Simulation with Process Mining & the Story of Apromore with Marcello La Rosa, CEO and Co-Founder of Apromore

- Mining your business podcast is back in the new year. It's such a festive time. Jakob, happy new year. - Thank you, Patrick. - This is, of course, a show all about process mining, data science, and advanced business analytics. Marcelo Lavrassa is joining us on the podcast to tell us what a digital twin is and what it does in your process. Amongst many other exciting topics. Marcelo Lavrassa is the CEO and co-founder of Applemore and a professor at the University of Melbourne. Let's get into it. (upbeat music) Hello there, process mining enthusiasts. Are you mining your business also in 2023? Because we certainly are. And I hope that you are as well. Because, you know, we are showing no signs of stopping and we are very excited for the next year of our show. Actually, this is a third attempt at this recording because work and illness got in the way before and I'm just so happy that we finally, we're able to bring our today's guest, Marcelo Lavrassa on the show. So Marcelo, welcome. Jakob. - Hi, everyone. - Marcelo, you are Italian living in Australia. Are you right now actually in Italy or are you in Australia? - I'm actually in Sicily, yes. - Ah, very, very nice. What brought you to Australia? - Oh, look, it's a long story. When I completed my master's thesis back at the university in Turin, I worked on process automation. Back then we were working on a language called B-Pel. It was a standard. And I came up with the idea of working with my supervisor on a paper to present that work from thesis. So we went to Nancy in France to attend, I believe, the third edition of the International Conference on Business Process Management. And there I met who would become my PhD supervisor, I guy, you might have heard of Marlum Duma. And Marlum was actually working on exactly the very same topic. So we decided then to come up with a PhD proposal. He pulled out out of the blue a scholarship, I applied, I got the scholarship at the month after I was in Brisbane working with him on my PhD. So yeah, that's what brought me in Australia. Then when I think, funnily enough during my PhD, Marlum left Australia to move to Estonia, but he kept supervising me from the distance. Then I completed my PhD. And then we kept working together since then. And that ultimately led to a promoted. - Marlum, so you've been living in Australia basically ever since, if I'm correct. And you, what I found super interesting that you have this background in academia, but now you are transitioning into an actually business world. And do you actually, as of today, identify yourself more of a CEO of a business company or as a professor at the university? - So it depends on who's going to listen to the podcast. - Everyone. - Look, I must admit that transition hasn't been simple. When I decided together with Marlum and other folks to spin out a promoted from the University of Melbourne, because in the meantime, I moved from Brisbane down to Marlum, it wasn't easy because bot Marlum and I had a lot of academic roles. For example, back then, I was deputy head of school. I was leader of the information systems group. I was driving the professional education arm on BPM. And I had a number of research projects. I was the lead CI for. So this transition has been incremental. At this stage now, I'm part time with the University of Melbourne. And I've relinquished all these leadership roles. So put it this way, I'm just a simple professor of process mining, which I love, by the way, I definitely love continuing teaching and research. But now, most of my time every day is spent working on the company. We are now in a very critical phase of a promoted, it's not anymore as a start-up. We are over 50 people, more than 70, if you include contractors. So that requires my full attention. - We will get to up for more in one second. But first, I want to just say that there's nothing as a simple professor of process mining, because this is still, well, I didn't even know that there were even universities or track that we're teaching it. And I just would like to stay a little bit longer on this piece of universities. And could you tell to our listeners how there's a track at university that focuses on both process mining, but let's say, probably more on the high level on business process management look like? - So first of all, I must admit that I've been very lucky because the university gave me a lot of space to focus on process mining, even more specifically than BPM, my position at the university in Brisbane was more broadly on BPM, but Melbourne Uni wanted to develop specific capabilities on process mining. So how does it look like a day in the life of a professor in BPM or process mining? So there are typically three components. There is teaching. So for example, I've been teaching a subject on business process management. Based on our book, the Fundamental So BPM, little promotion there. So we just walk through the various stages of the BPM life cycle from identification through to process monitoring. Then we have also a subject on process analytics, which is very focused on process mining and the link between mining and simulation, which we are gonna talk about later. And so that's basically teaching. Then there is research, right? So you supervise PhD students, you come up with fancy ideas, and then you basically seek scholarships to support the studies of the students. For example, at the moment, I have a student with whom we are working on using causal machine learning for prescriptive monitoring, right? Another one where we are exploring more from an organizational perspective, the uptake and challenges in using process mining for process compliance checking in a financial services sector. So you've got the PhD supervision, and that is also related to seeking funds, right? So applying for grants. For example, we recently got a large scale industry innovation hub, and one component of that project, which is