From AI Lab to Shopfloor: What It Really Takes to Deploy Industrial AI - with Christian Zillner
Show notes
Welcome to the Trend Detection podcast, brought to you by Senseye Predictive Maintenance – which gives you visibility and insights into all your assets, from single machines to full plants to help you reduce downtime, increase knowledge sharing and accelerate digital transformation across your organization.
In this episode of the Trend Detection podcast, we’re joined by Christian Zillner, who leads global AI deployment for Digital Industries Automation at Siemens, to explore what it really takes to scale industrial AI from experiments to real shop‑floor impact.
Drawing on hands‑on experience across industries, Christian shares practical lessons on what works, what doesn’t, and why many AI initiatives struggle to move beyond pilots, including:
What industrial AI deployment really means—going beyond algorithms to include business cases, ownership, services, and organisational change
Why many AI pilots fail to scale, from unrealistic expectations to non‑serviceable, custom architectures
The human side of IT/OT convergence, and how unclear roles and ownership can derail progress
How to choose between cloud, edge, or hybrid AI based on latency, security, cost, and operational constraints
The role of partners and ecosystems in taking AI from the lab to production
Where AI delivers real value today—and where expectations still need grounding
Why standardising the deployment platform early is critical to long‑term scalability
Practical advice for moving from experimentation to production with a small set of repeatable, high‑value use cases
A refreshingly realistic discussion on industrial AI for anyone responsible for digitalisation, automation, or AI strategy in manufacturing.
You can find out more about how Senseye Predictive Maintenance can reduce unplanned downtime and contribute towards improved sustainability within your manufacturing plants, by visiting: www.siemens.com/senseye-predictive-maintenance
Connect with Christian on LinkedIn
https://www.linkedin.com/in/christian-zillner/
Transcript
This is obviously one way where you try to solve a problem that is not solvable with classical automation. The nice thing is that you usually have a pretty defined business case already. The downside is that it's usually a pretty complex problem, otherwise they would have solved it already. Um. Then the other way around, if you go top down, um, like. Welcome to the trend detection podcast powered by Siemens. And now your host Nel Sullivan. Hello and welcome to the latest episode of the Trend detection podcast. Uh, I'm really pleased to be joined by a brand new guest, cos I often have a lot of recurring guests on this podcast, but a brand new guest and one of my, Um, Siemens colleagues, Christian Zillner is joining us today, and we're talking about the very important topic of industrial AI, but we're actually, I know it's one of these phrases which is often used, but we're gonna sort of ground it in some reality, particularly on how to deploy it and get real value from it ultimately is what we're looking at. Um, but before we sort of dive into that, I'll hand over to Christian. Just to introduce yourself and welcome to the podcast. Hi. Hi, everyone. Um, my name is Chris. Thanks for having me. As I already mentioned, I work for Siemens. I'm heading the global AI deployment team in DI automation. Means that I am in touch with a lot of projects and use cases where we deploy AI on the shop floor at our customer premises together with partners in this fast moving environment. It's a lot of fun, a lot of learnings, and I'm very happy to be here. I guess just a question sort of a little bit off topic for what we um sort of planned, but you mentioned all the different news cases. Is that a good thing to have lots of different news cases to focus on, or is it, you know, can it be quite challenging from that perspective, cos usually people say focus on things, right? Yeah, focus definitely is always a good idea. I think from my perspective we're just working in so many different verticals with different requirements. Some are very close to our shop floor where we're having what we call adaptive control. This means like AI really controls processes on the shop floor. They are very different requirements to, I don't know, some monitoring or predictive maintenance of pumps in a, I don't know, pipeline. Of gas or whatever, uh, you want to monitor. Uh, so this is the variety we have to handle on, on the siemen side, but it's also very interesting to always see the same learnings all over. Yeah. Our focus is always a good idea. Yeah, exactly. Maybe I already knew the answer to that question, I guess, but I thought it was interesting. Interesting to ask anyway, but, um, and, and I guess today we're actually looking at that journey from an AI lab, um, all the way through to the shop floor, and I guess the value sort of generated there. And we're also focusing, I mentioned in the intro, um, around deployment as well. So what, what does that actually mean to you, Christian, in practice? Um, Yeah, it's a very big word obviously, but for us what we realized and learned over time is that deployment really is working with our customers because usually an implementation of AI or scaling of AI on the shop floor where you need to bridge IT and OT interests is kind of a change process. We usually trigger a lot of responsibility processes, scoping