RAAIS 2017 - Michael Feindt, Founder of Blue Yonder

Yeah, thank you very much. So I gave my talk the title 'Making Humans Work Smarter with Vertical AI Solutions for Supply Chain and Pricing.' That's what we are doing. And actually, we also actually hire more people after they introduce AI systems, and they can't afford it because they are more effective and they do make money, and they can afford it and make it easy. Work with the humans in the companies also better, smarter, and better. Okay, so I'm the founder of Blue Yonder, and we bring digital innovation and disruption from Germany, so also there something happens. And the main background comes from... so I'm originally a particle physicist, and I'm a professor at the Karlsruhe Institute of Technology for physics. And have to get meaning out of that. Okay, what we are selling now is really value for specific tasks through data and completely scientific software, which is nowadays called artificial intelligence. Okay, what makes machine intelligent? So in my view, there are two main branches in artificial intelligence. The first one is to learn that a human can do well, but the computer cannot do easily: that is really seeing, recognizing images, understanding voice, text, things like that. And there are more to law and people move on networks have done, and all of you know that there's tremendous... concrete problem, but then also data and active machine learning in order to become as good as the best human or even better. So a good example for that is beating the world champion in Go, for example. And sometimes that is also not really profitable, but for commercial applications it's important it also has to be affordable. So you cannot always build and train huge networks if your dataset is already many billion individual items, too large. You also have to think about resources. So it's not only deep neural networks, also other algorithms. And to my important personal and professional once-in-a-lifetime decisions, they are done by gut feeling and the Swiss way. So this will be completely human always. I'm completely convinced that in a normal enterprise, the decisions that are taken there can be automated by artificial intelligence driven by actual data, and more than 99% of these decisions can be automated. And that's what we are actually doing. These are decisions, these are repeated decisions which come again and again in a similar setting. For example, if you're in a supermarket and you have 20,000 articles, which to order today? Which article to order today? And a million or two million long list. So what people actually are doing when they again and again have to decide, seeing that most of the times they do nothing, or it's the same as always, just before. Don't think, some use business rules which might be good or not be good. And really to think about a single decision: how many should I order? Oh, now the weather is so and there's a holiday and so on. This almost never happens, very, very seldom. And we have to also know how the human brain functions. Our decision-making is done mainly 99.9% by the first system, system one: it's a fast intuitive system, gut feeling, experience. We don't think really originally. And only very few... Prize winner Daniel Kahneman said that even experts make many, many wrong decisions because we are humans and our system one rules, not the rational system. And system one cannot speak statistics, for example. And that's very important to know. So what we are doing is predictive analytics and prescriptive analytics. I will show what it is in the next slide. And these are both disciplines of machine learning, and machine learning again of artificial intelligence. Many people who are not experts are completely confused and they think, 'Oh, now machine learning is out and artificial intelligence is in.' This is of course nonsense, but non-educated people don't understand these things. So I also always show transparencies... across the utility function: what does it cost me or how good is it for me if the truth will differ from my prediction on the low side or on the high side? I then do an optimization for a decision that I have to take. And if I have to do all this here very often, automate this. This is a recipe for better decisions in the age of AI. So when can we use predictive analytics? Almost very often you can. Most people think that the world is actually two or have a deterministic view of the world. So here, of course, an expert can calculate what will happen exactly, right? But on the other side, you have some drawing of lottery numbers and it's completely random. But every week the same is done, but every... deterministic, for example, how many of these apples will be sold tomorrow? Right, so it's not clear today, we don't know. But there's so many, this is also a chaotic system and we can never say for sure how many will be. But of course we can say something: it will depend on the weather, on the price, how they look, and so on. Right, and that is exactly what we are doing. So with many, many inputs, for example the weather, but also the weather forecast for tomorrow, the price, promotion, competitor prices, embargo, so many possible influences are there. And we have many individuals, what we call events. There are many items in the store, a supermarket chain has many stores, and we have... columns, right? So a lot of values of measured quantities. Okay, a prediction for a standard is just the number 250 apples. But we know it will not be 250; it might also be 260 or 280 or 100 only, right? For us, a prediction is a complete probability distribution for each of these possible futures. That gives more information, right? And once you have this, then you can also do a complete risk management of every single decision, right? And this contains all information that you can give. Now it should be as narrow as possible, that means as precise as possible, but not more than allowed by the laws of nature... on inside, right? And Blue Yonder software is there's a big difference. But now the probability distribution is difficult to handle, right? So nobody can do something with it apart from mathematicians. But what we can do to account for the best ordering decisions, we have to know the cost function, utility function, that depends on, for example, you don't want to be out of stock, you don't want to have lost sales, but also you don't want to throw away an article after the end of the shelf life. So depending on input on purchase prices and on shelf life and other quantities, one has individual cost... million, about 15 million such decisions per day, right? And the result is a recipe: order so and so much, right? But of course if you have so many, then you should also automate that. And that automation is really important. I show you on this slide here. And so this prescriptive analytics gives you a list with all what you should do. But then people still don't do it because they don't believe it. That's their system one: 'No, no, no, I know I should buy less, and more there.' And so this is real data from a client from us. So he had an out-of-stock rate of about 7.5% before we started. Then we started giving prescriptions, so it went down to about 5%. And then... were not accepted by the human factor, right? And then it states, and that was a decision at a constant overall stock level and waste rate. So you can model it, bring down the out-of-stock rate by a factor of 10. So what you can achieve, and this is also very important here. And so any