#15: Demand AI: Transforming supply chain forecasting for competitive edge

Søren Hammer Pedersen (00:10):

Hello everybody. Warm welcome to this S&OP MasterClass from Roima. My name is Søren Hammer-Pedersen. I’m the chief commercial officer for PERITO IBP within Roima and I’m a host for this podcast. The purpose of these S&OP Masterclasses is to dive into trending topics, hot topics within supply chain planning, give you our perspectives on these topics and hopefully some tricks and tricks that you can use at home in your supply chain planning.

(00:39):

Today’s topic is no different. We are looking into demand AI and have been really looking forward to this topic. And you’re in luck. I’m not alone in the studio. I brought my very good colleague, Benjamin Obling, here today to help shed a light on this exciting topic. So welcome Benjamin.

Benjamin Obling (00:54):

Thank you.

Søren Hammer Pedersen (00:55):

Before we go into this topic here, there are some people out there that might not have heard about Benjamin. Guess so please give a bit of an introduction. Who are Benjamin?

Benjamin Obling (01:07):

Yeah, so my name is Benjamin Obling. I’m overall responsible for onboarding of new clients on the PERITO IBP in Roima and the continuous operation with our clients. I’ve been working in the integrated business planning for the last 15 years, focusing on demand planning, inventory planning, supply planning, before that in financial financial planning.

Søren Hammer Pedersen (01:29):

Perfect. You sound like you had the right experience for this. Let’s get into it. Demand AI. Could you start by elaborating a bit, what is demand AI?

Benjamin Obling (01:40):

Yeah, so what we are focusing on with the demand AI here is creating the best possible forecast or prediction of the future sales in general. It could also be consumption for consumption-based companies, but in many cases it would be the sales. So that would be down to the single SKUs, the different warehouses or sales channels. That could be like countries or customers, et cetera. So predicting the future sales for the next two years into the future, for example, down on a monthly, weekly, or even daily basis, depending on the requirement of the business.

Søren Hammer Pedersen (02:14):

So taking it down to the ground level, it’s basically adding new functionality, new techniques into an existing process, especially around the forecasting to basically enhance what we are doing there?

Benjamin Obling (02:26):

Forecasting as such. There’s no new thing and that’s been around for ages. So this is really about getting the higher accuracy on a existing process. That can be a super manual process in some companies using Excel. Surprisingly many companies, you would say, doing that with the rolling a 12-month average or something like that. So really basic techniques. It could be like statistical time series-based forecasting and then the now advanced algorithms, using deep learning networks because we can see we get a higher accuracy on the demand forecast and we can also increase the automation a lot in that area.

Søren Hammer Pedersen (03:01):

Okay, exciting. I know that Roima and PERITO IBP just launched a new module regarding demand AI. What was the reason behind that? Why didn’t you pick all the other fruits out there that could be the next module? What was the inspiration behind that?

Benjamin Obling (03:17):

That is of course, that is because the demand forecast is really the fuel, the starting point for the full, integrated business planning. So if we have the wrong forecast, the demand forecast, everything else will be wrong. We can set up safety stocks, balancing the inventories, making our supply plan, looking at manning machinery capacity, all of that, looking at different sourcing flows and so on. But if we’re setting those sourcing flows up, balancing the inventories, making our revenue predictions, purchasing raw components based on a wrong forecast, then it’s all going to be wrong. So basically if you have garbage in, garbage out. And then we can see that there is now really strong methods, really strong techniques that can improve the accuracy. So for that reason, we dived into that some time ago.

Søren Hammer Pedersen (04:03):

Okay, and you say there’s really strong techniques, and we’re going to dive into those in a minute, but will this, in your opinion, be a game changer? Because coming from supply chain planning, of course we have been talking about demand planning for a long time. It’s a natural part of what we do, but will this actually change things in your opinion?

Benjamin Obling (04:24):

I think it is, you could say it’s an improvement in the accuracy and, as such, it will trickle down to all of the other areas. It’s not like it won’t fix all supply chain problems from day one to day two, so that’s not the case. And you’ll still have what we call random walks. So there will still be changes in the sales patterns. So when does the customers purchase the different goods? That will just be a matter of random walks.

