Artificial intelligence is now available to everyone
Over the last years, artificial intelligence (AI) has become a mainstream topic—seemingly omnipresent as a solution to complex problems. The availability of AI-based analytics and the many examples of the power of AI have contributed to this hype.
There are many examples of AI outperforming even well-trained humans. One of the most prominent examples is how AlphaGo and its successors have been able to gain dominance in Go, one of the most complex games to master. In 2016, AlphaGo was already able to beat 9-dan (highest achievable professional rank in Go) professional players and in 2017, AlphaGo Master beat the number one player worldwide, Ke Jie. In 2019, DeepMind released MuZero, a successor of AlphaGo that is able to learn new games without even knowing the rules and without any human teaching (“unsupervised learning”), which has proven to outperform other solutions in Go, chess, as well as other games.
The algorithmic basis of such programs is readily available to everyone, often as open-source libraries for well-established programming languages. This has resulted in many established IT providers incorporating AI into their solutions, as well as new players quickly entering the AI application field. The hype around AI is further pushed through the aggressive marketing efforts of IT solution providers that claim AI is the solution to most problems today.
At first glance, supply chain planning seems predestined to be a playing field for AI, as there typically is a wide variety of complex problems. While planners today spend a significant amount of time creating operational demand and supply plans, trying to predict what will happen in the future often based on simple statistics, and trying to find an optimum among competing objectives, many of the often still manual tasks could be taken over by systems. In some cases, AI could even improve accuracy and maximize margins.
Applications of AI in supply chain planning
AI is already being successfully applied in demand planning. Further use cases are being implemented in supply planning, inbound logistics, quality improvement of master data, and learning the behavior of parts of the value chain to provide input for a continuous improvement process.
Availability is key—but how to predict?
Let’s look at demand planning in the grocery retail industry to illustrate the power of AI. The goal of a retailer is to have all products available on shelf—this creates a visual impression that there is a wide selection available to the consumer. The challenge is to have the right stock, especially during evening hours; in the case of fresh products like fruits and vegetables, this might lead to wastage.
A perfect forecast helps increase availability and revenue while decreasing wastage. Unfortunately, several factors influence demand and the purchase behavior of the consumer. Factors like weather or competitor promotions can be important influencing factors.
Those external factors are often not considered properly. The typical forecast engine in grocery retail leverages historical demand data to predict the future. Let’s take barbecue as an example—shelf life is limited, and the consumer demand is highly dependent on weather— no one would buy steaks for a good barbecue on a rainy day. Forecasting demand based on historical sales would thus be completely misleading. What we often see in reality is manual change of proposed orders based on the experience of the planner. This takes time, is error prone, and not replicable: what happens if the best planner leaves the company?
Wouldn’t it be great to learn the experience of the planner? To take many influencing factors into account, without having to discern which are most important? This is where AI comes into play. A neural network “learns” the behavior of the consumer—the system is trained with massive data from previous weeks, months, and years. The observed correlation of input factors to demand is leveraged to come up with a forecast for the coming days and weeks. Building on the example of barbecue steaks: training the AI would, for example, show that more steaks are bought on sunny days, bank holidays, and for sporting events. How can this insight be used for more accurate forecasts? Bank holidays and sporting events are known well in advance, and the weather forecast is usually accurate for the coming couple of days. Leveraging this data can significantly improve the predictions and optimize the replenishment decisions on barbecue steaks for the next few days.
Wouldn’t it be great to know when a shipment arrives at your receiving dock?
Production and assembly lines rely heavily on suppliers delivering components on time. Due to issues in global logistics systems, as we often experienced during the COVID-19 pandemic, significant delays happen and lead to firefighting, expedition cost (for example, air freight), and even production stoppage. A reliable prediction of the expected time of arrival is crucial for smooth production flows, but not easy to estimate, as many external factors influence the delivery reliability and actual time of arrival.
This is where AI can help. Multiple supply chain software providers use AI to learn the complex environment of shipping, for example, based on past sailings, the reliability of the ship owner, weather conditions, and information on port congestion to accurately predict the arrival time of a shipment. This information can be leveraged in an integrated planning system to evaluate alternative routings, if required.
