The Rise of Intelligent Machines: Insights on AI and Unstructured Data from Wolfgang Kratsch

We are back yet again with another exciting episode of the Mininger Business podcast. The show all about process mining data science and advanced business analytics, joining me as always is my friend and my colleague Jakub, how are you today? Hello, Patrick. I'm doing fantastic. Nice. We're going to step outside of our comfort zone a little bit today, don't you think? Um, a little bit. Because today we will be talking about chatbonds, generative AIs and unstructured data with both on crutch, co-founder of Kredium and professor of AI at the University of Augsburg. Are you ready to talk AI? Hell yeah. Alright, let's get into it. Welcome to yet another episode of Mininger Business and today it's going to be a very interesting discussion because, you know, all around the internet you can read about these chat gbt bots that you input something, a question or a problem and this artificial intelligence AI can actually solve it for you. And there are so many interesting things that are going on the back on, you know, behind the scenes of this technology and there is also a very interesting intersection with what is happening in the world of business processes and, you know, process mining. But before we dive into the topic, Patrick, I got a question for you. Have you ever used chat gbt and if you did, what did you use it for? Oh, interesting. You asked, yes, I have and I used it to figure out some SQL joins and just to see if it could do them and it recognized the tables I was talking about joined the tables correctly in the right order. So yeah, I was a, had a bit of an existential crisis there for a little bit. Well, I hope our job is still secure, but let's see with the new developments how it goes. I actually also have an interesting story and that was when our colleague, Marta actually posted this super cool, a very nice song about our foosball we have in the office in German it's called kicker. And I will just read the first, the first note, which is a table kicker at process and where efficiency and strategy are at hand, amidst the flow charts and data streams, a game that makes the work day dreams. That's just beautiful. Marta, I would say good job, but it wasn't you, so anyhow, this topic really resonates through the internet and everybody is trying it out and there are so many interesting inputs and also outputs that you're getting from this technology. And now it obviously raises a lot of question. What can it do and well with it and to answer these questions, but also to put it a little bit into a context of what we are talking about in our podcast, which is process mining. We have invited today's guest, Wolfgang Krasch. Wolfie, welcome to our podcast. Thank you very much. Thanks for inviting me. Hello. Wolfie, I have, you know, I'll start with a very blunt question and that is, will AI take over the world? Yeah, so that's question I hear really, really often so. I think it's hard to give really an answer to that, so I was hoping you know, of course, now maybe somehow, but I think generally not. And what about our jobs, Patrick already mentioned that, you know, he was able to create some nice joins and logic for whatever he's supposed to do at work. And now, you know, an artificial intelligence can do it for him. Yeah. I think we have to make use of AI in a smart way to support us in our jobs to maybe get our work done in maybe each other time or have more time for maybe, maybe other topics, maybe more interesting topics that needs our human creativity. So I think that's the way we should look at all these AI topics. So more meetings. I'm not sure whether that's really, really, and the creative part of the work. I hear it's very creative to always meet and discuss an airline, I love the word alignment. I don't think I've ever heard it before I started working in consulting, but here we are. Wolfie, so could you tell us a bit about yourself, because when a person takes a look at your LinkedIn page, there is actually a lot of institutions that you're currently working for, you are even a co-founder of a company. And before I even get into those topics, I just wanted to also thank Laura Marcos, who connected us in the first place, because we got to meet in Salosphere last year. So Laura, if you're listening, thank you for that. And I hope your thesis or your research is almost done because you know, you're in line next for the guests. So Wolfie, back to you. What is it that you do currently? Yeah, so maybe I can give a short summary of all that organizations and stuff that is listed on LinkedIn. So, and to be honest, since last week, there is some new organization also, and in there that's not already listed at LinkedIn, so last week I'm a professor for a blind AI at the Technical University of Applied Science at Augsburg, so that's my current job. But I also serve at Traunhofer Fitt in the branch of Maximilian Röcklinger, I guess, you also had him at the podcast, yeah. And he was also my PhD supervisor, so there's no thing that's the circle at that point. And yeah, so that's my scientific part, and I'm also really applied researcher, that's why I'm really interested to bring things to work and in organizations and not just writing some papers in an diary tower, and that's why it also co-founder the startup that is really helping to digitize processes in some specific domain, that's the building, the building domain. And yeah, I think that's somehow a wrap-up of my person. You also have then here that you're the co-founder of Coretium, can you mention a little bit what that is? Yeah, so Medium, that's what I just talked about, the building, the building analytics startup. So the idea is that yeah, there are multiple data sources out there, for example, satellite imagery, but also official data from the