life sciences, even more extreme, because if your product has bad quality or creates a negative outcome, you can actually kill somebody. That’s not like a product defect. There’s a negative impact because that product is being consumed by a human.
And so when it comes to AI, we had to approach it through the same lens. It’s not the software, it’s the trust and credibility.
And when you break that problem down, there are two parts to it. Part A is that implicitly AI is what’s called a probabilistic system. It has different outcomes in different ways depending on what you prompt it. So how do you take a probabilistic system that can hallucinate and move it into a very deterministic setting? Manufacturing does, ideally, the same thing in a repeatable way over and over again. That is the nirvana state that you want to get to in a plant, that you have a very low defect rate, you have high throughput, and everything’s just working. So implicitly, there’s conflict here: probabilism with determinism, conflict.
And so in order to build that basis of trust and credibility, we recognized that we had to constrain the AI, we had to constrict the AI, and we had to make the AI extremely focused. And so this is where this system of agents and workflows comes into play. We didn’t want to just have it where you have a prompt and you get an answer and you can have a wide range of probabilities of what the answer could be.
That’s bad in manufacturing. Instead, we created these agents and workflows where the agent follows an extremely tight and constrained workflow, where that workflow has a series of steps. This will feel very familiar to people in manufacturing. That step has a different step type. It can be an instruction, it can be a connector, it can be a compliance gate, it can be an artifact that’s being generated.
Then the instruction set that you give to the agent can also be extremely narrow.
And so by creating a very narrow set of outcomes, a very narrow set of consistent steps that this AI can perform, we saw consistency shoot through the roof. And so it comes back to how we approached AI in general.
We created this eval system. I’m sure if people are familiar, they’ll see these charts online where OpenAI and Anthropic will always say, “I have the best,” and there’ll be these bar graphs, and you always wonder, like, what are the graphs for? Like, who created the graphs, man?
But what we did is we created our own eval system for manufacturing that specifically tests consistency. We need consistency in our world. And so it wasn’t until we constricted the AI in this way to force it to follow these steps, to have these result sets that are repeatable, that we then saw consistency shoot through the roof.
And so our hope for this is that when you have a consistent output, people can trust it. Credibility goes up, and then yes, over time a plant can adopt this technology in a way they feel comfortable with for their specific digital maturity level.
John Koetsier: I love it. You put guardrails around it. You’re bowling, but you put the bumpers up. That’s exactly right. The AI can’t go out and just create chaos. It has to run within its lane. Makes a ton of sense.
Where does the human fit in all this? How does the human in the loop, or how is the human operating in all this scenario?
Angelo Stracquatanio: So one of the other benefits of creating software for extremely regulated spaces, whether it’s life sciences, medical device, et cetera, is that having the human in the center of the process, having the human sign off on what the system is doing or what’s being performed, is critical.
And so we took that exact same human-in-the-loop concept, and you mentioned the word guardrail, so that’s exactly what we call it. Whenever the agent goes and does something that triggers either what we call a connector, so integrating with an existing system, reading to it, writing to it, creating an output, a customer can optionally put in these guardrails where the human has to accept or reject what the AI is about to do.
And we combine that with what we call sources, where we give you a guardrail. It says, “Hey, I’m going to go do this thing. This is what I’m about to go do.” But then we also show the full source history of what led to that decision so that the human can ultimately say, “Yes, I feel comfortable with you, agent, performing this action.”
And it doesn’t need to be overly burdensome. The customer can decide what levels of guardrails they want, but our very strong belief system for this is that humans still have to be accountable. I know this sounds so simple and maybe it’s a bit overly simplistic, or maybe even a bit romantic, that I just believe that humans still need to be deeply, deeply associated with the process.
But even the name Apprentice, our company, we called it Apprentice because we believe in supporting, not replacing, the human. So we took that same philosophy and built it into the product where we want the human to be in control. We want the human to understand what this thing is doing, but we’re going to let the agent do a lot of the things that would have been manual before, that would have taken an extreme amount of time.
But we just keep the human in the loop at all times. So this way the human is in control. The human is accountable because, at the end of the day, the product that’s being created, that’s what they’re responsible for.
John Koetsier: I’m going to read you a quote from an IBM training manual in literally 1979, and this may or may not be familiar to you, but the quote is, “A computer can never be held accountable. Therefore, a computer must not make a management decision.” Agree or disagree?
Angelo Stracquatanio: I agree strongly because this piece about human accountability, I don’t think, is going to change. A lot of people are asking a lot of philosophical questions about AI and how does it relate to our world? How does it connect to jobs, et cetera?
In my view, and it’s not just me, the broader company feels the same way. AI is meant to support, not replace. Humans still have to be accountable. They have to make decisions, they have to understand what these tools are doing. And again, maybe it’s a bit derivative, but we feel that AI is just a different tool, no different than our manufacturing execution system, which can do a lot of things in a highly automated way.