Learning how a complex system like a company makes decisions—and how to improve
Imagine a company with dozens of plants, thousands of suppliers, and a demanding customer base. Planning is done in an integrated way, but still with many manual decisions—using the experience of the planners. Some value chains in a company perform better than others, or outperform on all dimensions, like lower cost or higher service, and at the same time manage to maintain lower inventory, but it is unclear why. Insights would help improve the setup and parameters of the planning and execution systems—and AI can help.
Most companies sit on a huge amount of unprocessed data: historical transactions, master data, changes to forecast, production orders, purchase orders, and more. AI can uncover insights about why some countries, value chains, and planners are more successful than others – by learning the decisions and the complex environments they are made in. After understanding the “why,” the system is able to provide the “what” and “how,” guiding the planner in a much more efficient way to make the right decisions going forward.
One very important element is the quality of master data, such as capacity, lot size, lead time to replenish, and dimensions and weight of a product. Master data is often not updated for many years, while suppliers change their footprint or improve their processes, reducing or increasing the lead time to replenish. Getting the master data right is the single most important element to proper planning, correct proposals, and the acceptance of the organization.
AI can help increase the quality of master data by using the past to make improvements for the future. One company had long lead times as master data in the system, but did not want to update the information even though the supplier delivered quicker and components would therefore be kept on stock for a longer time. Analyzing the historical data and learning the probability distribution of the lead time allowed for a more targeted ordering with the supplier. This level of advanced planning also requires well-configured systems and the ability to work with scenarios, tracking individual lead times based on historical analysis.
What it takes for AI to generate value
AI has been a buzzword on every COO’s digitization agenda in the last decade. AI is proclaimed to be the next big thing, with hype around the terms AI, machine learning, and deep learning—all used interchangeably. But the question remains: how can AI work in a supply chain environment and generate value?
AI algorithms have been continuously developed since the creation of the term back in 1956 at the Dartmouth Conference, the birth of the field of AI. The modeling algorithms are not new and radical—neural networks, Bayesian algorithms, and gradient boost have been around for a while and have been continuously improved over time. Every second-year data science student knows where to go to find the standard modeling libraries for R and Python, where all the knowledge of the crowd is available at the click of a button. The pure modeling part of AI is therefore no longer the challenge.
We have seen in our countless AI-based projects that 90 percent of the work is not the model itself, but the underlying data availability and quality. It is about harnessing humanity’s collective intelligence to solve the world’s toughest problems. This includes identifying the long list of potential external and internal influencing variables, preselection, definition of data update frequency, design of data and output granularity requirements, evaluation of data sources and data quality, decisions on data and model scope, the development of the testing environment, the back-test to compare actual with simulated performance, and the close integration to the technical solutions. Understanding the relative importance of the variables driving the analyzed processes, for example, the product’s price gaps between key competitors, can be the most important drivers of sales volume. This includes intelligent training and selection of test sets, with the flexibility to incorporate manual input. Publicly available world knowledge can be combined with functional data enhancements, like open source repositories and curated libraries, to shift from purely descriptive to truly prescriptive AI use. It has historically been a challenge to provide the right level of availability of downstream data, for example, ship-to data, VMI feeds, and POS data. However, AI-powered research platforms that leverage web knowledge, like SparkBeyond, enable at-scale testing of millions of transformations on raw data to identify features that drive business outcomes by identifying variables with the highest prediction power. To train an AI algorithm, we need data—and a lot of it. But historical data is often not kept, and the data on storage is often on the wrong granularity (no details), without use of correlation analysis. If the granularity is not at the right level, the changes and drivers leading to numbers are not tracked. This is unnecessary, considering the cost of storage per month is below $0.01 per GB. Ten years of full order history for an average-sized company costs no more than $3. There is virtually no financial barrier to tracking data and no need to overwrite it to clear up storage space—so why not ensure the right data is stored and used?
In order to drive change management of the new data, processes, and algorithms, the “explainability” or interpretability of the underlying AI is key to turn the “black box” into a “white box.” Supply chain software providers with modern advanced planning systems have heavily invested in the interpretability of their algorithms. The graphical overlay of historical sales with the underlying influencing factors of demand, for instance, can make a world of a difference in the acceptance of an AI-based recommendation to a planner. Automatic root cause analysis for failures enables strong performance management dialogues.