States. And all the data is able to describe a single building and really detailed level. And if you have that to call a digital trend of a building, then you are able to digitize all the building-related processes, really easy. So for example, if you have some building insurance or some building financing processes, there are nowadays you extract all the relevant information out of multiple tons of papers, and there's listed all the properties and all the specifics about a building. And with our digital trend of the building, you can extract it directly in a digital way. And then you are really speed up your processes. So does it mean that it's basically serves as an input into some data-driven decisioning when let's say an insurance company is deciding what should be the price of their insurance for your little flat or your house and gives you the firepower to actually defend this? Yeah, exactly, so that could be one use case. And so for example, one customer we already have, it's in the insurance domain. And typically if someone wants to get an insurance for his building, then you have to input multiple data, for example, the living space of your apartment and where is it located and all that stuff, or the building age. And yeah, you have to do that as a manual input. And with our solution, you can just provide your address of the building. And then we match all the data we get from different data sources. And we have for sure also to be processed and also use somehow AI, because it's also some kind of unstructured data. But we can directly forward that to the company and then all the data is already there. And the customer just have to input his address and then the output is there. Yeah, I bet that every data nerd that's listening to our podcast is suddenly getting excited about all these options. I have our, however, a bit different question. And that is, how do you combine work in a private sector running a company with academia? Because knowing academics, they are pretty busy. Yeah, so I think you have to identify synergies between all of different areas, that the startup, but also the scientific work. And yeah, it has to contribute together. And so that's also the startup and the spin-off we created. Directly was an output of our research projects. And then we, yeah, it was quite close together the topics. And that's how also we do now projects together. So it's found over. It's some really, really applied scientific institution. But it stops when you want to realize something at a customer. And there our startup can jump in and be the realization partner for all that stuff. And that's why you can see that some synergies you have to make use of. And then it works out at the end. Now, where in all of this fits process mining? Yeah, that's also an interesting question. So I guess also from my research direction, I would say I'm situated in the intersection of data science and process science. So because I studied information systems, so it's also not too technical, but the technique is quite important. But you have still the human factor in it and the organization factor. And that's why I started with classical business process management research. And then realized, yeah, you have also always to analyze the context of your process. And their data is quite important to describe that context. And I think Redium is for that specific domain, the building analytics domain, is one part that can bring context to also process analytics. And because at the end, everything is a process. Also at the building insurance and that building financing and other players in that area. And that's why process mining is at the end. From me, the technology to analyze all the data points I gather. So you've mentioned insurance and finance with tracking a building and all the space and all the things that come along with this specific address. Do you foresee that this technology being used in other contexts outside of finance and insurance? Or one that you might be really excited about maybe trying? So you mean our technology at Redium, you're building up? Yeah, so at the end, we are not really domain specific, so we are just trying to build up some platform for analyzing buildings in an easy way, not to have all that data pre-processing hassle. And yeah, I think there are applications or use cases in several domains. For example, we have also the public sector with all the transformation of the building stock and the sustainability topics. And there we are also quite in huge context. But the first, the go-to market was easier in the classical domain at the building sector. Now, Wolfie, you mentioned that you became a professor of AI. First of all, congratulations, if it's a new feat. That sounds pretty cool. And my next question would be actually about this intersection. Where do you see AI can play a major role when applying process mining? Yeah, so I think that's in general, it's a really interesting question about the rest of the relationship of AI and process mining. I think that you will also get several answers if you ask again, the community is in the last episode about this intelligence process mining. We love diversity. Yeah, I can't believe that's kind of a similar topic, but I guess so also the research I do at process mining is a bit more about the use case is looking in the forage. So not in past, but seeing what is in for the future, if you analyze all the data you gather about the past, and there AI is really key to make good predictions. So I think that's now the tool to use when you want to make good predictions. But also, I think if you also want to make the scope of it broader, not just process mining, but BPM in total, then I think AI is also a tool to really bring process mining to other activities in the business process management, not just coming up or discovering a process or measuring