Artificial intelligence is everywhere in manufacturing right now. But for many plant teams, it still feels more like a concept than a practical tool. In this episode of Manufacturing Matters, Aaron Hand sits down with Alex Sandoval, founder and CEO of Allie, to break down what AI actually looks like on the factory floor.

Sandoval explains how AI co-pilots can connect to PLCs, quality systems, and maintenance data to act like a 24/7 expert supervisor — surfacing problems before they cause downtime and guiding operators toward faster, better decisions. But the technology itself isn’t the biggest hurdle. From disconnected machines and unclear data ownership to the cultural challenges of trust and fear, most manufacturers aren’t failing because of AI. They’re getting stuck before it can deliver value.

Aaron Hand: [00:00:07] Hello and welcome to this episode of Manufacturing Matters, where we talk about the technologies and trends that are affecting global manufacturing. I’m Aaron Hand with Tech B2B Marketing. I’m here today with Alex Sandoval, founder and CEO of Allie. Welcome, Alex.

Alex Sandoval: [00:00:24] Thank you so much, Aaron. Thanks for having us.

Aaron Hand: [00:00:27] No problem. So, before we get too far into this, what Allie does is develops AI copilots for manufacturing. So, connecting factory equipment and systems, turning real-time production data into decisions that improve productivity and quality. So, is that a fair assessment, Alex?

Alex Sandoval: [00:00:47] Yes. That is. And we like to think of ourselves as a system of intelligence. So, again, we connect to all types of data sources inside of facilities around the world, from PLC data, telemetry sensor data, like temperature sensors that are calling pressure or viscosity all the real-time metrics that matter. We also connect to adjacent datasets like quality datasets from the lab, from tests, maintenance logs. So, we have a really good 360 view of a manufacturing facility. And then we plug that into our AI agent architecture that essentially you can really think of it like specialized AI workers that are an expert in your plant and expert in your process. And then they are continuously making decisions with the objective of ultimately impacting productivity, quality, OE, all the metrics that plant managers really care about.

Aaron Hand: [00:01:49] Okay, great. Thank you. So, you know, we’re hearing about artificial intelligence everywhere in manufacturing right now, but a lot of it still feels pretty theoretical. How do you explain to a plant manager, for example, what Allie AI does in plain terms?

Alex Sandoval: [00:02:08] Yeah. Look, in plain terms, you have to really think of one of our AI agents like your best shift supervisor. Somebody who’s been working there for 20 years. The one who knows why a line slows down on Tuesday morning. The difference is this is a cloned version of that shift supervisor awake 24/7 across every line and every plant. And really, we don’t replace operators. We just give them a tool that augments their decision making. Because when you have so much data and so much information to really quantify and analyze, it’s much easier if an AI agent can surface what really matters towards you. So, what the agent does is it catches a problem. So, say there’s a temperature deviation in line three that you’re going over by 10 degrees, and in the next 20, 30 minutes, if that trend continues, you’re going to have a quality event where you’re going to have to stop the line and restart it, right? So, what the agent does is to send you a notification to say, Hey, the temperature is trending in a negative direction. You should really consider making this intervention, which is changing that temperature. And the operator will accept that intervention, will change that temperature, and then the line is back on track. So, it’s just a plant manager or an operator or a line supervisor having a tool with superintelligence of your line, of your process, of your machines to help you make decisions.

Aaron Hand: [00:03:47] So, of course, in order to be able to do all of that, to be able to make these decisions, you really have to have the data in place. You have to have good data. And AI-ready data, specifically. So, give us a practical look at what that actually means in factory operation. How do you get your data ready?