of business cases, financial decisions attached to it. And then obviously there's also the technical implementation, but it's just a part of this whole deployment cycle we're in. And obviously we always try to to do that in line with partners and internal teams and then customer teams. So it's quite a bit of a tricky, complex environment we're in there, but it's also very interesting, obviously. I noticed you've mentioned sort of partners a couple of times already. So what's a partner's role in these projects? What both their importance and, you know, what, what do they bring to the table? What we learned over time is that it is crucial if you want To really scale and I think you called it from the lab to the shop floor. If you really want to implement real industrial AI use cases, you need to have an owner of the application. What we mean by owner is that someone who services over the lifetime of the application, it's no different to any automation project you're probably used to, and those system integrators, um, machine builders. Global solution integrators are our partners on that journey to implement, and this is the one kind of partner. And then obviously we also have technology partners. We partner with camera vendors, we partner with software, other software players. We even partner with cloud providers where we try to standardize the deployment architecture. So there's ecosystem is really, really important to make it a successful implementation. I mean, you mentioned successful implementations, but I mean, quite often, and you'll be able to, you know, tell me and the listeners, but, better than anyone, you know, that a lot of industrial AI copilots, sorry, um, show a lot of promise, but, you know, don't scale at the end of it, unfortunately. I mean, so from what you've seen, you know, what, what are the key reasons behind that? You know, what, and maybe also on the flip side, what, you know, where do you see the success factors, what success factors are involved? Oh, yeah, um, we can confirm, like, um, as Siemens. A lot of own factories ourselves. I was responsible for implementing AI in our own factories as well, and we failed many times. My very subjective learnings there are that expectation management is key in combination with the business case because usually those use cases develop over time. So you need to have a good backing and quite a bit of a business case to finance an industrial AI use case. And a second, we bet in the, uh, in the past, a lot of custom architectures and we heavily regretted it afterwards in the scaling phase because if you're locked into custom solutions, it's really hard to enable operation services that are mentioned before for partners taking over the solution, um, doing MLOps services, um, providing it, uh, is really, really, really hard if you don't have any standards. And so what I'm hearing there is when these projects don't get off the ground, is it usually a technology problem or, or something else? You've sort of mentioned, I think, you know, with the architecture you mentioned there, but are there other factors at play as well? Yeah, Obviously the evolution of model availability in our space is just enormous, so we A lot more possibilities in the accessibility towards machine learning platforms to train models has evolved a lot in the last couple of years. So most of the predictive quality inspection vision use cases also robotic physical eye use cases we see and we implement. The model side is pretty much covered. The second bigger hurdle is obviously then data availability, and this connected to a standardized approach of how to collect data, how to provide it for model training, and then obviously connect them. Models on the shop floor with the real sensors is a challenge, because there's no standards for IoT platforms yet that standardize this whole process from cloud to edge across all verticals and all use cases we see that. So this is definitely quite a bit of a challenge. And then the third thing is the deployment platform, uh, where we, where we have, where we see great progress with GPUs availability. But there, this is connected then a lot with, uh, with the price point and the business case I mentioned before already, because you can spend quite a bit of on compute if you want, either in the cloud or on the edge. I mean, maybe we're leaning more into sort of the adoption of technology here, but are there cultural factors that complement what you, you were explaining there on the technology side? I think you have to have a look on who drives the business case or who drives the use case. Obviously the profile and the capabilities of an OT team or shop floor team is very different to maybe a data science IT minded team. And for an industrial AI project or a deployment or a scale, even a scaling, you need different profiles, and we always differentiate them on first, the domain expert that really knows the process well. He's usually responsible for the business case, seeing the implications, safety, quality. All of this. Then we have the IT personas in it. It's a lot about security, CRA compliance, cyber resilient Act compliance is a big, big issue these days. Then we have the data science team doing the data engineering, um, selecting the model, model validation, and then you need some kind of connectivity guy that bridges the gap from like data collection on the shop floor, bringing, bringing it to the cloud or to the computer environment where you