the performance of the process, but also maybe improvement of processes. And there AI can bring in new capabilities. Can you give us an example of how do you specifically mean that? Yeah, so for example, we wrote some papers about using generative AI, so same technology as ChatGVT, which is quite in discussion right now, to make use of all the event data to generate better processes. So that's the basic idea. And I think that can show the direction of where can it going to really support to come up with a better process model, and not just yeah, with some insight that there might be a bottleneck or there might be a problem, but directly come up with a suggestion for solution. So that's what we try to do with bringing more AI capabilities into process mining, process analytics, yeah. I think we will try to build up this bridge into this topic and where it actually means those parts. And I still want to get back a little bit on this ChatGVT, which we mentioned at the beginning. And I guess having you here, a professor of AI is probably the best person who can tell us and tell the audience what it actually is. And especially what I'm interested in, how do you personally see this usage in a relationship to business processes. So you mentioned already that they give your recommendations, but I'm sure it can do so much more. Yeah, absolutely. So maybe first to question what is it? Yeah, what is ChatGVT even? Yeah, so maybe if you just look at the name of this model, so ChatGVT, it stands for Generative pre-trained transformer. So Generative, I already talked about that's some new kind of AI. So not just coming up with predictions or coming up with some clustering of your existing data, but generating new data. So that's new type of AI. And pre-trained means it's really it's already usable out of the box. So you don't have to bring all your data, your annotated data that was a huge barrier of applying AI and in past. So you have now something that you directly can start with. So that's pre-trained. And transformer, that's kind of a specialized architecture of deep neural networks that allows you to allow the model to decide which data points are important and which are not at all important. So it can also neglect data points. And that's the huge power it gains. So can really understand what is important and what to focus on. So I mean chat bots have been around for a while. I mean the most basic ones respond to basic input that you are predefined and then gives you some predefined answer back. And now there has been some attempts to do some chat bots. Like I think Tay, Microsoft Tay or something was a couple of years ago on Twitter or something like that. And so it has been really a slow progress or a pretty fast progress actually. But what makes chat GBT so much different than all those that came before it? What makes it stand out? Why is this so popular all of a sudden? Yeah, I think it it combines somehow this generative AI component is pre-trained with a huge amount of data, tons of data with the conversation aspect. Because you already had these generative transform models before. So there was GBT3 for example. Last year was also quite popular but not so in a broader audience. It was just in technically audience. And now this conversational aspect that was added by the chat GBT, I think that brings it to a situation where it's quite applicable to multi-purpose use cases. So it's not a specific model build up for a specific problem. But you can just conversate with it in a natural language and solve multiple problems. And I think that combination of the really easy natural conversation with that bot and have all the power of these huge language models. I think that's something new. Being a professor and having all this knowledge about the inner workings of how this thing really works and how this operates, seeing a lot of these use cases on what people post on YouTube and Twitter and all the social media is about what all the things that they have gotten it to do. Does it at some point surprise you or is it just like yeah we knew that this was going to happen and it was just a matter of time and it's finally here or is it more that it's surprising you? I think I think the answer is twofold. So I think if you just think about that and what is what is what is based on chat GPT then it's quite clear to me that it has this power. But now I think what is new that is so the community is growing so fastly and that's why you have so so many use cases that are explored from from also non-technic people and I think that's maybe here also the point that yeah there are so there's the exploration is so stunned by so some multiple people from different angles and that's why it surprised me also at the end because I think some people are trying out totally different things than I do and I think that's maybe also the really fast progress we're doing and exploring that technology right now. Yeah well I usually ended up on I don't even know what I want to ask so that's my problem. Then Wolfgang speaking of chatbots in general I remember you know when I am using or interacting with a chatbot it's usually on some kind of FAQ or on a website with a mobile operator where I'm struggling with my turf because I didn't I didn't get charged or I got charged too much for one to cancel something and it's ridiculous because it's going through this you know this like a map where it just tries to point you towards a right end and then just connect you with the right operator who knows who you are and knows what problem you have. So I guess this is how we know the the bots from the past and possibly also from the present. Can you can you tell us about where we are currently with chatbots using being used in processes and why is even a chatbot such an interesting tool to have when executing a process. Yeah so I guess what you're