in the area of bioengineering processes, is to use process mining on this totally new green field space. So we're process mining has never been applied, like trying to mine a bioengineering process and identify opportunities for improvement, like how you can improve yield. And quality of the result. Life, for example, when synthesizing a vaccine. Very, very exciting, very, very exciting. And then there is engagement. Engagement means talking to industry. For example, before I mentioned this initiative related to professional training. So professional training, consultancy services, they all fall under the realm of industry engagement. These are the three key areas for an academic and in Australia, all three have an equal importance. And that's very interesting for an academic like me, who's very focused on application oriented research, ability to leverage industry contacts, to study real life problems, and then come up with ideas to solve these problems. - Now can I ask as a student at the University of Melbourne, if I wanted to learn about process mining, would this be more focused on the business process management part, so all the business processes and what you can do with process mining, or will I be learning about the causal machine learning models for prescriptive process mining, right? Those are fairly two different ends of the process mining spectrum. So what does that look like? - You can do both. So for example, we have a master's of information systems, where we have a pathway in business process management. You'll start by learning the fundamentals of VPM. There is also another subject, business analysis, digital business analysis, where you learn how to put VPM in context, right? Because VPM is just one piece of the puzzle for the overall organizational improvement, it's not the only methodology, right? And then you can delve more into the technicalities of the various approaches and techniques that are available in the various stages of the VPM life cycle. And one of these subjects is the one I was mentioning before, which is the process analytics one. It's an elective subject. So the other one, the VPM is a compulsory subject in the lower core of the master of information systems. So every single student out of the 300 students, we get a pair semester intake will go through that subject but the process analytics one is optional. And only those who wanna do it then have a chance to do that. And then you can do a project, like an industry based project on VPM or even more specifically on process mining. Did you say you had 300 students doing that? As a master, three, three, 50s sometimes, yes. - Oh, wow. How long have you been offering these courses? - Since I arrived in Melbourne, so 2018. - Oh, that's right, yes. And we saw that we were offering these subjects at QT and my colleagues, they're still offering this type of subjects. - So here you go, if you wanna go study process mining, well, Melbourne is one of the places that you can probably do that apart from a few European schools. And it's very exciting to hear that the business process management and process mining itself, they have their own track that you can study now. Now moving on a little, you said something interesting and that you really enjoyed this application part of research, where you are applying the research and ideas and the latest developments on actual or solving actual problems. And I guess that was also a centerpiece of founding your company. A promo, is that correct? - Correct, yes, it all started from our own research. A promo, it was a research project that we started about a decade ago. Back in 2009, it's actually now more than 10 years. And we attracted some funds, some industry sponsor funds, and we started working on what was to be an advanced process model repository. A promo here is an acronym that stands for advanced process model repository. - Ah, ah. - And you know who came up with this weird name? - Marlon, probably more. - No, it was a project funded by the Dutch government. So the chief investigator was Ville van der Alst. - Ah, of course, who else? - Yes, well, indeed. Invited me to spend some time in Aindhoven back then it was working there. As a postdoc, and so kind of I was working, you know, between the two countries, because I still had my position in Australia. I was a senior lecturer in Australia. But then I also started this sort of postdoc from the distance. Now it would be totally normal, but back in 2009 was kind of an exception in Aindhoven. And we started working on this idea and I was super excited to come up with a product. I mean, having a computer engineer in background, I love building stuff, right? So the idea was super fancy, but I wanted to build the product. And mindful of the success of PROM, which back then was, you know, the very first process mining tool. I wanted to follow suit and do something very similar with a PROMore. And then when I went back to Australia, to Brisbane, I started looking for fans. And we got this linkage project, which is a project that is sponsored by the Australian government and industry. Back then it was a large insurance institute. And we started building this product. Now over time, about 2012, 2013, our research focus pivoted from process modeling more specifically to process mining. And then the tool follows suit. So this prototype that we were building started hosting capabilities more specific to process mining besides the existing capabilities around modeling. And then in 2019, after a strong encouragement from the University of Melbourne, so back then had moved to Melbourne and relocated the whole APROMore initiative with my research group. You know, we were 13 folks, I believe. We moved from Brisbane to Melbourne and we brought a PROMore along. And then we started thinking about commercializing the software because it was strong interest from industry. We were using APROMore to deliver consultancy projects in process mining. And organizations were using the software looking at us using the