Staying abreast of changes in supply chain technology has become almost a full-time job. From robotics and automation to data analytics and the industrial Internet of Things, new technologies are emerging that have the potential to further improve how goods are shipped, handled, stored, and delivered. With all of these technologies competing for our attention, it can be difficult to know where to focus.
One new technology that does deserve a close look is artificial intelligence (AI). In the simplest terms, AI is the development of computer systems that can perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision making, and language translation. AI has been around since 1956,1 but humans typically have had to explicitly program intelligence into computers.
One type of AI called machine learning, which has become prominent in recent years, explores ways to enable computer programs to improve their output based on learning from data inputs. These programs can be embedded in machines, or they can operate on servers or in the cloud. Large technology companies such as Amazon, Google, Facebook, Microsoft, and others are already incorporating machine learning into their offerings,2 creating more intuitive Web searches, better image and voice recognition, and smarter devices.
There are some similarities between machine learning and data analytics, or the processes used to collect, transform, and analyze data. Both require a clean, diverse, and large data set to function effectively. The primary difference, however, is that data analytics allows users to draw conclusions from data but requires them to take the action to improve their supply chain. For the right types of problems, machine learning can automate the actions based on a "training data set," described in the discussion of supervised learning later in this article. For many supply chain executives, AI—and particularly machine learningis an important technology to consider because it allows tasks to be automated. Organizations that begin today to develop AI strategies that are relevant to the supply chain will be positioned to increase productivity, speed, and efficiency as the technology matures.
Yet most supply chain professionals don't work at companies like the technology giants mentioned earlier. They don't have hundreds of data scientists on staff, and they do not have large research and development budgets. Nor can they look to a standard definition of the role of AI in the supply chain. The goal of this article is to highlight what steps these companies can take to enable AI in an important part of the supply chain: the warehouse.
The current state of AI
AI is growing rapidly today because of the convergence of several factors. First is the rise in the amount of data being generated through increased connectivity and the advanced sensors that enable more aspects of our lives to be digitized. Second is the continued rise in computing power in everything from mobile devices to the cloud. As a result, machine-learning applications that are running on the latest computing hardware and have access to large, diverse, and high-quality data sets can now automate a wide range of tasks.
Here's an example that will be familiar to many consumers. If you have an iPhone and commute to work every morning, you may have noticed recently that when you get in the car in the morning, your phone, without prompting, issues a notification telling you how long it will take you to drive to work and the best route to take based on traffic conditions. The first time this happened, you may have thought, "How did my phone know I was going to work? That's cool—and a little creepy."
The phone knows because it has machine learning embedded in the device, allowing it to predict what you are going to do based on what you have done in the past. If you change jobs and start driving to a different destination, the device adjusts its predictions and gives you a notification based on your new destination. What's especially powerful about this example is that the device is getting more useful to the user without the user or the software developer having to take any action.
Another example is self-driving cars. The current generation of self-driving cars on the road today is being used to collect data that will lead to improvements in the next generation of autonomous vehicles. Whenever human operators override the vehicle controls, that data is pooled with data from other vehicles and analyzed to determine why the override was necessary. All vehicles become smarter based on that experience.
While it's easy to get swept up in the exciting developments associated with AI today, it's also important to understand its limitations. In a 2016 article in Harvard Business Review, "What Artificial Intelligence Can and Can't Do Right Now," Andrew Ng, former head of the Stanford Artificial Intelligence Laboratory and former chief scientist of multinational tech company Baidu's AI team, states clearly, "AI will transform many industries. But it's not magic."3
Ng stresses that while there are a wide variety of use cases for AI, most applications use a type of machine learning called supervised learning. In supervised learning, a training input data set is associated with the correct output decision. The machine-learning algorithm uses this training set to make decisions based on new input data. Some common applications for supervised learning are photo tagging, loan processing, and speech recognition. In each case, the system receives inputs—in the case of photo tagging, pictures—and makes decisions or responses based on what it has learned from its training data set.
