Industrial AI: From Pilot to Profit

This podcast is presented by NXAI, your partner for time series foundation models and physical AI. Hello everybody welcome to a new episode of our Industrial AI Podcasts And this is very special podcast. My name is Roy Viva and one of our biggest fans Biggest supporter from day One. Boris Scharinger is our guest.

Boris Welcome To The Podcast. Yay, thanks for having me. I'm so glad to be here and i'm happy too. welcome also my co-host Peter Sieberg.

peter Welcome to the podcast. good morning robert. Good Morning Boris. great to have you with us.

yeah boris You are here today not in your capacity as a Siemens employee but As an author of An important book And The Name Of The Book Is Industrial AI From Pilot To Profit Key Concepts Success Factors use cases and market mechanics. I'm very happy that you are here, and we'll discuss with us your book. Perfect! Yes.

But before we start can please introduce yourself maybe briefly in two sentences to the listeners? Two sentences...I started my career basically coming from IT management, IT service management which is an interesting place to be because there's already a area where innovation versus stability, right? IT must be robust and so on but you also wouldn't like to innovate.

And then I moved on...I did a little bit of work in data analytics for audit and then i moved into the area of industry. four dot zero focused pretty early long before there was this hype and momentum in the space of AI. I focused on analyzing technology Okay.

So and now this book, why is this book necessary to read? Yes because it's all about also the core topic of your podcast right AI for in industry. do not buy The Book. No you can't do both.

And both supplement each other exactly. It's All About Why & How AI Needs To Be Different from what we see in the consumer space, in a social media space to industrial applications. The additional diligence that is needed... the additional planning and- What do you see?

And it was key motivator for me writing this book. We really overhyped expectations coming form top management personal experiences of how easy it is to use JetGPT etc.. overhyped expectations and then the pressure that is put on people to get AI deployed in engineering environments, and on shop floor environment. And so everyone's hitting those walls.

we are all familiar with the statistics of how many proof-of concepts don't make it into productive deployments. That success rate is very low. This normal but not something is not influenceable, right? We can work on improving and increasing the success rate.

But then we have to diligently look at different aspects of what it takes to make AI industrial gate to deploy it diligently with impact on What Is My Proof Of Concepts? What Am I Looking At? What Are The Gaps Between Proof-Of-Concept Activities And What It Really Takes to deploy it in a productive, operational environment. And the more I understand those differences - the scope of proof-of-concept versus the scope for deployment - the more can work on details and address gaps….

…and communicate towards top management what realistic expectations are. In the end when we run AI initiatives in industrial enterprises... We need to set the expectations, right? Otherwise we will fail.

It would not be a successful AI initiative if you fail on very basic expectation setting and then your doom-to-fail is that it's no fun! And this why I decided write my book to inspire people what possible but also give tools at hand really prepare properly understand those gaps address these gaps help a little bit with formulating an AI strategy and in industrial context. And then last but not least, providing many examples of solutions that are commercially available partially from established players out there on the market.

Boris, in twenty-twenty that's six years ago Robert and I wrote a book called. what was it called? In German same thing car E. And the industry AI an industry we actually have or almost say We had the same publisher because we've actually been asked just a couple of days ago to discontinue our book.

So thank you very much for that, Boris. Yeah thanks, Boris! You're throwing us out. well there's huge differences...

We did our one in German six years ago. nevertheless what would you say? Assuming And maybe not so specifically on comparing the books. But what has happened, would you say in those six years?

In the area of industrial AI? it is interesting because twenty, twenty, nineteen was a period and time when I started to look into the technology area for industry. So prior to the generative high right at momentum end that has changed a lot. Different stakeholders also within the community, within Siemens.

Within Siemens sales. everyone said Boris our customers don't care what technology is used to solve their business problem at that point in time. so stay away with talking about AI and Stay Away Talking About The Differences Between Consumer AI And Industrial AI And AI Quality. No One Cares About What The Technology Ingredients Are that I used to solve a business challenge.

And then with twenty-twenty two and the arrival of chat GPT, and the momentum off generative AI that changed a lot That was turned upside down. top management tells their direct reports do something with AI? i don't care what use case it is but we really need to deploy this AI thing. We still see Focusing on the business problem, being absolutely technology agnostic.

