#215 Manuel Haug: How Celonis Uses AI to Optimize Business Processes

0:00The difference to traditional consulting, you know, like we don't go in and do interviews and try to find out, you know, like by like selected conversations what's going on, but we connect to the underlying. source systems and we are reconstructing from that this is how your company actually is running this is what we can see in your systems how fast you're answering to to the support request that you're getting how well on time do you deliver your product to your customer for example like we are reconstructing that in the end from the it systems and that's how you get started right like salonus gives you that transparency on what is going on in the company and then that's typically the starting point for starting to change how a company is running you know like this could be then like strategic that you look at maybe we want to open a new sales channel and we have to design this new new process that we want to add to the company or it's very operational we can detect that you for example are paying your customers twice sometimes right like you're invoicing duplicates and we can help you to actually set measurements in place to avoid such inefficiencies in the process and then operationalize parts of these inefficiencies that we can see in how our company is running okay so uh manuel it's great to talk to you Tell us about Salonis.

1:07Yes, of course. Nice to meet you as well, Greg. So if you think about Salonis, what we are doing is we are like the leading vendor for process intelligence platforms. So what we are building is an enterprise software that helps our customers in the end to understand and optimize their business processes in a company. And that is used by a huge set of the largest companies in the world to optimize their businesses, optimize how they serve a customer, optimize how they build their products, optimize how they purchase things. So that's what we help our customers with in the end at Zeromus. Yeah. And you have a platform that you work that has various tools on it.

1:58that you use or do your customers log on to this platform and then you work together? How does that work? Typically, it's a combination of both. So what we typically do with our customers is we help our customers to get started. So when a customer starts with Solonus, we are a partner of ours. So we have a very big partner network with the large and also small consulting companies that help our customers to get started with Solonis. And then what we typically see is that a customer starts very focused. So, for example, you look at how do you purchase goods in the company and you look at a certain aspect of that procurement process.

2:42And then you start actually sourcing the data for that. You set up Solonis as a system and you start to implement a first application that's powered by Solonis. And that in the very beginning, we typically do jointly or a partner of us is doing that jointly together with the customers. And while doing that, a customer typically sets up, you know, like a small center of excellence. Maybe after you've looked at your procurement process and one use case in that procurement process, you go for the second use case. And that's then gradually done more and more by the customers themselves. So they have done a small team of experts that build these applications and these solutions to improve the process.

3:17and then they also start to gradually hand it over more and more to the business side. So that actually the ownership of such an app sits in the procurement team itself. So the end user of Zolonis as a product is then always a mix of a builder building these solutions and an end user that's actually running the operational process in an interface provided by Zolonis or through an automation that's interfacing with a user that's geared and triggered by Zolonis. So that's typically how a customer gets started and how it then slowly also is maturing over time. Yeah. And that initial discovery, is that done, you know, team to team through conversation?

4:02or is there some sort of a, do you plug into their systems and have a piece of software build kind of a digital pathway that shows where data is flowing and that sort of thing? Yeah, I think you're hitting an interesting angle because I think that's maybe the answer to what makes the bonus different to traditional consulting. And that's maybe the answer to, you know, like, what is process mining doing in general? Like, what is the core technology that we are using? So like if you think about process mining in general, what we are doing at Salonis, we look at a company as a combination of different processes.

4:39And the process is, for example, you know, it could be, you know, like you're selling something to your customer. You're invoicing your customer. You are actually responding to a service request that your customer gives you. And if you now think about how are these things done at a company, all of these processes are done in digital systems or they are supported by a digital system. And so that means when you respond to a product request from your customer, you leave a digital footprint in your systems. And Solonus is a system you can imagine now like a big vacuum. Like we connect to these systems and we reconstruct what is happening when you are answering to the product request.

5:20And that's then the difference to traditional consulting. We don't go in and do interviews and try to find out by selected conversations what's going on, but we connect to the underlying source systems. And we are reconstructing from that, this is how your company actually is running. This is what we can see in your systems, how fast you're answering to the support request that you're getting. How well on time do you deliver your product to your customer, for example. So we are reconstructing that in the end from the IT systems. And that's, to your question, how you get started. Celonis gives you that transparency on what is going on in the company.