Alex Sandoval: [00:04:10] Well, the first thing is making sure your PLC data is open and there. A lot of times what we see in the plants that we visit is you already have a ton of data, but it’s just not being exploited. It’s not being used in the best way possible. So, if you look inside of a PLC, and for those of you outside of the industry, a PLC is the controller, right? Like the brain of the machine where it’s hosting a ton of datasets from everything the machine does, from set points to process variables to machine health variables. But it’s surprising when you open that and you open that PLC, you can have up to 10,000 tags that are running every single second, right? So, that’s a gold mine for AI. And really, the first thing we should look at is your PLC data, right? Now, the problem with that data is the machines actually don’t talk to each other. You can have a machine that you just bought, and it can have those 20,000 tags, but it’s not talking to the machine before, and it’s not talking to the machine after that. So, in a process industry where you’re transforming products and one machine really depends on each other, that part is, really important. Now, complementary to that rich PLC dataset, you have other datasets that are important. I would say maintenance is a really critical one. Where do you manage all your work orders to make sure that you are making the, or handling the jobs that the machine needs to be able to be maintained in the right way. So, that’s like an oil change or a change of component once it hits a certain number of cycles.

Alex Sandoval: [00:05:58] So, all of those things are really useful for AI to know if actually the problem that we’re seeing in the line maybe is deriving from mechanical failure or  something that has been missed in terms of maintenance. And then the other dataset I would say is quality is super important. So, typically, you have either a lab that is running batch testing, or you even have sometimes an in-line quality machine that is testing everything in real time and kicking the products out or rejecting the products if there’s any misalignment to the quality spec. So, all of these things are important. Now, I do want to say that 90% of the plants that we visit do not have AI-ready data. This is something that we actually work with the facilities, work with the plant managers, the IT/OT people to really figure out what’s the data that you actually need? Because sometimes, it’s not about connecting all of your datasets. It’s actually connecting to your most important ones, so you can start to deliver value, right? And you can pick a selection of what those datasets are, pull them together through a variety of techniques. You can do it in a server, you can do it in a data lake, you can do it in a database. So, wherever really that suits you. And then from that, you can start to run testing on how models ingest, are trained on those datasets to be able to deliver real value to your plan.

Aaron Hand: [00:07:37] Okay. Certainly, I think the problem is rarely a lack of data, right? I mean, I think back in the day, which probably wasn’t even that many years ago, I feel like we were writing about big data all the time. And we don’t call it big data anymore because it’s just the norm, right? I mean, I think they recognized how much data was flowing into the the manufacturing floor. And everybody knows there’s tons of data coming off all of those systems. But the trick is how to actually bring that data together in a meaningful way. So, is there a way that people’s efforts typically stall at that data layer? I mean, what sort of problems are you seeing the most when you walk into a factory?

Alex Sandoval: [00:08:29] Well, the first thing I would say is I don’t think a lot of companies agree on who owns the governance to this data, right? Because ultimately, when you look at manufacturing data, you generally have two types of datasets. You have data that comes from the IT layer and data that comes from the OT layer, right? And they’re very different datasets. And typically you have a team that manages the OT and a team that manages the IT. Now, in order for AI to work, IT and OT need to converge. They really need to work together in a single unified architecture. So, who owns that? Is that the IT system? Is that the OT system? Is that the manufacturing leaders? So, ultimately there needs to be a decision in terms of ownership and in terms of governance. Now governance in the AI world is really, really important, right? Because any change to architecture, any change to a dataset is really going to impact everything that you’re building on top of that. So, say like, Hey, I want to change something as simple as the amount of seconds that I am using to collect my speed metric for one of my lines. And just changing that, even though it sounds very minuscule, has a huge impact on your production targets, a huge impact on your efficiency metrics and your efficiency reporting. All of these things need to be figured out. So, one is ownership and two is governance.

Aaron Hand: [00:10:06] And you talk about that gap between IT and OT, which becomes an issue not just with AI, but I would think you’d have governance issues also between supplier and customer. If you have an integrator working on your system, or if it’s a particular company’s equipment, are there even more issues about who owns that data?

Alex Sandoval: [00:10:32] It depends on your contracts and the fine print between your OEM, your machine maker, and yourself. I think that adds a layer of complexity to the way that you are using and managing data. There are some clauses, and it’s really important. And I say this disclaimer to anyone that is negotiating a machine right now or buying a machine to make sure that they really include the data as part of that purchase. Because truthfully, as a machine owner, you should own the data that is generated from your machine. You should at least be able to have read access to that data, so you can store it. And data is, you know, back in the day, people talked about oil, but I think you’ve heard the phrase, data being the new oil. And it’s important that you are the owner of your most valuable asset. So, you know, sometimes you start using a machine and a year later, you want to run a project like this, and you read the small print of that contract and it says, Hey, actually you don’t own your data. Your OEM or the machine maker owns the data, and you have to pay some absurd fee to be able to gather the access to the data that you already have in your facility. So, that part is really important. And I would really encourage anyone who is negotiating a new facility, buying a new machine, to make sure that they own that data because then they’re going to regret it when they want to use it in different ways.