software and say, "Hey, can we also use it?" Our selfie looks pretty simple, pretty intuitive. And he said, "Well, you know, it's not commercially available, but it's open source, you can go down loaded." Of course, you know, that it doesn't work like that. So we had a few discussions with my head of school, my dean, they said, "Why don't you commercialize it?" You know, we will provide strong support in terms of seed funding, I mean, concretely. And the university really gave us a strong push to start commercializing the software. So fast forward, three and a half years from there, we've done the seed funding and we've done a serious say and more recently, a serious B round, which we close in November. Last year. - I saw this big splash actually around the serious B where you joined forces with Salesforce and GBTech, raised actually over a little over 10 million euros dollars. So first of all, congratulations as a CEO of a company that must be quite in achievement and a moment of pride because that's, you know, it not only allows you to further develop the product, but also, you know, makes you sure, supports you in this idea that what you're doing is probably really worth it and really useful. So once again, congratulations. What does it mean for the company going forward? Look, we are very proud of this result. GBTech came in already in the series A, but the series B welcomed Salesforce as a key investor. It's definitely an acknowledgement that we are on the right track by all means, especially because this is coming from a software giant. Yeah. And it's an organization whose customers have been demanding about process mining. You know, they might not call it process mining. They might call it improving service efficiency, improving service quality, right? Improving customer experience, monitoring risk and compliance, accelerating digital transformation. But at the end of the day, we know that we can answer these questions through data. And data is very rich in Salesforce. So that provides an enormous opportunity for us to support our joint customers by using process mining on top of Salesforce. In addition, there is a link with MuleSoft. In fact, it all started through a tech partnership with MuleSoft. We've been developing a connector. You know, MuleSoft is an integration middleware or is started as an integration middleware. Now their offer is more broad. But we can connect to systems within an organization by exploding, by leveraging this integration middleware infrastructure that the customer already has. Really a more connected for MuleSoft. And then we started exploring the link with RPA because MuleSoft now offers an RPA capability. So we can push, we can talk about this a bit more in that later. But we can identify automation opportunities and push them to their RPA manager. So you can record these opportunities directly there and start assessing which ones you want to implement with bots first. So it's a whole ecosystem of integration. It's not just process mining on top of Salesforce, which per se would be huge opportunity. But it goes beyond that, it goes into integration, it goes into RPA, there are a number of opportunities to work together that we started exploring. And what we notice is that customers are responding very well to those signals. They love this idea. We call it round three process mining. Round three process mining. Collect data, analyze and push out actions such as opportunities for automation that then are enacted by means of RPA bots. - Well, that's an idea for a name of the episode. - Yeah, perfect. So looking back at the journey of Appermur ever since it's inception to what you wanted to build and from that all the way to now raising over 10 million you as dollars with some major players and building your own ecosystem, what recommendation would you give to other people in similar shoes to what you were back then to how you get to where you are now? - The journey can be very bumpy and I think we have been experiencing the whole lot because we build the company through COVID. So as soon as we started, we had to switch to remote working. When COVID hit, we were only five. And we were in the middle of raising our serious say, you can imagine the impact COVID had on closing our serious say. Then we had the economic crisis, we are now experiencing, which of course is an easy on anyone. So we can say that throughout this journey we have been collecting a lot of experience. I think that perhaps what made the difference for us is to never lose trust on ourselves and on our ideas. We know we have a strong product, we know we have a strong strategy, we insist until we get there. And this is pretty much my key recommendation. Then I can tell you do this, but not that, but in your situation you may actually work the other way around. I mean, I don't have any statistical significance out of my experience. It's only my experience, but I can tell you that perhaps what really made the difference was to persist, being resilient and persist against really all odds. So when you say that we never lose trust, who is we in a way? It's Marlon, my co-founder Simon, myself, the whole executive team at Appromore, we have a strong corporate culture. People love working for Appromore. People took pay cuts to come and work for us, being a small company, of course. We are somehow limited in what we can offer, but people believe in the story, see that this is fresh, this is new, it's different from let me call it mainstream process mining. It's very different. And especially what really excites us is how customers respond to those signals. Now, Marcelo speaking of mainstream process mining, how do you set yourself apart from that, from this consensus of how big giants on the market are doing process mining? And what is your, let's say, main unique selling proposition compared to the others? So first of all, my deepest respect for the giants, they came first, they did a lot of things right, and they got where they are. So my deepest respect for where they