describing is totally the the the situation where we are at the moment with chatbots and processes so that's the typical the typical use case if you if you have some some service call for example and the first answering your phone is some bot really just rooting you through some some predefined rules and that's still still there or there yeah and but now what what will be possible with chatgbt for example is to really have some more a child bot in there that is able to dynamically get your problem and also also just react on what you are saying so so that's that's really the new thing about that and I think that's also the the the YouTube opportunity if you take this technology in the next trend process my or bpm bots to say then it should be able to self as a personal processes system to you so if you I think that's that's somehow our vision if you maybe maybe just remind yourself at these this excellent excellent approach in former times so I think windows xp you mean clippy exactly clippy so I think the idea of clippy wasn't bad but I think the technology behind was not that good to really assist you in a in a appropriate way but now we are at at this point that we have some some technology in place that is able to react in a child way on your input and I think that shows the direction of where it's can going to that you maybe don't have to interact with so many applications and stuff in your in your process but that there is an assistant that connects maybe the different applications and reacts to your input and yeah makes your life hopefully much easier. I just recently saw a very very funny meme which actually was featuring clippy there was this scupidou meme when the guy wants to take off the mask of the of the bad guy right and he's saying let's look at who this gpt chat gpt it really is and then it just pulls it out and he sees clippy from from yeah we'll definitely post it on our on our channel so you can laugh with us as well what I wanted to ask though is if we are looking at this typical process let's say the service customer service right you have a problem with your cell phone or with the with the tariff that you are preparing so it's still a process we could probably could mine it we could visualize it you would have the start when the customer you know calls and then there will be these points either being directed through this chatbot or then you know some interaction with the customer service and then something happens and the case is closed. How would you utilize this if you if you mind this process how would you utilize these data points and this info that you're getting utilizing process mining to feed the chatbot and improve it and how would it actually look like then yeah so I think what you can do and also what we did in a paper first to to to somehow develop an approach to make all the data that is generated by a conversation with the chatbot applicable to process mining so to to to to extract all the data points to have really the overview how the process is running at the end to be able again to to discover your your process model behind and then maybe also see how the chatbot is reacting to the user input because I think that's also quite important to understand if maybe your process also is changing yeah and then you can use also that information to redesign or improve your your your process at the end or maybe also your chatbot. So I was always wondering about user input and learning from user input for chatbots because I'm not sure about you guys but whenever I speak to someone on the phone I know basically no hey I'm going to introduce myself here's my problem and then they're going to ask me some question back and all these things I kind of know how that conversation is going to go but when it's a chatbot I am not certain what the best way is to feed it the things that I need to give it so it figures out my problem like is it just keywords that I can put in am I supposed to write in sentences am I supposed to say hello there Mr. chatbot and I am blah blah blah and like follow the same conversation style that I would a normal human right so is there some like I'm sure that the space of all the conversations that can happen in a chatbot is widely different to someone speaking on the phone yeah I think that's also depending on which chatbot technology you are using at the end so if we are using advanced chatbots such as chatgbd for example it's quite natural language model and that's why you can yeah inject with a chatbot quiet in a natural way as you would do it with a human being and but they are I totally got your point so they are also chatbots that are quite rule based and just react on some specific keywords and stuff like that and yeah I think that's really depending on technology behind you do you foresee that chatbots will be like will also have a voice element to it so you can instead of like typing something and just call somebody and it's basically a simulated voice because we have AIs for voices right now for singing in a whole bunch of other things as well right it's called that very Patrick okay yeah but it doesn't sound so robotic it could almost be a human being right yeah so I think I think that's quite quite an interesting topic and I think that's also building a bridge to the other applications or use cases of generative AI to really generate for example also the deep breaks I guess that's the negative part about that but I think that shows that you are able to really emulate voices that are really really similar to the to the real ones and that's why maybe it shows also that chatbots if you add some user interface of natural language and natural voice and on top of it then it can really really interact as yeah nearly nearly same to human being and I think that can also lower the barrier to really integrate that technology to as another process