And now we have a situation where everyone wants to introduce their technology and tries to find a proper use case. Yeah that's an interesting point because from my experience most companies fail To define pain in what they want to solve Exactly. And I would go even one step further. We have in industry a lot of AI initiatives that primarily dominated by the ambition to make the employee-based learn, right?

So do something with it because we all need to know and understand how it works! Then use cases like let's write an agent crawls through my emails and sorts my e-mails according to priority. Now, from a corporate perspective having this learning curve is great And it's good! I'm not saying don't do this learning Curve but This doesn't create any differentiation in the market Any mode.

It does NOT help with your core processes To become more competitive as a company. We see Use cases in procurement, we see use case is in accounting. We see use cases everywhere where individual employee productivity is marginally improved. now we can discuss whether ten percent productivity increase is marginal but we are far away from the strong differentiating use cases that we see form a couple of thought leaders out there That really kind changing market dynamics and structures by using AI.

Why? Because finding those differentiating use cases is not the moment. So there's no killer application in a moment when it comes to industrial AI. You can look at that way, I do believe and some of you also your listeners have met me And heard me talking about that before as strongly believed That area of design space exploration Is something we all should invest In.

You know, we have a solid chance to create pipelines AI and data-driven pipelines in the space of product development. Of product design, product engineering that really creates differentiation In The Market by Solidly And Quickly Improving The Quality Of Your Products In The market. I would Have One Or Two Examples If you Like. So one Of My Favorite Examples Also Known To The Community Many Of You Is Probably that Siemens energy used AI in simulation, to accelerate a simulation.

you even had podcast episode with Peter Kurt and also Ben Amnuri from Siemens Energy about it. So the gas turbine blade design was so significantly improved by this exercise of design space exploration with the help of simulation and AI, that Siemens Energy became a new performance leader in large gas turbines. And that changed for years - the market share that Siemen's energy could win in that space." Another example is naturally how Tesla approached automotive production.

most of you know about the giga presses, right? And the idea that uh...the Gigapress creates a car body or the underbody at least as one piece. where other automotive companies would have seventy different parts.

they need to be assembled and welded. So Tesla really substantially changed efficiency in car production by this Gigapresse approach. Now To Make The Gigapres Happen you have to find the right material, and that material must be found with two gold dimensions. The one gold dimension is...

the car must work! Yeah? The static of a car, the flexibility all of it must work. It has to be neutral to road chemicals That we use in winter for example To melt their eyes away.

All those attributes need to be fulfilled With materials like this. And on the other side, you need to find that material. That you can inject in a very short time into this giga press and then it cools down in a really short time because faster it cool's done The higher is throughput of the Giga Press? To find that Material mix that Tesla then after they have found actually has patented simulation and design space exploration.

I thought patents are for the week. That's a different discussion. and what he tells everyone, and what Tesla does. The Tesla engineers do that may be two different things but they really developed strong differentiating capability by the use of simulation in material discovery.

I had an instrumental role from my alter colleagues activities, they are completely different from you know individual employee productivity increases because now my AI writes the marketing blocks for LinkedIn as opposed to me sitting down and sketching every text by hand. Right? But podcast is that still the real borrower's? or Are you leaving there too?

No it's my digital twin. but It's interesting Because your two examples engineering topics, right? So no operational. No shop floor topics.

well that this is also part of my book and you know You said it's about key concepts and success factors in market mechanics. And here's my point with the shop floor. The fact that shop floor is about physics and machines impacts the way how that scales in need of physical AI to control them. So, you know the laws of physics must be somehow injected into AI and immediately at that idea of humanoid robots.

no this is not a point! The point is physical assets on a shop floor follow the depreciation cycles of physical asset. so if I have a brownfield environment If i have production line And In my production Line I Have Say Ten Machines And the data scientist comes and says, I have this great idea. Let's do this model.

it is going to help us improving yield of our production line." Then one these ten machines doesn't deliver or provide a data point that that data scientist needed. then scientists need trigger say an update version upgrade for machine replacement. And if he or she knocks on the door of the CFO saying, you know for my model.