Aaron Hand: [00:12:15] Allie is a strong player in the food and beverage industry, correct? CPG, as well, consumer packaged goods. So, let’s get a little deeper into there. We share some common ground there. So, that might be a language I understand a little bit better. Do you see any unique things going on with food and beverage or CPG that makes this problem harder than any other industries?

Alex Sandoval: [00:12:40] Yeah, 100%. So, we really started with the process industry. Manufacturing has process discreet and then hybrid manufacturing. And then within the process industry, you have industries like food, beverage, pharma, plastics. So, a ton of industries that, in essence, what they do is that they have this concept of a recipe. You have an incoming raw material that could be one of many different types of materials. And they go through this continuous process where you’re transforming those materials continuously. And in order for the next process to accept the incoming material, it must meet a very specific spec, right? It must meet a standard. Otherwise the process is broken and isn’t going to be able to finish at the correct spec of what that finished product needs to be. Now, if you think of a line in the process space, you can have up to 25, maybe 30 machines that are part of one line from start to finish. You can receive, maybe you can do some mixing, some cutting, some baking, some freezing. So, all these processes that happen in the middle. Now, what happens if you make a mistake in one of the processes? And what is making a mistake mean? Well, actually you’re controlling maybe 20 to 30 variables per process. So, multiply 20 variables times 40 stations. That is a lot of complexity to manage. And you cannot make a single mistake. Otherwise that mistake is going to cost you downtime, is going to cost you probably reprocessing and waste that, you know, you’re going to have product that you can’t even use. So, that level of complexity. And then one more thing to add. We were just talking about machines.

Alex Sandoval: [00:14:46] So, you have machines from different OEMs. You can have maybe 10 to 15 brands of different machines that are not talking to each other, that don’t share a common language, that don’t share a common code, even from the PLC, and that is like a third component that adds to this complexity. So, this is why, in my opinion, all this complexity is much better managed through a model that really does not have a limitation or gap of how many things it can process at the same time. So, a model can go and run all of these datasets, can figure out what’s in spec, figure out what’s out of spec, optimize what’s the best set point that a specific machine needs to run, and even predict what’s going to happen in the future, what’s going to happen in the next 20, 30 minutes. So, this is why AI and models really become the best at analyzing, processing, and looking at all of this information and then surfacing to that line supervisor what’s important. Because imagine a line supervisor having to manage all of these things at once, all the time, and at the same time they have a job to do, which is to report, to be able to hire people, to be able to make decisions on new lines, on new machines. So, we really try to make that easy for them so that we have this tool that is continuously, every single second, analyzing all this information and being able to surface what’s the highest value action you can take to be able to meet your objective. And that’s really what makes this segment unique and why it makes it even better prepared for AI.

Ad: [00:16:38] Everybody, I just want to take a quick second to thank our sponsor. Manufacturing Matters is sponsored by Tech B2B Marketing. They’re a full service public relations and marketing agency that focuses on technology companies, especially in the energy and automation markets. They provide full service–media relations, investor relations, employee relations, full end-to-end content development, video services, animation, full IT stack development from website integrating with ERP systems, CRM, marketing automation systems. And they bring a whole lot of knowledge and experience about technical markets since they’ve been servicing those markets for over 30 years. So, if you have any questions, you want to learn more, go to techb2b.com. And now let’s get back to the show.