actually at the end now one of the core ideas behind process mining and you know business process management in general is making your processes more efficient and interestingly I think more people who are listening to the podcast will be probably more familiar with processes such as purchase to pay accounts payable in the big four how I like to call them rather than well actually everybody probably called a mobile operator at some point however um how would you integrate a chatbot into what would be the questions or ideas if you wanted to integrate a chatbot to support your you know efficiency and support you with the correct business decisioning essentially behind every process that you implement I guess there has to be some core concept score question that you should ask at the beginning to be even able to design a support chatbot that can help you with a proper execution of I don't know purchase to pay accounts payable process yeah I think at the end you have to ask yourself that it's really reasonable to introduce a chatbot on that specific process because when the process is quite clear and quite modeled in a proper way and you don't need really that more open conversation compared to other user interfaces and then I think it makes no maybe it makes no sense to add a chatbot on that side of the process but maybe it would make sense to integrate it in a management of that process so if you if you ask yourself how can I improve maybe that that process and then you can use the chatbot and maybe maybe as input to use all your procedures and stuff that describing that process and ask the board for example ask chat GPT what are your suggestions to improve the process at the end so I think that's where the intelligent chatbot can come in also and that really standard processes to see I guess one of the biggest issues that we have or I would say anyone who wants to deploy such a solution would be unstructured data because it's not only a process data that you want to feed your chatbot with right it's going to be so much so much more and my question is probably first if you could tell us what unstructured data even is yeah so unstructured data I think there's also not really a really common definition of that so where is the where is the barrier of structure to unstructured data so I think but from my point of view the unstructured data there's not really a data scheme behind so the data is there in a really unstructured way for example image data or sound data so data data points and the data scheme make not really does not really make sense in the in the application domain so there are just some some some pixels for example some some pixel values and for example of image but you can make use of all the data because it describes maybe the context of something if you have for example a process and for example a production process in some production production environment and there you have some some camera filming all that production environment then each pixel of the camera image can somehow help you to to understand what is going on there and that's why we think that unstructured data also can help to add some context information that might be important to re-understand better process going so yeah Patrick go ahead so do you mean that as more as a metaphor or do you actually think like a specific pixel can help you figure out something in your process like if a sudden if the pixel is of all of a sudden orange that means there's a fire so my plant is burning down or yeah exactly so that's that's that's that's that's the point so I think in advance you you are not really aware of what this pixel is meaning so there's no no defined meaning in the data model so that that's the important part here that's why it's really unstructured can mean everything but if you bring it together with the structured process information then it can make sense to incorporate also that information and that's why one one stream of our research inquiry these unstructured data to really understand what is happening there right so if at any point something we we don't really know the connection there's a whole bunch of data it's seemingly unconnected but if whatever at some point this is orange that means our production goes down 5% or something maybe there maybe you can look into that but these things seem to be correlated right for example all you can also another example if you if you have go back to to maybe some some call centers or something like that then you have also that if you if you just take the recording of the phone call then you can extract the meaning with some NLP models and stuff like that some some yeah sentiment analysis or something yeah exactly and I think that's that's the interesting point that you can also extract from the from the from the voice of the person speaking maybe it's is the person angry or might it be a problem or something like that and that's also really important contextual information if you want to predict the outcome of the process yeah and that's why I think there's really really much value in the structured data and mining that unstructured data so what it seems like is almost as if we are doing the typical process mining implementation and we just read this ERP data from the system and we define those a few maybe dozens of activities we are still lacking quite a bit and there is still so much I would almost say gold hidden in those unstructured data that you're talking about that could relate but could also maybe doesn't mean anything really yeah to enhance the process and get even more data fed in into this you know decision link about it yeah exactly and that's what we what you call that there are so many blind spots in the system and you are for the moment you are mining mostly the ERP data and that's at the first glimpse it's really the most important data to to describe your