decision management decision is always a compromise between optimizing different KPIs, right? If you order too many, then you have waste; if you order too few, you have out-of-stock. So whatever you decide, you are somewhere on such a curve, right? But if you want to improve one, you are getting worse in the other. And by breaking down now this management decision to every... essentially you go to a more efficient frontier curve. Okay, so what we achieve in supply chain in retail is: 99% automation, less waste, better freshness. So the freshness of each single article in the store is on average longer lifetime, less capital binding, less out-of-stock, more turnover, more efficiency. It really works great. So one example here from the UK is we did it just last year for Morrisons, number four retailer. It had the largest growth rate of the classical retailers growth rate in this year, and its equity price went up again. And so this is very, very positive. And this now also you can ask: are people working at more than fine with that or not? Or do they feel bad? And we have just had in The Grocer this article which was very nice. Now what we are doing is actually we make our customers work smarter. And Martha became store of the week, and this is a quote: 'A new system, the replenishment system, has made a big difference not just in terms of maintaining great availability but on... high in the store.' And there was a mystery shopping, and the Morrisons one here completely against after same story. Tesco and Rachel's because of that, so very positive. And people like it. They don't want to sit every day in front of the screen and say 'yes, yes, yes, no, yes.' They don't want that. They want to really serve the end consumer. And that's a very more powerful for how far the German grows out. We do meat replenishment, the complete supply chain up to the production plant. And also that is very well, and it reduces write-off. And write-off for me needs also if you bring it down the complete supply chain that... is a completely data-driven company, and that gives a big advantage in everything. So for Otto, for example, we are also doing predictive shipping. They originally had five to seven days to deliver for their marketplace articles. What marketplace items? Now we reduced it to one to two days because we make already predictions of what will be ordered tomorrow, the day after tomorrow, and so on. And they pre-order from the suppliers and have it there just in time. And so that was also very, very successful, and it will be rolled out now to more than two million items. Pricing is another thing. And pricing is also... to distinguish between correlation and causal effect. We have to build the causal effect of price changing, given all the rest from historical data and actual measurement, and then to adjust prices depending on online. It may be very often offline in brick-and-mortar stores, it's very easy, very simple. And what we achieve again is something like 99% automation, more market share, more turnover, more profit. So these are things that usually cannot be achieved to maturity. And also if you make other prices higher, you get more market share but less profit, right? All our clients are very, very worried here, but that's very important. So this is accepted. And it was very interesting, for example, for promotion, no rest at the end of the season. So what it's called markdown pricing, one can do this with artificial intelligence way better than just keeping price constant all the time and in the last weeks you almost give it for free. So this is an example from my old life here. So artificial intelligence can actually also change the very intellectual work. And what we have done here is to actually change or automate a part of the work that the research physicist... and found that usually a PhD student has to do 72 decisions in such an analysis. And we automate it and optimize all of them, and run on the data: 10 years of data taken by the experiment on the new program. And the result was that it found more than two times as many B mesons that the task of the experiment correctly reconstructed than the 400 physicists in ten years, including myself, two before together. And yeah, so it was a work of artificial intelligence and three PhD students corresponds to about 500 normal PhD theses. And in the next generation of accelerator experiments... what's interesting is this experiment cost the taxpayer just the operation of ten years at 700 million euros, and we doubled the physics output of this with this work, right? So way more efficient. Okay, another thing that we are also doing for particle physics has not so much to do with that, but our decision algorithm, the neural basis, we also implement it on a chip because for the next generation of experiment there will be so much sensor data that it cannot be read out into computers anymore. Not because... so often collisions and so much data that impulses which will read out now for the time being at CERN about 30,000... put our decision on an FPGA and put this directly on the plate of sensors, so that it decides which part of the sensor should be read out at all. So could be interesting for Industry 4.0, of course. And this has now a world record of, we think, 8 billion decisions per second, intelligent decisions per second. Okay, that is where I think really an important task of AI will be: this vertical end-to-end solutions, so specialized providers as Blue Yonder combined expertise and experience solving complicated business problems by AI and machine learning, and combined expert domain knowledge with AI... that you showed before. And every enterprise can be optimized by AI. Yes, I completely agree, and that is exactly what we are doing here. And it has to be put into consumer the robust solutions, right, for specific tasks, and ready to deliver proven value and competitive advantages are fast to be accepted and to be used. Yeah, I'm ready. And the last transparency: to be ready and accepted and used, this is important that it is also consumable for non-experts. Most of the people in the world are not experts in AI, of course. And I also think we have to hide complexity and provide it as a service, and make and allow... anybody has any?

Okay, good question. And that is something that by production and experience goes down and down and down. So actually, from the first contact there to actually have it run on first tours, that took about three months. And to complete... company, and that is exactly where we work on that. This code is as smooth as possible and it's assisted as well as possible.

Yes, especially if one has no history for this article. So we do all this also for completely new articles. So that is an important thing. By just what we know: what the brand is, what does it make, and what is it, and how near it is to articles that have existed before. And then we learn. And in this case, of course, the width of the distribution is larger, right? So we are not as certain as if we already know how this article behaves really measured. That's another thing. But that's a good thing since we always have the complete probability distribution with the uncertainty. This is completely handled automatically. And all the... yes, we also put add external data from where it's good. For example, weather forecasts, historical weather forecasts, and so on. That's not so easy to get, right? And of course holidays and whatever, also based on the location that we just are. Alright, thanks so much, Michael.