Aaron Hand: [00:17:19] So, you talk about all of those variables, all the different points in a factory. And I think about the food industry in particular. Of course, pharma has faced demands for traceability for years. But the food industry that’s really ramping up with FSMA, the Food Safety Modernization, I can never say that word, Act. In the US, where the FDA rules, putting a lot more pressure on food companies to have that data easily accessible if, for example, there’s a recall. So, right now those rules don’t say, they don’t dictate that you need any particular technology, but what they do dictate, and I may be getting this wrong, the exact details wrong, but basically what it’s dictating is if you have a safety situation with the food, a recall situation, you need to be able to produce all your records within 24 hours. And sure, you can do that with Excel spreadsheets if you want, but good luck actually being able to produce the records then. So, you really do need to have all that data. Digital. Digital. There’s another word I can’t say, digitalized. Anyway, talk a bit about FSMA 204, which is basically a list of high-risk foods and the demands there for traceability. Where does AI fit into a company’s ability to react quickly in those situations?

Alex Sandoval: [00:18:57] Yeah, and not only food. We work with pharma companies as well. We’ve also visited auto companies, auto parts company, and they all have the same problem, right? You’re in an industry where there’s a lot of high risk to the end consumer if you make a mistake during your production, right? Imagine feeding your baby a specific type of food that has some sort of quality defect that can put their life in danger. And that happens the same in pharma because we’re putting this stuff into our body to be able to have some sort of effect in what we’re trying to achieve. And the same for cars. If one of these pieces has a mistake when you’re 100 miles per hour in the highway, that’s also going to be a big problem. So, a lot of regulators expect you to have every single decision and store it in the last five years of any decision or intervention that has been taken on the line, so that they can trace back, what was the source, what is the root cause of why this thing happened, right? And that means you have to have lot, location, time, and then specific characteristics of that batch queryable, on demand, whenever you need it.

Alex Sandoval: [00:20:20] And a lot of these types of companies can produce this data eventually, but regulation is demanding that it’s immediate. So, actually, in some way, regulation is doing good things in the sense that it’s almost putting demand on the manufacturers to have this data streaming and available in real time. But that has a secondary use case, which is not just fulfilling what the regulators want when they need it. You’re going to have all the data ready to go to be able to deliver value and apply models and train models from the same effort. So, beyond compliance, the same infrastructure that makes you FSMA ready is going to make you AI ready. So, don’t just treat it as a checkbox that you need to fulfill from your regulator. Treat it as an initiative that is then going to help you apply models on top of to be able to deliver more value to yourself.

Aaron Hand: [00:21:17] Let’s get into the AI agents specifically. What sorts of decisions are those agents making? And are they operating autonomously, or is that still hand in hand with the operator? Kind of describe what that looks like?

Alex Sandoval: [00:21:35] Well, the first thing that I want to note that is important is our solution, an AI agent, is not just a chatbot, and it’s not a dashboard, right? Because I think there’s a big misconception because of the usage of ChatGPT, the usage of Claude, that everybody seems to assume that the chatbot is the interface in which you’re interacting with AI, and AI is purely conversational. AI has been a thing for 20, 30 years, right? Because there’s a ton of techniques that are a part of AI, from machine learning techniques to vision to different types of data models to LLMs, which is kind of the newest thing that AI has really produced where it means AI went mainstream, right? But I just want to clarify that AI has been a part of the technical field for 20-, 30-plus years. Now, when you think of AI and an agent interacting with your datasets, what does that experience feel like? So, that can be something as simple as a notification on your app in your phone that says, Hey, I’m detecting the incoming batch to be off spec by, you know, the viscosity is not meeting, the viscosity, which is the density of the product, is not meeting its spec to go into the next product.

Alex Sandoval: [00:23:12] Please go and add two liters of water to that mix. And it seems very simple because it’s just a recommendation of something that you need to do. But behind it, there’s a ton of things that had to happen to be able to produce a recommendation of that quality and also a timely one that gives you an opportunity to be able to intervene with enough time to be able to cause an actual impact in that decision. That’s one way that we can interact with it. It’s just as simple as receiving a notification. In the AI world, there’s the concept called human-in-the-loop, which is you have a human who’s a part of that decision, who’s a part of that recommendation, verifying that the quality of that output that AI produces is good and is ready to take into the next step or to take into production. So, we do both mechanisms. We have some recommendations that have a human-in-the-loop that ultimately say, Hey, I approve this recommendation, or maybe it requires a human action, right? Like maybe you actually need to go and check and take pictures and do a certain type of intervention that AI doesn’t really have that capability or, we don’t have an autonomous robot that is going and doing these things.