6:01And then that's typically the starting point for starting to change how a company is running. This could be then strategic that you look at, maybe we want to open a new sales channel and we have to design this new process that we want to add to the company. Or it's very operational. we can detect that you for example are paying your customers twice sometimes right like you're invoicing duplicates very natural that you shouldn't do that right like and we can help you to actually set measurements in place to avoid such inefficiencies in the process and then operationalize parts of these these inefficiencies that we can see in our companies running yeah and And in that initial discovery, are you using, how much artificial intelligence are you using in that?

6:50Or is that changing? Are you adding more and more? I would say we are adding more and more. So like it's a combination, right? Like parts of this is very classic data analytics, connect the data, wire it up, be able to analyze it, to query it, to visualize it, to crunch the data. I think that's very classic techniques that we're using, but we're adding more and more AI. One example is understanding and adding unstructured data to these process data sets. So think about, maybe let's go back to the customer request, right? Like somebody writes a support ticket and opens that at your company. and then typically what you would put in at a portal at Amazon or like an e-commerce vendor, like you write a description, what's your problem?

7:43And then today you have that description sitting in an IT system. And AI and for example, large language models have opened up that you can now start to analyze this free text automatically. And you can actually, you can like parse it, enrich the process data with, you know, like, yeah, annotated data sets. So you can maybe classify the tickets up front already. You can maybe, you know, like add the intent, like a description of what is actually the intent of the ticket that got opened, you know, like helping the support team then to process the request faster, maybe to process it even completely automatically to some extent.

8:27And this is just one example. there are tons of these unstructured data sources in a typical process that you could also think about like if you if you purchase something at your at your supplier the descriptions of the materials that you're purchasing they are free text in the end you could think about you know like on the legal side and contracts that hold information about you like what are the terms when do you expect the deliveries of certain goods what are the penalties that you can expect when you don't deliver a good to your customers. For example, these things are often not stored in databases that you can just access that piece of information right away, but it's in documents.

9:09It's stored somewhere hidden in an unstructured data source. And AI has opened up very tremendously to add these kind of informations to the process data and helping, like in the end, AI fusing process intelligence if you want to. yeah is it fair to use the term digital twin in what you're building when you when you uh extract these business processes and and uh so that you can then look look at inefficiencies and ways to optimize them i think yes that that that is a good good good term but i think one thing where i always would like hint at is what we typically say is we are creating a digital twin of your digital processes so like we we are creating like a like a specialized version of a different why am i saying that if you just look at digital twins and very broadly they often also refer to a digital twin for example of of physical locations where you then try to model you know like a plan for example you know like the different machines in it and so on so like that's a little bit of different like direction where it's alone it's not like like deeply involved so long as it's focusing really on this, like how a company is running and building a digital twin of these business processes of a company in the end.

10:25Yeah. Um, you know, um, so, so once you've mapped this out and you're, you're then looking at ways to optimize, does Salonis have a series of, of products or solutions that you implement, uh, or is every system optimization bespoke? Again, it's a combination. So the core of Solonus is a platform where developers and citizen developers can build solutions on top of it. But in addition, we also offer like out of the box solutions for repeating problems. So, for example, if you go into, you know, like if you're a new customer and you start with Solonus and let's say you want to optimize your procurement process, Then what we offer is a pretty broad set of applications that you can install on Solonus and that are ready to be used.

11:21You then can use our builder tools to customize these solutions, to make them fit to the specifics of your company. But we offer a combination. So we have a platform where developers, our partners, but also our customers can develop solutions to certain business problems. But then we have a marketplace and an app store that gives you already applications for common problems that we find in different industries and different companies and different processes in the end. Yeah. I mean, there's so much talk right now about optimization and increased productivity as generative AI starts flowing through the economy.

