When it came out, ChatGPT seemed like magic. It has led supply chain vendors to discuss how they currently use artificial intelligence. Further, virtually every supplier of supply chain solutions is eager to explain the ongoing investments they are making in artificial intelligence.
Any device that can perceive its environment and can take actions that maximize its chance of success at some goal is engaged in some form of artificial intelligence. AI is not a new technology in the supply chain realm; it has been used in some cases for decades. More recently, many other cases have emerged.
Optimization is used in supply planning, factory scheduling, supply chain design, and transportation planning. In a broad sense, optimization refers to creating plans that help companies achieve service levels and other goals at the lowest cost. In mathematical terms, optimization is a mixed-integer or linear programming approach to finding the best combination of warehouses, factories, transportation flows, and other supply chain resources under real-world constraints.
Machine Learning occurs when a machine takes the output, observes its accuracy, and updates its model so that better outputs will occur. Demand planning engines have natural feedback loops that allow the forecast engine to learn. The forecast can be compared to what actually shipped or sold.
Since ML began being used in demand forecasting in the early 2000s, ML has helped greatly increase the breadth and depth of forecasting. Now, ML forecasting is not just monthly or quarterly; weekly and even daily forecasting is now possible. We have moved from product-level forecasts at a regional level to stock-keeping unit forecasts made at the store level. More recently, demand planning applications based on machine learning have improved forecasting by incorporating competitor pricing data, store traffic, and weather data.
We are no longer just forecasting demand but also when trucks and factory machinery are likely to break down (predictive maintenance), the optimal amount of inventory to hold and where it should be held (inventory optimization), and labor forecasting in the warehouse. This type of forecasting can forecast the number of employees required to perform estimated work down to the day, shift, job, and zone level. ML can also be used to generate labor standards for warehouse workers.
ML techniques like clustering, data similarity, and semantic tagging can automate master data management. Without accurate data, companies face the garbage in, garbage out problem.
In terms of supply planning, if key parameters (like supplier lead times) are no longer correct, then the planning becomes suboptimal. ML is being used to keep key parameters and policies up to date. It is also being used to predict whether an SKU believed to be in stock at a store is actually out of stock.
Supply chain risk solutions use ML and other forms of AI to predict which suppliers are included in a company’s multi-tier supply chain. This is becoming increasingly necessary as customs will hold up shipments at the port if it believes the shipment contains products made with slave labor from China, even if those components came from their supplier’s supplier’s supplier and represent a minuscule portion of the total cost of the product. Shippers’ end-to-end supply chain predictions are based on applying AI to OpenWeb searches, import/export records, data from sourcing platforms like ThomasNet, federal logistics records, and other data. These predictions accelerate a company’s ability to verify how its extended supply chain is constructed. Customs uses the same technology to determine which shipments should be denied entry.
Natural Language Processing is used to classify commodity classification for use in imports and exports and in real-time supply chain risk solutions.
The Harmonized System is a commodity classification coding taxonomy that forms the basis upon which all goods are identified for customs. It is used by customs authorities worldwide. Using the right product classification allows companies to pay the correct tariffs. Paying the right tariffs is necessary to avoid government fines and calculate the true landed cost of products. The problem is that there is an incredible gap between how products are described commercially and how they are expressed in the national customs tariff schedules. This has resulted in error rates as high as 30%. The combination of natural language processing and expert systems has been used to automate and significantly improve the classification process.
Real-time risk solutions also use natural language processing to read online publications and other data sources, make sense of what they read, contextualize the data into information, and report supply chain disruptions caused by weather, geopolitical events, and other hazards in near real-time. Every step in that value chain has search terms associated with it. The names of the suppliers, carriers, logistics service providers become search terms. Those search terms are paired with terms signaling a problem – those terms might be “bankruptcy,” “plant fire,” “port explosion,” “strike”, and many, many other terms. So, the term “Haiphong” when combined in an article with the phrase “port fire” would generate an alert.
Reinforcement Learning is a form of machine learning that lets AI models refine their decision-making process based on positive, neutral, and negative feedback. For example, if you want to train a vision system to recognize a dog’s image, you will start by using humans to look at tens of thousands of images of animals. The humans label the pictures as dog, not dog, or unclear. The computer is then presented with those images. The system would say, “this is a dog” or “this is not a dog” and it learns whether its conclusion was correct.
Drones use this form of AI to improve inventory accuracy in a warehouse. Reinforcement learning allows the drone to recognize warehouse racks, pallets, and cases and get close enough to inventory to scan the barcodes. Similarly, reinforcement learning has been applied to security camera footage in the warehouse to ensure workers are following standard operating procedures.
Simultaneous localization and mapping (SLAM) allows a vehicle to construct and update a map of an unknown environment while simultaneously keeping track of the vehicle’s location within it. This technology allows mobile robots to move autonomously through a warehouse.
Drones and autonomous mobile robots using SLAM are in an early adoption stage for last-mile deliveries. Autonomous trucks will revolutionize logistics.
Autonomous trucks are not yet feasible, but we are probably just a couple of years out from being able to transport goods from a distribution center to a retail facility autonomously.
Causal AI is a technique in artificial intelligence that builds a causal model and can make inferences using causality rather than just correlation. Cause-and-effect relationships in an extended supply chain can be an intricate web that is difficult to unravel, but these relationships govern business operations. A causal model graph represents a network of interconnected entities and relationships, enabling the system to understand how various factors influence each other to create an optimized outcome. By leveraging causal knowledge and data graphs, Causal AI can navigate complex business scenarios, anticipate outcomes, and recommend optimal courses of action. Georgia-Pacific has demonstrated an application of Causal AI to improve touchless commerce dramatically. The solution was used to detect and correct both common and uncommon order errors or discrepancies in near real-time.
GenerativeAI is the new kid on the block. GenAI can generate text, images, videos, or other data using generative models. Some warehouse management suppliers are exploring using GenAI to generate end-of-shift reports or talking points used at standup meetings at the beginning of a shift.
Several supply chain application vendors are investing in GenAI to improve their user interfaces. The idea is that a user will make a request, and the system will take them directly to the answer they seek. GenAI can also help interpret complex charts and planning outputs. If a planning system indicates that a plan shows high costs or an inability to achieve targeted service levels, GenAI can help explain the upstream constraints driving that outcome.
Planning vendors are also interested in using GenAI to solve the black box problem. The black box problem occurs when planners don’t understand how the planning engine produced the plan it did. If they don’t understand it, they don’t trust it, and they then produce a much less optimal plan using Excel.
In the longer term, GenAI will help some planning vendors generate autonomous plans. When disruptions constantly occur, there is no time to constantly create and analyze scenarios on how to react best. Autonomous planning can improve a company’s supply chain agility. However, it is worth noting that a few planning suppliers can already generate autonomous plans based on ML and attribute-based planning rather than having to rely on GenAI.