But when it’s about to do a thing that’s going to create a different outcome, a human still has to be there to make a decision. A human still has to be there to say, “Yes, I am confident and comfortable that we’re about to change our production process.”
And so coming back to the IBM quote, I think the difference here is that throughput’s just going to go way up. Humans will still be accountable, but when these agents can do way, way more, then it just means that the amount of throughput per person is just going to go up.
And so it’s the same thing when it was manual spreadsheets before and an MES got installed. It’s just now going to go way faster. But the human has to be at the center of all this.
John Koetsier: I think that’s a very comforting answer for executives of manufacturing companies because, as we see legally, we see companies like Meta, Facebook, others, Google, who are getting sued for actions that their AIs have done, including some of the chatbots and personalities that are out there for very lonely people, and are getting sued for harm, yet there’s no hiding. Legally, there are people who are legally responsible, so I think that makes a ton of sense.
How autonomous do you see this becoming? That kind of comes out of this conversation we’re just having about humans in charge. How autonomous can this become? Do you see this growing and growing and growing? I mean, ultimately humans are still going to be in charge. They’re setting the guardrails. They’re setting up the bumpers on the bowling lane. But how autonomous can a factory become?
I even start thinking about things like robots coming in. Robots are in manufacturing all over the place, especially in some very high-tech sectors. We’re seeing humanoid robots starting to come in as well. How do you see the factory of the future?
Angelo Stracquatanio: I always come back to this concept of the plant maturity model, and folks, if they’re not familiar with this, can Google it, but there are five different levels of digital or plant maturity.
It goes from a level one, or zero, forgive me, that’s basically paper. It’s all manual process. There’s no automation. You’re doing things either for the first time, you’re learning, or you simply don’t have any high levels of automation, all the way up to what’s called an adaptive plant. And this is where I see the world heading.
I don’t think that we’re there yet because it’s not just a software problem, as you pointed out. It’s a hardware problem, it’s a robotics problem. We’re only, on our side of the table, one piece of the puzzle. We’re in software. But I do think that over time you’ll see this movement toward an adaptive plant.
And so what does that mean? An adaptive plant is where you’re not just automating a process, which I think when most people think of plant maturity, they think of, “Hey, I have an existing process today. It’s maybe manual. I want to then automate it.” And even within that very simple objective, there are different levels of automation.
There is, “Hey, I want to have a robot do this the same way. I want to have a human do this the same way. I want to use software. I want to use different types of software.” And so within that whole automation sphere, this is where I see AI working today, where today, AI, in our opinion, these agents are meant to take manual workflows that exist today and improve the level of automation for the workflows that are happening manually today, whether those manual workflows are paper-driven, systems-driven, data-driven, whatever is happening manually. I think that’s where agents are perfect for today.
John Koetsier: Mm-hmm.
Angelo Stracquatanio: But I do think that where this is going is an adaptive plant because I think what’s going to happen too is that you’re going to see higher levels of new product introduction coming into the manufacturing facility because the upper end of the spectrum is also going to change manufacturing’s downstream.
But upstream, there’s going to be higher levels of design, higher levels of new product creation, and ultimately manufacturing plants that need to adapt to new products, new product variants, new SKUs, whatever it ends up being.
And so I think what’s going to happen over time is that AI will go from being this automation layer, basically a new automation layer, that’s the simplest way I can describe this, to then being this adaptive layer where it’s reacting to manufacturing events. Those manufacturing events, or existing events, new products, new workflows, who knows what’s going to happen in the future.
But I think one piece of the missing puzzle is how will equipment adapt to this? Will the AI set new set points on its own? Will the AI transform and create a more modular manufacturing process? I think that’s going to take new tooling. I think that’s going to take new equipment. It’s going to take new hardware.
But I do think over time the plant will morph from high autonomy, which I think we can get to in fairly short order, to high adaptability because that adaptability is going to be necessary when the sheer volume of new products will increase because the AI is also going to help in design.
And when it helps in design, you’re going to have to create new products. And I think that’s where the adaptability over time will become pretty interesting. I just don’t think we’re there just yet because there are so many different aspects of the chain that need to get solved for.
John Koetsier: What’s super interesting to me about that is that is future-proofing, right?
Because you look at manufacturing nirvana, if you just look at the manufacturing piece alone, separate from any market conditions whatsoever, any product conditions whatsoever, manufacturing alone: do the same thing, do it repeatedly, do it again and again, get extremely good at it, zero variability, super fast, super efficient, super effective. Boom. There you go.
In the real world, we’re having new products and innovations happening more and more rapidly, and people want personality, so they want something a little different than what somebody else wants. There’s some variation within a certain product class, and that’s just getting more and more and more diverse.