are, I admire them, I admire their story, and I certainly have to learn from that story. But the market is getting saturated, right? And we know, we know. And coming from an academic background, somehow I have a very good understanding of where we can get with process mining, which comes back to the initial point of where I promote a came from. It came from our own research. It came from over 10 years of research and innovation and leading universities. Melbourne, Brisbane, Tartu, kind of, and look, I'm not cheating. I'm not kidding. When I say that we probably only achieved 20% to 30% of the full potential of process mining. What is coming out of research is like 10 years ahead of what we are trialling in industry. And that is one key difference in our product strategy. We are spearheading innovation in process mining. So the product is the tradeoff between our desire to advance the state of the art in process mining directly informed from our research and immediate customer demand. I need this feature, I need that feature to move on with the project, right? So that is a distinguishing factor in a promote. We might not implement all, we might not support all 250 features that you might see in the market. We might support 100, 150, but in our advice, these are the critical ones you need to deliver value. And if we decide to implement a feature, we do it to its full-ext extent. We mentioned simulation before. We're going to hopefully touch back on this. That is an example of where we go really in depth. Another one is automated discovery. So the key point that I want to make is state of the art capabilities highly sophisticated, but at the same time, very simple to use. So we are spending quite a lot of time to try and break down a capability that on phase value might really be complex, might look complex like simulation or like predictive monitoring. What would be the easiest way to realize predictive monitoring capabilities in your software through a machine learning environment where your data science can write predictive, can build predictive models, e.g. in Python or in a Python dialect, and then deploy them. Well, that's not the way we're following. The target audience for process mining is not technical dudes. It's business analyst. It's process managers. It's a claims manager. It's a loan operator. It's an IT service desk, Clark. So this is citizen process mining. Now, breaking down a complex capability like predictive monitoring or simulation in a way that is easily consumable by a non-technical audience is a key challenge we've been dealing with at a performance. So the second distinguishing factor of our product besides having state-of-the-art sophisticated features is the way we expose these features in a simple, intuitive way. Through a sleek, easy-to-use interface. I'm not kidding when I say that you can master 80% of our product on a 15-hour training course. 15, 15, 1, 5, 1 training course, not 20, all right? Now, what does that mean concretely for a customer? That the time to value is significantly shorted. We are not talking about months, if not years, we are talking about weeks. We have customers that have set up an enterprise-wide initiative in process mining with a center of excellence in four to six weeks. That means you can start delivering value already in the first two, three months. Pay back to the business and leverage that to get consensus so you can scale your initiative, right? So that's another important point. And the third one that I wanted to make is the focus. Coming back to mainstream process mining. You know, process mining has been incredibly successful in P2P and not to see, you know, prefer to pay, order to cash, accounts payable, accounts receivable, that space, which is the SAP space, right? Is the core manufacturing space. We started looking elsewhere. You know, that's a quite crowded space. But at the same time, we believe that, again, coming back to trying to exploit the full potential of process mining, that process mining can deliver more value when you hit an organization right at its core. And P2P is not a core process. P2P is a support process. We need P2P for the business to run, but we don't make money out of P2P. We don't place our customers out of P2P. So I'm talking about service-intensive organizations and high-touch customer-facing processes. I'm talking about claims handling in an insurance institute. I'm talking about loan application in a bank. I'm talking about call center or customer onboarding in an organization. I'm talking about meter to cash in a utility company. So these are the processes we've been specializing on. So while our product, like process mining, is industry agnostic, and we had the luxury of working with a number of customers from different organizations, engineering, manufacturing, logistics, telcos, financial services, tertiary education, government, and so on. We have built capabilities in very specific verticals. So the whole BFSI banking financial services and insurance of which I mentioned a couple of examples, that's a vertical we have been developing deep expertise. Like you talk about loan application, we know exactly where to look into when it comes to loan application. I know what problems customer can face in this application to prove it. - Now, could you get specific about especially this banking process mining use cases because I don't think we've ever mentioned those, and I'll be honest, I haven't worked on a banking use case yet. How do you utilize process mining in such a process and what does this process really look like? What can I imagine when you say, okay, let's take a look at the banking process or claims process? So first of all, let me say that this comes with a risk because something is that you work on procure to pay. And if things go wrong, okay, fine. The business is not gonna crack. Something else is that you work with a bank on their loan origination process. That's how they make money. That's how they serve customers. That's heating straight into their top