(04:50):

So even with the best possible model, it’s not possible to predict because it is rolling a dice, whether a client, a customer of a company is calling on Friday or on Monday or et cetera. There are things here that you cannot control. You cannot predict it. There is basically not a model that can do that. So you’ll still need to have the inventory balancing, the supply chain balancing, leveling, et cetera. Being prepared for the unknown will still be relevant to do afterwards, but the fuel and the input will be better, but it won’t solve the rest of the chain.

Søren Hammer Pedersen (05:23):

Okay, exciting. Let’s take a deep dive into the techniques behind this because I know many of our listeners would love to know more. How do we do this? Could you start by saying what kind of techniques are we applying into these models, these techniques for using demand AI?

Benjamin Obling (05:42):

Yes. Yeah, happy to dive into this because it’s also something where there are a lot of buzzwords out there, a lot of talk about AI. So what do we actually do and which techniques are we doing? So let’s dive into that. So what we are using is neural networks. We’ve been investigating various different techniques like machine learning, XGBoost models, et cetera. So there are a range of existing models that you can use, profit or other techniques. What we’ve seen so far was that they really don’t enhance our demand forecast. We’ve been doing demand forecasts for our clients for the last 15 years. We have hundreds of thousands of objects, so that would be the lowest level, like item number, et cetera. So hundreds of thousands of objects that we predict every month, every week, every day, depending on the requirement. So you can say we had to beat that element, having our analysts looking at that. So you could say having more simple models, like what I would call in the machine learning models, like XGBoost, et cetera, for example, to be a bit nerdy here.

Søren Hammer Pedersen (06:45):

That’s okay.

Benjamin Obling (06:47):

That really didn’t do the trick. So we were not able to beat our own forecast with that. So for that reason, they were discarded. Another element that we found out was that, by default, AI models are black box models. So the way that it works, especially when you go into deep learning, is that it works like the human brain. So you have some inputs. Those inputs here would be the sales on the different items, on the different month or weeks and days, et cetera, in the past. That would be the input. Then you have some hidden layers and now it gets a bit technical, but those hidden layers are the neurons, just like the neurons in our brain.

(07:25):

They are then passing on. They’re multiplying that input, passing it on to the next, and then you’ll have an output, which will be what is our prediction of the sales. When we have those input, these weights are set and that is a bit like or actually completely when we are educating our kids that this is a hot plate. So when they start to put their hand on a hot plate, we say, ”No, don’t do that.” That will do what, in machine learning or AI models, is called backpropagation.

(07:53):

That’s the fancy word of actually saying now we need to change the weights in our brain so that when we are approaching a hot plate, we’re thinking, okay, now the weights are telling me that I shouldn’t put my hand there. It’s exactly the same. Use backpropagation in deep learning models, neural networks to say, ”Okay, when we have these sales patterns here, what is then the likely next output? Should I put my hand down or should I take my hand up? Should I predict 100 or 200 or 150 as the sales for the next month?” For example. So that’s the, you could say, the techniques behind the neural networks. The neural networks is the same technique that is used in large language models like ChatGPT, for example, where you’re predicting based on some input.

(08:39):

So it’s a bit deep into that, but you could say another important element that we also realized was that, so first of all, we need to use the deep learning models, so the more advanced neural networks like ChatGPT is using. That’s one element. And we also need to be able to separate the forecast into different elements because if we’re just asking the AI models, the deep learning models, to just predict the sales, sometimes we get something which is completely off the charts. So it’s completely random. It would be a bit like, using the ChatGPT reference again, it’s hallucinating. It’s saying something completely crap. And we can’t have that, of course. We can’t send that to our clients, we can’t send that into the supply chain, so that doesn’t fly. So what we did in order to do that was we put up some fences around, so how much can the AI algorithm actually predict? How much is it allowed to go? So it cannot go out of bounds.

Søren Hammer Pedersen (09:38):

Yeah, I think really, really interesting and a great overview on the techniques. Two points that I think are very interesting that might be interesting to people. It’s a learning curve, I can hear. It’s not just pressing a button, downloading something, going that. How long has that process been for you now, developing this new module?

