Transforming Engineering Process Intelligence — with Acceleer and MontBlanc AI

Episode Summary

The episode titled "Transforming Engineering Process Intelligence — with Acceleer and MontBlanc AI" delves into how artificial intelligence (AI) is revolutionizing industrial automation within process industries such as food, pharmaceuticals, and chemicals. Hosted by Michael Finocara from the AI Across the Product Lifecycle podcast, the discussion features David de Maire of Axelir and Marcus Guerster of MontBlanc AI. Both companies are pioneering the integration of advanced analytics and machine learning into industrial processes to enhance efficiency and decision-making.

Key insights shared during the episode include the necessity for industrial automation to transition to more modern development environments that support reliable AI applications, such as Siemens’ Beccos or Cortexys. Additionally, both guests emphasize the importance of working seamlessly with legacy systems without disrupting existing operations, thereby making their solutions accessible and user-friendly. They also highlight the role of platforms like LinkedIn and Slack in connecting with potential users and partners.

For PLM and engineering professionals, the episode underscores the transformative potential of AI in streamlining complex industrial processes and improving operational efficiency. By adopting these new tools, professionals can enhance their decision-making capabilities while maintaining compatibility with existing systems, ultimately driving innovation and productivity in process industries.

---

Full Transcript

Michael Finocchiaro

And we're live. Hi, this is Michael Finicara from the ⁓ AI Across the Product Lifecycle podcast. And I have the pleasure to be with David de Maire of Axelir and Marcus Guester of Montblanc AI, two exciting companies transforming process industries today with AI. ⁓ So, David, can you tell us a little about yourself and a little about Axelir?

Davy

Yeah. Hello, Michael. So, ⁓ I've been working for more than 20 years in, system integration, industrial automation, mostly focused on process industries. So food, pharma, chemicals, and I started, ⁓ I started in Belgium 20 years ago. I'm from Belgium, but I quickly, ⁓ that's the produce abroad. And because of that, also very early, I ended up in China. So in 2006, I already started doing projects in China. And then in 2012, actually I moved permanently to Shanghai ⁓ to build a team there from scratch ⁓ around automation and industrial software and MES. And for the past 12 years, I served ⁓ the multinational companies building green fields in China with our team, again, food, farm and chemicals. And then two years ago, ⁓ I decided to leave my employer because I got a little bit frustrated with how the entire engineering product lifecycle is, especially for industrial automation, because I thought it was so manual. So I left both my employer and also I also left China. I came back to Belgium to start my own company, which is a startup and it's called Axelir. So and that's what I'm doing today.

Michael Finocchiaro

Exciting. Thank you, Davy. How about you, Marcus? What's your story?

Markus Guerster

Yeah, hi Michael. Nice to be here. my story is really, they were originally from Germany. And when I was 15 or 16, I was working in a dairy. I was doing kind of all the, know, mundane works that you basically do there if you don't have much of education. Didn't think much about it. Went on to kind of study the classical engineering, sort of software development. education. Also studied basically machine learning and AI before it really became a hype in a way. Worked on a few things there. Similar to Davey, also went to industry, worked for company and also very similar like three years ago, kind of got a little bit tired and then went into starting one plant guy with a couple of colleagues, really with the idea to bring machine learning and advanced analytics into the process industry. And funny enough at that time, I was visiting the dairy again that I did my internship when I was 15. ⁓ And surprise, surprise, not much has changed in terms of data and dashboarding and so on. So yeah, super curious. And I think we're gonna touch on a few of those points today as well.

Michael Finocchiaro

Hmm. Awesome, thank you very much. Yeah, and it's interesting too because, you know, mostly I talk to people from discrete entities, right, the classic CAD simulations. So thank you very much because it's so interesting to hear about the side with formulation and materials and just a whole different set of concerns. And of course the processes themselves, which are interesting. So let's back up for a moment and look back at like 2022, 2023 when... Chat GPT 3 came out and I had never even heard of Chat GPT 1 or 2. And it was like, holy cow, what is this all about? What's going to happen? Marcus, were you like super motivated? You were like, yeah, this is the future. Let's go. Or you were a bit standoffish. Let's see what it does despite a lot of hype. What were you? you bullish or skeptical back then?

Markus Guerster

I was skeptical to be honest, because AI is not nothing new, right? It exists since the seventies. It's been there like pretty much forever. It has found a lot of applications as well as sort of the statistics, machine learning side, deep neural networks, et cetera, all of things were under coming, but they were always thought to be unreliable. Nobody really understands them.

Michael Finocchiaro

Mm-hmm.

Markus Guerster

That was really before like that launch that you said, right? That was kind of the thinking and the science community. It's all just, yeah, who knows what's going on? Which is still true today. But then with the launch, actually, well, it doesn't really matter. You don't need to understand what's going on. It provides so much value. And I think when the first time I tried it, my skepticism completely went away because I could see. what it can do, right? And then I think the numbers just explained themselves. The curve was crazy in terms of adoption.

Michael Finocchiaro

Yeah, we've never seen anything like that before. ⁓ David, did you have a similar feeling?