12:05How much of these solutions leverage generative AI? I think the honest answer is more and more. I would be lying if I would say every single solution. But what we see is that more and more of our customers are using Celonis to enable the AI solutions that they're trying to build. Maybe what you're hinting at and what I refer to reference to what we've seen the last one two years with a couple of customers a lot is there is obviously a lot of hype and also hope for ai how can it help me to run my my company better and actually every single customer we talk to has ai budget and they build ai solutions some like more mature some like more in a in a pilot phase and in in many cases so donors can add like a very valuable and missing piece to that to those solutions and that doesn't mean that you build the ai solution completely in solonis but you can right like sometimes customers literally build an ai agent or an app that's infused by ai in solonis sometimes we also just expose the the process intelligence via apis and our customers integrate what solonis can provide into existing applications and what we then can add is in the end this process context right like in the end give the ai this this like knowledge on what's going on in the company you know like you can think about like the different statuses of you know like your your processes but then also the context right like how should it run what does good look like how do we measure performance what do we do we optimize against like this is in the end the type of information you get from solonis for the for the ai solutions yeah i i would think uh yeah and so so generative ai there's a lot of pilot programs as you say but but it hasn't really impacted enterprise in the way that a lot of people believe it will.

14:09Traditionally, when you're working with enterprises, at what point do they bring Salonis in? I mean, do they have, you know, it's kind of like in France, the Napoleonic Code, and just you keep adding, adding, adding, and pretty soon it's this giant monster. is it like that that companies have been developing and growing and they end up with this tangle of business processes and they need someone to come in and comb through it and smooth it out? At what point in their growth do companies generally bring you in? so so like i think it's a good good a good good analogy that you did you found there like i think reality is the monster exists already in most cases so if you look at the larger companies like like the pepsi or bmw or any of the larger enterprises typically the it environment and how processes are run is pretty complex so like and you have always a like a conflict of interest to some extent right like you have on the one end you have the enterprise core that you ideally just want to keep running and make it cheaper.

15:30And then you have the, on the opposite side, you have the need to innovate, to optimize the efficiency, to be able to very quickly add alternative flavors to a process, for example. And those are a little bit in contrast because do you want to actually touch your legacy enterprise IT? and rebuild it completely just to add, you know, like a 1 % uplift in your margin, but risk that you shut down your entire ERP system landscape? Most likely not. And that's very often where we then, like, what we find at customers and where Solonis can actually kind of like add this like translation layer between the two.

16:14And you can, like, this is a little bit a different flavor of what I was explaining earlier. You can also think about Solonis as this, you know, like, common language between these enterprise core and then more business focused applications so that you can almost decouple them and then you can also you can win time for example if you're undergoing like larger system transformations maybe upgrading from certain like systems to to another but decouple like this like enterprise i.t renovation from from the business applications that you want to sit on put on top so for example you are setting up automation initiatives to you know like like get more agility in how you're running your process.

16:52And Celonus then gives you this abstraction layer in between. And maybe back to your question, I said that the monster already exists. Most of our customers are larger companies. So these are companies where processes are established usually. So we have smaller customers and smaller companies using Celonus. We are using us ourselves. So we are not at the size of BMW or Pepsi yet. but the majority of our customers are the large companies, the large businesses with well-established processes where this complexity already exists and where we help them to get a step change in how they're operating.

17:28Yeah, and then once you're in a large company, as you said, you work on one discrete business process and then once you've got that optimized, it's obvious then to move to related business processes and you kind of crawl through the company, does it end up being an ongoing relationship because the business at the same time is growing and adding business processes? Yes, 100%. And maybe there are two answers. One is I think processes don't follow system boundaries, for example, or department boundaries. When you talk about a business processes, how you purchase the goods for your product influences how well you can produce it.

18:19And those two influence how well you can serve your customers. How well you actually answer your support requests for the customers will influence how much do you sell. So there is a lot of interdependencies between these different processes. So what we see with customers is a very natural evolution that while you start very focused with one process, that you actually create a connected representation of these different processes. and that you then also cross optimize, right? Like maybe I can optimize some detail on how I purchase things, but I have to actually accept it and my quality goes down in the product I can produce and things like that, right?