Benjamin Obling (10:00):

I’d say that’s over the last years, actually two, three years then being intensively more and more, you could say intensifying in that area. So it certainly is, there is no silver bullet in that. You can ask ChatGPT, it will give you a model and it might give you a fair forecast, but you’d say not something that we would be ready to send to our clients, for sure.

Søren Hammer Pedersen (10:23):

How much is the need in these techniques to diversify between different types of companies, manufacturing companies, logistics companies, using this in there? Is there a need for a great different kind of approaches or can we get it to a level where it’s one fits all or something similar?

Benjamin Obling (10:44):

Yeah, it’s not a one fits all because it will be different parameters that goes into the analysis. And that’s one of the interesting part about the demand AI using deep learning models is that you can use additional indicators. So one is, of course, the historical sales on the item number. That will be a pretty good indication of what is going to happen in the future and that will stay like that of course. But the next part is, okay, how is the group of that performing?

(11:11):

So if we have a certain soft drink, how is the whole soft drink business performing? Or what is the purchase manager index saying? What is the GDP growth or other external indicators? And that will be different across the different companies, of course. So you could say if you’re an agriculture, it’s a different external indicator you would use, as compared to others. But the general model on saying, okay, so how do we predict the single SKUs and that we can take X number of variables into it, that can be generic. So there is a lot of scale effects in this, but you need to know the business and the industry.

Søren Hammer Pedersen (11:51):

Interesting. So we have a big core and then we adapt a bit to the individual. And I guess also, in terms of it can also spawn off individual, maybe smaller AI analysis that is needed for a specific company. If they have a need within that demand planning, then you could spin off different kind of AI analysis that would need for that specific company?

Benjamin Obling (12:13):

Yeah. Examples of that could be like in promotional planning, in tender planning, for example-

Søren Hammer Pedersen (12:20):

Yeah, pricing maybe.

Benjamin Obling (12:21):

You have pricing, et cetera. So others where you could say that’s, in general, you could say that would be what you could call a regression problem. So there you have a bit like, if you remember the regression analysis from school, it’s basically the same, but now it’s just using neural networks in order to do that, what you call sequential regression. The interesting part is you can add in a lot more indicators and it can be non-linear, so we can find some patterns that are not linear and that was not possible in the old days. And it will be a lot better. The AI models is a lot better or stronger in finding the correlations between, and also if you have different indicators that are correlating, that was a problem in the earlier days. Now the deep learning models is actually able to distinguish and handle that. So the feature engineering is done by the model itself, which is interesting because that was a very time-consuming exercise earlier.

Søren Hammer Pedersen (13:19):

Last point before we maybe dive a bit more into the business value is you mentioned fences that we have built into and that is needed. How does this fit into the planning ecosystem, especially on the demand side here? Is it just any new demands for the demand planners, the organization in general, that you have been becoming aware of, implementing this?

Benjamin Obling (13:46):

Yeah, so you could say it’s the same type of integration or interaction you have with the demand plan. You just get a better forecast in the beginning and then you still need to align that with the market expectations. Some of the market expectations can then be taken into account earlier because you have these indicators, et cetera. But then you still need to have the interaction afterwards with the market intelligence because there are things that the algorithms or the network cannot know. For example, new product introductions, so the promotion is coming or it’s not coming. It wouldn’t have a clue about whether, or a new product coming in. It doesn’t know that.

Søren Hammer Pedersen (14:29):

No, of course.

Benjamin Obling (14:29):

When you tell it it’s coming, a new product for example, and if you tell it, ”This would probably look like these four different items,” then it can make a good prediction for you. And you can say, ”Yeah, but we expect it to be a better version of these four, so the level will be here, the seasonality will be like this or the rhythm,” et cetera. So again, it can help the planners a lot.

Søren Hammer Pedersen (14:50):

Yeah. Okay. Interesting. Let’s talk value a bit here because we are doing this for a reason. It improves things. So could you give some, now we have been rolling this out and looking at different type of content. What kind of real-life example, what kind of benefits have you seen so far from these new methods?