Davy

⁓ Yeah, but I was very quickly convinced. So the moment it was released, I played continuously with it to push it as far as possible. And then I also told my team at that time, so my automation engineers and my software engineers, you have to use it because, know, especially in automation, it's often difficult to find the documentation and and how to do things. ⁓ So I quickly saw the potential, even though it was ⁓ still not so easy. But I noticed it's mainly from my team that it was mainly the software engineers who were able to use it. ⁓ For automation, there was more skepticism in my team. yeah, it was still difficult for them to use it. But I was convinced. I was already thinking, OK. And this is just the start because For me, was very clear this was just a point in a trend. And so I was already thinking, okay, what is now the next step going to be because they will be able to improve this. And I listened very carefully to the podcasts about people talking about it. So yeah, and it was the right moment because it was a bit, still a bit before, but around that time that I started thinking about building my own company. it's also. gave like, it was also like something I could talk to and the reason through things. So yeah, I went, I was very convinced from the start. And since then I really pushed to use it as much as possible.

Michael Finocchiaro

That's a great segue into the second question, which is more on how ⁓ AI is changing the way we develop software. I guys both are founding a big software company, software startups. ⁓ I mean, I think the low-hanging fruit is probably product requirements documents. I those seem to be completely delegated to AI these days. or that's what I'm hearing. How are your developers using AI in a daily basis? How has it transformed? For instance, has it changed the way you do it agile or waterfall? Has that come into question? I would just like to understand how the transformations in terms of software development. So either of you can pick that one up. I don't know if you want to start again, Marcus or Davy, you can do it.

Markus Guerster

I can't, yeah. it's actually, think that the way we use it mostly, ⁓ they, you said, like if you're starting a company, right, you have so many hats. So you use it as an assembly board, as an advisor, it's kind of all kind of the other things. ⁓ but in terms of development for the software itself, I think it also has evolved quite a lot. In the beginning, it was really quite narrow, right? You had to be, you know, send the bug there, sort of debugging. was kind of quite helpful, but not really writing full pieces of code. That's how we used it quite a lot. It quickly got into like the architecture pieces of how we actually design software and then get the bigger picture. Because then you broke it down and then you could work with AI again to figure out the smaller pieces. And I think ⁓ over probably the last weeks or months, I think with the whole Adjantic approach that seems to be much more reliable, like Codex or Claude or whatever you use, we certainly moved more into really delegating things to AI completely and let it write the code itself. It needs to be very thoughtful, right? We always think about it. It's a new engineer that you bring on. So you have to tell it how to do it, what are our guidelines, what are our standards, et cetera. ⁓ It works, I think, extremely well. The technology is maturing very, very quickly. One of the things, one of the challenges that we're running into is that the context windows of the models exceeds the context windows of us as a human. So what I mean by this, very quickly, the AI writes whole gigantic features, and in the end, nobody's able to actually digest and review and check again.

Michael Finocchiaro

Right.

Markus Guerster

Like they're always working. It's amazing. It never actually introduces any bugs or very narrowly, but the problem is how do you maintain such a code base going forward when you write gigantic things very, very quickly? So I think this is kind of the challenges that we are facing right now. But of course, as an AI company that sells AI, we also use it internally. ⁓

Michael Finocchiaro

Everything. Mm.

Markus Guerster

And we have followed the technology and to my surprise, was skeptical in the beginning a little bit when JetGBT came out. Also when agents came out that hype around there. But, ⁓ I think it actually always exceeds my expectations on how we're sure it's because.

Michael Finocchiaro

Nice. Is that your experience too, Davy?

Davy

⁓ yeah, so I, ⁓ I also started it using quite early for coding. ⁓ so if you, for the people who use this from the start, you know, we started with copying pasting. just had a chat, even though you asked some code and you copy, you copy paste it in your ⁓ IDE because we had nothing like cursor or ⁓ co-pilot. And then, ⁓ I think one of the first ones was then cursor.

Michael Finocchiaro

and visual code and yeah.

Davy

you actually had to download a copy of Visual Studio codes. They copied Visual Studio code, but they adjusted it. And then I started using that. And that was a lot of code completion. So you could prompt and it completed things and they have a special graphical user interface to help it make you easy. For a brief moment, I used to augment. They had an amazing context. So they were super fast to understand the complex code. code base and then Claude code was released and that was amazing. So Claude code ⁓ is basically they said, yeah, you should or you can program in the CLI. For me at that moment, it was crazy. How can you program in the CLI? You don't have a nice overview. But then also around that moment it clicked for me because the key point in agentic coding, so not vibe coding, but agentic coding is Git. ⁓

Michael Finocchiaro

It's a revolution, yeah.

