Deutsche Post AG (DHL) Earnings Call Transcript & Summary

Hello and welcome, everybody, out there to next edition of our IR Virtual Tutorial from Deutsche Post DHL in our effort to make an experience for you what we mean by Excellence Simply Delivered, our claim. For today's round, as you've seen from the invite, I'm glad to have with me today, Katrin. Katrin König, thanks very much for taking the time and making yourself available. I'm pretty sure it's going to be an interesting, I don't know, 60, 70 minutes, something like that. We have introduced this format not too long ago to have a chance to still educate you on specific topics in the group which we think are of interest for you. And I'm pretty sure you're going to agree with that also after today's session. Invites are already out for the next one with John Pearson and team on why e-commerce works for DHL Express. So looking forward to that one, and there's more to come. But let's now focus on today's topic. And you see here, our strategy pictorial -- group strategy, as announced last fall. And needless to say that over the past couple of months, we have plenty of opportunity to see that the strategic approach of our group is actually working out. And one element we want to focus on today is, of course, digitalization as one of the key enablers of the group's future success. So what do we mean by digitalization? Well obviously, in line with our 3 bottom lines, we want to come to a better customer experience with the help of digital tools. At the same time, there are plenty of opportunities to make work easier and more efficient for our employees and overall, increase operational efficiency. And that's where you guys come into play because that's going to typically spell some sort of financial benefit. So we will take a look today at how we are dealing with the theme of digitalization. We have been identifying 5 areas of digital improvements and initiatives where we, on a group level, run so-called Centers of Excellence, which are then sort of developing the right know-how and the tools and everything that's necessary, and then make sure that this is scaled up through all the divisions, which is clearly a big advantage over smaller players who only have one division and have to take the investment. And we can much better scale that. And you see the areas that we are covering with this approach. So today, quite obviously, we want to talk about data analytics. And I think what Katrin is going to show you is that, indeed, this approach is one of the key enablers really to digitalize our operations and bring it into life and fruition throughout the group. There are, as we will see, a multitude of solutions, whether that's benefiting the operational experience from an employee point of view or leading to a better customer experience or simply saving money, and we will share with you where we stand in rolling this out throughout the group. One more technical remark. For those of you who followed earlier tutorials, you know the procedure. You see on your screen, a Q&A field. [Operator Instructions] So -- and with that lengthy introduction, finally, over to you, Katrin, please.