Benjamin Obling (15:13):

Yeah, so one element is the raw prediction, you could say. So we have a higher accuracy, so that’s good and that will trickle down, as we discussed earlier. Of course, that will mean that you don’t need as much buffering in the inventory because you have a lower demand uncertainty, you have a better prediction in terms of new hires, new factories, production lines, et cetera, because the prediction is stronger. So that’s really the main benefit here.

(15:41):

Then you have the better alerts. So that means we can also use the models here to predict alerts saying, okay, when you are a demand planner, where do you need to guide your attention? So you have, let’s say, 10,000 item numbers. You get an automated prediction, but where, if I have two hours as demand planner, where should I spend my time? Where can I add value to the automation? There, we are using AI also to generate alerts. So to say here, something extraordinary happens with the sales. It could be it’s going up, it’s going down. For example, the sales is stopping, our market intelligence is out of bounds, seems unlikely. Is that absolutely certain that this is correct? Because you’re normally overestimating and now you’re doing it again, it looks. So type inputs like that.

(16:25):

Which can really, again, mean that the alerts and the demand planners or the market intelligence together with the algorithms and the AI networks here can then react faster because the network and the automated prediction will pick it up eventually anyways. If demand is increasing, it’ll also increase, but the question is do we increase it right away? If we have, let’s say, one month, four weeks of high demand, should we then increase it directly? It’s like a very concrete example. Yes, in some cases you should and that’s good, you should react fast. In other cases, maybe your competitor just went stock out, for example. So it will drop down to the normal level. So how fast should we react? And that is actually not an AI model that can tell you that. That’s more a business decision on how fast do we want to react to an opportunity because you risk having overstock if you react super fast and it’s actually a temporary thing.

Søren Hammer Pedersen (17:19):

I think one thing you didn’t mention that it’s one of my favorite topics. I talk to this with companies all the time when we talk demand planning, it’s season detection. So many times, I see the situation where we have too much in the low season and vice versa. How can we help companies using these techniques on that?

Benjamin Obling (17:42):

Yeah, so one of the elements we do in our prediction using the networks is that, instead of just the next sales, we are actually splitting it up into a seasonal element, into a trend element and a level element and that season element. So we’d separate that out and then we detect that either on the lowest level on a higher level, and there we use the neural networks again to set the parameters. And the good thing about that is that it’s actually doing some of the data cleansing. So if you have a very high peak in your high season, but it’s more extreme, it will actually remove that and figure out what is the normal season and capture the season really good, actually surprisingly good. It’s one of the areas where we can really see a huge benefit, together with the level where you can then say, okay, what is the level of the future demand?

(18:35):

So capturing the seasonality, I would say right now I have a hard time seeing actually how we would do that better, what we see now, because it is pretty good. But you could say another advantage of splitting it into these actually traditional elements of seasonality level and trend is that then we can control the output. Because again, back to the hallucination thing, sometimes it will just say, ”Okay, it will go exponentially up,” or it will jump to a crazy level because it’s picking up something. It’s put some weights, the back propagation, the hot plate on the hand went wrong somehow, or there was a certain combination and we can’t really find out why because it’s a black box. It would be a weight set on network number on level four in the hidden layer of the neuron, blah, blah, impossible to find. So here we set up rules to say, ”Okay, but you cannot go out of bounds in this way.” So then we know what type of forecast. We’re certain what we are getting out of it, that it will not go crazy. That’s pretty key.

Søren Hammer Pedersen (19:38):

That’s pretty key. So crystal clear benefits here. We will improve the accuracy, we’ll prove the way we detect things, we will improve the input we give to the rest of the supply planning, basically, on this. So no-brainer, we should go forward with this.

Benjamin Obling (19:57):

Yeah, and then automation that I didn’t mention before. That’s also, of course, a key driver. You could say companies that are doing this in Excel, for example, extracting data, making VLOOKUPs, making a model more or less advanced, trying to find the seasonality, maybe doing a good job in some cases, but spending a lot of time on it. So what we normally see is that, when that’s the case, when that’s what you’re coming from, we can both increase the accuracy and we can reduce the load a lot or actually remove the prediction work. So the raw prediction work is completely removed.

