Where are we on the AI/ML adoption curve in the food supply chain—fresh, dry, frozen?
To answer this question, it is useful to restate the business objective regardless of the technology which is keeping the right products on the shelves, ensuring our deliveries come at the best time, forecasting the supply chain simultaneously for inventory, demand and supply, minimizing both out-of-stocks and food waste, and targeting efficient ratios between the number of employees in a store and sales volume generated.
Further, distribution centers must deliver on perfect order fulfillment, wholesalers are must provide an overall demand forecast, delivery must engage in route optimization, resource scheduling needs to be coordinated, and overall order fulfillment must be efficient. In a recent survey, more than 40 percent of grocers are challenged by a lack of real-time inventory visibility, insufficient forecasting technology, and the technical inability to keep up with business demands.
Given that AI has been around for a while, and has already become pervasive in many of our daily activities in terms of search, investing, healthcare, marketing, apparel, etc., one would expect that all companies need to understand the future implications to their competitive environment and any related opportunities/risks.
However, embedding AI/ML capabilities in supply chain solutions requires a significant amount of industry knowledge mixed with an equal amount of technical depth in an area which has evolved significantly in just the past couple years. This results in many “hobbyists” saying they can provide AI/ML for the food supply chain, but, in reality, the industrial strength capabilities we need to drive value are difficult to find (despite all the fancy tv commercials claiming otherwise). So while we are clearly beginning to see the potential opportunities, movement to this point has been largely around Proof of Concept (PoC) and thus little true progress has been made.
One of the big issues we face in the food supply chain is where could we possibly squeeze any additional benefit out of a supply chain that has been squeezed pretty hard to date? We know by using AI/ML technology we create an advantage through improved efficiency and speed, but how does this manifest itself in terms of value creation? Companies are looking for advantages in terms of additional revenues, higher margins and lower costs—all while actually increasing efficiencies in the food supply chain—is getting harder to achieve. However, we must remember that the supply chain optimization/planning/execution that has been accomplished to date was done using yesterday’s technology solutions and that a new frontier is now open.
For example, when I filed my original supply chain patent years ago, it was based on the fact that a new frontier had been created by an advancement in technology. Until that point we had to design systems in multiple modules since computers weren’t powerful enough to solve the problem holistically. That is why ERP systems were designed in multiple modules. Problem was when you reassembled the pieces to create the overall supply/production plan, it wasn’t very optimal as a whole given all the local decision making/optimization that existed within the various ERP modules. Then came a significant technology advancement—in-memory processing—that gave me the ability to solve the problem holistically—demand/material/capacity—and suddenly a whole new software segment was born, SCM, delivering significant value upside as compared to ERP. As I crossed the globe and presented this new capability at industry conferences and to multiple analysts, at first they didn’t understand and tried to shove it under the old ERP umbrella. However, as we know SCM and SCE took off in a big way given the tremendous value available to those who adopted the new solutions.
In much the same way AI/ML is a new technology that is opening new frontiers in business value generation and there are direct applications for the food supply chain. Machine Learning relies on a continual process of technological learning from being fed relevant data and gaining experience by understanding the outcomes related to that data. This is why ML technology is so effective at answering the more complex questions that lead to a higher value.
Rich and powerful data
Complexity in the food supply chain is dealt with on a daily basis. Food manufacturers and distributors interact with thousands of customers and products, all producing various types, structures, and levels of data. Machine Learning is proving much more efficient at unraveling complex data quickly and meaningfully than any technology we have ever experienced. For instance, manufacturers and distributors might need to know which products are better to deliver last-minute; or which baseline product should regularly be in stock.
From a demand planning perspective, with retailers being dependent on promotions for twenty percent plus in sales, using machine learning to determine the promotion mix becomes a top priority. Machine learning algorithms can learn to be more effective over time at clustering promotions based on looking at many more variables than is otherwise possible using traditional, linear-based forecasting techniques. This technique can also drive significant reductions in churn, in some cases reducing product level churn as much as 30 percent.
Continuing with the downstream portion of the supply chain, we then move to create baseline forecasts to determine which particular products should be stocked, the assortments, the placement, and the timing. ML-powered algorithms learn from a multitude of factors that are likely to influence buyer behavior – including promotions, social media, or the weather—which then are used to more accurately manage inventory levels and replenishment further upstream in the supply chain. Not only will such advanced technology know when shelves are empty, but more importantly, it will predict what will happen next using ML-based algorithms especially since many companies are now building forecasts at the product/store level in daily, weekly, and monthly time frames. For a product being forecast daily at the store level, the ML approach may be an algorithm applied to the point of sale data stream. Further upstream in the supply chain, forecasting that same product at the warehouse on a monthly basis would require a different ML algorithm using warehouse shipment history and warehouse ordering patterns.
