Listen to the latest episode of Bramasol’s Insights to Action Podcast Series. In this episode, Birgit Starmanns, Global Head for Thought Leadership Strategy and Programs at SAP’s Global Center of Excellence for Finance and Risk, discusses Machine Learning – what it is and how it can be applied to finance and risk management.

Also, below is a transcript of the podcast episode:

Jim Hunt: Hello, this is Jim Hunt for Bramasol’s Insights to Action podcast series. Today. I’m really happy to have a return visit from Birgit Starmanns, who is the Global Head for Thought Leadership, Strategy and Programs in the Center of Excellence for Finance and Risk. Birgit’s functional experience is in finance and management accounting. She’s got lots of experience in SAP S/4HANA finance as well as core SAP ERP and EPM systems. She has over 30 years of experience across center of excellence issues and solution marketing, solution management, strategic customer communities, and management consulting organizations. So today we’re going to have a real interesting discussion on machine learning and specifically how it applies to finance and risk teams. So Birgit, it’s really great to have you again. Thanks for joining us.

Birgit Starmanns: Thanks Jim, for having me.

Jim Hunt: So, Birgit, can you start by giving us an overview of machine learning?

Birgit Starmanns: Sure. Thanks Jim. So machine learning is actually a part of the overall AI category or artificial intelligence, so that when we think of that, the first category we normally think about is the rules based. But that’s basically going in and doing your configuration of these are the particular rules, and we’re going to keep executing them. Unfortunately, most companies after an implementation don’t ever go back and take a look at those rules. And as the business environment changes, then the rules no longer apply or they apply in fewer percent of the cases. So that means basically the finance and risk teams actually have to spend more time on exceptions, because again, the rules are very rarely revisited when it comes to the post implementation, just getting the system live. And yeah, it’s great. Now we’re going to continue to go, but that also means that as time goes by, those rules are less valid.

Birgit Starmanns: So this is one of the reasons for using machine learning that it’s not necessary to revisit them. Another category is the whole robotics automation. We’re seeing more of a move towards bots, but basically that’s more like a think of it like a, an Excel sheet macro, where you’ve got certain rules that you just execute periodically and it’s a whole set of rules. So it’s really a package of information. What machine learning actually does is it learns. So in a way it’s a little bit of a black box because it learns from prior transactions that took place. So whether it’s accounts payable or accounts receivable, these are the ways that I would clear an open item. And sometimes it’s two way matching. Sometimes it’s three way matching. And a lot of times, the finance team spent a lot of time just looking at those various transactions and doing that manually.

And even if there’s an automated job, again, going back to the idea of rules over time, become less applicable than they’re doing more and more exceptions. So what machine learning does, it takes a look at those rules, but then it also takes a look to see how a human person went and handled those exceptions. And it then learns from that. So that does not mean that it creates rules, that it puts into a configuration table, but basically it’s part of the algorithm that it has learned from these things and applies the actions that a human being actually took for those exceptions. And then it can basically take that into account and clear more items and therefore reducing the number of exceptions that finance and risk teams have to deal with.

Jim Hunt: So it kind of starts with that base of rules, but then it shadows the expert and mimics the changes they make and then incorporates those. And that’s how it learns?

Birgit Starmanns: Exactly. It’s almost like a human learning, right? So you get certain parameters when you’re a kid don’t touch the hot stove ever. And then yeah, as you learn, okay, well, I just turned it on. It’s probably not that bad if I have to remove a pan or do something. And then later on, it gets hotter then you know, you really shouldn’t touch it. So there’s basically kind of a pattern of progression. And then as you grow up, you start learning what you can and can’t do. It’s the same thing with machine learning. As human beings we don’t have a set of rules that we consult every time we’re confronted with that same stove. So, the machine learns in the same way and yeah, it doesn’t create a checklist. I mean, it does, I guess, internally, but not, not something that I can go back and configure necessarily.

Jim Hunt: And, uh, if it’s able to essentially take that learning and apply it to forward-looking new situations. Here, your hot stove, example is intriguing. If a kid learns not to touch the hot stove, that’s pretty transferable to don’t touch the hot radiator or the hot barbecue and that kind of thing.