Søren Hammer Pedersen (20:33):

Yeah. Cool. So if there’s a check mark with the value, and I hear there is the next question, of course, people are thinking out there, okay, I want this, how easy is it to get started and implement if you want to go this direction with demand AI?

Benjamin Obling (20:55):

Yeah. So you need to set up different ways. You need to set up the model, you need to have these learnings. So it took us quite a long time, also longer than I expected and hoped, actually, because the first predictions using the normal machine learning techniques or what is standard, being very nerdy, down to the deep dive, installing Python, using some Python libraries and so on, you’ll get something out of it. But again, it’s not good enough. You can’t control it, et cetera. You need to set up, you could say, fences around it, having the logic, the market intelligence on top and so on.

(21:34):

So you could say there, the easy answer is, of course you installed PERITO IBP. But otherwise, I would really recommend work with these fences, work with the rules on top of the AI because you need to fence it. You need to put it in a, say, lock it somehow. Otherwise, you really need to do validations afterwards on the prediction that you get out of it, so you ensure that it doesn’t go crazy and hallucinate.

Søren Hammer Pedersen (22:04):

Okay. So you need to have, as an organization and, of course, as a supply chain director, you need to have a healthy respect for the work that needs to go into this. If you, as an organization, want to go into this, you need to draw in the right experts that knows about it. It’s not just downloading ChatGPT or going on that website here.

Benjamin Obling (22:25):

No. No, it’s not.

Søren Hammer Pedersen (22:26):

No. Then let’s play with the thought that it could be, of course, using PERITO IBP, but also, as a company, actually doing the effort there and have the demand working now. One thing that I would be worried about or at least very aware of, especially on the change management side of implementing this, is how do you build trust in this? Because the black box has not gone become smaller, using AI in all of this. So how on earth do we actually get trust into the organization that they should actually use this data-driven input into their planning?

Benjamin Obling (23:09):

Yeah, because if you just compare it to, let’s say, a rolling 12-month average, that would be-

Søren Hammer Pedersen (23:14):

Excel.

Benjamin Obling (23:14):

That’s the super simple way of predicting, can actually be a fair model in some cases. And if you’re not managing your AI model, deep learning model, I might even suggest that you start by using a rolling 12-month average, just to say. Then you don’t capture the seasonality, you don’t capture the trend level changes, et cetera, but you’re sure that it won’t go crazy. But you could say building the trust is then putting up these fences, of course, making sure that the proposals are good and they are managed, making the alerts so that it’s stating clearly, okay, where do you need to look as a demand planner? This is out of the ordinary. I’ve made a prediction. It’s the best possible prediction with the neural network. I can’t explain it because that’s how neural networks work, but I can guide you towards saying, ”Okay, this might be crazy, but it might be correct. Please give a market intelligence input.”

(24:12):

And I think it’s important to separate between there is a raw prediction, which is the model work, the rolling 12-months average and the super basic, the deep learning models, or you could say the controlled deep learning models that we are talking about. That is one exercise. The next exercise is then putting in the market intelligence on top of that. So having the market expectations, new products, changes in the market, overall expectations and so on. You need a way to show, say, the sales organizations for example, where we would like you to spend the time evaluating the prediction we’ve made. Unless you have a super stable demand, then it’s all good and perfect, but very few companies have that. Changes in the portfolio, changes in markets, customers, et cetera-

Søren Hammer Pedersen (25:02):

Of course.

Benjamin Obling (25:03):

That’s the problem.

Søren Hammer Pedersen (25:04):

But I also hear you’re saying that you need, as an organization, to allow time for the proof of the data. Basically when we implement this, people need to see with their own eyes that this actually works. So don’t expect to be 100% up and running or 100% automated the first month. People need to see and build the trust over time, I guess, also in the data here?

Benjamin Obling (25:29):

Yes, absolutely. Run it as a parallel forecast in the beginning, et cetera. So make sure that you don’t have 2,000 bits and pieces coming in with a container in the warehouse. That’s the way you experienced that it went crazy and it was hallucinating, but the way you find out is when you look at it before you release it into the MRP planning in the ERP system and before you ship off all the goods in the wrong direction.