process so that's totally clear but there might be also something in between so maybe there are also some activities that are totally manually or that are other systems that are not not recording some event data and there are so many data points that are that are falling behind and but if you at the end want to really have some good prediction of the process there's some some some yeah where is it going in the future and that might be really the important data points that you need then to to really to analyze and I think that's that's the point where you need more research I mean at some point I think that was kind of I mean in classical process mining we also have then the task mining stuff where we look at the and non ERP related data that is the actual clicks and strokes of the keyboard and all that stuff but that still tends to be more structured data than right yeah so I think it depends so if you if you just have some click stream data then it can be also somehow unstructured because I think there are also various ways of of of filling some tasks there are multiple multiple buttons that are doing the same and maybe you have also some yeah some other some other activities at at at a desktop that are not related to to through your task at the end but I think it's it's a similar similar approach but more in the physical world so to because it takes all takes not all all no not all takes place at your desktop at the end you have also some larger areas and production halls or some some logistic processes or something like that and there you might be interested in in model data points so it's almost if you were saying if this person goes to have a meeting for 30 minutes and then makes a coffee in between a good receipt and posting of invoice the posting of invoice gets usually delayed by two days because the person forgets about this exactly so you might be also interested in luck of the coffee machine oh GDPR I'm not so certain we can do that anyway yeah but this uh generally it feels extremely overwhelming because to enrich the process with these amounts of data and you could literally feel then that everything um first of all you have to you you have to get the data somewhere then you have to also process it and then you can also get a lot of noise into into the data how do you even look at this sure yeah so that's why maybe I think that's a point why it's not really really in practice at the moment um and still still point for for research because we we we need some concepts to to abstract the data to really um yeah bringing that data to a to the business level because we want to we want to correlate that at the end with our event data we we got from ERP systems and typically you it's on a totally different level if you if you if you take some sensor data or some some other unstructured data and that's why a lot of research is going on there to really abstract the data and also um I think as a information system researcher it's also um the the question of is it really worthwhile to integrate certain data source because it's also comes with some costs some investment and I think you always have to answer that question make it makes it really sense to integrate the data in my specific use case and I think that's that's also really important to do research there um would you maybe have an example where you could say that an external system that gives you this this perspective of unstructured data actually enriches the process in um statistically significant way yeah so I think um the processes are all the the the the main theory is it is um most uh no let me start again so I guess in the in the domains of um automotive product for example there you already see in the quality assurance that they are using a lot of camera systems to to track the pieces they are producing and do some quality assurance and they already um output data also in an in a structured way at the end and you can already make use of that if you if you combine that with your process data you get at the ERP site then you already see which instances are at the end produced in the right way and which instances might have a quality problem and I think that's um there you can get the idea how powerful it could be to to bring these types of data together um how do you technically then process it how do you uh actually insert this this data into process because we all know how process mining event look looks like it's something you have you know you have your case you have uh you know the time that you then had an event happened and then you suddenly have some camera um yeah it's uh I can't imagine how difficult it is to be to even connect this uh image to a correct item that is probably now being pushed through uh an assembler line yeah I think that's also at at the end the the correlation of of the data data data points I think that's the most um the hardest part because all the computer vision capabilities they are out there so you can you can track all the objects in really a good way really accurate way and there are powerful models out there that you can make use of maybe you have to train then somehow on on the specific environment but at the end you can get all the event data out of the video and then you have to um yeah first find your traces so where does the does the does the trace start and where does it end and what's the new instance at the end so I think that's the first hard part and then really to to um correlate that and connect that with your your pdata that's the second part and you might have some some timestamps to to connect it so that's that's one part is possible and there might be also some other informations some other sensors some RID technology for example that's also make sense to connect but um yeah I think that's quite quite um a difficult topic and it's tackled by by research still yeah I mean a lot