Given a sufficiently large data set of inputs that is annotated with the appropriate human response (or output)—for example, this picture is a face—it's possible to build an AI application that allows a computer system to receive new input data and make decisions on its own. This allows processes to be automated that couldn't easily be automated in the past and, ultimately, will enable warehouses to operate with greater effectiveness. The key to unlocking the potential benefits of supervised learning is the size, quality, and diversity of the data set used to make decisions. The larger and more diverse the training input data set, the better the decisions that will be made by the machine-learning algorithm.
Choosing a use case
As you consider opportunities to apply AI in the supply chain, it may be tempting to start with the technology and seek out an application. However, more useful applications are likely to emerge if you evaluate the business drivers that represent the greatest challenges or opportunities for your company, and then apply an appropriate understanding of AI technology's capabilities to those issues.
In relation to the warehouse, AI applications should be guided by the key performance indicators (KPIs) a particular organization is trying to optimize, such as order accuracy, safety, productivity, fulfillment time, facility damage, or inventory accuracy. Warehouses typically already have a wealth of data that is related to their KPIs and could be used by an AI application to automate tasks or decisions. However, this data typically is in a form that is not conducive to using AI techniques, and it often is spread across various warehouse systems. As a result, many AI applications will likely require information to be aggregated across various information systems in the warehouse before it can be used.
The following examples illustrate the potential for AI in the warehouse. Each of them is focused on a KPI: productivity, equipment utilization, or efficiency. While the examples may not be applicable to every warehouse, they do show how companies can take available data and fit that data into a form in which machine-learning techniques can be applied.
Productivity. When it comes to picking orders, all warehouses experience a range of productivity, from their highest-performing order pickers to their average performers. However, those warehouses that do not use system-directed picking often experience a greater range of productivity than warehouses that do use it.
For those warehouses that do not use system-directed picking, machine learning offers an opportunity to leverage the experience of their most productive order pickers and move toward a system-directed solution for all order pickers. If you think in terms of the supervised learning described above, the input data for the AI application would be the pick lists of the selected operators with the highest productivity, and the output data would be the sequence in which they picked the products on those lists. The output data would be based on bar-code scans or other available information. In addition to shortest overall travel distance, avoiding congestion can often be a significant factor in maximizing picking productivity. Since the best order pickers probably consider both of these factors in their pick sequences, the data sets should contain this information.
With this properly annotated data set, a machine-learning algorithm could receive new orders and sort them in the best order to be picked. In this way, the algorithm can replicate the choices that the most productive order pickers are making and enable all order pickers to improve their productivity.
Equipment utilization. There is a relationship between the number of cases or pallets a particular warehouse needs to move in a day and the amount of material handling equipment required to support that goal. In most cases this is estimated as a linear relationship. However, there may be additional factors that contribute to the amount of equipment needed, such as the skill level of the operators and the mix of stock-keeping units (SKU).
In this case, the input would be all available data that could impact equipment requirements, including the detailed order list of what needs to be shipped from the warehouse management system (WMS) and the productivity level of the operators from the labor management system (LMS). The output data would be the material handling equipment utilization data from the lift truck fleet management system.
With this properly annotated data set, a machine-learning algorithm could receive a forecast of orders for the coming weeks or months together with data about the current skill level of the operators, and then provide an estimate of the material handling equipment needed. The lift truck fleet manager would then be in a good position to work with the equipment provider to ensure that the required equipment will be available through short-term rentals or new equipment purchases.
Efficiency. A good slotting strategy seeks to optimize the location of high-velocity SKUs while also spreading them out enough across the pickface to minimize congestion and improve picking efficiency. But with demand changing constantly and the number of SKUs in some warehouses in the thousands, it can be difficult and time-consuming for a human to keep SKUs in the optimum locations based on their velocity. Some warehouse operators use slotting software products that assist in keeping the SKUs slotted in the optimum positions. These slotting products typically provide an interface that allows the user to include operating rules for the warehouse. When given past sales history or a forecast of expected future sales, the slotting products can then provide a recommended slotting strategy. However, it is common for the people in charge of slotting to make adjustments to the slotting strategy based on their own knowledge of the warehouse that is not reflected in the operating rules.
In this case, the input data would be the initial slotting strategy as recommended by the slotting product. The output data would be the final slotting strategy as executed. A machine-learning algorithm could be incorporated into a slotting product, which could then learn over time the preferences of the person implementing the final slotting strategy and make these adjustments automatically.