For me activities I really need this machine to be replaced The CFO goes like Are You Crazy? This Machine is depreciated over seven years Over ten Years and now after four years you want To replace it. that's not commercially possible. That's the reality Of working with physical assets.

The data scientist goes into the suspension mode for three years, hoping that after four years that machine gets replaced and he can continue with his idea in use case and project. And this is a huge part why... the AI adoption on a shop floor is slowed down! That's the reality of a machine park you have to work with or live with then using soft sensors use retrofit sensors to collect the data points that maybe the machine doesn't provide, but we all know this is tricky and isn't an ideal way of doing it.

But that's reality! So product engineering... you said your examples are really in a space for product engineering? Yes AI scales better on the product design & engineering side.

Why? because this is digital. You don't need to replace a machine to collect another data point. And this is why the potential of using AI and increasing productivity by AI in product design engineering, it's a lot higher scales better than that actually on the shop floor side of life.

So who are you targeting then specifically? Who have our listeners should consider In addition to continuing listening to our podcast also say oh well That could be interesting for me because And I was thinking also, as Robert suggested and you now talked about engineering a recall turbine design. Ben Ham he joined us in Wurzburg AI at the monastery. for those of you that recall He Was There As Well.

But then You Have A Lot Wide Arrange At The Same Time Of Use Cases. So specifically who would you Who Did You Have In Mind When You Were Writing The Book? A couple of different personas, this book is also extremely helpful at least. That's first feedback that I have received for startups in the space of industrial AI because they understand a bit more on their robustness and reliability needs and requirements of larger corporations in industry.

In a nutshell i would describe This Book Is Meant For Anyone Who Needs To Balance The Fragility brought by AI and innovation on one side, like how can we disrupt things in processes? And the stability needs that you find an industry. Whether your are a startup working with a corporate whether you're production line manager and data scientist knocks on your door shockingly or whether you are data scientists and plan to knock of your production line managers Or engineer who works product design simulation.

This book will help you to understand the other side better. Right, yeah if I may extend the question... The book is a bit. how many pages it?

Three hundred fifty. There we go. that's big difference from what Robert and we had about one hundred... You were lazy, lazy office!

Yeah, lazy boss I just look at one specific area, it's on a very small point. I see data centers, AI semiconductor energy and geopoliticals. so if i may extend the question number one are you thinking of an international global audience. this book can be read by Chinese person as well as by an American, and a Dutch or European.

And then secondly maybe you can talk about depending on how the answer is going to be where is Europe? Where's Germany in exactly this area of data center semiconductor etc.? Yeah, so of course I wrote a little bit about the geopolitical situation. About the ingredients that you need for AI whether it's talent to talent side Whether its data center and semiconductor site.

It is not major part but provides context. And that context important because if look at export control restrictions For example of AI semiconductors You are machine tool builder and you have to make up your mind, how do I sell my machines with AI functionality? To China versus How Do I Sell My Machines With AI Functionality And AI Acceleration Into the US. Can i use one technology stack or for geopolitical reasons ?

I Have To Create Product Variants. Then This Provides Context For Decision Makers. Again, they need to balance the topics of innovation versus reliability and robustness because two kind of spin that thread further. Funny thing if you manage quality of AI You will find out that a Huawei chip does floating point calculations differently from an NVIDIA chip.

so your models who are way accelerator, AI accelerator for your machine in China will actually deliver different results compared. Yeah funny thing. not so funny if you're the product manager or a quality engineer for that solution. but you have to incorporate that into your thinking and testing concepts which by the way is never done in a proof of concept.

So this exactly one example where In the proof of concept, you say no. The model works. that's fine. it works perfectly on my laptop with a GPU card from NVIDIA.

Fine let's deploy It. and then You hit the wall finding out That the Model behaves differently On A different AI Accelerator And you haven't planned for the test Concepts covering them right. So you can see immediately There is a link between what's happening on the geopolitical side With some rivalry Between the blocks export control regulations impacting your technology roadmap, and adding additional need for quality management. so AI becomes industrial grade or stays industrial.

Robert if I may just leave it to you because my feeling is that we've now come I'm not sure. You need to confirm if that is the big message you are handing over because for us, believe Robert and me... you have been Mr Industrial Grade AI. Maybe it was Jay Lee who came up with the term industrial AI first.