train the model, and then. Reverse it back and bounce the model back to the shop floor then and the fifth obviously is once you've delivered the project is then service where you need guys to take care of the envelope services. So it's quite a bit of a cultural aspect in it because you have different personas. Um, and we try to, to reduce the complexity by a lot with UX or UI. So G AI also helps us a lot to reduce complexity in the interaction of engineering, but we're still not there in the whole life cycle or MLlos life cycle. Yeah, maybe let's dive insight a little bit. You said generative AI is helping to make the complex less complex. in, in, in what ways? What, what examples can you provide? Usually an OT engineer doesn't have any Python programming skills. So if we have a copilot kind of interface where he can do intent-based engineering for, I don't know, some kind of a, just making up an example, a vision solution where he wants to detect scratches on a metal sheet or whatever. Then intent-based engineering could be a key to enable an automation guy to set up and, and sketch and develop an industrial AI solution and train a model without any data science, uh, know-how. Uh, so this enables us to, to scale and the, the cool thing is, If we could do that, we could reintegrate in the already existing service processes because automation knowledge is there. So talking again about machine learning operations, if some guy or if an industrial AI solution only requires automation knowledge, we are a lot further down the road in AI adoption. And I know we're sort of jumping back to what you were saying before, those, those different personas, is it, it's interesting because with so many different, you know, personas involved in a, in a deployment. Is there someone who actually fully owns the deployment from your perspective, or is it a shared ownership model? What, how would you sort of describe it? It obviously always depends a lot on the use case, but we always say that IT owns the standards and the security. And OT owns the use case because they, the intensivation is quite different. IT is Incentivized on security and OT is incentivized on efficiency, output, up time, all of those. So this is how we usually split it and then if we translate what you just asked is that the use case usually stays with the OT, which means that the solution has to be serviced by the OT department or business unit. And is that ownership always clear from the start within an organization? I know it depends, of course, between organizations and what use case, but, you know, is it always clear straight away, this, these are the defined roles, or can it get a bit blurred sometimes? Ah, it's very blurry. Usually we have to kind of, we have those C-level things where two C-level guys say, oh, we want to do something with AI on the shop floor. Um, or we have those bottom-up, uh, use cases where a guy tried everything from classical automation to smart sensors, smart cameras, whatever, and then ends up at AI vision, for example, because he could not find another way to solve his problem, but it's a significant problem, so he wants to, to throw some money at it or it's significant business wise for him. So those are two very different directions and, and either of them has different responsibilities, budget constraints, and all of it. But as part of our deployments, we usually develop racing matrix uh with our, with our customers to see the responsibilities, accountabilities within the shop floor deployment and then obviously services is one of them. Um, and I, but I guess to sort of shift gears a little bit, you know, another thing I wanted to cover today is sort of the difference, if we're looking at industrial AI in terms of the deployment of these projects, you know, cloud versus edge, and where you see things at the minute is, you know, is it leaning towards one or the other, or, you know, are organizations thinking, yeah, in, in a certain way around this topic? Maybe you could go into that a little bit. Yeah, I think AI, the adoption is heavily structured in three maturity processes. One is model availability. There we did already quite a huge progress in the last couple of years. Then we have data availability. This is heavily connected to Aure cloud. And then we have the deployment or the inference. Maybe you heard of that buzzword a lot like where to deploy my AI on. And data and inference is a different thing. So I think you need quite a bit of a hybrid approach to collect data and then push it to the engineering environment where you have the possibility to train your models, validate all of it. This is one thing, and then you need to choose a deployment environment. So where you want to do the influencing of your AI model. And this is very dependent on your, on your use case. If you have very high latency and very high um security uh requirements, uh, you usually go edge only. If you have less latency requirements, like if you do monitoring of a pipeline, just going back to my, to my example, where you maybe do one or you have to do one prediction a minute, then you can easily also go cloud. A second factor is the data format. If you want to push all your data to the cloud, the cloud providers will love it because it would have been enormously expensive. If you talk about vision data. We have use cases where you have 16 cameras to connect it to. One AI model, so it will be large amount of, of data streams, um, that has to be