AI in Intralogistics: Customer Benefit is Decisive

Helmut Prieschenk from Witron (pictured) and Franziskos Kyriakopoulos, founder of 7LYTIX from Linz, Austria, have been discussing ChatGPT, machine learning in logistics, and demand forecasting for food retailers. Both agree – AI technology offers a wide range of optimization potential for optimizing processes in the distribution centre as well as the entire supply chain. But high data quality is not the only crucial factor. Equally important for the data models are the experiences of people and the requirements of consumers.

“And then overnight everyone was an AI influencer,” joked Prieschenk, Managing Director of Witron. He wanted to talk about industrial AI, demand forecasting, and a bit about ChatGPT. Kyriakopoulos and his team develop machine learning solutions for the retail and industry sector. He is physicist, while Prieschenk is a mathematician. “That’s a dangerous mixture.” Prieschenk warned. “Of course, we have already dealt with LLMs (Large Language Models) at Witron. However, I would ask for a certain serenity. The world will not come to an end through their use – and we are continuously verifying whether such tools are suitable to reasonably help our customers or our developers with the implementation of concrete customer requirements.”

Kyriakopoulos agreed, but already outlines applications. “LLMs are good at processing sequences – orders, debits, sales, or customer communications. That can be used in intra-logistics as well. There’s a lot of hype, a lot of influencers running around spreading half-truths.” Witron has already experienced this, Prieschenk says. Competitors to the OPM system were advertising AI in the stacking algorithm. “But the results can’t beat the functionalities of our Witron OPM. These weren’t developed with AI, but with a great deal of human intelligence, based on solid software development, intensive communication with the users, and years of practical experience. We always have to take a sober approach. Our customers are basically not looking for a new tool. They have a problem and need a working solution that optimizes the logistics process in the distribution centre or in the supply chain, that works stable in practical use, and can be usefully integrated into a grown structure.”

But isn’t this soberness holding us back in Germany and Europe? “I certainly need a ROI”, Prieschenk strongly emphasizes. “LLM developers have a burn rate of $500 million per year and need another few billion”, said Kyriakopoulos. “That would be inconceivable in Germany or in Austria.”

Are we taking too few risks? Prieschenk is sceptical. “I don’t think so. When I look at the investments in Q-commerce, for example, I get dizzy. That’s where a lot of investors took a full risk. But the market has developed into a completely different direction. Predicted growth rates failed to appear. In the meantime, consolidation is taking place. Investors have moved on. Our retailers want AI and are investing in the technology. But we and our customers need AI tools, such as sample or image identification, that are transparent to then solve problems that we couldn’t solve before or could only solve with a lot of effort.”

The 7LYTIX developers work with LLMs, but the focus is on demand forecasting. “We can provide added values, but some companies often don’t understand at the beginning what the added value of the model will be. More sales through better communication with the customer or lost sales? Many people can’t calculate that. That’s where they need help from us”, stated Kyriakopoulos. Prieschenk adds: “Our Witron customers can calculate very well and have perfected their business over decades. But I understand what Mr. Kyriakopoulous means: First, we need to clarify what is to be optimized. The retailers ask themselves whether they want to optimize the supply chain network, the warehouse size, whether they want to be closer to the customer, whether to reduce throughput times, change delivery cycles, reduce food waste and stock-out, or have less stock in the warehouse. In this respect, we have learned a lot together with our customers from different parts of the world. We also learned that the requirements for bank holidays in Finland are different from those in the U.S., or that a Monday holds different requirements than a Thursday.” Kyriakopoulos agrees. “We need a requirement first and then a corresponding AI tool. And we don’t need deep learning all-around.”

How much accuracy is required?

How does his demand forecasting work? “First, we need to obtain an overview of the data. This is laborious work for many retailers. It’s not only about stored goods, but also about the amount of goods in the store, how much was sold, which influencing factors such as promotions exist, how many lost sales are in the store, and much more”, explained Kyriakopoulos. In addition, there are customer cards, seasons, the location of the store or special offers. “And we need to know what’s in the distribution centre, in the back room of the store, in the trucks on the road, because optimization does not end in the store. It is also important to avoid cross-company or cross-divisional restrictions as well as data lakes. A major part of the required data is mostly known, but different departments unfortunately pursue different interests.” Prieschenk agreed: “Even holistic logistics design should not only focus on the distribution centre or the key interests of individual logistics areas, or process-influencing departments such as purchasing or shipping. It’s important to include the entire supply chain into the optimization process – both internally and externally – and to avoid silos as much as possible, both physically and in terms of IT.”

“The data flow into very simple models”, continued Kyriakopoulos. “The baseline is the people’s experiences. That’s not AI yet. We talk about regressions. Then we ask ourselves if we became better. This is followed by time series analyses and first machine learning methods. We always have to look at how much accuracy we can achieve through the next level versus how much is the added value for the customer and user.”

And Witron? “We have to make sure that the mechanics fit the model. Because physics must work in the same way. Do we supply cases or pieces? Or one item with both options? How often is a store delivered? What happens when the product range changes?” answered Prieschenk. WITRON logistics centres create flexibility for both the store and e-commerce. The key to successful implementation, however, is to think the process backwards throughout all channels – from the consumer to the distribution centre and, if necessary, even further back, all the way to the supplier. He sees a challenge especially in the explainability of the model. “We experience push and pull systems with our customers. Some work better than others.”

Will store managers let an AI model specify their orders in the future? Kyriakopoulos knows the argument from the fashion industry. “If someone has been shopping there for 20 years, then it’s difficult to immediately explain the added value or to convince the consumer that this model might be better. But we make it transparent – we say which factors we use, how we weight them, and where the respective factor applies.”

The human being has the control

The experts from Austria can look 18 months into the future. They use interfaces to connect the model to the existing systems of the retailer, the steel manufacturer, or the shoe retailer. “I do not want to tear everything down to use an AI model”, Kyriakopoulos laughed. “This is the right way – the integration into existing architectures”, confirmed Prieschenk.

But how robust is the model? Keyword: Covid 19. “We weren’t able to see that either,” explained the Austrian expert. “We were working with the model in frozen logistics at the time. The short-term forecast wasn’t good at the beginning, but after one week, the model worked again. After two weeks, it was stable. But the forecast alone is not enough. The customer has to work with it – for example strengthen marketing channels, running promotions, or adjusting prices, if necessary.”

“That’s crucial,” Prieschenk said. “This is when people take over control. Never underestimate the gut feeling of a logistics manager, service technician, or store operator. People’s experiences and a well-functioning data model are the basis for making intelligent – i.e., right decisions in the long-term. In the distribution centre, this also applies to the implementation of maintenance strategies or the ‘correct operation’ of the system. And importantly, the models, tools, and solutions have to be stable and prove themselves in practical use, delivering real added values in day-to-day business.”

AI provides information, the person in charge decides and continues to have control over the process. “We revolutionized physics in the logistics centre over 20 years ago. With the OPM solution, we have managed that goods are automatically stacked onto pallets and roll containers without errors and in a store-friendly manner. Now we are taking the next step and opting for data and end-to-end logistics models. And I am sure that I will still experience an end-to-end Witron AI model for the warehouse,” predicted Prieschenk.