Like most everything in 2020, global supply chains were thrown into disarray during the pandemic. Demand plans and equipment maintenance schedules went out the window with U.S. e-commerce volume growing 44% as consumers stayed home to buy everything they needed to sustain life. In response, many supply chain managers increased safety stocks to hedge against increased volatility. Now, as we begin to emerge from lockdown, everyone is looking for ways to boost supply chain velocity and efficiency.
Artificial intelligence (AI) and machine learning (ML) technologies seem a natural fit for helping wring out greater efficiency and better decision-making in this area. Steve Banker, vice president of supply chain services at ARC Advisory Group, wrote recently about a host of AI-driven supply chain use cases, ranging from those that are still hype-stage to those with established return on investment. Banker cites, in order from most hypothetical to most mature: blockchain, autonomous trucking, ML for warehouse management, robotic shuttle optimization, ML for transportation, ML for demand planning, real-time location services, and IoT for transportation.
Vendors are rushing to add AI/ML capabilities to their software. “It’s the arms race of enterprise software,” says Shaun Phillips, director of product management for QAD DynaSys. But there is real value to be had—now—for use cases like demand planning and warehouse management.
Demand planning, the low-hanging fruit
For many companies, demand planning is a good place to begin with AI/ML. According to Bill Panak, vice president of data sciences at Logility, there’s overwhelming proof that ML algorithms outperform the classic models that were used for building a forecast. ML enables auto-tuning, or automatic adjustment, says Panak, that is especially useful in so-called “black swan” events like the pandemic. The ML capabilities inherent in Logility’s demand-planning application can spot patterns and trends in all types of data (structured and unstructured, internal, partner, published sources, public) long before a human planner would. ML is extremely useful for helping optimize pricing and promotions, too, he says.
QAD also tackled demand planning as its most immediately useful AI/ML use case, working with clients starting in 2018 to understand what its clients really needed. “We sat down with customers at our user event,” says Phillips. “Some of our customers said it’s all hype,” recalls Phillips. But four emerged with interesting potential use cases, and they were willing to share their strategy with QAD.
The first use case was cluster analysis for demand planning. Since companies can’t do a demand plan for every SKU, they would choose an attribute and aggregate it up to all the SKUs with the attribute for planning purposes. But making these decisions at the aggregate level was not giving the right decision on a more granular level. “We analyzed millions of SKUs—when did they peak, when did they trough? What was the customer service level? Was it an esoteric item? What were the different types of raw materials used?” says Phillips.
The QAD DynaSys team built clusters based on shared common sales behavior. “They were getting smarter forecasts for less work. When they introduce a new product, they would align by the brand and product size and we put it into a cluster.” The result was a much more accurate plan. This application was based exclusively on internal data, including sales history, service level, unit cost, price point, and bills of material.
The next use case was using ML to analyze complex sales behavior for three customers, including a retailer and an alcoholic beverage maker. For the retailer, the application leveraged external data sources to analyze the sales of complex products that contributed greatly to total profit but were only sold three months per year—things like sunscreen or Christmas trees. “We focused on trying to predict the sales each month to handle demand outside the key sales periods (summer and the holidays, respectively, for the products just mentioned).
For the beverage company, the application analyzed public customs data to see when a certain type of glass bottle entered the country. If these bottles came in at higher-than-expected quantities, that was a signal that their competitor was gearing up to meet a peak demand. The three companies are still benefitting from these applications.
QAD DynaSys also developed an app for daily forecasting, which was useful for companies that make products that expire quickly, such as fresh meats, produce, and newspapers. “You can’t sell Monday’s demand on Tuesday because it has expired,” says Phillips. “We give them weekly data broken down into days.”
QAD is now into the second phase of its AI/ML journey, planning to offer adaptive supply chain planning in its flagship ERP application. (The company has not yet announced when it expects to release these capabilities.) Phillips calls this application a “supply chain digital twin.”
