While machine learning and artificial intelligence (AI) have been used in supply chain applications for some time, there is an ongoing arms race to more effectively leverage both machine learning and artificial intelligence in demand planning solutions in new ways.
This is not surprising. Demand planning is one of the key applications in supply chain planning (SCP) suites. In ARC's recent global market study on this market, demand applications account for just under a third of a $2 billion plus market. And these applications are often the wedge purchase; the SCP solution that is first implemented by a company that then goes on to purchase other solutions in the suite.
Machine learning works by taking the output of an application (for example, a forecast), examining that output against some measure of the truth, and then adjusting the parameters or math involved in generating the output (forecast), and seeing if the adjustments lead to more accurate outputs. It can truly be said that the machine is learning from experience.
The definition of artificial intelligence (AI) is broader. Any device that can perceive its environment and takes actions that maximize its chance of success at some goal is engaged in some form of AI. AI includes expert systems that read and interpret written language or natural language processing applications like the iPhone's Siri.
Before any demand management solution is implemented, the master data needs to be cleaned up. This process of cleaning and normalizing the data takes far longer, costs more, and requires more ongoing effort than almost all companies anticipate. AI is being used to help normalize the data. So, for example, demand planning applications can be used to forecast how much of a product will be shipped to a large customer. In the database, Procter & Gamble, Porter & Gambel (sp.), Procter and Gamble, and so forth all refer to the same customer. But digital applications need precision; All these names must be spelled and punctuated identically. It is an easy task, but a time consuming one, for a human to do. With AI, the task can be greatly sped up. LLamasoft is one solution provider that is using AI in this manner.
Cyrus Hadavi, the CEO at Adexa, said that data can be cleaned in another way. Forecasts often involve the inputs of sales people. "What we are doing is data testing. We look at past data and look for patterns. Which sales managers have been consistently too optimistic? Which one consistently too pessimistic? We build this analysis into the software."
JDA and Logility have been using machine learning for close to twenty years to help classify stock keeping units (SKUs) in demand planning applications. Different algorithms should be used for forecasting different types of products or groups. By properly classifying the SKU, better forecasts resulted. This has become common among supply chain planning suppliers.
Over time, more data inputs were introduced into the demand planning process, and many companies are doing far more forecasts over more planning horizons (daily, weekly, monthly, etc.) and ship to locations. E2open began to use this technology not to classify an SKU, but to find the downstream data sources that increase forecast accuracy for each ship to location and planning horizon. In short, E2open is using machine learning in more of an end-to-end black box type solution.
Other forecasts are possible. Sujit Singh, the Chief of Operations at Arkieva, told me that in the last three years customers have begun asking them to make sales predictions. The VP of Sales is saying, "For these key customers who have a sales order placement pattern, study their data to know when to expect the next order. Use this sales forecast to ask a customer that is 3 weeks late placing an order, What's up? Why haven't you placed the order?" Further, using a variation of market-basket analysis for the B2B world, sales teams are able to create demand for other products.Â? This can not only improve the demand forecast, it can help the sales organization close business.
Demand planning solutions improved when downstream data such as point of sale became available. But there are many, many more data streams that have not been tapped. ToolsGroup and Relex Solutions have done trials to examine whether using social media and weather data will improve forecasts. Eventually, these may turn into enhancements that move into their standard solutions.
But many data sources will not be amenable to becoming part of a standard solution. A large paper distributor, for example, might be able to improve their forecast by using paper exports from Russia, but almost no other company could productively use that data. LLamasoft has introduced a demand modeling tool called Demand Guru to help companies develop more customized demand planning solutions. This tool allows a company to more easily clean, blend and then ingest time series data and then test whether that data set improves forecasting over some forecast period or ship to location. Historical third-party data can be imported into the tool and then the company can examine whether their past forecasts could have been improved by using that data set. The data could be anything from GDP data, demographic data, gas prices, housing starts, to the afore mentioned paper exports from Russia. It really is any time series data series that a forecaster believes might improve their forecast accuracy.
A few supply chain planning software companies are exploring the use of natural language processing to improve the user experience. Karin Bursa, an executive vice president at Logility, foresees a day in the near future when a demand planner asks the application at the beginning of the day, "What are the top three issues I need to address today?" Or, perhaps, "what are the top three promotions I should run to address declining sales in the southeast?"
But it is not only demand planning applications that will use AI. It will increasingly be used across the entire SCP suite. But that is a topic for another day.