In an earlier blog, I talked about how Artificial Intelligence is already transforming supply chain planning. It is fundamentally changing the competitive landscape in supply chain operations, and those leading the pack have experimented in a few distinct areas. I also challenged you to start exploring where investment could drive top and bottom line impact across your organization, as there are broad opportunities for machine learning solutions in supply chain planning to drive value across an enterprise. This blog, therefore, is a continuation that shares some practical examples of the application of machine learning in supply chain planning to support your journey.
Determining Use Cases
To begin, what are the machine learning use cases in your business? What tasks take a lot of effort that can be automated? Here are a few areas to start your exploration:
Root cause analytics – why did we run out of stock; why did we miss a delivery date? This effort starts with a lot of data engineering and descriptive analytics.
Anomaly detection in supply chain – stock-outs, bad data, et cetera.
Forecasting – not just of future demand, but also end of life and new product introduction
Supply planning – replacing some of the more mundane tasks & planner intuition
Data cleaning – there are tools that do this now driving autonomic master data management
Production planning – using sensor information to feed back to your planning processes; think dynamic replenishment
Codification of customer intelligence – building in knowledge of promo phasing, buying patterns, etc. instead of human knowledge transfer challenged by attrition
Applications to Consider
Improving Demand Planning. The power of machine learning in demand planning is the combination of (1) more algorithms to test your data against your performance, and more importantly (2) recognizing the set of features that drives sales in your business. We worked with a CPG company using data from one of their global business units to understand how the performance of machine learning or deep learning algorithms in forecasting would compare to their status quo forecast from their advanced planning solution (see Figure 1). More than 75% of the time, a machine learning algorithm delivered a better forecast than their current solution. And this was consistent across the portfolio – fast moving, volatile and seasonal. We also observed an improvement of at least 25% in MAPE (Mean Absolute Percentage Error) in more than a third of the SKUs. This is due to the combination of new features, which is not feasible in statistical time series, and new algorithms. It is important to note, however, that machine learning solutions can complement but not necessarily replace your current systems. You can feed a portion of your portfolio that will have the largest benefit through an open source algorithm and route it back through SAP APO as an example; there is no need to replace your current systems.
Figure 1. Performance comparison of ML & DL against status quo in a CPG company
Sensor Analytics in Supply & Production Planning. The idea is to ensure the yield is good and can catch any potential issues ahead of time through preventative maintenance. Affordable sensors and the ability to store and process that data in a cost-effective way allows companies to add new data in understanding machine failure. We worked with data that an industrial manufacturer was collecting from sensors installed in the field across the state of Illinois. Their problem was that they would deploy field service to replace a failed device, only to find that, days later, additional devices were failing in a similar service area. We developed a model to predict which devices were at risk of imminent failure on the days following a planned visit such that the teams could take preventative action forward to replace the devices at risk of failure while already out servicing a failure. This allowed them to optimize resources and costs associated with future field visits.
Inventory Management. You can apply new machine learning techniques to detect bad data (finding outlier data before it impacts Multi-Echelon Inventory Optimization results), automatically clean data, automate seasonality detection & get the inventory parameters right, to better understand variability through the supply chain, and to predict stock-out’s at DC or store level. We worked with a CPG company to deliver a tailored inventory optimization solution using machine learning powered parameters and the results were stunning (see Figure 2). The planners who adopted the solution delivered a 19% inventory reduction with sustained improvement trend without impacting service levels. This is a space ripe for machine learning application.
Figure 2. Comparison of ML-powered inventory optimization and IPFA on inventory level of a CPG company
Automating Root Cause Analysis. RCA is not a new concept, but machine learning has revolutionized the way of thinking about it, taking it to a whole new level. You can take many disparate data sources and bring them together to piece together a story of what is happening, and when, to eliminate the human bias in assigning root cause. Root cause analytics can apply in many areas of operations – order processing, warehouse operations, customer delivery, and vendor management, just to name a few.
Let’s take order processing as an example. You can think of each order cycle consisting of a sequence of events – order drop, fraud check, order routing, etc. with time stamps and an eventual outcome (successful or not successful order). First you have to put this data together, then ensure that history is labelled. Labelling is nothing more than telling the model if the order cycle was successful or not. Once you have this data, you can understand correlations and patterns, and then set-up models that will call out irregularities proactively. We worked with a Fortune 100 online retailer to improve on-time delivery within their digital supply chain. Leveraging root cause analytics on orders across almost 50 countries and analyzing ~80k orders every day was not an easy effort. But root cause analytics kick-started their journey to achieving 60% reduction in late deliveries, driving a 4% revenue lift in their largest region. Exercises such as this require investment, but can result in healthy returns.
Segmentation. Unsupervised learning is useful when there is a lot of data and you are trying to find the most meaningful patterns in the data. The power of machine learning is to be able to segment not with just one or two features, but with dozens of features. In exploring segmentation with dozens of features across years of data with a Quick Service Restaurant (QSR) client, we segmented more than 2000 suppliers into 6 distinct segments (see Figure 3). This helped in easily drawing attention to the declining performers and the worst performers to tailor supplier relationship management programs that would cater to these distinct segments. Utilizing machine learning algorithms allowed them to automate this classification and understand the moves between segments over time.
Figure 3. Segmentation of a QSR’s supplier performance
Leverage machine learning to drive more efficiency through process automation in the repetitive, manual and transactional tasks within planning. You will also see an effectiveness benefit through the reduction of human error and bias. Freeing up resources consumed by non-value adding activity will allow you to pivot them into more strategic problem solving and analytical work. And algorithms will help enable planners to make better decisions. If your company is not already on a journey to disrupt itself, you need to identify opportunities to accelerate. And perhaps some of the use cases mentioned in this blog have some application in your organization to help you leapfrog your competition.