Davy

And ⁓ Git becomes a becomes a ⁓ tool to ⁓ keep your agent encoding under control. ⁓ you have around that moment, I figured out, okay, you need to start from a clean commit. You ask the agent what to do and then you review in Git. You don't necessarily have to have everything integrated like cursor. And so that's really, so that's the start of. Claude code, which is now around one year ago that it came with this principle. then, yeah, from there, I experimented with a lot of different approaches, elaborate PRDs. But I stopped with this because I think it's a little bit like what Marcus says. We have a problem with our own context window. These agents can write so much code already one year ago. You could not keep up if you write. big PRDs. So actually I went back to ⁓ small features, tried to give individual features, tried them, test them. Of course, ⁓ actually it is a very new application. There's nothing on the markets that's similar. ⁓ And it's at the intersection of software and automation engineering, which is something I believe will change a lot, but it also requires a lot of experimentation in the user and developer experience. So for the automation engineers. And so it's not something you can just say to the agents, now build me a huge features because you need to also have a tight feedback loop with the people who are using it. so I came back, I came back to do more smaller features, but at a much higher velocity. if teams ask new features, often the day after they have the features ready. tested. And the testing is also very important. I also saw, so these agents can build their own unit tests. So if you give them a feature, can tell them do test-driven development. First, make a failing test. Then implement the code, and then test your own tests. And at a certain moment, the amount of tests in the code base went exponential, because it's becoming so easy for agents to also add the testing. You cannot trust them.

Michael Finocchiaro

Hmm.

Davy

you have to check are they actually testing, not taking bypasses at a certain moment that was an issue. But actually today I have a very lean, ⁓ very agile workflow that's also very trustable because everything goes through Git, everything is reviewed, everything is tested. And so it's at a point where it's reliable, very high speed. ⁓

Michael Finocchiaro

⁓ Yeah.

Davy

And now the limitation I think is because every time you have to start from a clean context. So every time the agent forgets what it has done before. So I think if they can solve this, that the agents have long-term memory, then it will be crazy.

Michael Finocchiaro

I played around with that a bit with antigravity and now with GSD instead of cloud code. I if you guys have messed with that, but it gives you a bit more memory and bit more ways of structuring. ⁓ But just another question on the development side. Has AI and agentic AI changed the philosophy of the way multiple programmers work together? Because already you're a team between you and the AI, right? What about the other humans and the other developers? it changed the way you have implemented Agile with the other developers or it's still the same thing? just have this, everybody has this other pal next to him that comes to the Scrum meeting and he's just called, you know, Davies agent and Marcus's agent. Either of you can pick that up or neither of you, whatever.

Markus Guerster

swap the ordina, the EVP on time.

Davy

Well, that also requires quite some effort to think about how you collaborate. So one of the latest things that we like to do now is to put, for example, to extract the chat with the agent and store it in the repo ⁓ as a commit. So basically, when you go back,

Michael Finocchiaro

All right, that's great.

Davy

when people make comments, can actually look back at what has actually led to, which conversation has led to ⁓ this change in the code. ⁓ I'm still not convinced if it's the right way, but that's, for example, one of the experiments that's happening now. Some people say, that's a good approach. Some people say, no, you should actually not put this in your Git repo. ⁓

Michael Finocchiaro

Great idea.

Davy

But it has some advantages because you have like your entire reasoning workflow between the human and the agent stored in a certain commit. So if you later...

Michael Finocchiaro

What's sort of like the design intent in CAD? You're making sure it gets written down, right?

Davy

Yes. Yeah, so you automatically have the reasoning behind it. And these agents are also very proficient with Git. So actually, if you have a new session with the agent, you can tell them, oh, we changed something in the past, and I have no idea why it has been changed. So then you have the code changes, and you have the intent together, matter who collaborated with the agent. So yeah. Still, it's not so easy because also people adjust at a different speed ⁓ to the agents. It's also changing some fasts. Some people have different preferences for how to do things. ⁓ it's bringing a lot of impact. And I think we just at the start of seeing how teams ⁓ are collaborating with it.

Michael Finocchiaro

Do you have anything to add, Marcus?

Markus Guerster

I think it raises a quite a interesting dilemma because the direction where everything is going is certainly towards you centralized knowledge base to some degree. You centralized memory, right? Michael, said it, long-term memory. also mentioned that you put it into the Git commit to have some sort of trace history of what's going on and the... The memory exchange pipeline of agents is going to be million times better than the communication between people. So I expect in the future that everybody has the agents, right? And the agents communicate with each other. And this is where the majority of information flows. Like yes, I implemented it. The other developer implemented it like this. This is the pattern we're using. And that whole thing between the agents that every developer is using, it's going to be. outpacing by many, many numbers what people are interacting with, because that's a lot of the actual development flow that we have today, right? We jump on the dailies. We have all the demos, the retrospective. We have all those patterns to make sure that developers are communicating and exchanging information. Now, if this is all kind of replaced a little bit by this gigantic pipeline, I think it's restructuring the whole flow a little bit.

Michael Finocchiaro

So like what I was saying earlier, I think it's changing agile. Maybe you're going to have a scrum with just agents at some point. No human involved. It could happen, right?

Markus Guerster

That's probably

Davy

So.

Markus Guerster

the direction that things are going.

Davy

So if you look at engineering, like in process industry, the most extreme of one of the most extreme things is pharmaceutical engineering. You really have the V model and today you have no choice. It is waterfall because you have to start from the design and then everything is... Yes. And so it makes it a very wide, a very long engineering...

Michael Finocchiaro

Yep. because of the regulatory compliance, because of traceability. Richard.

Davy

timeline, but I think what's agentic engineering will allow, not only agentic engineering, basically digitizing the entire engineering workflow is you will be able to do this. You will be able to bring the execution, the development very much closer to the definition. And so we will crush by just making the whole loop much, much tighter. And so it will be, it will be waterfall, but so fast, it's almost agile.