Søren Hammer Pedersen (25:55):

So there is a role still out there for the demand planner. We are not subsidizing them out here. We are improving the input they get to do their work in a smarter way and hopefully a faster way to improve the input into the organization.

Benjamin Obling (26:09):

Yeah, we are. Certainly reducing the time. This is certainly reducing the time they need to spend on it, removing a lot of the work and then focusing the time to something that makes more value. So instead of extracting data from an ERP system, calculating alpha, beta, gamma values, like in the good old days with time series analysis and so on, you would now spend time on, okay, what is the prediction for Germany, as such? So how will the purchase manager index decrease or new tariffs, for example, how will that impact demand overall? How does that look? Where do I need to look for changes, et cetera? So guided towards that. So a lot more, you could say, business-oriented, management direction, strategic direction, instead of nitty-gritty pushing data around.

Søren Hammer Pedersen (26:58):

But I guess that’s also why this area is both a great opportunity for companies and a threat, in my opinion, in the sense that this will become a competitive advantage. Things are moving very, very fast. You mentioned let’s just use the rolling 12 months, blah, blah, blah on that one, but if you do that in three years time and all your competitors has gone down this road ...

Benjamin Obling (27:24):

Yeah. Then you’re going to be in trouble because you are going to ship goods in the wrong direction and make wrong purchase orders and wrong productions, and you’re going to have disappointed customers because the goods weren’t where they should be. So agree. I mean it’s certainly something that you should do. And also there are so many derived costs in this, in terms of transportation and scrap and unsatisfied customers, loss of gross profit, et cetera, that being good in this area really pays off, again, unless you have a very, very stable business, which very few have.

Søren Hammer Pedersen (27:59):

Yeah, and I guess it’s almost like back to the threat part. It’s almost like a silent or a killer that’s just coming, sneaking up on you. Because if your competitors is doing this, you won’t notice because they won’t say it on their website or anything. You’ll see it once their delivery performance increases, their inventories, they can have more in their inventories, things like that, that actually hurt your ability to compete in different markets.

Benjamin Obling (28:29):

Yeah. You could say you would learn it by hearing more and more from customer service that the customers are complaining that they don’t get their goods. On the other hand, you have the CFO and the owner complaining, ”Why are you spending so much working capital and not being able to deliver?” Then of course, again, making sure or here keeping in mind that we’ll never remove the uncertainty completely. There is an element of true random walk uncertainty in the demand plan, in the demand for us. We’ll not have 100% accuracy. That’s the nature. That’s the nature of it. And in some companies, 50% accuracy, for example, can be really good, as good as it gets. There is, let’s say, a hidden or an unknown max accuracy you can get because there is this element of randomness in the world.

Søren Hammer Pedersen (29:18):

But I hear a clear recommendation from you that this is an area that you should look into.

Benjamin Obling (29:22):

Absolutely, absolutely.

Søren Hammer Pedersen (29:24):

Yeah. Maybe the final thing, because time is also running here, is now we have been talking about demand AI as a very interesting area, but of course AI is not stopping there, I guess. So looking down the road here and the road ahead, how do you see this play out within supply chain planning? What are the next steps that will become massive?

Benjamin Obling (29:49):

Yeah. I could say in general it’s about predicting results. So other things that are interesting to predict is, for example, what is the lead time of a vendor, for example? So we have seen in the past that this is the tendency, et cetera. So how do we predict that standard lead time, for example, would be another element that is interesting. When we look at the single shipments, you have inbound and you have a certain date for that. What is the prediction that that date is actually going to hold? Because you might want to do something else if the prediction is different. So you could say predicting all of the inbound in the supply chain area is certainly also super interesting.

(30:29):

Also in inventory balancing, there are some very interesting techniques about reinforced learning and so on. That seems very interesting. Again, it’s about making sure that we can control it. So having target customer service levels, et cetera, is important compared to just getting a figure out. So being able to control it. So those are some of the super interesting elements as well. You could say in general, it’s about predicting wherever we have uncertainty and we can predict it there, it’s relevant. It can also be in classification as another, you could say, deep learning technique is classification, saying these demand patterns are normal or these supply patterns are normal, so we classify it as normal. These are out of the ordinary. Okay, we classify it as that. This is where we should spend our time manually.