And whether we get to the nirvana of 3D-printing everything, the cornucopia that just makes everything custom on demand for whoever wants it whenever they want it, probably not so much. It’s not very efficient, let’s put it that way, at manufacturing. More variability is certainly a market demand, right?
And so being able to rapidly reconfigure, rejig your manufacturing capacity to create 10 times as many products, a hundred times as many products, can you imagine a thousand times as many products with the same sort of physical space or hardware? I mean, the mind boggles, right? But this sort of thing is kind of a necessary precondition to be able to enable that future if we’re going there.
Angelo Stracquatanio: That’s exactly right. And interestingly enough, we have customers doing this today. So a portion of our customer base is creating personalized medicine.
So think about this. On one end of the spectrum, you have additive manufacturing with 3D printing or subtractive, where you’re creating these one-off parts. You can maybe create your initial assembly, and there’s some variability there. Maybe you’re tweaking the features, the specs, whatever it is.
On the other end, you’re creating an individual medicine for a person based on their DNA and their blood cells because, if folks are familiar with CRISPR or have heard about CRISPR technology, for the first time in human history, we have the ability to edit our own genes. And when we can edit our own genes, we can not just treat disease, we can cure disease. But to cure disease, this happens on an individual patient basis.
So if you have a patient who’s very sick, they’re in the hospital, they’ll actually take their white blood cells, their DNA, and ship it to a manufacturing facility. This happens today. Currently, Apprentice is producing with our manufacturing execution system two thirds of the commercially approved gene therapies in the world. So this is our core foundation. We’ve been doing this for the better part of half a decade now because these drugs are pretty new, but that manufacturing happens on a one-off, individual-per-person basis.
Yes, is it expensive? Yes. Is it extremely complicated? But when you look at the process and what goes into creating individual medicines for individual people, there are some patterns, and those patterns that we think can start to be extrapolated out to other industries, other processes, to create higher levels of personalization.
The personalization could be on a per-product basis, or it can be completely different products over time. And so I do think, and it does not in my head seem crazy because it’s happening today for medicine, why can’t it also happen for many other products?
And so while yes, the core basis of manufacturing design and nirvana is repeatable, consistent, high-quality products, then it’s just a transference to high-quality, repeatable, consistent products just on an individual basis.
John Koetsier: Mm-hmm.
Angelo Stracquatanio: And I do think that is where this world is moving to. Will it happen this year? No. But I do think that over time this technology enables that type of approach that’s already happening today just with medicine.
John Koetsier: Wow. The craft economy meets the Industrial Revolution. Very, very interesting.
Let’s talk about some use cases, some examples. You’ve been working on this for a while. You’ve had it in with some customers, some beta testers, and some people are using it. What are they experiencing? What are they getting? Give us some numbers, benefits, improvements, speed-up, quality improvement, whatever it might be.
Angelo Stracquatanio: So we’ve been working with our core customers now that have already had our manufacturing execution system for many years, and we’ve had AI embedded into that product for two years now.
We are doing things like authoring recipes. So instead of manually authoring a recipe, the agent will author a recipe for you. We are doing it for quality reviews. So instead of manually reviewing your quality process, they would do it for you. And we applied this to this new agent, our A1 agent. Now, not only could it do it within our system, it can cut across many systems.
And so the first immediate benefit that our customers saw using A1, not just within our MES but cutting across their various systems, whether it’s ERP, QMS, SCADA, PLC, we even connected it down to an IoT-level network where it’s trending and actually capturing the data in real time, the first biggest area of benefit that we saw was any manual process that they’re doing today dramatically compressed the cycle time.
So let me give some examples. Troubleshooting is a big area of focus because anytime that the line is impacted and it’s down, the time to resolution is extremely high value for our customers. And so just very simply getting time to resolution is a big benefit because now they’re not going through and doing any manual review. It’s then getting them to their process that they need to perform, and then the humans go and perform it.
Quality review: some customers have 40% of their cost of goods sold, 40%, that’s focused on quality review time. Now granted, that’s for a very specialized product. It’s not for your commoditized products, but for industries that have high quality or review time, pressing that down by a third or more has not only a direct impact on time to market, but it has a direct impact on cost of goods sold.
And so this is where the agent, that’s not really doing anything wildly sophisticated, it’s just helping the human review and investigate faster. There’s direct one-to-one correlation with cost of goods sold.
Conversely, then you move up the ladder and you get to more advanced analytics, and you’re starting to identify, well, what are my trends that I’m seeing within manufacturing? And this is where then, instead of creating custom visualizations, custom integrations, custom data store mapping, the agent can do those associations for you, put it together, and create an analysis.
And then the last area is around alarm triage. This is by far the most popular use case that we see because right now, to be frank, most manufacturing teams are ignoring a lot of their alarms. It’s just the reality, because you have these things publishing throughout the day. Some of them are extraneous alarms. It’s just a low-severity thing. They’re more like info publishing versus actually something severe.