Michael Finocchiaro

Okay Wow, that's a good one. I like that. That sort of nicely leads into the next subject, is when we're using AXELIA or Montbon AI, where are we using AI in the application? Is it part of something that the user sees when they log in and they're using it? Or is it part of the plumbing because you're using AL or some advanced algorithms on the bottom? Where are the touch points? ⁓ And maybe part of that, you can mention One of the things we sort of talked about is the fact that AI is probabilistic. And I think that in the process industries, you need something that's deterministic, right? You need to be very sure of the exact quantity of this chemical at this temperature, at this pressure, mixed with this. I mean, all of that stuff has to be extremely precise. Otherwise, you get garbage at the end, right? You don't get the thing you're looking for. So how does that work in terms of your products?

Markus Guerster

I can start, think we're using it on a variety of different, different layers. think the most important thing, and we're not talking about AI here, it's you have to have the foundation right. Right. And the, think very much overlooked thing in the process industry and probably not a piece of manufacturing is as well. Everybody's saying, yeah, we don't have data, but that's incorrect. There is data, it's just not accessible, right? It's lying on the PLCs. It's lying there and it's probably the highest quality data there is in this world because PLCs, are a hundred percent correct or 99.99 % correct. So it's just a matter of getting that ground together first because that addresses your point, Michael, where you said, yeah, AI needs to be grounded, right? And this is the grounding.

Michael Finocchiaro

is locked up in the silos, right? Mmm.

Markus Guerster

⁓ So you have that data foundation. And then what we do in our products first to have a very deterministic or mostly deterministic statistics layer on top of this machine learning layer, right? That crunches that huge volumes of data because you cannot feed that directly into an LLM. So that crunches the data ⁓ that is very deterministic. And that allows the agent then to interact and do analysis with either the process data or without the documents. So the way this shows up to the user is first of all, of course, the co-pilot view, right? Did you have all the time, like in pretty much any product or coming into almost any product, but we also have it running on the background and constantly generating those notches. The things that keep you becoming better. Are your processes running correctly? Is any machine about to fail? Has there been any major changes? Is there any new documentation that you need to be aware of? Maybe we send you a piece of information every week or so. All of those background things are very, very heavily driven by our agents.

Michael Finocchiaro

Thank you. That's really a good answer. Thanks, Marcus. Davy, do you have anything to add?

Davy

⁓ Yes, so your question was partly about the deterministic nature that we need in industry. So Axelir is an application that generates code.

Michael Finocchiaro

And also how... ⁓

Davy

Go ahead. So actually, there's an application that generates the code for the PLC, which is, of course, very, very critical. ⁓ And actually, at the core, doesn't use AI to generate this code. So it's deterministic code generation. And so ⁓ we actually take the existing workflow where ⁓ people had to make functional descriptions and then they program by hand. So we transform this so

Michael Finocchiaro

Go ahead. No, it's fine.

Davy

people can automatically generate the code from the functional descriptions. And so actually, for the users, instead of letting the AI help with the code generation, because we have this for a big part solved, we actually allow people to use AI to generate the functional descriptions. So actually, the design intent, which is much more easy for people to collaborate with. ⁓ Coding, asking the agent to code is very, very tough.

Michael Finocchiaro

Mm. Okay.

Davy

You have to review what the agent is doing. You really need to take care of the entire workflow, especially if it's a critical application. But if you ask the agents to collaborate on the design, ⁓ it's much easier. It's less taxing for your mind because you can just read a human language what the design intent is. And so you can ask the agent to generate it ⁓ in a very human readable format. And that's what actually it is doing defining this human readable format a bit as a standard. And then the engineers can decide themselves if they do totally by hand or if they start using the agent. And so how we implement it, we actually allow people to just sync the entire design in textual format in markdown files, let the agents collaborate with them and I think back to the platform. So you can just use cloth code or or Gemini or Codex or whatever. So you just bring your own agent.

Michael Finocchiaro

⁓ Thanks. There's actually a really good audience today. We've got around between 16 and 20 people. That's just great. Thank you everybody that's out there. ⁓ My friend Brian Carroll, who will be on my podcast at five o'clock today on change management, has a question. said that ⁓ since there wasn't really a formal response about adding team members, how can any initiative work across the manufacturing lines with all the IoT and edge server data being connected and adjusted into AI? So I how do you, I guess how do you take this generic idea and then ensure that each of the lines that you're managing with the software is consistent.

Markus Guerster

Yeah, I think that that's what goes back to the fundamental ⁓ things that have to sort out before you even sort of work on the AI part. ⁓ The way we approach it ⁓ is really having a long list of connectors available, To connect to all your variety of different IoT devices, PLCs, sensors, SQL databases, CSV files. Because the landscape is heterogeneous. Like nobody always has just one single one. Otherwise we wouldn't be sitting here and discussing this. ⁓ So that's the first step that has to be sorted out. The nice thing though about having an Atlantic AI within our product is that you don't need to be perfect in sort of schema alignment.

Michael Finocchiaro

Right.

