#26: Agentic AI in Operations Management – from insight to intervention

Rafael Amaral (00:00):
If you are a plant manager and you have those few hundred, 300 employees and they're running shifts and you're producing nonstop, the amount of noise, the amount of request for your attention to fix something, it's a bit of a firefighting scene.

(00:19):
Once we started using Aura against the TilliT database, it was absolutely mind-blowing the result because suddenly you start to correlate the issues that you're having with your machines with the process that the operators are doing or not doing.

(00:36):
If there is some sort of a power that allows you to see things that you're not seeing, that is untapped value by definition. You read the report, it makes sense to you, but it had never come to you. You never thought about that. That's untapped value.

Søren Hammer Pedersen (00:55):
When you come in Monday morning, what is it you get in your hand?

(01:10):
Hello everybody. A warm welcome to this S&OP Masterclass from Roima. My name is Søren Hammer Pedersen and I'll be your host for this session here today. The purpose of these masterclasses is that we dive into trending topics within supply chain planning and try to give you our perspective on why this is important to you and hopefully give you something that you can use in your daily working life. Today's topic is no different. We are recapping back to our last podcast about agentic AI and we are going into the specifics now. Last time we talked a lot about agentic AI and the technology Aura, and we are going to look at what does that look like for real in manufacturing within the supply chain.

(01:54):
And again, I'm joined in the studio by Rafael, which are helping me to give you our perspective on this today.

(02:02):
Welcome again, Rafael.

Rafael Amaral (02:03):
Thanks, Søren. Great to be here.

Søren Hammer Pedersen (02:07):
Just to kick things off, Rafael, not all our listeners necessarily heard the last podcast. So just maybe a brief introduction. Who you are Rafael?

Rafael Amaral (02:16):
Yes. My name is Rafael Amaral. I've been working on manufacturing technology for around 20 years now. Half of my career was in supply chain planning and the other half I went into manufacturing execution systems. I currently head the engineering team in Roima that works with TilliT stack and that's the cloud native MES application. Yeah. We started working with AI for a while now within TilliT and Aura is a new set of tool sets that is really changing the game for us.

Søren Hammer Pedersen (02:58):
Yeah. Perfect. I think you qualify for this session today. And we are of course diving into the manufacturing going into the shop floor in the production sites. But before we get there, just again, a bit of a recap in terms of this area of agentic AI and what the Aura technology is.

Rafael Amaral (03:18):
Yeah. What we did with Aura, we've built a tool that gives a little bit more power to the AI technology so that it can read the data that we have available and it can come up with its own insights to its own analysis and everything based on goals and goals being essentially KPIs.

(03:46):
The history of Aura came from a bit of a general frustration around us building systems of record like an MES or ERP that are essentially holding a lot of data. And even though everybody would say that data is gold, then where's my shovel? So the reality is that most of our customers that had all that data set would only ever report on a very thin slice of a dashboard, a Pareto chart, extremely simple views of that whole data and not able to leverage from the value that we could take out of it. So Aura came from that gap.

Søren Hammer Pedersen (04:40):
Yeah. Okay. But let's dive into the topic and go into the manufacturing side. What is the problem that we are trying to solve here just to frame it before we go into the specific technology?

Rafael Amaral (04:57):
Yeah. Like I said, we have all these systems of records. They're all in silos, but let's focus a little bit on the manufacturing execution system. Manufacturing execution system is a combination of a bunch of functionalities that are geared towards managing a factory. A factory is a very complicated environment. It has machines, it has people, it has to do with raw material, it has to do with inter-logistics. And you're always trying to get the best out of it. The highest amount of efficiency, the lowest amount of cost. So the amount of data yet the connectivity between all those elements, because it's one factory and you're making one product, you could be talking to somebody that's managing the quality elements of the production, yet you can be talking to a maintenance that's working on that machine and there is zero interconnectivity right now between these people from the data set perspective. But because you're talking about one factory, there are physical connectivity. If you think about the level of complexity that a factory holds and all the people that are managing the factory having access to just a couple of dashboards, maybe one focusing on downtime, the other one focused on SPC and the quality trends, there's a lot that we are not understanding of the things that we could be doing to that factory in order to improve it. You're just missing out.

(06:41):
So Aura is essentially as if you were hiring a team of say five or six data scientists with access to all the data that you have available. And they would on a day-to-day come up with suggestions like, "Why don't you focus your time on this? Because this is causing that that is impacting your bottom line." So why do we need this? Because at least my customers, the people that I work with, they don't have available in their team all these data scientists. They might have one or two business analysts that are running pretty simple reports and coming up with some of these analysis, but it's really not as deep as it could be.