Yes. Thanks, Martin. And welcome also from my side. I'm really glad I have the chance to talk about that exciting topic today. And let me start the whole topic with a short teaser. So this teaser is an example for one of the data science solutions that we developed, and I think it nicely shows what data science can do for us as a company and what benefit it brings. So what is it about? Well it's about a problem that you probably all know from your personal lives as well. So sometimes you need to go to a place you have never been before, it's an address, and you need to, yes, find this address on the map. It could be the hotel of your next vacation destination or something like that. And I think all of this -- all of us to do this use external map services like Google, Bing or HERE to solve that. When we look at our company, we actually face this problem millions of times every day. Because we get the shipments from our customers with some address information which can be more or less cleaned or structured, and we need to locate that as well somewhere on the map, on the earth. And in order to find those addresses, we actually came up with our own internal approach because as you will see in a minute, this works for us much better than the existing services. So I will show you now how we leverage our unique internal data assets together with machine learning algorithms to solve this problem of geocoding. So let me quickly just share our geocoding engine with you. So this is our geocoding engine in the front end that we developed, and I think that shows nicely what we can do here. What you see on the left-hand side here is a delivery address from The Netherlands as we typically get them from our customers. So now, and this looks similar to the typical search engines you know, we want to know where this address is. So let's search for it. And we see, it's somewhere in the area of Tilburg. And if we zoom in further, we see that in this case, it's a hospital, and we get the geocode on a very specific point in that hospital area. Now I want to explain you quickly, first, how we do this; and then secondly, why this works for us much better than the existing services that are out there. Well I mentioned already, we have millions of shipments that we deliver every day. And whenever we do that, we actually capture the geocode of those locations. And this is what you see here now on the map. So all those blue pins are historical deliveries in that area. And this is the base for our solution. So we use that data and we apply machine learning algorithms to it to structure the data, to match it and then to determine the final geocode. Now why is this internal approach with our own data so much stronger? Well the advantage is that the information that we use is much more relevant and specific for our purposes. So this is a hospital, as I just said. And very often, we need to deliver to large buildings like hospitals or malls or so. And you can see here that our data knows exactly the location where the courier needs to go to. So it wouldn't point us to the main entrance or the center of the hospital, but really to the courier entrance, so to say. And an external service, as you see it here with the red and yellow flag, just cannot know this information. So this is just much more powerful. And you can imagine in countries where we don't have a structure, an address system in place, this becomes even more beneficial for us. So this was a quick example, so -- and a teaser. So let's get back and start with the actual, yes, content and look a bit more into the topic. And we want to do that based on deep dives in 4 areas today. So we want to start to look into the, yes, big opportunities that we see and that we have in the area of data analytics. Second, we then want to share with you our understanding and our approach to analytics with a typical project cycle. Then we want to share further use cases and especially talk about the big initiatives that we are running currently in the area of analytics and the benefits that we get from those. And then finally, we will share a bit more about our approach on how to scale a topic like analytics into a large organization like ours. So starting with the big opportunities. So what you see on the next page I think is not new to you. So those are our 4 -- our 3 bottom lines. And I think the geocoding has already shown you that analytics can bring benefits in many different ways. And when we look at those bottom lines, analytics can clearly support all of them. And let me just go through them based on the example of that geocoding that we just saw. So clearly, if we have more precise geocodes, we can do a better route planning and we end up with shorter routes. So this leads to clear cost savings but, of course, also reduces our CO2 emissions that we produce. But also when we look at our couriers, right? They have a better tool support, they get more precise routes, they have to do less detours and they can deliver a better service. And this, finally, when we look at our customer, pays off for the customers because they experience this better service quality that we can deliver to them. And we will go through further examples along the presentation, and you will see that all of them will address at least one of our bottom lines. But before we do that, I want to have a bit of a closer look into the financial aspect and the investment of choice bottom line. And what you see on that page are our main 2 cost blocks that we have, so the staff costs and the transport costs. And when we had a look at all the analytics initiatives and projects that we ran, we saw that almost all of them address at least one of those buckets. And I can maybe quickly highlight that on one very concrete example, and this example is about volume prediction. So if we have the data-driven -- no, not that. If we have a data-driven forecast of the volumes that we expect to handle in our sites and the volumes that we want to transport on our routes, we can optimize the resource planning and also the transport routes based on that. And we've seen in the projects that we ran that, that can lead to significant reductions in both areas. And this is also why, and I think Martin said it in the beginning, analytics is one of the substantial pillars in our digitalization strategy and will have a significant contribution to it. And I mean you do see big numbers on that chart. But today, it will not be about sharing with you concrete numbers on savings use case by use case, but it's much more about understanding how we really enable an analytics culture, how we create a mindset shift because this will really enable our organization to manage, get the company better in many different ways and aspects. All right. And yes, to illustrate that to you, I think I will now talk a bit about, yes, our view on analytics and really on our approach to the topic. And to do that, let me maybe start by sharing with you our understanding of analytics because I think that's quite important. Because what I often see is that there is an impression that analytics is purely about taking huge amounts of data, throwing that into some magic tool, running some fancy algorithms and then automatically, you get exciting insights that are, of course, business-relevant and value adding. Well our understanding and our view of the reality is a bit different. And actually, in our understanding, we start the other way around. And you see that illustrated on the left-hand side. So for us, analytics always starts with a very concrete business problem and a question you want to solve. Taking the volume prediction, I want to know how much volume I need to solve in my sorting center tomorrow. Only then you start collecting data, data which can help you address this question. And that can come from different sources, can be in different formats. And then you start the whole process of cleaning, processing the data. And you run your mathematical models, but you always have the business question in mind, and you always combine it with the business knowledge and the expertise that you have because only then you can really come up with valuable insights in the end. And we will go through an example, a bit more in depth in a minute. But I would also quickly like to, yes, show you a bit what is inside this machine that you see on the left here. And on the right-hand side of the screen, you see one way of classifying data analytics. And this is a quite common framework so you might have seen it before, but let me quickly go through it. So it starts at the bottom left with descriptive analytics. So this is really about understanding the past and understanding what has happened, so the typical business intelligence or dashboard type of activities. When we then start looking into the future, trying to make guesses about what will happen, we are then in the predictive area. And you can make all sorts of predictions, right? So you can predict volumes, you can predict customer behavior, demand or you can even predict the geocodes, as we've seen it in the beginning. The next step after predictive is then the prescriptive analytics. So this is about optimizing your processes based on those predictions. So this might sound complex, but that's a simple example. So we have a prediction of our volumes in our transport network. Well based on that, we optimize the transport routes for the next day. And the descriptive area is an equally important part for us. But in the following, when I will talk about data analytics and data science, we will refer to the predictive and really the prescriptive analytics applications. All right. So now we have a common understanding of analytics. What I would like to share with you now a bit more in detail is our approach to a typical analytics project because I think it has some quite interesting insights that we see along that process. And those general 6 steps that you see here, they are applicable to more or less any analytics project that we run. But I want to make this again based on a concrete example, which is the operational volume prediction at Express.