line, straight. Talk to an insurance company, claims handling, cost center heating straight on their bottom line. High sense of urgency, top priority for the bank or the insurance company for these two processes to work spot on. Now, let me take an example, loan application. So loan application and specifically, one where we've been working a lot is mortgage lending. So home loan application, right? So this is part of consumer finance. And we look at the whole value chain as you know, process mining provides the best benefits if you're able to discover and analyze the end-to-end process. So it starts with origination of the loan, through brokers going out and seeking new loan applications like new customers. Then it goes into lodgement, triage and verification, assessment, negotiation and settlement. These are typically the key stages of this application to approval process. Now, each stage, typically different area of focus for the organization, meaning different manager, meaning different KPIs. Take, for example, the whole lodgement and triage. There is a key KPI there, which is called time to write. - Time to write history. - Never heard of it. - Can you tell us what lodgement and triage even means in the first place? 'Cause I think this is gonna be-- - Yeah, so triage means once you get an application, like a loan application, you need to channel it through the right verification. - Right. - You know, and triage may be based on the amount of the loan application, on the type of property, on the type of loan, on the demographics of the customer. But also when you triage, you do an initial verification to make sure all the key elements are in place. For example, is there a declaration? And is there a declaration signed? Are all the details of the property that we wanna buy, that the customer wants to buy there, and all of that, because otherwise you push it back. - Got it. - So time to write or time to first, is a KPI that basically measures, and it is in fact a key SLA for a loan application process in banks, is the time that it takes from the lodgement of the loan application to the dispatchment of the first loan offer. You know, you might provide multiple loan offers, you know, because then you enter into some sort of back and forth with the customer. But what does it take? How long does it take to respond with the first offer? That's critical. And why is it critical? Because that's the first place in this process where customers may drop off. You don't provide me an offer within 10 business days. I'll be shopping around with your competitors. - Right. - So time to write is essential, is an essential KPI in a loan application process. Now in a promoter, we can measure SLA adherence at different levels, at the level of individual activity, how long does it take to assess a loan application, at the level of the end-to-end process, how long does it take until you settle the loan from the time the application was lodged, but also between milestones, which is the case for time to write, between the lodgement and the first instance of this activity provide offer, this batch offer. Now you don't wanna do that by writing a script. So I'm going back to this point of being simple. Totally no code. No code means there is really no way you can try and sneak code into a promoter. - But no code means no code. - I'll stop you here. Like there has to be something in the bag that has been either coded or programmed. And if you are ingesting the data that you might not be familiar with, you still have to process them and create something on top of it to run this now. Or this is how I imagine or always these type of things. - So first of all, to answer the first parts of your questions, there are hundreds of queries that we run in the background. The challenge is how you can expose these queries through icons, buttons, and simple user flows. That's where we believe the tool is strong. The second is the part about data preparation and transformation. This is also no code. It's also through a graphical interface. You can import in a promoter two or more event logs, for example, in CSV or parquet formats. I find they don't need to be event logs. They just need to be data, data files, right? Not yet with the key ingredients of an event log, right? Like, you know, activity timestamps is identified. Then you can join them or do a vertical union. Again, we'll do this with drag and drop. Then you move to the transformation phase where you can start applying operations like you can concatenate to attributes. I don't know, date and time to come up with timestamps, right? Or you may, for example, one thing that we do in loan application is to calculate the loan to value ratio, which is the ratio between the amount of the loan and the amount or the value of the property you want to buy, right? And then we calculate the mortgage lender insurance. So basically in Australia, for example, you need to pay an insurance if the loan to value ratio is more than 80%. So basically, you need to provide a deposit that is at least 20% of the total value. Otherwise, you're going to pay this mortgage lender insurance. Now, these are two attributes that we can add in this data transformation phase. And you can define the rules in a totally no code style. You have the list of operators. You have a building block for the rule. And then you just drag and drop, select and build the rule. Now, am I going to fulfill 100% of my transformation needs? No, of course. Because I mean, it's not as expressive as just writing code. That's clear. But the question, again, is who is your target audience? Because if I want to do an incredibly complex transformation-- and by the way, this day, we can concatenate these operators. We can concatenate TTL pipelines. We can actually build some quite complex stuff, still in a no code environment, but there will always be that operation that is not yet supported. Well, in that case, this should be better off done by your data engineers. And if you're talking to a company, they