Developing a strategy
After identifying a warehouse-related area that could benefit from AI, it's important to set a strategy that will prepare your company for implementing the application. In his Harvard Business Review article, Andrew Ng makes some helpful observations about how executives should think about their AI strategy. The key to developing a successful strategy, he writes, is "understanding where value is created and what's hard to copy."
AI researchers, Ng points out, publish and share ideas frequently and open-source their code so there is ready access to the latest thinking. Instead, the "scarce resources" that allow an organization to develop an AI strategy that delivers competitive advantage are data and talent. It is much easier to replicate software than to get access to data sources, especially data sources that have been annotated with the correct output. So, people who have the expertise to identify and acquire high-value data, and to customize software in order to get the value from that data, become the truly differentiating component of an AI strategy. In other words, as they pave the way for artificial intelligence in the warehouse, organizations should focus on improving the quality of their data and talent.
The key question to address regarding data is, what data that is unique to your company can be used to improve the KPIs that are most important to the business? Once that has been determined, it is important to take steps to improve the quality of the data that is in your warehouse information systems. Commonly referred to as data governance, this is important for ensuring that there is "one source of truth" for the data elements that you use to run your supply chain.
For example, forklift operator information can be stored in multiple systems in a warehouse, including the human resource system, LMS, WMS, and forklift fleet management system. If all of this data has been keyed in separately, it is possible that the names and identification numbers for the same employee might not match across systems. For instance, an individual could be identified as Jo Smith, #01425 in the WMS; Joanne Smith, #1425 in the LMS; and Joanne Smith, with no ID number in the fleet management system.
For those machine-learning use cases that are aggregating data across multiple systems, it is imperative that the operator data be clean. Organizations with good data governance would recognize one of these systems as having the master data records and would have an API (application programming interface) that exports this identical data into any other systems where it is needed.
If you have selected a use case that requires aggregating data from multiple systems, the next challenge will be integration; that is, ensuring that data from the various systems that run the warehouse can be combined into a form that can be used for machine learning. It is important to work with your providers to understand their capabilities and the potential for combining data from the various systems, such as fleet management, labor management, warehouse management, and enterprise resource planning (ERP) systems. This lays the foundation for a digital infrastructure that supports data analytics and artificial intelligence initiatives customized to your business. This can be technically challenging, but the APIs designed into many systems simplify this task.
A bigger challenge may be in the area of talent. How many people in your organization are dedicated to governing, integrating, and capturing value from the data that is being created? If the answer is "not enough," then you should recruit an executive sponsor—someone who sits at the board level and can be an effective advocate for building competitive advantage from the company's data assets.
This high level of advocacy can then be leveraged to begin the process of determining how your company wants to build capability in this area. For most companies, this will probably be accomplished through a mix of internal staff and external consultants. There are even crowdsourced machine-learning platforms, such as Kaggle or Experfy, that can be used to connect you and your data challenge with experts across the world. Building your data capabilities is an important priority because today's data has the potential to teach tomorrow's machine-learning applications. Many larger organizations have already begun building internal teams to guide their AI and data analytics efforts,4 and there is significant competition for specialists in this area.
While supply chain managers have myriad technologies to evaluate and technology-based changes to navigate, artificial intelligence should not be ignored. Neither should it be viewed as a panacea that will magically transform the supply chain. Instead, AI should be viewed as a tool capable of driving improvements in the KPIs that are critical to the success of your organization. It isn't necessary to become an AI expert to leverage this tool, but you do need to make sure your organization has in place the three fundamental requirements discussed above: define high-value use cases that are important for driving improvements to your business, create a digital infrastructure that enables high-quality data to be aggregated from multiple systems, and begin to build a team of data experts both inside and outside of your organization.
1. Tanya Lewis, "A Brief History of Artificial Intelligence," LiveScience (December 4, 2014)
2. Christina Mercer, "Tech giants investing in artificial intelligence," TechWorld (February 8, 2018)
3. Andrew Ng, "What Artificial Intelligence Can and Can't Do Right Now," Harvard Business Review (November 9, 2016)
4. Mike Faden, "Using AI to Solve Complex Global Supply Chain Management Challenges," American Express online (undated)