Which is actually referring to the daily challenge that our guys in the global express network have to deploy the right capacity on a given flight from up to somewhere in the regional network to have an idea of what was about the volume we are expecting on the next day, on a given day and then what type of capacity planning do we need to come to, right?

Exactly. So it's -- for the next couple of days, and it helps us to steer this capacity in 2 ways: on the one hand side, we can make sure we have enough capacity in the network; on the other hand, we can also sell excess capacity. So it's really an important input for all planning that follows then.

Okay. Selling excess capacity, not the usual problem these days, but in normal times. Perfect.

Yes.

Okay. Let's get through it.

Good. So we start with the first step, the problem definition. I mentioned already earlier, right, that we always start with the business problem. And this is what this first step is all about, really getting a very good understanding. And yes, Martin already shared a bit what it is about, right? So it's really about predicting the volume on each flight in our flight network for the near future with the 2 clear objectives that we have, steering our capacity for quality and for selling excess capacity. And this is -- so really the very first step, having an understanding of what it is about. Second step is then the process understanding. So here, we really want to understand how the model or the algorithm that we build will fit into the processes because only then, if it's fully integrated in the end, we will see benefits from the work that we are doing. And here, the data scientists, typically, they join the operations. They talk to users of the tool and so on. So it's a quite important step. And here, for example, we learned that there was an existing tool and a process already in place, so a tool with a rather simple forecast. But that was really good news because we knew that there was not big change in process or also change management with the people that we had to do. So that was quite helpful in that case. So then we can get to the third step, which is data collection. And here, you can really get creative and think about all potential data sources that could help you answer this question. And to clarify that already upfront, there's always a lot of ideas about fancy data sources. And it's good to think about them. But what we've seen in this case, but also in many other projects, the most important data is really your own historical data. So you want to predict your volumes? Well then you need to have your historical volumes as granular and as far back as you can have. But of course, there are other important data sources like information on holidays or the flight plan that you collect in that phase. Now that brings us to the first -- fourth step, the data cleansing and preprocessing. And this is quite an important one because here, the data scientist really sits very closely with the business experts. And here, it's about trying to understand patterns in the data. And for example, you see an outlier in your data series, then you want to understand the business reason behind so that you know whether you would take out this outlier, whether you average it out, whether you leave it in. So this is really very important, very interactive between, well, data science and business. And just to give you one concrete example again for the Express case, heavyweights were a big topic in that part, right, understanding whether they are part of the historical data, whether we want to take them out, whether we want to forecast them maybe separately to avoid distortions. So this is what this is all about in that phase. All right. Then the modeling step counts. And I think this is quite interesting because when you think of data science, you think it's all about modeling and all about what this guy here on the picture is doing. But actually, it's only step #5. So there are a lot of important steps before that. And why? Well in this case -- in this step then, the data scientist really combines his methodological knowledge with all the business information and all the business knowledge that he or she gained so far. And then a quite creative process starts because data science -- well it's called science, but sometimes it's more an art because there's not a textbook solution to it. There are a lot of algorithms that you can try out, play around with. And it's all about integrating this business knowledge that you found out in a smart way into your model. And again, I can make a quick example here. I talked about holidays. And I think it's quite obvious that holidays do have quite an impact on our volumes. But what we found out, for example, is that holidays in Australia do not only impact the volumes in and out of Sydney, but they also have an impact on our flights from Hong Kong to Leipzig, for example. And those complex interdependencies need to be somehow taken into account in the model. So this is a little bit a glimpse into what happens there. Now once we are done -- well you're never done with modeling but once we found the model we are happy with that has promising results, we then get into the last step. So that's the implementation and really the business process integration that takes place here then. And I already mentioned that we were quite lucky in this case because we only had to integrate our algorithm into the existing tool. And what we saw then was, well, on the positive side that the improvements that we saw on the theoretical side when we looked into the modeling also turned out to realize in the practical or in the live environment. So we saw improvement of 3 percentage points in the forecast