Birgit Starmanns: Exactly. Exactly. And there are so many situations on a transactional level where you have to basically clear an open item and yeah, I used to kind of joke how many ways can you clear an open item? Well, the answer is there’s a lot and there are a lot of different things that can go into that. Just on a transactional level. So basically do I have to deal with currency conversion? Is that an exception or is a customer paying for three different invoices or four different invoices or just one and how do I match that up? So those are usually the things that a person has to deal with, but if the machine understands those rules and has learned how to apply those rules over time, then that basically reduces the exceptions. And then the finance and risk teams actually have more time for dealing with the true exceptions such as collections and disputes, and not just on a transactional level and they can add more value to the business that way.

Jim Hunt: So, so what are the benefits? The key benefits, I assume speed is a big one, but I bet there are a lot of other ones too.

Birgit Starmanns: Yeah, there are definitely a lot of other benefits. So basically also reduced error rates. Because if you look at manual transactions, that doesn’t mean they’re error free because we’re all humans and sometimes mistakes happen. So things are not only faster and more automated, but we’re actually reducing the number of errors. The other really cool thing is that it’s not just the transactions and the associated master data, but machine learning can also take into account other items such as non-structured data. For example, a PDF comes in as part of payment advice, or there’s a PDF of an email or the email itself comes in. So being able to take that, for example, in the payment advice situation and apply that as well. So just because somebody sent back a PDF with part of their payment doesn’t mean that, that becomes an exception that has to be dealt with manually.

So basically it deals with that as well. And we’re seeing more and more applications of this. I mentioned earlier, accounts receivable, accounts payable, one of the big ticket items as GR/IR or goods, receipt, invoice, receipt, reconciliation that normally takes a really long time. And thanks to SAP HANA, in the backend, it does not normally take that long anymore. We can actually do that in a matter of minutes, but at the same time, we don’t want a lot of exceptions to fall out because of things that could have been learned and could have been dealt with ahead of time. So we’re seeing actually really good results from our customers when it comes to not only accelerating the processes, but also reducing the error rate.

Jim Hunt: So for instance, in the GR/IR matching situation, if a company name had an abbreviation or something different that was treated as an exception in the matching, the machine learning system learns that. And it’s no longer an exception.

Birgit Starmanns: Exactly. Because it learns it because normally we don’t create the rules for that many different variations of the master data. And of course, there’ve been a lot of situations where in terms of your comprehensive master data, there might be periods between, you know, S dot A dot P dot, and then there’s SAP and then maybe writing it out. And then there’s the SE and somebody might still have the old version of AIG. So at some point though, the machine learns that. So that’s more of the learning episode scenario, not so much, I’m going to go and recreate those rules to deal with that, but that’s something that a machine would learn, but that’s a really good example, Jim.

Jim Hunt: Alright. And those are kind of in the moment process flow, but you mentioned when we were setting this up, that machine learning can be predictive as well. Can you expand on that?

Birgit Starmanns: Yeah, definitely. Because when we look at predictive, this is really in a large part, part of the whole planning and budgeting process or making a decision for the company. So, a lot of times in predictive, we have in analytics a lot of great tools such as what if analysis and simulations and things like that. But the beauty of machine learning is that it can actually help uncover hidden trends. And we see that also in things like a business integrity screening, where we’re looking at fraud and what are some of the different scenarios that can happen. But when it comes to predictive items, for example, we expect certain revenue to come in and we might have that in a system where that could be processed in a sales order, but a sales order doesn’t have a financial entry yet it doesn’t have financial entries until there’s a goods issue. And until there’s an invoice to the customer, but looking at the sales orders and contracts that are out there, the system can actually see what the anticipated or predictive revenue would be. The same is true on the payable side, what is my predictive spend and actually, how does it break down to different levels, such as cost center levels? So I can get actually get predictive balance sheets and P&L’s based on some of the information that’s already in the system. Then seeing some of the trends that are behind that, you know, is there a monthly trend? Is there a seasonal trend for some retail organizations? Um, right now with a shelter at home, there’s probably a whole different trend in the way that customers purchase and services are provided. Now we’re all using more services at home than we would previously, so taking those kinds of seasonalities and some of the trends into consideration, it really has a lot of value.