18:57Like these are influencing factors and that's why our customers typically go, you know, like they go broader and broader, but then they constantly have to optimize like these different aspects together. And then I think the other aspect, I think that's what we see with almost all customers that we have. processes are a living living thing right like they you you you have always influences from outside the company you have always influences from you like the purchasing behavior from your customers for example there are macroeconomic factors that are changing right that so like like this like i think if you look maybe 10 20 years back before we started there was all sometimes this tendency to say okay we just design a process now and then let it run like that Reality is a little bit different, right?

19:41Like you cannot just design the process and hope that it continues to stay the same, even if you have optimized it with Solonus. So that's why our really successful customers don't see Solonus as a one-time project, but as a continuous system that helps you to run your company in a better way over time. In today's fast-paced world, ensuring your AI systems are not only compliant, but also accurate and robust is key. How can you achieve reliable AI outcomes while managing risks effectively? If you're facing challenges with AI risk management, having a solution that ensures accuracy and robustness is invaluable.

20:30Citrus X AI offers cutting-edge technology for you to detect and mitigate vulnerabilities, biases, and errors while ensuring accuracy, robustness, and compliance. Their solution is the only one that offers observability capabilities while helping you meet regulatory requirements. Citrus X. AI ensures your AI models are not just performing well, but navigating risk safely. Elevate your AI model management with CitrusX. Visit email.citrusx.ai slash ionai to book your free demo today and learn more about how their solution can benefit you. That's email, E-M-A-I-L dot citrusx, C-I-T-R-U-S-X dot A-I slash IonAI.

21:33IonAI, all run together, E-Y-E-O-N-A-I, to book your free demo today. Yep. You know, I talk to a lot of people. I mean, this is an AI podcast, so I talk to a lot of people about the effects of generative AI on productivity. And there are these wonderful projections of, you know, 20 % productivity gains across the board. And now people are talking about agentic workflows, about building AI agents that will handle discrete processes within the business workflow. and then you chain them together so they're talking to each other and really create kind of a base administration layer in a company that relieves a lot of that work from humans and increases productivity as a result.

22:53I would think that Salonis is, you know, working really hard to get on top of that, to be an expert on that. I mean, it's so new, but how do you feel about the promise of agents, AI agents, AI ensembles of agents, you know, in doing handling business processes? Yeah, I think it's maybe like, I mean, you're right, right? Like, this is one of the aspects where we work very closely with our customers and also like, like actually bring it into reality, right? Like from like the vision to how can we really run part of the process with AI actions and kind of go into detail in one or two examples on that.

23:41But in general, I think it's on the one hand side, I think in the short term, I think it's almost like it's overhyped, like that you can kind of like immediately get all of these efficiency gains like tomorrow or the next month i think that's not realistic like like on the one hand side i think there's like it's a combination right like the technology is not ready for that if you're really honest but also the companies are not ready for that like you like just plugging in such a system into a company acquire requires a lot of change right like and and these companies have existing like ways of working and and you will very gradually actually go through the maturity curve from where you at the beginning have the existing teams and the existing process to function better then you maybe can replace and augment some of these processes and then you're going into okay now we're going into the direction of what you can call maybe this autonomous enterprise and that then you like the company or parts of the company start to autonomously like work and interact with each other and I think in that aspect it's almost underestimated right like there you could make the claim that maybe 20 is a little bit low on how much efficiency gains you can really gain if such a vision becomes true.

24:54Yeah, what kind of efficiency gains, I don't know if that's the right way to frame the metric, but do you see generally when you are working on a business process or a series of business processes in a company is there is there sort of a ballpark number that yeah we're going to easily see 10 gain or uh i mean i it would be hard for me to say like a concrete percentage number across the board for every every customer it's like like you know like the processes are very bespoke i think like i can give you a very concrete example though right like this is like an example from a you know like from the automotive industry where a major fan manufacturer was experiencing interruptions in in the engine production line and and the reason for that is it's called failed material calls and what that means is the material you to build your engine is not available at the production line when it's needed and using solonis if going back to what we talked about at the beginning we can gift a company end-to-end visibility on this process.