(31:21):

So say classification and predictions could also be a customer churn. So what is the likelihood across our 2,000 customers? Where do we need to look? So if we then use classification and then say, ”Okay, how do we classify them into certain customers, we know they won’t churn, and others where we predict here there is a likelihood that they might churn? Okay, what do we do?” And this could be looking at customer service activity, their sales, how broad is it, how many order lines do they have? Which region are they in? What is the economic activity, blah, blah? Then saying, ”Okay, what is then the likelihood of that churn?” And then we can do something proactively about it.

Søren Hammer Pedersen (31:57):

Really, really interesting perspective, and I can hear the journey is not stopping here. The one thing I really like what you say or I have a great interest in is, of course, the prediction of master data in the sense that I would say every other meeting I have with companies in Europe, at some point in that meeting we talk about master data and we talk about that it’s no good in their company. It’s a huge problem. We can’t do anything. So the road on predicting and improving master data, is that going to be huge, I guess?

Benjamin Obling (32:34):

Absolutely, absolutely. And I think you can include the US and Asian as well. But you can say certainly, and everybody are embarrassed that the master data is not perfect, and that’s the case with everybody. Of course, there are different levels of maturity in this, but here you can certainly also use the predictions because it’s both the classification and also the prediction value in itself. So what is the MOQ? What is the lead time? What are the roundings, et cetera? The stability on the vendor and so on, the bill of material, are we actually using that? What is the scrap rate on the bill of material? How much time does it take routing, setup time, processing time, et cetera? Again, predictions, alert, classification. So neural networks would be a tremendous help in that.

(33:26):

And then guiding it. One element is the prediction and then you could just copy that into the ERP system. Wouldn’t recommend that, because it, again, might predict something that is off. But you could say, again here, what we do is then to say, ”Okay, let’s look at all the master data points that are super important in the planning. Let’s look at the ones that are really off. Okay, now we have that list. Then let’s look at the ones that we are going to use actually in the next period.” So if we have a master data point that is completely off, we have a better prediction with the AI model. But if we’re not going to use it for the next year, okay, why bother right now?

(34:03):

Okay, so what are we going to use right now? And then finally, what is, then, the economic impact of that being wrong? So if we need a manual validation, let’s do that based on the biggest impact. And those are the one where we are in doubt and then you’ll have the larger, the bulk or the lion’s share of it where we’ll say, ”Okay, here, we’ll actually just copy the prediction in because we are pretty sure that this is correct.” So we have maybe used multiple models to predict it and it’s coming up with the same, so let’s change the MOQ from 50 to 100. We’re pretty sure. We’ll just do that right off the bat, automatically. And the rest, let’s do it based on priority.

Søren Hammer Pedersen (34:44):

It’s so interesting. Time is running Benjamin, but so I think we have to wrap up this exciting topic, but before I do, any final thoughts from your side? Any recommendations for people looking into this area?

Benjamin Obling (35:02):

Yeah, jump into it. Dive into it. It’s super, super exciting and certainly a game changer in many ways. And it could really, really improve also the robustness and, of course, having the financial outcome. Get some learnings. Do it, fix it, try it, yeah, get started.

Søren Hammer Pedersen (35:19):

So if one of the first things you think of in the morning is supply chain planning and how can we improve, go this way?

Benjamin Obling (35:25):

Yeah.

Søren Hammer Pedersen (35:25):

Perfect. Thank you so much, Benjamin, for your-

Benjamin Obling (35:28):

Yeah, thank you.

Søren Hammer Pedersen (35:28):

Good input here today. And also thank you out there to you who watched and listened to this podcast here today. I hope you got some valuable input into this exciting area. We would love to discuss more with you, so if you have an interest in this, go into the Roima website or reach out to Benjamin and I. We would love to talk to you online, so feel free to do so. Thank you for your time and hope to see you next time as well.