Improve productivity and eliminate waste
Machine Learning has already made inroads into more intelligent forecasting which then drives our ability to operate an optimized supply chain. With our previous generation of technology, customers and trading partners participating in the food supply chain had periods of high error rate when forecasting the quantities of products to order to keep shelves fully stocked given they were using some type of time series forecast being fed historical sales data bolstered by out-of-date inventory information. Today, ML can develop a much more accurate picture of exactly what types of products are likely to sell, by looking at multiple scenarios in real time and using an expanded data set (POS, pricing, loyalty, store door, assortment, placement, logistics, sell in, sell out, sell through, ship to, supplier, consumer behavior, social media, competitive intelligence, and seasonality along with other external variables around events and weather). This means demand planning/forecasting, which should be a single version of the truth for all trading partners, can become much more accurate. The payoff is real in that today companies are attaining double-digit improvements in forecast error rates, demand planning productivity, cost reductions and on-time shipments using machine learning.
Digging deeper, supply chain challenges are mostly related to time, cost and resource constraints, which is right in the machine learning wheelhouse in terms of developing a highly accurate prediction capability. A great example of a multi-variable ML solution is DHL relying on ML to enable their Predictive Network Management system which analyzes 58 different parameters of internal data in order to identify the top factors influencing shipment delays.
The opportunities across the food supply chain and between the various trading partners, logistics is the glue holding the network together. Machine Learning techniques are the foundation of a broad spectrum of next-generation logistics and supply chain technologies. Key here would be the ability to anticipate anomalies in logistics costs and performance before they occur. The available data spectrum in this area is significant including sensors, telematics, intelligent transport systems, traffic data, and through systems like the AFS Electronic Proof of Delivery (ePoD) and Track & Trace ERP capabilities we are able to generate value out of all this data.
Procurement planning is a focal point given that inventory is a key value driver in the supply chain. With our G2 Analytics capabilities AFS provide an environment which provides full visibility to all KPI’s and enables the user to make adjustments through the AFS MBE or Management by Exception capabilities. “What if” planning scenarios are a valuable part of the workbench so that various potential outcomes can be evaluated. Recently the G2 application environment has been enhanced as a fully enabled cloud-based artificial intelligence and machine learning platform, taking analysis and predictive capabilities to even higher levels of performance and accuracy. Now any issues related to potential stockouts, purchase orders, requisitions, or material movements/transfers are immediately presented and corrected with real time connectivity back into the ERP system.
At AFS we have always been the dominant provider of sales, order management, and forecasting capabilities in a mobile framework. This allows all field sales and order management personnel to have more efficient conversations and related transactions.
Using ML in the logistics chain to reduce food waste is certainly one of the key objectives here. With the increased technical modeling capability we no longer have to live with tenants like all produce harvested on the same date will have the same amount of freshness and shelf-life. With an ML-based approach, we can determine the maximum shelf-life of the produce. This establishes your learning baseline for the algorithm. For example, you may determine that a particular lot of berries has a maximum freshness capacity of 12 days from time of harvest under a certain set of conditions which would include features like temperature and rate/range of cool down. Once the baseline is set additional variables related to the condition and handling of the produce at the pallet level from the time it’s harvested would be vectorized as part of the ML algorithm. This all leads to being able to more accurately predict how long that produce will last. Because the system utilizes machine learning, more data gathered over time from more pallets will drive more accurate predictions.
Within the warehouse itself we are working to leverage ML to shift with changing conditions and priorities. Pick density may be the initial WMS priority, but as carrier cut-off times approach, meeting order SLAs will take precedence. Machine learning can be used to predict the time required to complete the work. An optimization algorithm then uses those results to balance competing requirements while optimally utilizing available capacity.
The payoff can be significant with forecast errors being reduced up to 50 percent. Lost sales due to products not being available are being reduced by up to 65 percent through the use of Machine Learning-based planning and optimization techniques. Inventory reductions of 20 to 50 percent are also being achieved today when Machine Learning-based supply chain management systems are used.
Supplier quality continues to be an area of focus. Supplier performance is measured in many different ways given all the factors involved. Understanding the best patterns and worst patterns that lead to performance issues can drive huge value between trading partners. Using machine learning we can discover who is best/worst suppliers are, along with insight as to how errors are being captured and managed. Transparency through dashboards and workbench’s allows Manufacturers, Distributors, and their trading partners to understand and improve supplier quality, delivery, and consistency.