26:08And that made it very easy and just transparent, for example, to identify where are these problems coming from. You know, like that could be that you have material shortages. It could also be that you have wrong master data in your systems that cause these, you know, like wrong material assignments. It could also be that you have scheduling conflicts. Maybe you have the material, but it's just not at the right point in time. or you have other reasons that are causing these bottlenecks. And this example, this customer has actually realized 30 % reductions in these failed material calls, which is massive.

26:49If you think about it, the worst case scenario is that a production line is just standing still and not producing the car or the engine. And these are the kinds of efficiency gains we can see from many of our customers. Yeah. Have you guys experimented at all yet with AI agents? Maybe not directly with customers, but in your development? No, we have. We also have already a couple of them live with customers even. And it's like maybe to what I said at the beginning, it's like maturity. like that, like they are becoming more and more complex. And that's what we also see with the customers, right?

27:33Like at the beginning, one example we have live with the customer is an agent that's improving the cash conversion cycle and the productivity where like usually humans review order blocks, like credit blocks on an order. What happens there typically is, for example, you have maybe a note in your system still that there was something wrong with the license agreement with the customer. So you are blocking the order. So what usually would happen is that a human goes in, sees that there is a credit block, tries to investigate where is it coming from. So there's typically a type attached to it, then checks on the comments or any other information in the system.

28:13Okay, here was a comment from the licensing team why they had that block. Then you phone call your colleagues and ask them, is this still valid? I see you have said that actually five months ago. so like is there any any progress on on this on on this issue and then you get an answer yes no maybe and uh based on that on that on that response you can either release the block so you can actually continue with the process or you actually validate that you still don't pay the customer for example the supplier for example and we have set up an agent for this you know like investigation so the like you know like this the fact that you have this credit block at the point in time when you actually should be paying your customer that's the process inside that we can make very easily calculate and and and make transparent with our technology and so that's the point in time when we can hand over a work item to an ai agent and then we can also from the from the process context that we have in our data like how are typically these issues solved we can create instructions and prompts for the for the for the agent and so that agent is then actually generating an email it's reaching out to the look to a shared inbox of the pricing team in the example I just was making.

29:24It's actually like waiting for a response from the pricing team. It's processing these responses, and then it's annotating the order block. And it's recommending to the team that's responsible for the payments, can you remove it or can you not remove it? And they then have like a prioritized inbox, so to say, that they can work through. And the next step, very obviously, is, okay, now let's automate that end-to-end, right? Like, why have a human in the loop? But that's then the part of the maturity curve that you, like, from an agent that's helping and taking away some of the manual work they have to do to fully automating and fully running a part of your process.

30:02And that's one example, right? Like, we have been working pretty, pretty closely with many of our customers in the last year to build similar use cases in Salonis. And that's, yeah, that's fascinating. Is that agent built in-house by Salonis? do you adapt an off-the-shelf agent because there are now marketplaces of agents? So in general, it's a combination. So this concrete example is built in Celonis. So in Celonis, we offer as part of our portfolio for improving processes. We offer capabilities to build agents yourself, but we also integrate with the big vendors out there. So we have customers where we integrate with Microsoft Copilot Studio, for example, is one one one of the big ones we have um ibm what's next is another one amazon bedrock like the typically big players you you can integrate with telonis and it's then more it's more like a customer specific um setup right some some customers just have like certain tools already established in their in their enterprise i.t and then we usually integrate with those to actually just facilitate building these agents but many of our customers don't have an established tool yet So we also offer parts of this as the Solonus portfolio to build it in Solonus itself.