But the takeaway is that manufacturing teams are drowning in a lot of these system alarms and system events. When you can have the agent do the first-pass triage, even if it’s not going to full resolution, you still have your team doing resolution, if it’s just doing the first pass alone, it’s compressing the time that’s taking these teams to identify what do I even focus on versus swimming in these alarms in a given day.
And so whether it’s any of these different use cases, the biggest area that we’re seeing, number one, is just very simply time compression. And time compression does matter, and it has a really high ROI.
But over time you can graduate to these much more advanced use cases where the ROI can draw one-to-one comparison with either cost of goods sold, throughput, or quality. But that’s, again, on a per-team basis, how much they decide to implement AI, or they just get started simply and say, “I just want to do some basic paperwork reduction.”
Frankly, in and of itself, the quality-of-life improvement is high on its own.
John Koetsier: Super interesting. And I can imagine that this flood of alerts, which are somewhat ignorable, cascade and eventually do things that you cannot ignore.
I want to maybe end here, and this is a very interesting class of products for a very interesting job geopolitically right now because if you look at manufacturing, some massive fraction of the world’s manufacturing is overseas in China. It just is.
And I’ve talked to some other companies as well who are starting to apply AI to manufacturing, and they view it as a chance to not only level up manufacturing, but also be able to have manufacturing even in high-cost-of-living areas: the United States, Western Europe, other places like that, and to take it to the next level in terms of quality and cost, as you mentioned, and efficiency and effectiveness. How do you view what you do through that lens?
Angelo Stracquatanio: So it comes back to what is AI ultimately doing for these folks in manufacturing? And it’s taking your existing labor base and increasing the throughput of your existing labor base.
And so when you’re thinking about, well, how do I adapt, how do I compete, because there’s a lot of pressure right now. The tariffs are causing a lot of onshoring. The tariffs are causing a lot of cost recapturing in the sense of cost optimization, meaning, hey, if I can’t move my plant, I still need to be able to compete effectively.
Well, I have to look at my margins and say, where is there opportunity where previously maybe I didn’t want to or wasn’t motivated enough to focus on margin optimization? Well, this is where AI can come in because when you look at what goes into the plant and what goes into cost of goods sold, you have your raw material, which for right now, I don’t think AI is going to touch raw material. That’s fine.
It’s not going to change your CapEx investment just yet because the actual hardware is going to have to change over time, and you’re going to have to bring in either new equipment or change how the plant operates. But this is where AI can really benefit, in this third layer, which is labor and equipment and asset utilization.
So for the labor piece, it’s very obvious. There is the manual work that a human does with their two hands, and no, AI is not going to move their hands. I’m not some totally disconnected dude that thinks that, like, oh yeah, AI is going to just magically make Bob move his hands faster. No, that’s obviously not the case.
But Bob does a lot of other things beyond just moving his hands. Bob does a lot of things when it comes to reviewing paperwork, summaries, analysis, troubleshooting data. There’s so many other things that are part of Bob’s job beyond simply, Bob moves his hands to do this thing.
And so I think AI is going to help Bob do way more when it comes to the non-hand-movement stuff because ultimately that aspect of labor utilization is still an aspect of labor utilization that can be automated a lot more with AI so that Bob can focus on what he is really great at.
Conversely, then there’s asset utilization, and this is where I do think AI can be a great benefit to be able to better understand how can I optimize my line to either increase throughput, improve quality, look for issues or problems that I didn’t even know were existing on that given line, and help optimize that line.
Because ultimately, if you can drive higher throughput, maintain the same level of quality, get more out of your labor today to drive down, ultimately, the averaged cost of goods sold, all of a sudden your economics do start to change. And that’s my strong belief system, that if you apply AI in the right areas for the existing assets you have today, both labor, equipment, et cetera, you will be able to find clever ways, by automating those workflows, to then compress your costs of goods sold.
And if you can compress your costs of goods sold, then ultimately what’s the competitive advantage if it’s being created in other parts of the world? If you have a great labor base, if you have great equipment, you have high know-how, then if you just optimize your COGS, you are getting into ranges that do work.
And so yeah, that’s my kind of—
John Koetsier: I think that makes perfect sense. I absolutely love that. And you stack on top of that the ability to maybe rapidly reconfigure your manufacturing capacity over time as you get into higher-cost or higher-value goods that are more customized, various sectors, even personalized for individual people, and onshoring or nearshoring becomes very, very critical because it just makes everything easier.
So I think that’s super, super interesting. This has been fascinating as a whole. Really appreciate the time.
Angelo Stracquatanio: John, this was an absolute pleasure. Thank you so much for having me on the pod here, and I’m very much looking forward to seeing how this thing evolves in the next couple years.
Because we’re all learning together.