orchestrated and they usually we, we go to Edge and then just have like data collectors that, that collect or sort the data that you want to push to the cloud and later on. So that, so that's the, the way around, you know, sort of rising, the rising costs of cloud or that could potentially be from that particular use case you mentioned. Yeah. Take a step back a bit because I know we've talked about use cases and clearly that's the important part, because before you deploy AI you need to know sort of what use case you or problem you're looking to solve. So, what does that look like in your, in your world? Sort of, what's the process for defining use cases with a, with a customer? I think it's 22 ways. Um, the one way is, uh, if you have like a, a shop floor problem that you want to solve because it's immensely important for The factory output or and you go and basically bottom up. This is obviously one way where you try to solve a problem that is not solvable with classical automation. The nice thing is that you usually have a pretty defined business case already. The downside is that it's usually a pretty complex problem, otherwise they would have solved it already. Um. Then the other way around, if you go top down, um, like headquarter down, it's obviously the, the scaling is a lot different. Um, if we, if we, for example, take, take a vertical like consumer packaged goods, you have a lot of standardization. We usually look at very standardized processes and try to get into those standard architectures, or line setups, the, um, the AI or the industrial AI solution or Use case already implemented and standardized because then you have a copy and paste approach while scaling. Keep in mind here to choose the architecture upfront, because then if you want to scale and it's custom and not serviceable, it's obviously the cost of failure is a lot higher. Yeah, I guess you, you touched upon it then there a little bit, Christian, but maybe we could dive into that a bit more. So we've talked about the, you know, the problems of getting these projects sort of off the ground, you know, maybe in a pilot phase, but Equally, at scale, you know, clearly success has been achieved on one level, but what sort of issues become clear, and you mentioned architecture could be one, but what other sort of challenges sort of appear as, um, these projects scale upwards? I think the first thing is always use case definition, um, connected with the business case, but this is no different to any other automation project. So don't be too scared here. I think it's, it's the same, just do your job and scoping the project and defining success metrics, uh, scope of work, all of that. I think this is equally important for an AI project or even more important. Second, obviously is data availability. Um, there's shit in, shit out. If you have unreliable, uh, data that is not standardized, it will be costly or even not reliable in the end. Third is model development. Obviously, you need a good data science partner or you do it yourself. It, it's your choice as a, as a solution owner if you want to own it and see the model development as your core competency or you want to have it in-house, uh, or your, your partner here. And then 4th is the deployment environment as, as mentioned before. Um, here, I think it's important, the one side is the data connectivity to standardize and productize this. And I would strongly recommend you to partner there, um, because connectivity is a key issue for shopflow data. And as it is critical, it's very important. And then the other bit is the inference environment you want to choose, whether it's in the cloud or on the shop floor with Edge devices locally hosted or then in the in the cloud. And it, it's interesting that you, you were sort of balancing, you know, whether to own stuff for the organization to own, own, own it or outsource that potentially. And particularly on the, the model side. I mean, do you have a recommendation either way, or is it very much down to the, the use case, the organization, how they want to approach things? There's no right or wrong, it's just what's the right approach for them, maybe. I think the right approach is what you can service. If you have global availability of data science services within your corporation. You can do it in-house if not partner. Yeah, that, that keeps it very simple, right? When, you know, the, I guess the danger is with some of this, well, a lot of this is to overthink it, but that keeps it very simple, so that's good. That's very good. And I guess going back to the, the scale question as well, because you've mentioned the importance, uh, of, you know, consistency, particularly with architecture, and I'm sure there's other, um, you know, other areas too. But I guess there's also got to be elements of flexibility as well. So I'm just interested in how you sort of balance those two very contrasting things. In a, in a project when scaling in particular. I think this is strongly connected. In our case, we have the approach of partnering. This has always been our business also in automation, where we have system integrators, line builders as partners, and it's the same for AI solutions. We try to have a operator of like 80% product and 20% solution, and this makes it also an attractive business for our partners and we are developing. What we call verticalized use case offerings there together with our partners where we try to productize as much as possible. …