“We will digitally simulate the supply chain, automatically refreshing it with data like bills of material, the run rates, the yields, procurement and supplier lead times to be able to make real-time adjustments to plans and inventory levels. “Say I buy from this supplier every week and they ship on a 14-day lead time, so I always keep enough stock for 14 days,” he says. But if you can take your lead times from 14 days down to seven, you can make your supply chain much more agile, reducing the amount of inventory and thereby holding onto more cash.
Optimizing warehouse management
Manhattan Associates’ focus is applying ML algorithms to warehouse management applications via its Manhattan Active Warehouse Management product. Order streaming involves coordinating and optimizing all the aspects of the distribution center—the people and the automation equipment, as well as the shipment windows and requirements for fulfilling orders. “I have all these moving parts in the distribution center. Order streaming optimizes across all of those, adapting over time,” says Adam Kline, senior director of product management.
The core of order streaming is the work release engine, which gets inputs from all over the distribution center (DC). “It understands the layout of the DC. You may have 50 aisles full of storage locations, one-way aisles, bi-directional aisles. It understands how to build the most efficient pick path,” says Kline. The ML capabilities predict how long the work will take, which helps determine the sequence of work and who it should be delivered to.
It’s still early days for adoption, says Kline, with about 12 customers currently using Manhattan Active Warehouse. An apparel company and a home décor brand are currently using order streaming to drive lower order cycle times, better picking and packing efficiency, and an elimination of the number of orders that need to be sent with upgraded shipping, which takes a bite out of profit.
Beckhoff uses AI to enable more accurate predictive maintenance of equipment used in order and parcel fulfillment, another application that became more important with the e-commerce boom last year.
“Distribution centers typically have their peak season between November and the end of the year,” says Doug Schuchart, material handling and intralogistics manager at Beckhoff. “But there has been a peak since early spring of last year and it hasn't slowed down since.”
Throughout the pandemic, says Schuchart, AI has been instrumental in turbo-charging predictive maintenance applications. Companies need to know when the components are going to fail so they can replace them in advance of downtime. With no end to the peak season in sight, it’s challenging for maintenance crews to schedule operational downtime to do scheduled maintenance. Instead they can optionally apply AI theories to maintenance, replacing equipment prior to failure. Equipment sensor data (such as vibration and temperature) along with other data including hours of operation and scheduled maintenance) is used to build the AI model via MatLab, Python, or other open data science platforms. Once the model is trained, it is put directly on the machine controller so it is processed locally in real time, according to Schuchart.
“We don’t have to send so much data to the cloud or to the enterprise systems. We can do some initial analytics locally,” which minimizes the bandwidth of data sent and reduces cloud storage costs,” he says.
Beckhoff customers are also using the AI capabilities to optimize energy usage as part of their sustainability initiatives. This is important, as the ongoing peak demand in the fulfillment centers has radically increased energy usage.
Beckhoff uses AI/ML to fine-tune performance of equipment within automated storage and retrieval systems, for example. Shuttles numbering in the dozens and sometimes hundreds run back and forth to their target pick locations, potentially 24x7. “If we apply AI, we can look at the equipment profiles and analyze the time it takes to get the shuttle to its intended location. We can optimize the jerk and acceleration of the equipment, so it runs using less energy and minimizes the wear on the shuttle,” says Schuchart, two benefits for the price of one. “We can optimize the acceleration on any drive and motor in this way.”
Coming out of the pandemic, no one really knows how e-commerce demand will be affected. But it’s safe to say that online ordering will never go back to its pre-COVID-19 levels. E-commerce demand will continue to reshape supply chains, making efficiency and optimization more important than ever. Toward that end, AI/ML will play an increasing role.
Of course, what really drives supply chain value is the data. As companies fix quality issues and build up data as the foundation for AI/ML projects, they tend to realize just how limitless the prospects are with these technologies. “You can improve things like demand planning, but what really drives value is when you realize what data you have,” says Panak of Logility. “Investing in that data is just as important as what you want to do with it.”