of the times I mean if you put your luggage into like conveyor belt at the airport you know it scans the barcodes and all these things so it pretty much knows exactly or it should yeah exactly tell you where your luggage is at all time but I know that's not not always the case yeah so at the airport that's for example also a quiet um good application domain for for bringing video analytics into process mining because there is for example also the plane if if if it's standing at parking position then you can track if if the door is open and and all that stuff and then really connected to the data that is gathered by the systems in the airport at the end and that's um yeah just one one example where it can go to mm-hmm I still find it fascinating that we are actually able to tell whatever is happening just through the pictures and analyzing those pictures and videos um anyhow um what I also wanted to ask and maybe going a little bit to those uh chatbot examples and uh and uh the AI is that suddenly if let's say we move a few years ahead we have this process which is enriched by these vast amounts of data of unstructured uh um form and suddenly there is some AI giving us hints and tips on how to execute these processes to be more efficient to prevent certain mistakes to prevent bottlenecks or reworks um if you say this to a person who has uh you know no grasps of the technology that is behind it and probably even to a person who actually has a decent grasp of it and it's just afraid of it um well how do you even persuade them that uh receiving this sort of information and acting in accordance to this uh recommendation is actually a way to go um sorry can you can you repeat your question yeah it's uh it's a it's a tough one uh basically what I'm asking is that suddenly you create this um almost a black box which is of a lot of data structured and unstructured and um you know you have this chatbot and you ask um what should I do about order uh exit on that uh in order to receive the goods in time and then suddenly you get some some form of output which uh can tell you you should change the color of the box because suddenly you know this is the the sort of information that the system extracted then you are you are completely speechless because oh how does color of the box what does it yeah yeah i think um how do you even go and start trusting it and how do you even process the human being in this type of information i see yeah i guess that's also maybe the biggest problem of technologies like ChatGPT at the moment so i think we we we we we have several examples also posted linked in and somewhere that it fails also somehow and but it's really really um uh really optimistic about the answer so it really um it's self-confident and it it sounds like yeah it must be true but at the end it isn't true yeah and that's really really a problem and at the end it's a black box and that's maybe what what you would need to really integrate that in a in a productive system to provide also measures about the uncertainty of of um of the generated output to really um yeah make the the user understand um to yeah what's the what's the uncertainty of of the answer and can i trust it or is it not not trustworthy so i think that's the most important part and just but you also what would be also good to really um yeah to to unlock the black box at the end to really um provide a feeling how and why the result is generated in that way so i think that's also the possibility of of supporting there but at the moment i would i would agree that it's really difficult to to apply that technology in a productive way i mean i think that would be very interesting to see um if you ask or it gave you a suggestion like you should cancel this PO and you ask well why and you said well i saw three pixels that were blue so that's how you know right so i think um um with this uh this event of all this unstructured data and all these things where all these decisions could be coming from um i think it might raise more questions um then it would be answering um but my question was going to be um do um is there research done in in in your field about biased data and specifically for example if you if we were to look at the the call center example again right seeing how angry someone is right you have to first categorize that and everybody feels emotions differently so the categories might be dependent on the people that are listening right so exactly so that's also a quite important point to to um provide answers or informations about that the bias of the training data and that's also a problem of of of these um big black box models because we know there they they they they have been trained on huge amount of data but we do not really exactly know what data is behind and maybe um some classes or some some um i have to say some some space of domains or maybe under represented in the training data and that's why in a generated output they will be also under represented and that's a huge problem and um yeah there's a lot of research going on also in that sphere but um yeah i think that's that's also necessary if you really want to use all the technology with uh with uh um yeah um with the good papers at the end I'm assuming you can ask jgbt how biased it is and it will now yeah you can ask but i think the answer day wouldn't be so yeah so i could yeah good the AI even uh go the next step and like go crazy completely like uh behaving unpredictably uh without any reason and i know this is probably more of a question for a philosophy but uh i still had to ask yeah um i think that's yeah we we have to to ask the question and i think especially in if you have some converse conversation model if you can interact with the model um as with jgbt for example then there's always the possibility that it turns into wrong direction and i think we we read about