31:21Yeah, I mean, it's an exciting time in this world. How do you feel about the development of, I mean, as you said, things get overhyped. But on the other hand, you're at the sharp end of the spear. you're seeing it integrated into enterprises. Talk about how you see agentic workflows and generative AI working into business processes. Do you think it's something that's going to take 10 years? Do you think it's something that's going to, two years from now, there are going to be networks of agents handling a lot of the grunt work in these organizations. I think maybe as a disclaimer before that, I think the honest answer is there are some things we know and many things we don't know at the moment.

32:29Also, if you just look how far things are moving, even the jumps that the foundational models are making, I think it's literally very hard to predict exactly when this will arrive, especially as an enterprise cosmos. Enterprises are traditionally a little behind to what you can see on the consumer side. Also, very understandably, because it can be a lot more critical if you run your way, how you produce certain pharmaceuticals based on agentic web flows. compared to I just asked Chpt about drafting a letter to my uncle for me so like very different severity levels so like just that as a disclaimer but I think I think it will like it won't take 10 years until we see that but I think there are also levels to that so I think there will be the need to you know like to to be very specific what do you hand over to an AI how much control do you need to put in place and how do I actually govern it what it's doing so like I don't believe that there is one fits all solution to all of the business processes and how companies are running in general so maybe some of your processes you can run very autonomously and the way how the problem is solved is generated on the fly and I think you can see first elements of that working in reality now but i think there's also the other side of the spectrum maybe going back to you know like pharmaceuticals and so on where you want to have very tight control on what's actually going on where then ai is more you know like is maybe embedded in more classical workflows where you have tight control over uh control flows what's allowed what's not allowed what are the guarantees i want to give and i um and i i i where i'm more flexible for for adjustments and i think this is like a hybrid hybrid word for like an extended period of time that you will see there yeah yeah uh in your work um the uh can you talk a little bit about i was looking at some of your material process intelligence graph what what is that concept is that the visualization or the initial discovery that you were talking about when you plug into a system uh and yeah yeah this is this is in the end the the core data set that we work with so maybe i give you a little bit of context why like what where is this coming from like process mining what we are doing like this transparency and calculating the react the ds is process how it's running in a company at the beginning and also what the loners did a couple of years ago is you had traditionally looked at one single process so you actually look for example at you know like the maybe the delivery to your customer so you picked the deliveries and you actually looked at that single process how do you how do we deliver goods to to the customer for example and then when you wanted to analyze in addition to that okay how are we how are we producing our products you pick the product the production items and you looked at the process for the production items so you could then combine that by looking at them side by side but the process mining in its core was looking at individual processes um with the process intelligence graph what we what we started to do is combining these so combining the all of the processes in one data set so and and what we ultimately do is we look at the digital digital documents like the objects in a company so that's maybe the invoice that's maybe the production light item it might be a customer service request ticket um those are the digital documents that we find in the source systems.

36:25And then the second point, data point we get, we reconstruct from the systems, it's what's happening to these documents, what's happening to these objects. So you open an order, you cancel an order, you change a price on an order, you book an invoice, you send out a delivery. These are events that this is what's going on. And then we, like one event has a connection to multiple objects. So, for example, when you actually think about Amazon, right, like your basket and you click on checkout. When you click checkout, you have like an invoice that's created and sent to you as a customer. At the same time, in the background, the packaging is starting, right?

37:05Like the Amazon employees, they start to put stuff into a package and then like hand it over to delivery and so on, right? So this one event is touching multiple things in the background. And that creates like a graph. So that creates a graph of things that are happening that are touching multiple things in a company. And that allows us to stitch together the different processes in the end. So we can then see how does my purchasing influence my production? How does my production influence my customer satisfaction and how I deliver to my customers and so on. And that's what you asked earlier. This is what we call the digital twin of the processes of a company.

37:44Yeah. and then like maybe one more like this is the process data and then we add one more aspect to the process intelligence graph which is how should actually things run and we call that business context so we also put into the system how do we measure performance that can be as simple as a kpi but we also put into the system target process behavior so like maybe certain flows that that I expected and how the process got designed or statement of works like descriptions of how a certain step should be executed by the teams or by the systems. So that's what what is called the process intelligence graph.