Alex Sandoval: [00:24:37] So, it’s going to surface that to the human, to the operator that is reading that, and then they approve. And then eventually for certain types of decisions that are sort of lower risk, we’re going to have what we call closed-loop decisioning, which is AI doing this all on itself. So, AI detecting a deviation. And through that same PLC, instead of reading the information, it’s going to write on the PLC and make that change. I don’t think that we are there yet, so that’s definitely something that is still a few years away, I would say. And it’s not really, because the technology’s there, but there’s a lot of, because the technology currently does have this capability, but there is a lot of trust and change management that needs to happen on the way that these decisions are done. And it’s about maybe restructuring or reorganizing your workforce to be able to adapt to these things in the best way possible. And I think that’s what’s going to take some time before we let AI really run loose on manufacturing systems.

Aaron Hand: [00:25:54] That’s part of what I want to get into next is where we are in the journey. But I think, this example is so old, but it’s still so relevant is just how we all interact as drivers with map programs. I’m somebody who just gives my blind faith to the data, and I’ll just go wherever it tells me to go. And it doesn’t matter where I’m going in town. There’s three or four different ways. And I’m going to use it every time. And I look ridiculous because they’re like, you don’t know your way to the grocery store, but it’s going to tell me the fastest way. But they’re always going to be drivers that say, that’s absurd. Why would it tell me to go that way? I always go this way, which is, by the way, my wife sitting next to me in the car. Why are you going this way? I go this way to work every day. So, I think you’ve got the operators facing that with what AI is telling them to do. So, where we might be better off, I would guess, to take the human out of the loop, and let the AI use the data it has and not necessarily be so questioned. Can you give us a better idea of where we are on that journey? Is there still a lot of cultural aspect that we need to get over to let the technology do what it does best?

Alex Sandoval: [00:27:20] Yeah, I think it goes back to the concept of, as you said, trust, right? And like any relationship, trust is not built in the first minute of you getting to know someone, right? It takes a long time before you trust your girlfriend, your boyfriend, your partner, your roommate, your boss, your colleagues. And I think that same principle applies to AI. How do you build trust? So, building trust like you would build with any other human takes you ensuring that their values, their decision making, is aligned to yours. And it’s something that makes you like them more, makes you trust their decisions, makes you trust the way that they are operating. So, I think it’s really a time thing. It’s about letting people interact with AI, letting people see if the decisions that they’re surfacing and recommending are ones that they would take themselves. Then, that would make them much more aligned then, Oh, AI is thinking the same as me, like a process engineer, and is putting forth the same recommendation as I would do. And then once you start to get predictable and repeated output that meets that same quality, then operators are going to start trusting it. I think the second thing that we need to overcome beyond the trust thing is fear. There’s a lot of fear in AI, and fear is stemming out of . . . whenever there’s something that comes along that is so transformational in the way that it transforms the way that we work, transforms the way that we communicate, transforms the way that we operate, it’s a really scary thing.

Alex Sandoval: [00:29:12] Especially if you as an engineer have been working in this line, working in the system, and you have gained so much knowledge that being an engineer in a manufacturing line almost becomes your identity, right? And being able to trust the system to come in and start to help you in taking these decisions is a scary thing for an engineer because then they think, Oh wow, what if it starts taking some aspects of my job and taking some aspects of my decision? What am I going to do? Am I still an engineer? It’s even challenging your own identity. And I would say that I don’t think we need to shy away from the fact that there’s a lot of mundane and repetitive work that is going to be automated through AI. But when I talk to a lot of engineers, I asked them like, Hey, do you enjoy sitting and looking at data for hours, and putting together a ton of spreadsheets, and trying to manage that complexity so that you can produce a report? Or would you rather do it in three minutes with the aid of an AI tool? And everybody says, Of course, I’d rather do it in three minutes.

Alex Sandoval: [00:30:32] So, when you start to see it as, okay, this is not coming after me, but this is actually something I can use to my benefit, then you start to trust it more, and you start to trust it integrating in your workflows. I will say that the workers of tomorrow are AI-enabled workers. And if you’re someone who is not adopting, is not being enabled, and is pushing away these technologies, I would say maybe not today, but in a few months, your job will be at risk because the people whose job won’t be at risk are the ones who are using it, are the ones where getting the productivity gains, are the ones where being much smarter, who can take a lot more decisions at the same time because they’re aided by the technology that is taking away a cap on the things that they can do at once. So, I would encourage everybody who’s really watching this to go and try this out and to go and interact with it because it’s going to make you a lot more powerful at work.