Markus Guerster

because the AI agent is going to be able to figure it out. If things are slightly renamed, right, if you call a temperature in one ⁓ space and the other one is called temp, and the other one you misspelled it and called temperatur, whatever it is, right, then the agent knows, okay, probably they meant exactly the same. So let me just connect all of those three pieces together in the historical way without taking any sort of a chat again behind it, that would be a nightmare.

Michael Finocchiaro

Hmm.

Markus Guerster

Right. Because you have to change all of the schemas. Now, God forbid somebody renames it. Then you have to change the whole thing again. Probably re-micro data and like the whole drama. This is, it's a never ending story. And this is exactly why currently process manufacturers, it's all over the place. Exactly because nobody wants to have that effort to standardize everything, which is a never ending effort anyways.

Michael Finocchiaro

Again, yeah. Awesome. Davy, any thing to add?

Davy

Yes, so actually it is not trying to solve that part of the workflow. So having the agents in production, but what these agents need, they need a lot of context. So of course they have the real-time information, the IoT data. And I remember at the previous question I said that we allow agentic engineering based on the functional design specifications so the agents can engineer them. But at the end of the engineering phase, we also allowed to extract everything, again, in Markdown files in a kind of structured format, Markdown CSVs. And that's actually the design of the plants, the automation design of the plants. But the automation design basically says how the plant has to run. And so then you can pull this in, in AI copilots, and you give it together with all the IoT data. And then the agent really understands everything. It sees the sensors, even like Marcus said, even if they are not perfectly labeled. ⁓ It's very good at making this cross-reference. And if then you give proper documentation, like what is coming from the design ⁓ intent, then it becomes really, really powerful. And that's the focus ⁓ of actually on the output side. So we don't really want to make the agents for the production, but we want to provide the context as a very strong basis.

Michael Finocchiaro

And do you guys think that thanks to Accélère and Montblanc, that maybe processed industries will catch up or surpass discrete in terms of being modern?

Markus Guerster

That's our goal, David, I think, right there.

Davy

But yeah, I call my company Accelerate to accelerate the process industry. ⁓ I think process industry has a lot of opportunities, but the mindset is really behind. So if you look at the automation layer, so you know the PLCs, but in process industry, we have the DCS systems. So distributed control systems, and they are very closed.

Michael Finocchiaro

the digital transformation.

Davy

The PLCs are relatively open. And we are very far from, for example, just the engineering, the DCS engineering. With PLCs, certain PLC systems, can already do agentic engineering. DCS systems today, it's almost impossible. And so the entire mindset and process industry, it has to change. And it really has to accelerate. So if you, for example, look at pharmaceutical manufacturing, the plants, the entire engineering of the plant, it's way too slow from the moment that people have research in medicine to the moment that they can bring it to production, to the market. It's too slow. You saw it during COVID and people said, okay, we need to take shortcuts to bring the vaccines to the market. But it's the same for every medicine. And now we are back to this wide V waterfall. ⁓ project workflow and this has to speed up. So the mindset really has to change and AI can help with this, but the big bottleneck is the people that need to say, okay, we really want to move forward.

Michael Finocchiaro

I think we're loop. Okay, you're catching up there. Your connection dropped off for a second. ⁓ Let's talk a minute then about ⁓ now that you guys have, we've all lived through these first three, four years of the AI revolution, if we call it that. ⁓ And now we've had this insane Claude bot, molt book thing, whatever that was. ⁓ Where do you guys stand on? mean, you guys both started out relatively skeptical. And then rather convinced. now, you know, four years later into the hype cycle, where do you sit now in 2026? What do you think about where we're going and what do you think is going to be the next breakthrough? Dave, you start this.

Davy

Yeah, I didn't try the Cloudbot. I understand what it's doing. I think it could be interesting, but I don't see any immediate application for what I'm doing because everything needs to be so ⁓ well under control. And of course, Cloudbot is the opposite of control at this moment. ⁓ But I do believe that we are going to go to a situation where this will become good enough to... ⁓ to basically use it ⁓ as real assistants that are always available, that can do big parts of the jobs, of the tasks. That's also what I said at the start of the talk, if the long-term memory can be solved and you combine it with ⁓ something like how cloud works today, that will transform everything. And then actually the focus of people, the shift from actually doing the... the boring tasks to really steering the agents to really say, okay, we need to build this and we need to work towards this and let these agents maybe work for one day or two days and then let them come back to you to discuss the next step.

Michael Finocchiaro

I just wonder if there's a danger that we're all going to get so lazy that we won't even think about it. We're just going to be delegating. mean, already, you know, every time it says, would you like me to do this? you know, you're like, OK, fine. But wait a second. Where did it? Where did I start with this question? You know, sorry, Marcus, how do you want to what do you think about it? The way we're going and where it's going to land after a year or so.

Markus Guerster

Yeah, I think the direction is quite clear, right? Definitely replacing tasks and automating. think that increasing efficiency overall, think the direction is very, very clear. The thing that I'm thinking, that's more like bigger picture at the outset of what Davy and me and Michael, we were doing instead of the manufacturing space. But how are we actually, what are the, you know, the numbers? How do they work out for this big model providers and where's that direction going? Because the commodization of those models can go also in the wrong direction. If suddenly we have like it's been happening to Facebook and Instagram and all social platforms, if they're termed ads, that would destroy everything. Because suddenly I cannot use ChetGBT to brainstorm anymore because it's not neutral.