Supply chain planners are another ingredient in the recipe for successful AI. Traditional skill sets do not suffice to remain competitive anymore; over 50 percent of workforce activities could be automated with currently proven technologies. The demand management organization of the future will combine strong market expertise with the latest data science knowledge for forecasting:
The market expert is the link to the commercial organization and is at home in both worlds—commercial and supply chain. This role provides all market and customer-related input, for example, promotions and new POSs, which are then integrated in the forecast. This person must be highly skilled in applying market intelligence, have commercial expertise, and understand competitor dynamics.
The demand planning data scientist combines internal and external data input and analyzes it through AI to form the high-quality basis for future demand. The capability needs for this role include experience in coding statistical computer languages (R, Python, SLQ), a deep knowledge of a variety of machine learning techniques (clustering, decision tree learning, artificial neural networks), experience querying databases, a solid demand planning understanding, and a strong continuous improvement mindset.
The demand manager defines the demand planning segmentation, handles exceptions where human input and validation is required, for example, new product introductions or end of life, and facilitates demand consensus discussion in the S&OP/IBP process.
The combination of these different roles and capabilities is what will make AI successful.
What is required to be successful in supply chain jobs will change fundamentally. A workforce transition is needed, as workers require new skills and must continually learn and adapt to evolving industries. The key question with regard to capabilities is whether to build these in-house, hire, or rent/externalize the AI skill set of the future.
It is our conviction that AI-oriented skills are a core competency for the supply chain of the future and cannot be successfully externalized. Companies first have to quantify the skill gaps between the current and future states to then design a portfolio of initiatives to build the target state capabilities, for example, through hiring, reskilling, contracting and operationalizing the initiatives, and creating a roadmap and governance to start building the target capabilities quickly.
In addition, the identification of the right use case for AI in the end-to-end planning process is key. For supply chain planning, this must be centered on the core KPIs around cost, service, and capital, with a clear rationale of where the AI use case will drive value for the business and how it will scale across the organization. AI has shown tremendous impact improving out-of-stock situations (and therefore top-line sales), driving inventory optimization, and utilization increase. A 10 percent improvement in forecast accuracy can help capture up to 2 to 3 percent of profit margin improvement.
It is also worth noting that AI cannot be a standalone use case—it must be integrated into the entire process landscape to fully achieve bottom-line impact. If you look at the common example of AI for demand planning, for instance, using a gradient boost to improve forecast accuracy and reduce bias, it is key that the demand planning process fully leverages the input of the AI. A seamless integration between planning and execution has to be given, for example, the ability to replenish at the same, high frequency that a demand sensing-based demand signal would provide to the supply chain. We have seen countless examples where the best statistical base forecast was overwritten multiple times across the demand review cycle, fully deteriorating the accuracy of the original forecast, without measuring the added value of the activities that make up the consensus demand.
Use of AI in demand planning
Demand planning has been one of the first and most successful areas for AI in supply chain planning. This often creates the misperception that purely through the use of AI, forecasts can always be improved to a very reasonable level. In reality, where and how AI can add value needs to be better understood.
One of the first criteria is that the use of AI requires reasonable data as potential influencing factors. Weather is always quoted as the prime example of an external influencing factor to be considered, but this tends to be more complex than it may seem: future weather is unknown, so forecasts have to be used as a proxy. Weather forecasts tend to be quite accurate in the very near future, but very inaccurate a couple of weeks in advance. Interestingly, for mid-term forecasting, using calendar information is oftentimes more reliable than using weather forecasts. The relevant data also needs to exist. For example, in businesses with many one-off or make-to-order sales, like specially printed packaging for unannounced promotions, AI would not create useful forecasts.
Therefore, the level of the forecasting and relevant influencing factors should be determined first. In many cases, SKU-level forecasting is not even needed, and a forecast at production line level or at product group level is sufficient for capacity planning or raw material planning. The feature discovery to identify the relevant influencing factors should always consider if the features make sense from a business perspective versus testing random variables. There is a risk of overfitting to random variables with AI, and this also needs to be managed.