Søren Hammer Pedersen (07:30):
No. It's really about there's all these investigations you can call them that could be very, very fruitful within a factory, but we don't have the time, only a fraction of them actually get carried out on a weekly basis.

Rafael Amaral (07:46):
That's it. I mean, there's a lot that we don't know that why things goes wrong. Some of them are obvious, but some of them you just are not paying attention. And there's another element to this, which is if you are a plant manager and you have those 100, 200, 300 employees and they're running shifts and you're producing nonstop, the amount of noise, the amount of request for your attention to fix something a lot of the factories I go through, it's a bit of a firefighting thing. Because you've got issues with machines, you've got issues with quality, you've got issues with material. So if you are a plant manager there, one of your roles is to prioritize. And where do we put investment? What is the business case to do this improvement? Should I replace this machine or not? Should I improve some training or should I hire more people here?

(08:51):
The tool that we've built helping a lot is on that level of manager that can now look at facts and data on top of what they already have. But with these deeper facts and data, they can actually drive the right decisions.

Søren Hammer Pedersen (09:09):
Yeah. So it's all about finding the problem, qualifying the decisions we have, but also just increasing the sheer resources that we have to find things.

Rafael Amaral (09:21):
Yeah, exactly. So if you had a extra set of hands that could be trying to figure out where should we improve? Well, that is highly welcome anywhere you go. So that's what Aura is doing.

Søren Hammer Pedersen (09:36):
Okay. But then let's get a bit practical here going into the production side again here. When we apply this technology Aura to the manufacturing, how does it actually work? What it is that it does?

Rafael Amaral (09:54):
Again, it feeds on the data available. Let's take TilliT for example. TilliT is a complete package that allows you to monitor the machine's performance. We connect the machines to counters and states, know if your machine is running and whether it should be running and how much it's producing so we know the speed. Imagine you're working with, let's say, a beer factory because I do a lot of beer, I do a lot of wines, which I love. And then you have a beer factory that needs to bottle, bottle glasses and cans. These are expensive pieces of machines that are running at 60,000 cans an hour or even more. And they are complex pieces of equipment and they break. They stop. Whether it's because something went wrong with the machine, whether somebody made a mistake. A factory is a really complicated environment, period.

(11:01):
So with a system like TilliT, you'll start to monitoring the machine and then you start to know when the machine should be running and when it stopped. But with TilliT, you're not just doing OE because OE gives you a fraction of what you need to know on the holistic view. So you're also doing the processes, the quality, because even though you have that very sophisticated piece of equipment, you have a lot of people around it to feed in raw materials, to do quality checks, to fix problems. So then you have the quality and the process. With TilliT, we're digitizing all elements of the factory. If there is an activity that an employee has to do that we know. If the activity was successful, we know. If it was bad, we know as well.

(11:48):
And the other thing that's really interesting with TilliT is that because it's such a tool for the operator, we have all these places where the operator types in some comments. And why is that important? Because the comments from the operator is a way for the digital system to be aware of what's happening in the factory because the operator is saying something went wrong here and I'm having issues here.

(12:13):
Okay. So now we have after the few months of using TilliT, you have a database full of information there. But the reality is that we humans are not capable of digesting all the information. So we end up building reports that were showing, for instance, how did we do last shift? What was our performance? What was our output? What was our down times? And where were the top five reasons? That's kind of it. We can't really take much more.

(12:43):
Once we started using Aura against the TilliT database, it was absolutely mind-blowing the result because suddenly you start to correlate the issues that you're having with your machines with the process that the operators are doing or not doing. Which from a physical point of view, it's obvious. You go in a factory, you see this is connected to that obviously. But from our ways of managing point of view, not so obvious.

(13:14):
One of the plant managers that I'm talking to, he told me it was really interesting. They get so obsessed and no guilt here, right? But they get so obsessed with the OE, with that point of view of the efficiency that all their processes and reporting solutions, they can't connect with the holistic point of view. So suddenly you have an AI agent swarm. There's multiple agents with multiple KPIs and with the holistic approach of looking at the factory end to end. And what you get, you get the connections and then you show the connection to the plant manager and say, "Oh my God, I've never seen my factory with this lens."

(14:00):
And then they take that to the team and they act on it because it becomes obvious that they should act on it. It's a wake-up call to say, "Hey guys, this is so complicated that we have never connected the dots of that and that. But here it is. It's fact-based, all the proof. Let's change our process in order to fix that problem." And then we go to the next one. So Aura runs as if there was a team of analysts pointing you to places that you should work and improve.

Søren Hammer Pedersen (14:38):
The swarm of agent, it's really about them having different assignments you can say, but it's working towards the same goal of finding the improvement?