Aaron Hand: [00:31:43] Right. Absolutely. And I’m a broken record, but I just keep going back to the quote of, You won’t be replaced by AI. You’ll be replaced by somebody who knows how to use AI, right? So, as much as we hear about AI and manufacturing, we also hear plenty about the failures of AI in manufacturing. Do you think it comes back to that cultural issue, or what are some of the common mistakes that companies make when they’re trying to implement AI?

Alex Sandoval: [00:32:15] This is probably one of the things that we hear the most because there’s a lot of studies right now. I think MIT recently published one, less than 10% of manufacturers have adopted AI at scale. And I think there’s 70 or 80% that have tried it. So, there’s a huge gap between running an AI pilot, which everybody’s doing, and actually getting it to work and getting it to scale and grow in your organization. And that’s where everybody’s getting stuck. What are the problems that we see? We see a few common patterns. The first one is AI not being championed by somebody with board-level or real decisioning power in that work. And if AI is being championed on its own by a line supervisor in one facility, and it doesn’t really have that visibility or championship by someone who can say tomorrow, Hey, all right, let’s go and let’s deploy this. We see that process break the scale-up decision very quickly. So, whenever you start a project, make sure that it’s not starting in isolation in a local facility without the championship of somebody who really has authority and decision-making power within that org at a central role, right? Somebody who’s at the HQ of that operation. The second one is really trying to do too much with AI. So, thinking that AI is this panacea, know it all, solve it all, and it’s going to come in and solve every single one of your problems.

Alex Sandoval: [00:34:12] You know, that really not the case. I think what a lot of companies get right is making sure you bring that down to like one, two, three use cases that are very, very specific, very, very niche. And you start with that and getting them really right. Making sure that you have the data for it. And you can narrow it down to the use cases that are really going to be the most valuable for you. And then I would say the third thing has to do with data quality. So, it is running an AI project with really bad data. As you, as you say, and as a lot of people that have worked with AI know, your AI output is going to be the same as the quality of the data you put in. So, if you have really bad data and really bad data means you have manual inputs, you have data that is not organized in one place, that’s not clean, that has a lot of mistakes, then that’s the same thing that you’re going to get on the other side of the model. So, I would say those are the three things why people are having trouble scaling.

Aaron Hand: [00:35:32] When you talk about not trying to solve everything at once, start with a particular project, can you give an example from one of your customers? I know you can’t necessarily name the customer, but can you talk about what they were trying to do and maybe the struggles they were having and how AI helped out with that?

Alex Sandoval: [00:35:54] Yeah, absolutely. So, I’m going to give you an example. I’m going to give you two examples of common things. We do a lot of work with customers that have filling processes. So, that’s a filler. You’re putting liquid or semi-liquid into a bottle, a jar, a can, for whatever use case it can be, it can range from makeup to beer to water to juice to yogurt. So, all of these processes use a filler. Now, in the filler, the filler is a quite a complex machine. You know, it has valves and it has discs, and it has a number of components that are breaking down all the time. Typically, more advanced fillers have automated fault codes that tell you sort of a bit of an indication to what’s going on. But it’s really hard for an operator to become an expert in all of the filler components. If you read a manual from a filler, it’s like 400 to 500 pages of information about how a filler needs to be run, how it needs to be maintained, how all of these components and systems work together. And then not only that, it’s knowing how to solve them.

Alex Sandoval: [00:37:16] And what’s the intervention? What’s the routine that you actually need to do once a problem hits? So, we have a use case that it’s like a maintenance-augmented operator that uses AI to be able to solve problems quicker. So, typically there’s a problem. And that can be, your component is failing, which means your line stops. And in order for you to solve it, you need to do all these tests. Maybe check some components, check a filling level, or you have these certain types of metrics you need to check so that you can narrow down the root cause. And think of it like a doctor, right? Like a doctor is measuring your heart rate, and measuring your temperature, and measuring your pressure, and all these things and then asking you some questions. So, it’s the same thing for an engineer who’s at the filler. They need to go and test all these symptoms and ask these things before they get to the root cause. So, that process takes maybe one, two hours, right? From when you have a problem doing that diagnosis, doing that root cause analysis, and then figuring out what that intervention is. And that takes a team of two, three people.