probably already have a data engineering practice. They have pipelines in place. So that's the infrastructure they're going to use. Take, for example, Milso, right? We go to a company already as Milso, as integration middleware, they're going to use that. But now, talk to the analyst. Talk to the analyst needs to put together four or five files to start to analyze their loan application process. They grab data from our from an engineering system, from a couple of custom in-house systems, et cetera. They can do that in a premore in a totally no code way. They can even automate this pipeline and do delta ingestions periodically. So there has been a delta, historically, a delta quite large between doing data transformation through scripting and code and doing that in a no code UI based approach. Now, we have been shortening this delta over time, right? I'm not saying that this delta will disappear. That's a dream. There will always be a delta. But there's already quite a lot of operations you can do in a no code environment. Now, when I was in ICPM conference in Bolzano, a couple of months back, where you also actually participated. I was doing or supporting one of Marlon Dumas students in doing, let's say, a research paper or something on how should the user interface for a prescriptive process mining look like? And there was exactly this use case for the loans, where the end user, the business user, who approves the loans, gets this information that is somehow processed in your tool. And he just gets this very basic recommendation on what the he or she should do based on these data ingestion. Is this a part of your product? Like, are you imagining it in a way that at the end of the day, you'll just have this platform where the clerk or the risk manager, or whoever approves the loans, just sees the recommendation, sees the probable path of this specific loan and how it would evolve over time. And he or she just approves or rejects. And then the process goes on. Yes. So that's part of our product strategy. So in the last couple of years, we've addressed the foundations of process mining, what we call descriptive process mining. You know, automated discovery, performance analysis, conformance checking, variant analysis. Then we moved to one layer above what we call predictive process monitoring. Marlon and I actually started this stream of research within process mining about seven, eight years ago. We had two PhD students working on predictive process monitoring, both of them, by the way, got an award. So that type of work spiked a lot of interest in the community. And then we started looking at what's going to come next. You know, once we equip users with predictive analytics, what can we do with those analytics? And of course, the natural answer is recommendations. Recommendations that are provided through prescriptive analytics. And we take actually two sides to this approach. There is, let's say, operational support side, which is probably the use case we think of most commonly. You know, I'm running my own applications. I'm assessing my claims. How can I make sure that I minimize the SLA violations? For example, I have a 30-day SLA violation for handling a motor vehicle claim. What actions can I do to make sure that I reduce the overall number of SLA violations at 30 days? You know, one action could be to relocate some FTEs, some claims handlers from claims that are pretty safe, not any chance to violate the SLA's anytime soon. To those claims, there are a high likelihood of violating this 30-day SLA, right? So that's a concrete example of how this could be applied. In practice. Now, what we have been-- and then there is another area, which is more tactical. It's not operational support. It's not working on top of open cases, like cases that are unfolding, like claims I'm working on. It's actually looking at historical data. Why identify interventions that can be performed in the mid to long term, not just in real time or quasi-real time. And we call that automated process improvement, in a promo, and that is also related to simulations. So basically the ability to test different interventions and then provide recommendations on the basis of a set of constraints for this process and your KPI targets. I want to reduce cycle time by 20%. And the idea is for a promo to identify the top two, three interventions you can implement under your set of constraints to be able to achieve that target. And that may mean, for example, allocating a couple of bots for this activity and assisting a human on that activity with other bots and re-sequentializing some activity. So what I'm trying to say is that this area of prescriptive analytics for process mining or recommender systems goes beyond operational support. It also works on a practical context, right? So how does the prescriptive process mining work with the low-code user being, as user-friendly, or the interface being as user-friendly as possible? I mean, me being a user of the tool, if I say you should, and the tool tells me, you should probably do X instead of Y. My first question is, well, why? Why does the tool tell me to do this? And do you find that there's some sort of gap between the answering that question as to why you should do something without going into, hey, it's some black box, voodoo magic type of machine money thing that we have in the background somewhere? So I think you really touched upon the most important challenge in the key challenge in prescriptive analytics, which is related to the topic of transport AI. Can I trust this recommendation? It's so on the basis of what? Now, so far, the techniques for prescriptive monitoring that you've seen around are based on the predictive analytics. So they're based on correlation, they're based on statistics. But there is no causal inference to tell you that if you do this, this is going to be the outcome. This is going to be the result that you respect. And this is precisely what we are working on with this PhD student I was mentioning before. The use of