Artificial intelligence (AI) continues to draw a lot of attention as companies and technology vendors look at how machine learning could improve supply chain operations. In particular demand planning, understood here as the process of developing forecasts that will drive operational supply chain decisions, is being touted as the next potential field for innovation. Technology giants like Amazon and Microsoft have announced AI tools for improving demand planning, and several consulting companies are promoting their skills to bring AI to companies' demand planning processes. In fact, a recent survey by the Institute of Business Forecasting and Planning (IBF) identified AI as the technology that will have the largest impact on demand planning in the next seven years.1
It's not hard to see the fit between AI and demand planning. Demand planning involves lots of number crunching and data analytics, and it is repeated cycle after cycle. Given the nature of the activity, it is tempting to imagine that a self-learning AI application could do at least as good a job as a human planner at forecasting demand.
A closer look, however, reveals that there are some serious challenges to AI successfully penetrating the demand planning market. These challenges are not so much technical as they are managerial. Even if AI does not become a significant contributor to demand planning accuracy, addressing these challenges can only improve a company's demand planning performance.
The need for data and digital savviness
The most striking challenge that companies face as they apply AI to demand planning is the availability and accuracy of data. The more data that is provided to an AI application, the more robust the resulting conclusions are, making data availability an essential foundation to a successful AI implementation. Internally, companies already struggle to maintain accurate data, even for the most basic of elements such as product code. Ever-accelerating product launches and shrinking product lifecycles mean more product churn than ever. One corporate head of planning that we spoke to said: "Let's show we can correctly link product codes in substitutions (where one product transitions into replacing another) before thinking about AI."
In addition to internal data, a good demand plan also requires external data in the form of market intelligence, such as competitor actions, customer behaviors, and trade disruptions like price changes and sell-out data.
Furthermore, all of this data needs to be interpreted correctly. For example, in order to build a correct demand plan, an accurate baseline for demand must be established. One-off events, such as service issues and one-time promotions, have to be identified and accounted for, otherwise they may skew the planner's understanding of the underlying demand. This initial step of cleansing the data for statistical treatment is often a critical source of error, as it requires a clean view of the history of past activity of the product. One demand planning expert we spoke to claimed that from his experience this step of trying to define a clean baseline accounts for 60 percent of demand planning errors. In order for an AI application to learn from these one-off events they would need to be fully understood and coded, which is no small effort.
In addition to these data challenges, many companies today struggle with their digital culture and level of savviness. We spoke to many large multinationals that have made serious investments in demand planning tools, and almost all of them face the same struggle: Their planners prefer to build demand plans in Excel first, and then upload them into the expensive, integrated demand planning tools they must use to propagate their demand plans. The usual explanation for this resistance is that the tools don't have enough of the internal and external contextual data to build pertinent statistical plans.
A recent survey by Arizona State University, Colorado State University, and Competitive Insights revealed that Excel is by far the most common analytical tool used by supply chain planners, with advanced tools like supply chain control towers used by about 60 percent of companies.2 This matches our anecdotal observation that only about half of companies use an advanced planning system (APS).
The absence of data, resistance to using the existing suite of statistical tools, and low level of digital savviness represent non-negligible challenges to the deployment of AI-enabled demand planning.
The need for one set of numbers
Demand planning is a critical activity in the sales and operations planning (S&OP) process. The objective of S&OP is to obtain alignment from all actors in the company, ideally ensuring that operations mobilizes its resources to supply what the business needs to meet its financial goals, while also ensuring that the financial goals account for the current operational constraints.
A fundamental pillar of the S&OP process is the notion of "one set of numbers," which means that operations and finance are working off a shared understanding of the forward planned activity for the business. The primary drivers for this goal are that no market opportunities are missed due to asupply/demand imbalance, and operations is focused on the true business needs rather than on an inflated demand that acts as a supplemental safety stock.
When a company ties its financial plan to the operational plan, general managers are driven to involve themselves—along with their commercial and marketing teams—in the demand planning process in order to have the most viable demand plan possible. Their involvement in the planning process is critical, as they can provide the demand planners with the valuable external market intelligence mentioned earlier. Just as importantly, from a managerial perspective, having one set of numbers means that any effort by general managers to manipulate the demand plan would also change the financial pan, which they are loath to do as it constitutes their commitment to executive leadership.
When AI is used to generate a demand plan, that demand plan becomes part of the "one set of numbers." Otherwise general managers would be tempted to return to old reflexes such as considering the demand plan outside their sphere of interest, not being as committed to providing demand planners with the necessary external data and market intelligence, and perhaps once again adjusting the numbers to their subjective tastes. But maintaining the tie between the AI-generated demand plan and the financial plan would require asking general managers to allow their financial projections to be generated by the AI application. This would be a consequential management hurdle for supply chains to overcome.
That's because the introduction of AI-generated demand plans would bring with it what is termed the "explainability problem" of AI.3 This term describes the reluctance managers have to using AI applications that seem like a "black box," where the reasoning and logic used to obtain the results are difficult to explain, even if they are of high quality. The explainability problem is currently a tangible hurdle for successful AI deployments and is even driving some AI proponents to suggest solutions that may be less accurate but more explainable to the target business community.4
Our research suggests very few companies today have truly achieved a "one set of numbers" in practice.5 Having a more accurate, AI-enabled demand plan at the expense of placing a serious obstacle to implementing S&OP (due to the explainability problem) does not seem like a necessarily winning trade off. In other words, are companies better off having a (perhaps) highly accurate AI-generated demand plan that does not reflect the true business activity due to lack of alignment, or a slightly less accurate non-AI generated one that is aligned with the business ambitions?
The explainability problem doesn't preclude the use of AI for demand planning, but it does suggest that it be considered only for companies that have achieved very high S&OP maturity and integration between the operational and financial planning activities. This maturity would likely correspond with both more digitally savvy demand planners and a higher confidence of general managers in the ability of the demand planners to provide an AI-generated demand plan that represents the most accurate view of the forward business activity.
AI as accelerator
The challenges to applying AI to demand planning shouldn't be seen as insurmountable hurdles. Rather, AI could be an accelerator that pushes companies to confront these data and managerial issues head-on. Indeed, even if AI does not become a significant contributor to demand planning accuracy, addressing these challenges—data availability and accuracy and a willingness to use sophisticated analytics tools—can only improve a company's demand planning performance.
A sound, participatory S&OP process that assembles and leverages robust and accurate internal and external data to reach a consensus number for both operations and finance should be the target for all companies. If the IBF survey prediction is correct that the coming years will see deep contributions from AI in demand planning, these fundamentals of data management of managerial processes will have made it possible.
1. Eric Wilson, "IBF Survey: 70% of You Say AI is Future of Demand Planning," Demand-Planning.com (June 11, 2018), https://demand-planning.com/2018/06/11/ibf-survey-results-ai-demand-planning/
2. Zac Rogers, Lisa Harrington, Toby Gooley, Dale Rogers, Tami Kitajima, and Richard Sharpe, "Big data analytics in supply chain: Tackling the tidal wave," CSCMP's Supply Chain Quarterly (Quarter 3, 2017), https://www.supplychainquarterly.com/topics/Technology/20171030-big-data-analytics-in-supply-chain-tackling-the-tidal-wave/
3. Sanjay Srivastava, "Explainable AI: We need to know why algorithms make their decisions," Data Driven Investor (Nov. 7, 2018), https://medium.com/datadriveninvestor/explainable-ai-we-need-to-know-why-algorithms-make-their-decisions-1f5d45cccc84