Alex Sandoval: [00:38:32] And in those two hours, you probably stop bottling 150,000 units, right, because you’re talking about a high-speed filling machine. So, if you could do that in 10 minutes or 5 minutes, because you have AI quickly correlating all of the variables and all the things that happened in the last 30 minutes. It’s doing pattern matching. So, all right, so when this event happened or this fault code happened, what were the things that happened right before it, right? And it’s doing all of it in seconds, you know, in five seconds. And it can go and say, all right, I read the manual, which it can also ingest and interpret in seconds and do a root cause analysis in seconds and then produce an action plan in seconds. So, that same thing that took two hours and three people now can be done by one person in one minute, at least to diagnose and do the root cause and then another few minutes to do that intervention. So, it’s a really powerful thing to move from signal to action and compress that from hours to really minutes.

Aaron Hand: [00:39:45] Great example. One of the things I’m trying to get my head around too is just where AI or the AI platform fits into the whole manufacturing stack with ERP, MES, SCADA, not to throw too much alphabet soup out there. So, enterprise resource planning, manufacturing and execution and maybe your supervisory control, how does it fit in within those existing software systems?

Alex Sandoval: [00:40:16] Absolutely. Well, two things. We sit above a SCADA, and I say above only because we are ingesting all of the data that a SCADA system is collecting. I would say we we sit alongside an MES, a manufacturing execution system, and below an ERP. We’re not replacing any of them. So, this is a new system, and we call it a system of intelligence. You know, we have, I think of your ERP, your MES, your SCADA, your systems of record, and perhaps your systems of execution. And then this I would call a system of intelligence. Why do I call it a system of intelligence? Because we’re unifying signals, running models that understand correlation, that understand causality, and then expose those decisions to the user via agents. Existing systems keep doing what they do. They’re collecting production data. You’re doing production planning on there. You’re collecting signals. So, this is connected to the systems via API or via MCP or via any sort of integration mechanism where we’re able to read these signals in real time, understand them, run causality, run correlation, and then be able to expose those decisions, which can happen directly to the operator or even back to the system. Sometimes we write back to the ERP, we write back to the MES, and that’s something that’s very natural.

Aaron Hand: [00:41:53] So, we’re taking a lot of your time already. I don’t want to take too much more, but for manufacturers who are tuning in and they’re thinking about AI in their operations, what is your advice for the first practical steps that they should take?

Alex Sandoval: [00:42:14] First practical step is start small. So, don’t try to do this in five plants, 10 lines. Just pick one. Pick one plant. Pick one or two lines. Make sure you instrument it properly and you have all the datasets you need. Prove the loop and the model is is working predictably, accurately to the quality that you’re expecting. How fast did someone make a decision through the model is really the KPI that you should be tracking. And obviously, did some of the key performance indicators of your line increase? I think really that’s the main decision that you need to take. And build the right team around this, right? Don’t try to do it yourself. Make sure that you involve your IT person, your OT person, the plant manager. You get a high-level executive to at least know that the project is happening. And then, once everybody’s aligned that this is something you want to do, then go for it. And don’t stop until you get something that you can scale.

Aaron Hand: [00:43:31] Well, thank you, Alex, I really appreciate you taking the time to to talk with us today.

Alex Sandoval: [00:43:38] No, thank you so much, Aaron, for the time. And it was a great convo.

Aaron Hand: [00:43:43] All right. Thank you. And thanks to our viewers too for for giving us your time and joining us on Manufacturing Matters. If anyone has any questions for Alex, you can put them in the comments below, on LinkedIn or YouTube or wherever you might be watching. You can also access this podcast and any of our past podcasts on our website at manufacturing-matters.com or your favorite podcast platform. So, for now, do us a favor, hit like and subscribe and keep tuning in.