causal machine learning for prescriptive analytics, prescriptive process analytics. So this is really a paradigm shift in the whole area of prescriptive analytics because you identify those treatments, we call them treatments, that have a causal effect on a given part of the process. Part of the process could be a group of cases. So under certain conditions, if you apply this treatment, if you make one more loan offer on a price range with this price range, then you're going to increase the likelihood of leading to an acceptance of the offer by 20%. This increase is called uplift and is mathematically measured. So we can basically show you an uplift tree that really provides all the dependencies as to, you know, what is the root cause to explain? What is the key reason to explain why this treatment has this effect? And by doing so, we basically want to provide an explanation as to where these recommendations are coming from so that the end user can build trust in this recommendation. Then there is another part of your question very quickly, which is about how we expose these results, right? And this is more about, you go, going back to who is the target audience, and what type of analytics and how they can consume. You know, you can consume these recommendations through alerts. You know, we can send us luck message, we can send an email, and then the personal link and they see the dashboard and then the uplift tree, right? So it doesn't have to be complex. You do not need to expose and code for them to consume the results. So what I'm hearing is that we want to move away from this predictive models where we just take a bunch of data and we look at correlations because if we had a data based on, you know, let's say purchasing process or sales process in a COVID time, it would be quite a lot of skewed towards just nonsense values and nonsense predictions. And what you want to do is take out this root cause, or these root causes that are directly affecting your process in a certain way, and ultimately make a simulation on top of that. And now we can finally get into this simulation. And does this mean that if you are, let's say successful in identifying these pain points? And what you could do then is basically tell the user that if they do a certain action in the process, this is likely, or with this likelihood, this is how your process will look like in the future. - Correct, correct, yes. And simulation is part of the predictive process optimization layer, the second layer that I mentioned before, you know, the one that builds on top of the foundations. That's something that our customers are using systematically, now in projects, it's a tactical technique. So it allows you to articulate interventions and test their impact on the mid to long-term. So not something that you're going to do to fix the claims you're currently working on, or something that is going to help you for any new claim that is going to be lodged from now on, once you implement that change. For us, simulation is the natural next step to get process mining from analysis to action. And it is in fact the missing bit between analysis and an actual intervention. So automation is basically, you know, allowing you to intervene on a process and, you know, get value straight away. But the key problem is where to automate and how, and this is often the key cause for these automation projects like RPA projects to fail, where simulation fills that gap. You do your discovery and analysis with process mining. You come up with insights into your process structure, performance and compliance. Now, this insights foster the ideation of interventions. Oh, I find a bottleneck. The claim officer is a capacity bottleneck here. I don't have sufficient FTEs. And I can see that my claims queue up at the doorstep or says claim, because the claim officer is just overutilized, right? I move to simulation and I can test the removal of that bottleneck in different ways. So these are the possible interventions. One could be to simply allocate more FTEs. That's easy, but we all know that that comes at cost, you know, resources can be expensive. So an alternative could be to allocate a bot. So you can test their hypotheses with simulation, the tool will simulate this scenario and calculate for you the impact of that intervention on time, on cost, and on resource utilization. But the problem of simulation and why it hasn't picked up, basically, in the last couple of decades, is how to come up with the simulation model in the first place. I mean, one can question the utility of simulation. But when you talk to practitioners, I say, oh, that's great. But, you know, how on earth am I going to come up with the simulation parameters? And even before the simulation parameters, the very starting point is the model, right? The model of the process that I want to simulate. So that model, in a Promore, we can automatically extract in BPMN. It's a rich BPMN model automatically from the log. So you are starting with the actual process as it has been executed. On top of that, a Promore automatically discovers all the simulation parameters that are needed to simulate that very BPMN model, that as is model. Simulation parameters like arrival rate, statistics around the distribution of activity durations, branching probabilities, branching conditions, waiting times, resource allocation, FTs, all of that is automatically mined. And this BPMN model plus the simulation parameters is what everyone calls but hardly achieve the digital twin of your process. So what is a digital twin of a process? It's nothing else than a BPMN model with the simulation parameters. The challenge is how to get it automatically, because that's a moon day in task. I'm consuming error-prone. Where on earth am I going to fetch 350 simulation parameters? I mean, I'm talking about a real life process. Lone origination, 150 plus activities. How am I going to fetch the statistics about the duration of 150 activities? Now, you can do that. With a Promore, we can automatically