Michael Finocchiaro

And it's not objective.

Markus Guerster

Of course, but now I know it's actively steering me into a certain direction to buy a certain product. That would be a nightmare. Right. If it goes into that direction, that will destroy everything. think. but of course, then they also have to gain to earn the money. So the question really becomes where does the compute cost go? Like how much does it go down? How much more efficient do the models become? ⁓ And in the end, does it really pay off to replace a person or a person's hour with an agent? Because there has to be a significant cost benefit in the end, right? That is a factor of two is not enough. has to be a factor of 10 or a hundred or something like this to really adopt. And I'm very uncertain where it is, where it is cost. ⁓ benefit ratio is going to sit because right now the market is obviously distorted. Right. Everybody's just throwing out the models and investing billions and trillions into everything that I can possibly get to get market share. But where do we end up once all of this kind of investment cycle, everything is done. ⁓ So that's kind of the bigger, larger question that I have in mind that drives the SFU of the things that the coming years will bring.

Michael Finocchiaro

We pick up an interesting point with the upcoming advertising and the open AI stuff. And of course, he also had that insane comment about adult content in there, which is mind-blowingly stupid. ⁓ Won't that push us to be developing our own models when it becomes economically feasible? Could you envision there being a large ⁓ operation, large OT model, a large process model so you don't have that dependency anymore. Or maybe there's a split, there's a model which is just, here's all this OT data about process industries. And then here is how a language, I don't know, I'm just trying to, because at one point, if it goes that direction, if Anthropic, despite their ads in the Superbowl saying, know, making fun of OpenAI, if everybody goes that direction, we're going to go back to LM Studio and Olamma and stuff like that.

Markus Guerster

And some of the bigger models do exactly that.

Davy

Yeah, my-

Michael Finocchiaro

So, Mark, do want to finish your thought and then Dave, you can jump in.

Markus Guerster

Some of the bigger models do exactly this, right? They split themselves into experts models. So that could be a feasible approach that suddenly you can deploy yourself or at least lower the cost to run a model yourself. That would of course distort some of those things. Yeah. That's it. Sorry, David.

Michael Finocchiaro

Go ahead, David.

Davy

Yeah. So if, if certain common companies start with advertisements, um, and then other companies will not do advertisement and just ask you to pay and not to have the. I mean, it's very tough to have the leading models, but you can see open source is ⁓ so near to the close models. Okay, they are a couple of months behind, but these models are so powerful. So it's not so difficult to, even if the close models would go this direction, then open source would become even more attractive. And on your question, well, we have specialized... ⁓ OT models. Well, I hear sometimes Siemens is talking about this, but the thing is the moment that a model becomes openly available and has certain specific data insights, then the other models distill it. So they basically query the model. They extract all the information, they put it in their dataset, and they also put it in their own model. So these models are so hungry for data. And now people are really looking around for very specialized data. It's very difficult to make a model that has certain amount of knowledge that the other models don't have. Unless if you keep the totally closed. And you could do this if you have a fully proprietary data set and you only use it internally. But the moment that you make it available to the market, it's going to be very difficult to protect this. don't think for OT we will, I think it will be very difficult to not have the leading models. to take a big part of the knowledge. And can already see it today. If you query ChadiPT or Claude or Jim and I about really technical ⁓ industrial topics, they know a lot of things. And it's gonna be very difficult to keep up even for a company like Siemens to build a specialized model. They can maybe lead for a couple of months and then they will be catched up.

Michael Finocchiaro

So just before we switch gears and talk about digital maturity, which is a really interesting topic, particularly in process industries, in the demographics of the people watching this podcast, I do have some younger kind of entry level, entry career people. And ⁓ I'm wondering, you guys both have a lot of field experience. ⁓ Maybe some of the people listening are chemical engineers or electrical engineers that are interested in ⁓ that area. What kinds of things should they be? working on, getting experience on so that AI doesn't take their job, know, that anxiety everybody has because everything, you know, they're 30 seconds or less. Do have any tips for them on what they need to look at or focus on?

Davy

But they need to, first of all, they need to go out in the field. It's industry, you need to understand factories. If you're young, really get out there as soon as possible, spend time, especially doable when you're young, because it's more difficult when you have the family. So really get this experience. It's really, really important. And therefore the mindset, really have a mindset that you want to build things. You don't want to do these tasks. You don't care about individual tasks. You care about the results.

Michael Finocchiaro

Mm-hmm.

Davy

And AI is just a tool that will become more and more powerful. So you have to learn the tools. You have to learn AI. You have to use it. And over time, you will see you will just be able to do bigger and more complex things with this. Don't ignore it. You have to use it.

Michael Finocchiaro

Marcus?

Markus Guerster

Yeah, I think my take on this would be invest in relationships with people. I think that's the one thing that AI will hopefully never take away. ⁓ And invest in skills over knowledge. Because I think learning something by heart, I mean, once Google came up, was already kind of the value of that went lower. And now with AI, it's even lower than this, right? So really invest in your skills and your thinking and how you approach problems and how you break them down and how you handle complexity and sort of those patterns more over knowledge. And I think that's the more future proof.