security, which is huge in that area. But of course, there, we also found in the initial phase some smaller pain points, which we then, of course, took back to the model, fine-tuned the model to get rid of those. But yes, now our model is up and running and integrated into the business. So I think this hopefully gave you a bit of an understanding how we see and how we approach data analytics, which I think is quite important for the following part. And I already shared some use case examples along the way. But I want to spend a bit more time on further use cases and especially on really the big initiatives that we're running in that area. And yes, what you see on the next page, those are really the 6 key analytics solutions that we are developing and working on at the moment. And of course, there are a lot of other very specific projects that are driven from different teams. But those topics that you see here are really the ones that have the highest relevance and the highest potential across our divisions. So this is why we develop those really into group-wide solutions with a very high degree of scalability. And given they are really important for us, I would quickly like to go through them. And you can split them into 2 groups. The first 3 topics that you see here, they are really addressing our core processes and optimizing them. So the first one, the operational volume prediction. So we've just seen the example of the Express sector forecast, which is exactly this. So this is just key for our business to plan our resources in our sites to plan our transport assets. And every percentage point of improvement that we see here just have -- has a huge benefit for all the subsequent planning processes. And the vision that we have here is really that each planner in all sites and all facilities has such a data-driven forecast at hand that supports them in their planning steps. The second topic, the staff scheduling. This is kind of the natural next step, I would say, because we do know our expected volumes from the prediction and we also have a rough shift plan in place. And now it's about the short-term planning, the exact shift scheduling, determining who works where, when, considering sicknesses and breaks and all those things. And this is a super complex optimization problems. And our staff dispatchers in our sites, they spend hours with their tasks every day. What we did here is we developed an interactive optimization tool that supports the dispatches in that process. So that gives them more optimal solutions, but also helps them to run this process much faster than they could before. And again, we want to have all our dispatchers using such a tool support. So that brings me to the third topic, which is routing optimization. And yes, routing is, of course, not new, right? That's the core of our business, and we have been doing that for years and decades. But what we see here is that there are really new trends coming up because somehow, yes, our world is changing faster. And we see that the requirements of our customers but also from our operations are changing faster, and we want to change our routes more dynamically. And what we've seen here is that existing software that we used and tools that were in place were often not able to cater for those requirements and flexibility needs. And on the other hand, we saw that, well, internal solutions, they have that flexibility in that they can be really tailored to those needs. And let me go a bit more into the details of routing. In the page that you see here, we could actually draw for any of our solutions. So typically, there are a lot of application areas and a lot of projects below each topic. So it's not one project but a multitude of application areas. The picture for routing, you see here. So we have application areas and many we use. A lot of them really focusing on last mile, of course, because this has a huge cost levers, huge cost driver for us. Our vision for routing is really that we leverage our business expertise that we have and our method knowledge to develop truly customized tools for our specific problems and that through that, we are able to really address this flexibility needs that we see. Now this might now still sound a bit abstract, which is why I brought a very concrete project from the space of routing, and this is the RaptOR tool that we developed for freight. So the RaptOR tool supports our dispatchers in our freight terminals in their daily work. Those dispatchers every day get 100 -- or hundreds of orders, which they need to assign to tools and delivery vehicles. And when they do this, and you see that in some of those yellow boxes, there are a lot of constraints and a lot of things they need to take into account: vehicle sizes, capacities, stackability rules, driving regulations. So the option space for them simply explodes, and it's impossible to solve that for a human being in an optimal way. So this is what RaptOR now does for them. So the algorithm takes all those constraints into account and comes up with an optimal delivery schedule that dispatchers can then use. And RaptOR does that in seconds. So that brings me back to the bottom lines we talked about earlier. So you see, first of all, that this is a huge support for our dispatchers, right? They save a lot of time that they can spend on other tasks like finding even better deals for the routes that we came up with.

And this solution is being derived in seconds, however, using like the process that you described earlier, all the input that you have from previous observations, what roles -- what a role does the local holiday play, how are Wednesdays compared to Fridays, what's the current traffic situation, all these sort of things, right?