several um points that open AI or the Microsoft they try to prevent that and it's i think with jgbt it's yeah they they did a quite good job um but yeah it can be always some some new development some some new way to to interact with that to to break also this um kind of of of of um conventions day they they did yeah and um yeah i think that's really really um tough point to to check out um now i'm wealthy uh looking into the future um war what are the topics that are getting you the most excited about what do you what are you uh what will you focus on your research and uh what is the topic that you want to uncover the most yeah so maybe maybe um concerning the the the world of chatbots um i think that's that's twofold so i think we are really interested to to have possibilities to integrate all that new newly generated data by conversely by by conversing with with the chatbots and to to integrate that in in your um information or process analytics so that's one part but second also to make use of that technology in um the classical business process management to support all the business process management life cycle all the activities you have there and i think um yeah there is also much much to do because it's quite traditional field at the moment there's process mining somehow a bit uh yeah interrupting that also interacting with the traditional management field but um in general i think there is a lot to do and that's why i'm also quite um yeah quite optimistic to to contribute also there with AI and chatbots um so if we go back to our our favorite um guy in the whole world clippy um so back in the day clippy was fairly rudimentary and the adoption was low because it was mostly a pest rather than an actual help um and now it has been taken from that it has evolved so much that it's a AI is now doing so many things that we didn't even dream of things that it could do um so do you think looking into the future in five ten years do you think we as people in business or it just generally are ready for this new advent of highly generative AI's yeah i think that's um also a bit uh kind of philosophy this question um so i guess we in in a different context we we we have we already have some some kind of these systems if you take for example theory or if you take um amazon lexar or something like that then you are already interacting with such systems in your in your um in your daily daily not daily work but more in a private field yeah and um if you for example also um see how test lies is um building up their cars that there's also a lot of um a lot of actions triggered by just a voice system so there's no no um no button at the end to to open to to open your um your friend or something like that yeah just voice control and um i think that will be also um be used more and more in the business but i think it it takes time and at the end it's always some yeah question of change management also of how to to really um interact with human beings it's just not really replacing them i think that shouldn't be the target and i think that's also not really a good way but to integrate with human beings in the best way um i think it's really interesting because the two examples that you mentioned were examples of when we tell AI to do something but like with these um like hey why how can I make this process more efficient it's AI telling us to do a specific thing so the roles are a little bit reversed yeah it's really interesting yeah that's that's true so the generative part to really to really um to really accept answers or um yeah also just output from an AI i think that's um kind of interesting how how we really accept that also in now business or i think i think we already have it somehow for example outlook some some somehow suggest me to to answer in in some short way just okay or i i read it or i think that there have been something out there and i think that's just the beginning and maybe we will see that it helps us to really get done our routine task our standard tasks and then maybe we will accept more and more and maybe also integrate it in yeah more creative work at the end or creative tasks um but yeah that will be quite interesting how it's really look like in ten years um we are almost out of the time um last question i have for you where can people find you if they want to uh chat more about chat bots with you so um for sure i'm really re-opened to connect via LinkedIn i think that's that's uh now a day is the best and the fastest way to connect and then um yeah for sure really re-opened to have some some some sessions also discussion sessions with some some some conferences i i i try and or i think we we met at cello sphere for example Jakob i think that's and and or the ICPM i think we will also be there so um that's possibilities to meet in person but um to connect just LinkedIn and reach out yeah i think next ICPM is in Rome and um well Patrick i'm not sure about you but i don't think i'll let this one slip yeah i can't let it go no no yeah i can't let it go um bothy thank you very very much for joining us in our podcast it's been a pleasure and also a lot of fun and quite an inspired insightful discussion into how chat bots really work like and uh what um uh advantages and uh pros we can um extract from uh incorporating them into our business process management so thank you for that thank you very much for having me um i also thank you dear listeners for tuning in on another episode of mining your business podcast as usually are um eager to know what you think and how you like the show so leave us a comment um interact with us on LinkedIn or just send us an email on uh mining your business podcast at gmail.com um we will be looking forward and two weeks time with um special episode with um well i will probably leave this for later so um see you or talk to you in two weeks thank you bye bye