Michael Finocchiaro

Excellent. Thank you. I like that one a lot. That's the first one. Nobody's come up with that one. That's a good one. ⁓ So let's talk a bit ⁓ before we close about digital maturity. I always think of digital maturity on a scale of one to five. Like one is we're still using Excel for everything. and we exchange by email and it's super slow and it's very, very serial. And the other one would be fully agentic digital twins, my factory runs itself and it's all fully adaptive. Nobody is at that point, particularly not in process industries for all the reasons that especially Davy you mentioned. So when you guys go to customers and... especially in these industries, are they closer to three, closer to two, underneath one? mean, where in general do you feel that your customers sit on that spectrum?

Davy

Well, I like customers to come to me. That's also why I post so much on LinkedIn. So on average, they're already a bit further along the curve. So probably they already experimented to realize it's difficult. So for me on average, I would say maybe three, because what I do notice is almost ⁓ a it's minority who already realizes what AI can do, what agents can do.

Michael Finocchiaro

Yeah.

Davy

And I also don't see the management opening the possibility to use these agents. And these agents are expensive. They are 100, 200 US dollars or euros per month. It's not super expensive, but you need to make the budget available for the people, encourage them, tell them, look, the budget is there, just use it. On average, it's especially in the automation world, in the industrial automation world. It's single digit percentage of the people who are using it. And almost every automation engineer can benefit this. So this should actually be more than 90%. I mean, less than 10 % of people who are using it today.

Michael Finocchiaro

Thank you, Davy. Marcus?

Markus Guerster

Yeah, it's also to me sometimes quite shocking, ⁓ how little people have ever used actually JetDBT, you know, to maybe tried it once when it came out, but then it just disappeared. Like they're not having as part of the daily workflow. I think Davey, know, Michael, we're talking here about it's our daily tool. Like if it's down, we stop working basically, right? Like it's so elemental to what we're doing. ⁓ It's hard to think away that sometimes it's just shocking to see that somebody actually doesn't use it, where you see like such a clear use case to save half of their day easily with just having chetchbt open. So that's sometimes a bit shocking to me. ⁓ I don't know exactly the reasons for this. It's just, I think. It's becoming clear that the gap is also widening and widening. ⁓ Unfortunately, when AI and JetGBT came out, I was like, my God, this is amazing, right? Because it tries to close the gap a little bit, right? Like everybody has a much stronger tool. ⁓ But it doesn't seem to be the case. It seems to be that innovative people adopt the tool ⁓ and it's exponentially growing that gap, ⁓ larger and larger and larger. ⁓ so yes, I would say Michael, to the answer, would see more on like the one on average, probably there's some outliers. There's some outliers who, as I mentioned that then you get this exponential thing, right? If you have someone, the sea level who is a super fan from this, he pushes it through people realize it. Then you get like the snowball effect. ⁓ and then the whole organization is kind of on fire around this topic. but once you never get that starting point, you kind of it. a one and you're not moving much around the one. You're just scared, right? You're scared sitting in the one, unfortunately. and nobody has the, has the, I guess, yeah, just a desire to change, to get to a two. I think that's something also what's, what's driving a lot of the things, right? It's too comfortable. ⁓ It's more complaining than there's not enough pressure that people need to get from a one to two. That's sometimes my impression as well.

Michael Finocchiaro

So, but when people start using Mont Blanc AI or they're implementing Exelir, is there not a sort of a moment or an epiphany or sort of maybe a slow ripple effect that they start seeing the benefits and they realize data governance, breaking data silos, data ⁓ exchange, ⁓ collaboration between teams that all these things are contributing to more efficiencies, so more higher quality, less breakdowns. Are you seeing that already? Is it something that is very concrete and that we could say if you use AI startups and not just the stuff you've been using for 30 years, you can see these kind of improvements much faster.

Markus Guerster

Yeah, a hundred percent. mean, it has this noble effect. That's why I think the scale is not, you don't have the normal distribution. You have a lot sitting at the one and that's the ones that are sitting at the four. They're getting very quickly to five, six, seven, eight, nine. Right. That's kind of the, the snowball effect. But I think for some people, it seems like an impossible jump to get from one to four, which I think is. a completely wrong mindset because they be your product. know yours, ours, like the adoption that the initial barrier is so low. Like it's crazy low, but I think there's just so much brand marked from like past projects from big ERP implementations, from big MES implementations, from other big projects that if they hear another software, they think of years and months instead of, okay, one day. There you go. Then you get into that snowball effect.

Davy

Yeah, I think the barrier is, it's not so low because ⁓ it is a bit higher. I notice people sometimes don't know how to get started because there's so much information. Like now everybody's talking about cloud bots, but that's not what you should do. You should take a serious agent. So if people scroll their social media or they check on YouTube, it's actually the advice is too complex.

Michael Finocchiaro

David?

Davy

you should actually just take one of the common coding agents and get started with this. It doesn't matter if it's Copilot or Codex or Cloud. So what I noticed is because the initial step does take ⁓ some, yeah, a click. They need to click. And then once they are over this, then suddenly ⁓ it keeps going. And how to get over this? Of course you need to have the mindset, but it can help to have somebody. that knows how to do it. ⁓ Find who already knows how to do these things. Sit together for half hour, for one hour. And that already helps a lot.