discover this digital twin of a process. So we put the analyst in a position to simulate the assist and use that as a baseline to test the impact of any interventions they do. So I have my baseline. I've simulated, I've got my statistics. And they are in line with those of the assist process that I mind, right? So that's a way to validate the goodness of your simulation model. Now, you use that as a starting point to start modifying the simulation model. And there are two ways in which you can modify. One way is to act on the simulation parameters, for example, what if we launch a new marketing campaign to promote a new loan product, and we expect that to lead to a 20% increase in the arrival rate. So I'm going to change the arrival rate to put it 20% higher. How is that contextual change in my process going to impact on my SLA? Am I still going to be able to meet my 10 days time to write or not? So you can test the hypothesis. The second is an active intervention where you change the process, you resequentialize activities, you parallelize, you remove stats, or you replace going back to the initial point, a human being with a bot, or you assist that human being with a bot. So this is another intervention you can test. So the key point that I want to make is before jumping into automation, you can now test your hypothesis with simulation, assess mathematically what is the impact on cost, time, and resource utilization, and come up with a recipe that tells you if you do this, this is the impact you expect to achieve. Now, as you can see, that is going to give you a very rich context, a very rich set of information upon which you can base your unjustified your decision. I'm going to do this first, because I know that if I fix this bot on neck with five bots, this is going to be an impact, rather than that or that, other alternatives. That's in essence simulation. Well, if I learned one thing from this episode is that I might apply for the course in a University of Melbourne to listen a little more into this, because this is exciting stuff. And I wish in the podcast we could go more into the depth, but it is what it is. The time is slowly, slowly coming up. But I have one final question for you, Marcelo, and that is, at the beginning of the episode, you mentioned that process mining, as a market, as a-- us practitioners are utilizing, let's say, 20% to 30% of the capability that process mining offers. One of the things that might come in the near future, and it's already here, is the simulation, digital twin problematics, that you are just talking about. Could you maybe give us a glimpse of something else that we could really get excited about, and that will probably come in next five to 10 years? One thing that we started playing with at a Promore is the use of conversational votes. And so basically-- Yeah, all right, go on, go on. You can finish any picture. I see, at ICPM, we even provided a demo on that. So in some of these conversational vote capabilities are already available in a Promore. Like you can ask Siri about the cases of your claims that are non-compliant and violate a four-eye principle. And Siri replies, he gives you the list of cases and asks you if you want that to be sent over email. Now, we want to interact with a Promore in a totally touchless way, in a conversational way. But that goes beyond asking questions. As an answer, you can derive from the content on the screen. We want the tool to reason along with us. So, hey, Siri, what are the top three interventions I can make on my long-engination process in order to make sure that I keep the time to write at a most then working days? Wow. So that's where we want to go. It might take a while, but we are working on the research side of it at the university. Well, that is very exciting. And now, the final question, where could people, first of all, find you, but maybe there are some listeners who's hard-started pounding just by listening to your automation and simulation capabilities. And we would like to support you on your journey, during your research, or in a Promore. Where can people go and contact you and find you? So, the best way is to just approach me on LinkedIn. You can just search for my name on LinkedIn and start chatting with me. And yes, we are looking for talents and people who are excited to work with us, but on research as well as on the software level, as a product manager, as a software engineer, as a sales or marketing person. I mean, we are growing, as I was saying in a Promore, we are in the scale up phase, we are expanding globally, and we're always in search for great talents. I'll make a side note here as well. And if you got excited about prescriptive and predictive process mining, definitely go on and listen to our previous episode with a colleague of Marcelo with Marlon Duma, I think really, really good stuff there, as well. And Marcelo, what can I say? Just thank you very much for coming to our show, for me, it was a real pleasure. You, dear listeners, cannot see it, but I saw the spark in Marcelo's eyes when he started the going on simulation, which makes me wonder, maybe we should do this also in video next time. Oh, for sure. Marcelo, thank you very much. Thank you for coming here. And I wish you and all your team in up for more and also in your research all the best, because I hope that at some point we can get a glimpse of the technology and maybe implement it ourselves as well. Pleasure, thanks a lot for the invitation. And thanks, everyone, for listening to this podcast. All right, so if you enjoyed the episode, just text us a LinkedIn. We are very active there as well. You can also reach out to me and Patrick personally. We also have an email [email protected] always open to ideas, feedback and recommendations. And yeah, if you like us, leave us a review, leave us some good ratings. And we will be looking forward to hear from you and to talk to you in the next episode of miningyourbusinesspodcast. Thank you very much and bye-bye. Thank you. Bye. [MUSIC PLAYING]