And one of my favorite examples, and I think I use this more than once is the whole M&A scenario, the mergers and acquisitions, because it would be almost impossible for a human to go through and say, well, all right, these are the different scenarios that are possible, and there might be 50 or 60, and they have certain drivers that they can manipulate, but to do a traditional non machine learning oriented what if analysis then, you know, the drivers have to be changed every single time. And there is not often the bandwidth of the teams to actually take all of those different steps. So then they end up doing a pre-selection. So of these 50 things, I can only really evaluate 10 to 12 in depth while with machine learning, by automating how those drivers might be considered, I can actually evaluate all of these different scenarios. And that way there’s no opportunity loss because I preselected the wrong 10 to 12 items, but the machine can actually look at various ones. And this is definitely just as input to a human decision. So we’re not trying to say that this has to be posted automatically, but there’s nothing that can replace the human element when it comes to that. So there’s more information and better and more supported information to make a decision, but what is that financial impact of buying company A versus buying company B or buying company C versus making my own products and what are the costs associated with that? So really being able to look at all the scenarios, not just some preselected ones is actually a really key component of this. And then again, taking a look at the trends and applying those to those drivers as well.

Jim Hunt: That’s really fascinating. And I assume that you can also like narrow the, um, the gap between the prediction and the actual so that you can model the various scenarios kind of widening and narrowing your scope.

Birgit Starmanns: Definitely, definitely. It’s always, you know, those parameters. Yeah. If you have a certain amount of things, you can also always set the parameters or you can have it evaluate everything. And just actually from a human standpoint, say, okay, well, these are outliers anyway, or here’s my low risk versus high risk scenario. So, and that depends on how those drivers are set. So you’re very right about that.

Jim Hunt: That’s great. Well, so as we near the end of our session today, I mean, I’ve got a question. Why isn’t everybody already implementing machine learning?

Birgit Starmanns: There’s a fear of the black box. Seriously. I think sometimes there’s hesitation because there’s a question, well, what is it doing? And I’ve configured these rules. Um, but yeah, going back to that example, I used earlier as humans, we learn a certain way and we don’t usually go back to a checklist as we grow up and we kind of grow into the experiences that we all have as human beings, but there’s a fear of the black box. So when it comes to the transactional level, what a lot of our customers have been doing is they’ll run machine learning, but they won’t let it post anything. They want to take a look to see what would happen. And if you run something like cash application, here areall the exceptions I would have had. If I run machine learning on this, here are the amount of exceptions that I still have to deal with, but these are all the ones that I don’t have to deal with.

Birgit Starmanns: And then as they see that those are actually the decisions that a human being would have made. They start allowing within a certain tolerance level to allow the machine to post it automatically. So most customers start out by wanting to see what it would do, and then going on and saying, okay, I’m just going to post this automatically because I trust the system enough. Um, and then also as scenarios and business situations change, there’s a question also that comes up. Do I have to retrain it? Well? No, because it’s experiential learning for the machine. So you don’t have to retrain it because as different parameters take place, for example, sheltering, um, things change. And the way that exception handles are changed and the machine continues to learn. Now, if you buy a whole new company, you might want to just start from scratch and retrain the whole thing, but you don’t need to keep retraining it.

Birgit Starmanns: It is actually continuing education where it learns as time goes by. And then in predictive scenarios, you certainly don’t want it to basically just post something. So I think there’s more of a comfort factor if it sees, Oh, okay, well, this driver was changed in this way to come up with this outcome. So being able to take a look to see how those drivers have changed is really helpful also. But again, there’s nothing like the human element to actually evaluate that. We wouldn’t want a machine saying, well, you are now buying company D that’s not going to happen. I don’t think that ever will or should happen actually, but it’s really supporting the teams in making those decisions.