Michael Finocchiaro

I guess finding the champion in the organization that actually understands it is massively important. So do you guys have tips on how do you find that guy? You're looking on LinkedIn, maybe they're not easy to find or maybe they are.

Davy

Yeah, yes, yeah. for companies and their organization. Well, if you're a manager, should know your people. You should know who's the guy who...

Michael Finocchiaro

No, I'm saying for your customers, how do you and your customer, how do you find the champion that gets it? Right.

Davy

I have, I, so that's why I post every day on LinkedIn. just, I just let them come to me.

Michael Finocchiaro

Smoke. I bet you mocha is same.

Markus Guerster

The champion I think is easy to find, easy to find. That's not the biggest challenge. think helping the champion to convince internally to get that jump, that little jump of confidence, that has been the kind of larger challenge, I would say.

Michael Finocchiaro

And how, in both of your cases, because we still have a few more minutes, I would like to ask like how, is I think another concern is people think, oh, well, it's either or. if I go in with this AI stuff, have to sunset all the legacy and that's going to be super expensive and I'm to lose stuff. How do you guys cohabit with the legacy stuff that is going to stick around for a while?

Davy

Well, I, there's two things. actually for Axelir, you don't need to sunset. Axelir builds around legacy. But the problem is AI. If you bring in AI, a lot of like in PLC and especially DCS, ⁓ the major packages today cannot deal with AI. And the main reason is because they lock all the code inside their own applications. So AI requires you to have ⁓ open access to your data. And if you.

Michael Finocchiaro

Okay.

Davy

to your data or your code. And if you don't give this to the AI, or you give it, but in a very difficult way, that still requires a lot of manual steps or slow APIs, it doesn't work. So in industrial automation, if people want to use AI, they need to switch to the next generation, the latest generation of ⁓ development environments. And Siemens has this, ⁓ Beccos have this, Cortexys has this. I know Rockwell is working on something, but you need to start preparing to switch. That's yeah, unfortunately in industry it has an impact.

Michael Finocchiaro

Yeah, that same input, Marcus.

Markus Guerster

think we're a different product. So we certainly started from the beginning on to make sure we just connect to legacy. Like even if it's 20 years old, because that equipment is 20 years old and why would you change it if it's running? ⁓ Dave, you're talking about this development bum and it's different, but the hardware aspect, we're kind of not touching. So we work with legacy actually very, very well. ⁓ And we also not replacing any software legacy systems. You know, it's really just setting something on top of it that maybe replaces your paper and Excel spreadsheets around and Power BI and your 20 different screens that you pull up in your 8 a.m. meeting in the morning to see how everything is running. Those are the things that we're kind of replacing or renewing, but not the rest of the legacy systems.

Davy

Yeah, I for me, for me, the biggest competitor to Axel here is Excel.

Michael Finocchiaro

Well, thank you. Yeah, of course. It's the same in the PLM world. It's Excel as a competition. So how do we get past it? How do we defeat Excel? How do we make, I guess it's around the user experience. It's about the ease of use. It's about accessibility too, that it's not hard to get to the tool, right?

Davy

Yeah, made onboarding as easy as possible. Focus on the development experience.

Michael Finocchiaro

⁓ That's been a fantastic talk. I had a great time talking to you guys. Thank you, Marcus and Davy. So other than of course the threaded conference in Warwick on the 25th of March or the threaded conference in Miami on 30th of April, where can folks come and meet you in 2026?

Davy

For me it's LinkedIn.

Michael Finocchiaro

Okay, no conferences.

Davy

Nothing planned, nothing planned in the near future. I'm based in Brussels. If people live in the neighborhoods, always welcome. But yeah, most people I meet are through LinkedIn. It's very efficient. And of course, SASE, of course you have the SASE Slack. So there's a Slack community of the automation engineers who talk about the conversions of software and automation. So it's S-A-S-E.

Michael Finocchiaro

How about you, Marcus? ⁓ Okay.

Davy

⁓ It's a Slack where we discuss about where everything is going. ⁓ So yeah, come join this. If you're not a Mason engineer and you want to discuss about the future, come join the Slack.

Michael Finocchiaro

Thank you, David. How about you, Marcus?

Markus Guerster

Very similar, we're probably going to be at one conference and the IFT first in Chicago, but mostly also LinkedIn. It's certainly the easiest way to get in touch.

Michael Finocchiaro

Okay, well maybe I'll be able to drag you guys to one of my threaded events in the future, you It's been great. I certainly learned a lot about process industries and the transformation. I think you guys are an amazing job. Thank you to the audience. We had a really good audience for this one, so that was awesome. Please feel free after the call to post your questions on LinkedIn. And if you wait a couple of days, I'll have the polished version of this up on... YouTube. with that, thank you very much. And ⁓ I'll be back Thursday with another episode more in the simulation space. But it's been it's been great. Thank you very much, Davy and Marcus.

Davy

Okay, thank you, Michael.

Markus Guerster

Thanks, Michael.

Michael Finocchiaro

⁓ stop