All these sort of things are taken into account in the algorithms. And of course, the development was not done within seconds, but now the algorithm really runs in seconds. Whenever you give a new input, you get new solutions for that.

Okay. And there's probably also something like a learning curve for the further fine-tuning of the algorithm in the course?

Yes. Definitely. And I think one of the strengths of the algorithm is that we can really integrate or include new constraints rather quickly. And I can give you one example for that for one of our customers. So we typically roll it out customer by customer. So we had one customer with a transport network that had a lot of shipments from Eastern Europe into Germany. So what would RaptOR do? Calculate the shorter stops with all the constraints. And it would travel, of course, also through non-EU countries. Then we learned from the customer that they absolutely want to avoid that because there's customs processes, there's effort, there's time. So then this, for us, was a new constraint, so the team took that back and could include that into the algorithm within a few days. And this is what I meant when I talked about this flexibility that we need because you won't find a software that has all and everything included. And with the internal capabilities, we can just quickly adapt it, and that's been truly customized for our customers.

Okay. So your team is then also in constant exchange with the actual users of the tool itself?

Yes. Yes. So the team spent a lot of time on site really with the dispatchers. And in the initial development phase, it was more or less like a constant line that they had. So dispatch has around something. They said, "This doesn't make sense." So we looked into it and iterated through it. And I think this also helped the dispatchers to really gain trust into the tool and into the work that we are to support.

Quite important, yes.

Yes. All right. So now a quick look at the other 3 solutions, which are also called advanced data services. So here, we really leverage our internal data to generate new benefits or new insights. And the first topic we've seen in the very beginning, the geocoding. So here, we leverage our internal delivery data to come up with geocodes for new addresses. And the second topic, the product classification for customs, that's based on a pretty similar idea. So yes, what is it about? Whenever we send a shipment cross-border, and you can imagine, we have a lot of those shipments, we need to determine the customs code for this shipment. So this is a code that specifies exactly what's inside the shipment. So that could be a cotton T-shirt for men in red or something like that. And this process today literally is still done either manually by the customs agents or externally by our brokers. And as you can imagine, both there's quite some effort in terms of, yes, time or money. But again, here, we do have data assets that we can leverage because we have a huge database of historical shipments. And those shipments, they all have information like a description and value and other information. And of course, for the historical shipments, we also do have the customs code attached. So we can learn from this data, again, apply our machine learning algorithms and then predict for each new shipment what is the most likely customs code. And if you -- yes, if you remember the process as it has been in the past, this has a huge potential for automation, of course. Now the last topic in that area is the invoice overdue risk prediction. So that's a topic from the finance area. And again, I want to go a bit more into detail here. So again, we do leverage here data that we have. In this case, it's our internal accounts receivables data, and we use it really on an invoice level. And we come up with algorithms to predict them for any new invoice the risk or the likelihood that it will be paid late on time or even early. Now this prioritization in the past has been done either manually by collections expert or based on a set of fixed rules. What we now have is a machine learning algorithm that takes into account hundreds of statistical features, and it's constantly learning and adapting. And we've seen that this is a really valuable input for the collectors and the collections processes. And since the I/O ASCO, as we call it, is in place, we've also seen huge benefits for our cash inflow processes and steering. And of course, also during the corona situation, this has been a very valuable input.

So you do -- you're filtering out those invoices that we have sent out which are running at relatively higher risk for not being paid on time but not simply by only taking past payment patterns, but also taking into account state of the customer industry and other factors, yes? So it's not only about how well the specific account is doing, right?

Yes. Exactly. It's as it says here, it's hundreds of really, yes, variables and characteristics about the invoice, the customer, the industry, past behavior not only looking back a few weeks, but really looking at trends and doing extrapolations. This is all what happens in this, yes, quite sophisticated algorithm and when it's out, it's really for an invoice risk flag: is it high, is it medium or is it low?

So that means all the sales guys out there who have taken care of their customers have a clear list -- focus list which customers to contact and call and make sure that they are aware of that and don't waste their time on more secure cases, right?