Rafael Amaral (14:46):
That's it. We built Aura as if it was a mini factory. You have the plant manager agent and the plant manager agent, the role is to look at the end to end. I remember when I started doing TilliT, I was connected to the plant manager our first customer. And it was clear to me that his role, his full goal and objective was to get really good product out the door at the most efficient way. You can't be more end-to-end than that because you have to go through all the little steps. If you're a maintenance person, you might be only concerned about is this machine running or not? If you're a quality person, you just might be worried about, is this the right process variable quality check? But as a plant manager, you have a holistic view always. Always have the holistic view.

(15:48):
And then you have the different point of views because you're still trying to find issues that maintenance needs to solve. You still need to find issues that are quality related. But because of how we build Aura is this agent org chart and they can talk to each other. You start to connect the dots and that's really novel and super powerful.

Søren Hammer Pedersen (16:10):
Yeah. Perfect. Again, keeping the picture of all these agents, but let's go into practicalities. How does Aura or the agents work in these discovery cycles that we run in the plants?

Rafael Amaral (16:26):
Yeah. Through the journey of TilliT especially and its AI capability, we started helping out how to configure having little chatbots. We started doing a little bit of a shift handover process. So it summarizes everything that happened in the shift. These were pretty simple functionalities that took a piece of data and gave it to the AI and the AI replied that back with English in readable format. Now with Aura it's a bit different. Because what we wanted to do was give the maximum amount of value possible we could take. And for that, we had to really change the game by having Aura write its own code. And it brings all the technical challenge behind it to make sure that it's safe and all that. But basically in a simple format, we feed Aura and its swarm of agents that are protecting a little container and we have it run deep analysis on that full of data science capability tool set.

Søren Hammer Pedersen (17:53):
Yeah. It's all about, of course, we're giving it the goal to achieve, but then it runs a cycle around, okay, let's scan the data, the problem.

Rafael Amaral (18:05):
That's it.

Søren Hammer Pedersen (18:05):
Let's find the hypothesis that we want to check out. Let's validate them and so forth.

Rafael Amaral (18:11):
That's correct. That plant manager looks at the KPIs and tries to find a hotspot and delegates to the specialist agent to go deep into the detail and say, "Okay, this is the issue. It's connected to that." It might store lessons learned. It might talk to the other agent to say, "Is there a connection between quality and maintenance here that we have to be aware of?" And the result is human-readable insights. And that human-readable insights is taken to physical meetings, your daily huddles, your weekly meetings, and they're discussed. And then the final action is a plant manager changing a process or getting some maintenance done on a specific area or anything related to the improvement of the factory. Yeah.

Søren Hammer Pedersen (19:08):
But I guess the key point for the value creation of this is that how do we present it? Because one, of course, the technology runs. You're not that involved in that. But when you come in Monday morning, what do you meet then? What is it you get in your hand here?

Rafael Amaral (19:27):
You get in the user interface, you get a board with insights. It's your newspaper that has articles. So each insight will have a summary that is a COO level explanation of what is happening in your factory that you should be aware of. It will have a recommendation. And then it will have all the evidence to that. So it'll give you all the reports detailed of why did it reach that conclusion. So it is really a tool for a manager to just read that document and act on it. That's as simple as that. You'd go from raw data straight into the final assessment that needs action on. So we skip the dashboard, we skip the report.

Søren Hammer Pedersen (20:22):
Yeah. Could you give a concrete example of a finding that you have seen in a plant?

Rafael Amaral (20:28):
Yeah. I mean, a couple of examples was a typical one. Why running similar products with the night shift take 20% more changeover time than the day shift? Why? Yeah. Or why would all the quality checks pass even though we know that we have an issue with a machine that's causing a bad input of quality? So you start to look at areas of the factory and question what's happening in the factory that needs to be looked at and what do we do in order to improve that?

Søren Hammer Pedersen (21:14):
Yeah. And is it generic in the sense that it will work on any plants or?

Rafael Amaral (21:19):
100%. It's all based on goals and KPIs. That means what are the KPIs that we're looking at? We're looking at OE, we're looking at waste, pass fails. And from there it drives the analysis. And you see the flow of each agent. It goes through what a typical business analyst would do. Goes through that Perito. It goes through that top 10 worst candidates, pick the one drill-downs per month, see trends and start filtering. But it does that at a lightning speed in parallel runs over the weekend.

Søren Hammer Pedersen (22:02):
Yeah. I think that's interesting. Really over the weekend is a key point for me here that I really like that picture of coming in Monday morning and then you just have analysts ready for you. You don't have to start, "Okay, what happened actually?" It's there.