But for us, since a couple of years you have been the person who and we talked about that again. That's now week or two ago if you want to listen to that day listener was that. I think it was called The Harness Engineer, the AI harness engineer? And when we talked about that, I suggested well If they would've asked Boris he might have called engineers.

So Alice, a little bit more about that because you were the first one from the beginning to say in an industrial environment things are different in whatever consumer world I guess? In one way yes i started looking into this but when we look back at the famous infamous research paper by Google on hidden technical depths of machine learning solutions You could already see a very clear indication by the Google colleagues, you know how small the machine learning model part in the overall environment.

In the overall system that uses AI is and if there's a famous figure You could almost see visually from that figure, maybe the AI component or model itself is two percent of the overall system. When we all heard about this code leakage in the cloud environment and everyone was looking at it finding out how one point five percent was actually about the model and. And all the rest, uh, ninety eight dot five percent. where the harness what they now call the harness right?

Which is basically all sorts of you know API routers and quality mechanisms and checking mechanisms and off course all the billing and configuration and so on of the overall system and services belong to that harness. So Harness has been out a new term for the overarching system or a system that surrounds the AI model, but the message is still the same. The AI model itself it's just a fraction of overall solution and all of this including the harness is desperately needed to manage quality what the model delivers in terms results.

I want come back to your topic engineering because everybody talking about embodied AI, physical AI. And now you came around the corner and say yeah but this is too difficult. forget it concentrate on engineering AI or am I wrong? I'm at least saying that scaling productivity also maybe developing something that differentiates from others.

as a company The area of product design A product engineering is a very rewarding area to focus on. But there are lot of pains when it comes to the shop floor that can be solved with AI. so, There're lots. cost factors can be solve by...

So do you say okay if want build something new? You should focus on engineering and If you want to solve pain go in the shop. for was AI or I mean, we have also two flavors on the shop floor. We have of course green field projects where someone says you know let's build up this new factory and then we can really start from scratch.

We can select our machines and machine models in our suppliers for machines according to their contribution To your design data driven capabilities and AI-driven optimization capabilities. In a Greenfield You can really go full Monty, so to say. Brownfield environments as we all know are Really difficult and this is why the adoption And scaling speed Is simply slower? I'm not saying don't do it.

i am just saying let's manage The expectations right because It will be slower. if top management understands that then you Can run an AI initiative being successful because the target setting and goal-setting was realistic as opposed to you know, the overhyped expectations. And then we come back to humanoid robots... The overhype expectations.

that very sure I'm very sure they can't be met and everyone is disappointed in end of today. But what's your opinion on embodied AI? Because everybody talking now about embodied AI Let me put it that way because if we are talking about embodied AI, We're immediately talking about autonomous systems and to a certain extent Autonomous robots. And then we can discuss form factors.

The automotive industry spent at least two trillion dollars in the last decade on developing the autonomous vehicle And everyone spent a lot of money. and then if we look at the results, where are we? Well. We have a couple of technology leaders in that space but not many.

end. I still have to drive to work and i'm Not being driven to work Level autonomous driving level three not even four. right way mode does for. So there has been an incredible amount of money invested into that area and the degree Of truly your true autonomy in this area is really limited as an output.

It doesn't pay back, it's just lost investment if you will. now let's compare quickly and autonomous car and a humanoid robot. An autonomous car basically has four degrees of freedom that you need to be able to control You drive left ,you drive right, accelerate or brake That kind. Yeah, you can go reverse.

Yes? Okay maybe we add a fifth one. so this is the. these are dimensions that you need to manage.

in a sense if you look at an autonomous robot You know I mean immediately In the dozens of dimensions it Is right now autonomous car is focused and developed To not touch anything. Autonomous robots and humanoids don't make sense if they don't touch anything, because you know that need to perform work. They need be able do machine tending. so the problem size and complexity is by many factors higher in this space of autonomous robots And full appreciation for the venture capital goes into those topics.