Exactly. So it's, yes, a very crucial input now for the collectors and for -- yes, for prioritizing the day. And maybe that's a good moment to also show you how that looks in practice, right, because I talk about machine learning model, statistical features that sounds, well, abstract but also fancy. But I can quickly show you how that really looks like in our daily work. And to do that, let me quickly share my screen again. So what you see here is a dashboard that a collector at DGF is using to plan and prioritize the day. And as you can imagine, the data that we show here is no true and real-life data, but it's test data that we set up for training purposes of our collectors. And what you see here is a list of customers for one of those collectors with all sorts of information. This is what they had already in the past. What is now new is this colored column here, the IOR column, that assigns this risk flex, so medium, high or low risk of, yes, paying late. And you see, it's really straightforward. It has a color coding. It has 3 categories. So at the first glance, it already helps you to -- yes, to prioritize where to look at first. Now if we look at one of those customers a bit closer. So let's take the second one here, the customer with a high-risk flag, you get even more detail. So you see here all the invoices that are attached to their customers, again, all with a score here. This is really a high-risk customer because each single invoice gets a -- yes, a high-risk flag. But you can also see here then the actions that are triggered based on that. Some of them are automatically; some of them, of course, are derived by the collectors. And yes, you might now be a bit, I don't know, disappointed, but you might have expected to see something more fancy because the algorithms behind are quite sophisticated. But what you see here is really doing exactly what it is supposed to do because we have those algorithms and the results of those algorithms are really fully integrated into our business processes.

So I think what our specific audience does find fancy is, what is the actual real-life impact? Have we seen any improvement or any positive change to the collection success?

Yes. We have. So for DGF, in the cash inflow, we've seen a positive impact of a double-digit million euro figure. And also, on the EBIT side, this leaded -- led so far to a single-digit million improvement that we've seen. So that on the financial side. But of course, this comes with a lot more of automation potential. So we, for example, or DGF, they now installed bots that are -- so the so-called digital twins that are treating some of those customers with a very clear pattern automatically. So this is a huge efficiency gain on top. We could reduce our bad debts because we really identify problematic customers earlier. So there are a lot of benefits to that.

Right. Very tangible and relevant.

Yes.

Very good.

Yes. Yes. But I mean maybe one comment also to make on that. Of course, it led -- or it had to come with quite some change management and training we had to do because, of course, we are now, well, providing at first glance like box to our collectors. And we had to build up trust with them because they should rely on something they didn't know what it was about. And we did a lot of measures in that area. One nice one was that we took a set of new invoices. We asked the collections expert to classify them into high, medium, low. And we let the algorithm do the same. So in the end, the algorithm was better. And this was one way that really had to overcome the skepticism. And now the collectors really see that as an -- yes, as a support, which it should be, right, because it helps them really focus on the key things.

So often, you have to overcome this first wall of resistance because these guys, these collectors, I mean they've been doing that for years. And obviously, their starting point is, well I know my customers best, right?

Exactly. So this is often -- and I think that's a valid skepticism that you can have in the beginning. And if the machine cannot beat them, well then it also shouldn't. And then very often, it's -- I mean in this case, it was then clear. But very often, it's also not purely about, well replacing what they have been doing in the past, but providing input to them so that they could -- can really focus on the things that the machine doesn't know on the local expertise that they have. So it should -- very often, it's more valuable input that they can use to then do the last 10%, 15%, even better.

All right. Good.

Okay. So those were I think quite some examples that showed you, yes, how we leverage analytics in our daily work and how we integrate it into the business. So yes, let me now finish by talking a bit about our approach to really scale this topic of analytics into our organization, which is not as trivial as it might seem. Starting with the organizational setup. So what we have established here is a hub-and-spoke approach. Because for us, it really combines the advantages of a central and a decentral approach. So we have our hub, which is our central data analytics Center of Excellence. So in this team, one of the main things we do is we are developing those group-wide solutions, so the 6 topics I just talked about that are applied in every division. But besides that, we also act as a talent pool for the organization. We really bring in the latest methods, best practices. And of course, we try to connect the data scientist community across the company. The central team is then complemented by a divisional or functional spokes. So those are smaller data scientist teams that sit within the business. So that could be an ops analytics team within P&P or it could be a dedicated finance team. And those teams, of course, they are much closer to the business, and they bring in the subject matter knowledge. So just looking at the invoice overdue risk topic that was so valuable to have a dedicated finance data scientist there who knew the finance processes, the finance terms rather than just the data scientist who has never deal with finance topics before. And this is really the benefit that we see. And very often, we combine also projects with data centers from the hub and the spoke to really get the best of the 2 worlds out of that.