Jim Hunt: Interesting. You know, it might be a bit of a stretch to use this analogy, but as you were describing that, it occurred to me, it’s a lot like, um, autonomous self driving vehicles where we might buy one, but then we get in and clamp our white knuckle grip on the steering wheel because we’re afraid of what it might do. And then it takes a while to learn that it knows how to navigate and we’ll let it go little by little until we get confident that it’ll take good care of us.

Birgit Starmanns: It’s actually not that much of a stretch. And since I live in Mountain View, I see a lot of self driving cars. I mean, actually in all of Silicon Valley, I see a lot, but they all still have drivers behind the wheel. Um, that being said, yeah, there are more and more stories of somebody really doing emails and keeping their hands off the wheel as they become more comfortable. But I’m seeing a lot of that. And actually, this is a great little story. I was out at lunch yesterday in Mountain View, and they’ve closed certain streets off to traffic, just so restaurants can expand into the, into the street for outdoor seating as they reopened. And there’s this little machine it’s a little bit of a bot that’s just like chugging along. Well, it turns out it’s a delivery bot for certain restaurants. Um, and they’re expanding the number of restaurants that do that for at home delivery. And it’s actually a machine that does that. And it’s actually really cute to see it react to even pedestrians because it’ll slow down or stop if a pedestrian gets in front of it, cause they didn’t notice it. Or I saw it crossing the street and it would actually hold off, um, when it sensed the car coming and it would stop before it would keep going. So it was actually really fascinating to see, um, kind of on a smaller scale than the car to actually see it be aware of its surroundings. So that was actually really cool.

Jim Hunt: That’s a great story. You know, it’s really two sides to the coin. We talk about machine learning, but people have to learn how to deal with those smarter machines and systems as well.

Birgit Starmanns: That’s very true. And it’s a comfort factor, right? I mean, years ago, when I started consulting back in 1990, we didn’t have cell phones, right. So you agreed to meet somebody and, uh, yeah, you better be there because there’s no way to reach them. Right. Um, and so initially people just even looked at mobile phones in a funny way and look at us now we’re all attached to them. So it takes a while of getting used to it, but then you start to see the benefits and then you become more and more comfortable with it. And then of course you end up not being able to live without it. That does not mean machines are gonna take over the world, but yeah, they definitely help.

Jim Hunt: And you know, just one last question, we can, maybe you can do a minute or two on this, but I’m wondering, you know, our audience, if they’re interested in moving their company toward implementing machine learning, what are some of the first steps that they should take in order to get into it?

Birgit Starmanns: I would say as with most implementation’s start with one scenario, because what you don’t want to do is say, okay, well, these are the 10 scenarios that I’m going to implement right now. Um, but then if you don’t understand it and you set things up a certain way, then all of a sudden, if you decide to change it, then you have to do it over 10 different scenarios. So I would say the first step is to decide where you’ve got your biggest pain points and probably start with a transactional and maybe a very simple predictive, right? So I’m going to deal with these open items, let’s say cash app for receivables, or GR/IR and keep it very specific to that because on that first project, as with any implementation, you learn a lot. So even if you’re an old hat at SAP implementations, because this works a little bit differently than, okay, I’m going to go find the rule and just change that configuration.

Birgit Starmanns: It’s a little bit different. Um, yeah. And again, you can always start from scratch and retrain the machine during, during the testing phase. So it’s not that you’re stuck with anything. Um, but start with one, just to get that learning before you expand into others. And then for predictive also start with something very easy. Maybe predicting cost center spend, instead of going whole hog into the here’s the corporate balance sheet prediction, this is what it’s going to look like, but yeah, let’s start at the cost center level and start with that and look at spend, right. And then maybe start with profit centers and look at the potential predictive revenue and the costs associated with that. So again, I would say low hanging fruit, um, I think right now nobody has an appetite for huge implementations anyway, but you learn a lot and you can show a quicker benefits, um, you know, of reduced error rates of quicker automation of freeing up your people to do more value added activities. But yeah, take it one step at a time would be my advice.

Jim Hunt: Great advice. Birgit, I never fail to learn new things whenever I talk to you and I really appreciate it. Thank you for being here with us today.

Birgit Starmanns: Thanks so much for having me, Jim. I enjoyed the conversation.

You may also be interested in…