The automotive industry has spent a lot more on autonomous driving. And I wouldn't say they haven't arrived anywhere, but the results are very sobering and financial results are more than sobering for many of the automotive companies that have invested heavily in these types of technologies. Your point is not please machine builders. do not integrate AI functionality in your machine, right?

Absolutely. No and I think that there is a mandate we really need to be on the mission. every one of us needs to be out of mission To use AI as an optimization tool to improve production. We know if you look at it from a grandchild compatibility perspective Business.

do I perform business in a way that my grandkids. Will appreciate and like what i did? we need to invest the lot into making production more efficient consume less resources, less energy right? And applies to any machine using one favorite examples of mine and certain extent is still unresolved.

instead many cases machines are so precise When they are operational when there are warmed up That you would never switch them off like a CNC machine of the certain type. You wouldn't ever switched him off during off times. So if your shift, you still have a full-powered Machine because after machine cools down it will lose precision right. so we're consuming a lot of energy despite the machine not doing anything Because the recalibration process of the machine regaining precision is come by some.

yeah you know let's use AI and maybe create calibration processes if possible that allow us to switch machines off when they are not in use. Maybe we can do it, this alone is an area where By the way, autonomous driving was invented forty years ago. Right? I know five hundred meters from where i'm sitting at The Runway here south of Munich but that's a different topic for those of you who are interested.

it was by Professor Dieter Ernst Dijkmans professor Here At The University another professor. his name is Will van der Alst Dutch like myself. he's professor at the Aachen University, and also the chief scientist at Ceylonis. And he has written forward to your book now for Robert and myself.

Will is known as I would call him The Process Mining Pope. You could say maybe better as well. so the question Play in what you have seen the last. You started two thousand nineteen so to say, The Last Five Years or whatever.

and What role does it play at the same time also? Process mining is a very interesting discipline closely related To A data-driven approach of business process reengineering right Business process improvements no matter what business process. And process mining had a great run. The first generation of process mining activities was really in, you know very workflow heavy processes like in financial services where the creation of a financial service is offering say an insurance policy was heavily dominated by these types of call center interaction work flow management tools and so on and so forth right?

In industry to a large extent we always said that It will do the job in one way or other, right? And then we build all these workflows around SAP to manage our business. Also ticketing processes for example was really easy to analyze and improve using process mining. The precondition of doing process mining is structured workflow information time stamps, activities and tasks being executed by certain users.

And I would collect all of these log files and then use process mining to understand task waiting times, task issues, iterations unnecessary and so on and so forth. Now AI comes into play for my perspective. Process mining is a flavor of analytics but with AI and generative AI We can even go into pre-instructured information, we can go in to emails and understand e-mails. And my favorite example is...

Don't ask from what company I do gain that experience. but when you order something a purchase create or purchase order ...we have an approver determinism algorithm in SAP and it determines the approver! The approver never knows What the purchase order request is about and we're not gonna ask you again what company work for.

no, No. And then after that approver has been determined by the SAP system There's an email sent out. You know there's a task for you to approve this purchase order. Please have a look at it and so on in support.

and Then an email based process starts of one week two weeks six weeks where people communicate to each other, you know what is this purchase order about? What does the project context why it doesn't have to be approved and so on. And all of that happens by email! After that clarification communication coordination has taken place The approver goes into the SAP system and says I now approve.

Now when we just look in to the log files of SAP You'll see Approver request created. Six weeks waiting time. Then after six weeks It's been approved. Now, with generative AI you can even use email communication if it's allowed and approved by our councils.

To look into the communication flow. how are people interacting to solve that stuff? So this is an example of how AI and process mining now create a more powerful analytical tool for both business processing engineering because we can analyze them Approver determinism algorithm in our ERP system. But now something else is on the horizon, there has been a research paper quite a couple of years ago where they did experiments with using process mining to validate if a task sequence by a robotic environment yeah?

A robot has being given a task. so real robot or bot? no that was a real robot okay. So we are now back on the shop floor.

We're in a task sequence planning and it has been done by some sort of algorithm, and then process mining was actually used to verify or validate whether the tasks' sequence is good quality. And that brings up our thought which if you look into an agentic AI environment that they are dominated by an AI orchestration layer, and that orchestration-layer interacts with the user gets a complex task and breaks the complex tasks down into several different smaller subtasks for which it would talk to agents.