Rafael Amaral (22:15):
These reports are not new. We have business analysts working on factories. We have those analysts done in Excel. Most of the times, even when we have systems of record, you would see a lot of the times analysts downloading data to Excel, running some comparisons and then going to management and say, "Hey, I found out this." So that's happened. We're just boosting that. We're boosting it to be wider, meaning that analyst has even the best of them, they're still humans with that kind of limitations that we have. So we're boosting it horizontally in terms of I have a wider specter.

(23:01):
That's what I hear from the actual analyst that I'm helping is it compares data that I don't usually compare. For instance, you would always look at the downtime reason report, but now I'm correlating the downtime reason report with what's happening in my CIL, Cleaning Inspection Lubrication process or what's happening in my continuous improvement or my RCA. It's really hard to do that as a human, especially when you don't know you have to do that. And it's doing deeper. So even if you do have an issue that you know about, the fact that you have a deeper and stronger argument, it might actually move the needle in terms of the investment requirement to fix that issue.

(23:46):
Coming back to the bottom line is we're helping the plant manager deal with a prioritization by filtering all the noise, interpreting all the noise, including pass fails, downtime reasons, comments from the operators, process execution, all that gigabytes of data and into a very easy to read report to say, "Okay, if you do this, this is what you tend to win."

(24:15):
And I forgot to mention in the beginning, we also have the CFO agent where if you give it the right unit economics, your labor costs, your utility costs, it will build a report with the financials attached to it, which saves you a bit of time to calculate them.

Søren Hammer Pedersen (24:33):
Yeah. I like the fact that of course we are building on what's already working in there, but I think it's also empowering both the plan manager and the planners basically to improve the quality of the decisions or quality prioritization, but also how much can we improve? I guess that goes up quite a bit because when we did three things in a week, six months ago, maybe we do 20.

Rafael Amaral (24:59):
Yeah, exactly. So remember that we're trying to improve things against the data that we have available against the analysis we have available. So if there is some sort of a power that allows you to see things that you're not seeing, that is untapped value by definition. The system tells you that if you fix this problem, you have a knock-on effect on that problem. You read the report, it makes sense to you, but it had never come to you. You never thought about that. That's untapped value.

Søren Hammer Pedersen (25:37):
Yeah. Excellent. I think it makes extremely good sense. And I think also the listeners out there can hear that this is highly interesting. And some of them are might thinking, okay, this sounds interesting for us, this technology. How hard is it to get going on this if you haven't done anything? You just have a plan.

Rafael Amaral (25:59):
If you're a Roima customer, the good thing is that we build Aura in order that it's connected to the different Roima customers. So the start and get go is immediately. It's very quick. For non-Roima customers, you can still use it. All you've got to do is get access to that data. Now, like I said in the beginning, because we're giving so much power to this tool, there is a strong security aspect to it. So there are rules that we have to abide that we have to get the data into the right shape. But once we have that connectivity, we can self-discover the data. We can understand the schema. So virtually we can work with any data set.

Søren Hammer Pedersen (26:50):
Okay. Quite easy. I think the security part is something that we hear a lot. Of course, there's AI policies coming out in many companies and in general. So the access to the data is something that you of course have to be quite aware of.

Rafael Amaral (27:03):
That's it. And this is where the whole solution was built to be extremely blocked and protected because obviously, like I said, we're giving a lot of power to this AI. So you don't want it to be doing things that you don't want it to do. That means that we need to basically bring the data into this environment. But once it's in, then you let the AI swarm loose and then you just get the result out of it.

Søren Hammer Pedersen (27:32):
Yeah. I like the picture of the AI swarm. And I think it's also a general recommendation. I think one thing is Aura. Of course, we love to talk more on that, but there's also a timing issue. We talked in last podcast about the second wave of AI coming into supply chain, and this is a crisp example of exactly that. So either way, either tool, this is something that you as a supply chain professional should look into.

Rafael Amaral (27:58):
Absolutely. I think that the amount of value that's there for the taking, if you're not going after, your competitors will. And yeah, you don't want to miss out.

Søren Hammer Pedersen (28:10):
Perfect. Crazy interesting, Rafael. Thank you so much for your time here today to give us your perspective on this and hopefully we will see each other in the studio again.

Rafael Amaral (28:21):
Absolutely. It was a pleasure. Thank you so much.

Søren Hammer Pedersen (28:23):
And also thank you out there listening in to our podcast here today. I hope you find it very useful, this area around agentic AI and especially how it looks in manufacturing. As always, if this is something that interests you, you are more than welcome to reach out to either Rafael or me. Check out the Roima website and see what's going on. A lot of interesting stuff coming out there. Other than that, we hope you join us again for the next podcast. And on that, have a great day out there.