AI will also not be able to fully replace manual input to the demand planning: customers might communicate changes to their ordering behavior, for instance, that they will discontinue ordering a product, switch to a different one, or buy more or less going forward. Such explicit customer information needs to be captured in the forecasts. At the same time, AI can add value by checking the quality of the manual input of planners by understanding when planners improved the forecast accuracies through manual changes versus when they created poorer results; this input could then be automatically processed accordingly.
Secondly, the role of the planner is not only to create a demand forecast, but to improve the demand plans by addressing the root causes of variability. A planner with many years in the same role knows the environment, customer, suppliers, and is familiar with all situations, such as end-of-month discounts to reduce inventory or increase revenue. A lot of these decisions seem irrational, as they do not optimize the system, just small parts of it. If AI were to learn the system’s behavior, it would find a strong correlation between increased demand toward the end of the quarter and a steep decline in the following month. Unfortunately, this increase is often pushed by the companies’ own sales and marketing teams, not by the customer. An experienced planner would try to avoid this behavior, as it is bad for the system and instead actively work on remedying the root causes.
No matter how powerful the analytics are, the results also need to be trusted and not be continuously overridden by planners. This is still one of the top reasons why planners like to use standard statistics: the causal relationships are completely transparent. This has become a major aspect in further pushing AI adoption: to create a necessary level of “explainability.” Some solution providers incorporate elements to show the significance of influencing factors and to simulate how forecasts would change in certain scenarios. It is important to get away from pure “black box” approaches as long as there is not yet sufficient trust in the outcomes.
What’s next on the path toward smart and autonomous supply chains?
If we are able to predict the future to a large extent, why would we need manual interventions of the planner to change dates, volume, or allocations?
Since the early days of supply chain management, the community has discussed the vision of a system with all data in the right quality and in real-time, leveraging advanced algorithms to plan ahead, make trade-off decisions, and execute automatically. All of this is possible in theory, and a lot is possible with the currently available technology. So why do we not see more advanced applications of AI? Unfortunately, reality is much more complex—poor master data, irrational decision making, and optimizing for local objectives hold us back from achieving the vision. The future will involve many years of a combined approach: using AI to set the groundwork, do the heavy lifting, and analyze data, while leveraging the experience of the planner to further improve the outcome.
Smart contracts will change the way we operate our logistics systems—imagine a container that knows the load, promised delivery dates, and conditions. This container would be able to interact with a freight forwarder or directly with a vessel to select the best route and price. All could be done autonomously, with no interaction with anyone in the port (for instance, custom clearance, negotiation with freight forwarder). AI plays an important role in these scenarios, as we need to foresee the accuracy of the predictions of the container transit time.
If we push this thinking further, we might see digital twins of complex global supply chains, enabling scenario analysis, simulation, and optimization. AI combined with experienced supply chain professionals can optimize the system: the AI provides insights, scenarios, and correlations and proposes improvement opportunities, while the supply chain professional guides the analysis and simulation based on experience.
The future of supply chain will be as no-touch and autonomous as possible—a hybrid system that incorporates AI and human intelligence to optimize the setup and operation of complex systems.
About the authors
Knut Alicke is a Partner in McKinsey’s Stuttgart office and a leader of the Supply Chain
Management Practice in Europe. He focuses on supply chain planning and execution,
especially related to Supply Chain 4.0 and the digital supply chain. Knut holds a PhD in
logistics and an advanced degree in supply chain management from the University of
Karlsruhe. He is still active as a professor at the universities of Cologne and Karlsruhe.
Prior to joining McKinsey, he worked with ICON/e2open.
Julian Fischer is an Expert Associate Partner at McKinsey’s Munich office. He has spent
the last decade in supply chain, leading transformation programs with global companies
to improve service, productivity, and capital. His expertise includes advanced analytics for
sales forecasting, integrated planning, and S&OP, improving supply planning processes, as
well as implementing the right organizational setup for success. Prior to joining McKinsey,
he worked in the automotive industry.
Jürgen Rachor is a Senior Expert in McKinsey’s Supply Chain Management Practice,
located in the Frankfurt office. He is part of the practice’s leadership team and heads
our global Supply Chain Executive Academy. Jürgen’s work focuses on supply chain
transformation, end-to-end supply chain planning, as well as supply chain digital and
analytics. Prior to joining McKinsey, he worked on strategic topics at SAP.