And say you know agent one please start now. agent three You can take over once that Agent One has finished. Agent two do something in parallel. Agent four merge all together and provide back to the orchestration layers.

All of these agents, hopefully in an industrial-grade AI environment would lock their activities. So you could audit it and then use process mining once again to say thank you AI orchestration layer for your task sequence planning. but actually I found a flaw. Agent Force work started prior to agent two having finished its.

Actually, that's a sequence we don't want to allow because we are really keen on Agent Two finishing fully. so Agent Four can work in quality input. So process mining you could write a must-sequence or required sequence of task activities into a prompt but then it is static right? You do not want the AI orchestration layer to have dynamic capabilities to use several agents in a certain sequence, you don't want that sequence into a prompt.

And process mining could become an important governance and quality assurance mechanism for agentic AI environments from our perspective. So your book the whole AI ecosystem involves quickly. how will it stay up-to-date Boris? What is I had a fantastic call this morning with Sylvia Hasselbach from the Hansa Verlag and we're discussing already what would be at good time pace to update the book, have revision two or three.

From our current perspective you see so much dynamics in AI space but also in startup landscape that will work towards an almost annual update of... That's what we didn't do, Robert. Maybe that is why I'm out of the market now. Lazy author!

But we need to make a podcast weekly Peter. so... Yes? that was our excuse.

Yeah So i think there isn´t way in actual landscape also in changes on regulation In geopolitical aspects of whats happening. There are no other ways than keeping it updated at fairly short pace Right, so you're saying even if people wouldn't buy it now but later. You would try to make sure that they get like an up-to date version. which brings me to my final question combination.

probably and of course as always I'm going ask for some kind of outlook. Now, I'm sure that is in your book. there's hundreds thousands of topics. i have one specific one but please extend it with the two three things that you think even more important.

The thing is that Robert and I talked to weeks ago about this crazy word jopo-calypse And I suggested that nobody really knows how big be, although it's only two percent of the code as we just heard. But I think at least we all agree. It's rather powerful but nobody knows exactly what's happening that maybe you know. What is your perspective?

Do You want to share now with young people listening To us or also Maybe let say middle-aged generation That for whatever reason Is considering if they should continue what They're doing? Well, what I can certainly say is that the jobs with very repetitive activities will decline. So we see already significant job loss in shared services centers In near shore locations right around Krakow? In Poland and other places in Eastern Europe We have near-shore centers executing Very repetitive workflows And we can see the automation capabilities in those areas are really peaking, so that has job impact.

We have other areas where more creativity is needed and I'm looking at maybe software developers for example and vibe coding and all of that stuff... And then you could look at it and yes may be your not writing code anymore as a software developer in Python but than you're writing the prompts to generate. So maybe the way how you program changes, but one of the bigger venture capitalists in North America said looking at the fear of losing jobs in the space of software development.

The ambition of mankind have always been bigger than its capabilities. so we shouldn't worry. job profiles will change yes? But there will be enough work to be done by humans because our ambitions always have been, you know let's fly to Mars and once that we've done this maybe we'll fly to Jupiter.

I don't know but i'm looking at it optimistically with the individual duty to stay tuned on what is happening out here use trainings and other means, be ready and prepared to do. maybe a switch of job holds as opposed to learning a job once then staying forty years within the job profile. That is definitely needed. that it's accelerating requires more mental agility by our employment base than ever before.

Perfect Boris, it was a pleasure. Thank you very much once again. the title of the book industrial AI from pilot to profit key concept success factors use cases and market mechanics. we will share The link in the show notes.

We wish you all the best with your book annual update. That's our way to go. You have a lot to do I think And we keep off fingers crossed for you for your book. Thank you too, if I may add Peter and Robert without your podcast.

And with out the community around your podcasts just mentioning that Marco Huber did an excellent job in reviewing my draft manuscript. Without debt this book wouldn't have been possible. So thank you from very bottom of my heart to both of you. Boris, thank you very much.

Also from my side Robert and I will stay clear with another book for the next five years. it's your mark And then we'll talk again in six years. perfect Peter. thanks a lot bye-bye.