I have just completed a global market study on the supply chain planning software market. This gave me the opportunity to talk to executives across the industry.
Machine learning is hot. Solution providers in supply chain planning (SCP) tell me customers want to know how these technologies will be used in future SCP solutions. But machine learning is just one form of intelligence that can be embedded in SCP applications. The growing intelligence of these solutions ranges from better integration frameworks all the way up to fully automated planning.
Better integration frameworks
Integration frameworks allow data from multiple sources and networks to be pulled into planning solutions much more easily. Logility’s Karin Bursa, an executive vice president, points out that “many companies have multiple ERP systems.” She sees faster integration with better certainty and master data management, as a key differentiator for Logility. The master data logic understands the range of data that is appropriate for a particular field and can track and highlight when inappropriate data gets entered. Logility’s solution also uses net change logic. In other words, their system only looks at data elements that have been updated or changed. This makes same day or inter-day data updates more efficient.
Vikash Goyal, a vice president of product strategy at Oracle, gave a different example of this. One source of useful data comes from supply chain network solutions like those offered by E2open, GT Nexus, the Ariba Network, or other providers. If a company wants to collaborate on order forecast or purchase orders with a trading partner who is on a certain supply-chain network, for example GT Nexus, they can select that network provider from a drop-down box and specify the trading partner along with collaboration parameters. It is all automated.
Robust role-based views
This is not a new area of investment; it has been going on for several years. Many suppliers have invested in easier to use interfaces, particularly Excel-style interfaces. These interfaces have workflows that allow planners to tackle the most important planning problems in order of importance. Demand planners may want to view forecasts in units by week at ship to locations. Financial planners may want to see monthly views of revenues by business unit. Many suppliers offer integrated business planning (IBP) modules, sometimes called supply-chain control towers or cockpits, that allow for a variety of views by the wide variety of actors in a corporation involved in balancing supply with demand in ways that maximize the company’s strategic objectives. Those objectives might differ by product or customer and can include things like profit maximization, achieving revenue targets, gaining market share, and other things as well.
Anaplan’s Vivek Soneja, the global head of the supply chain management line of business, points out that having robust capabilities in this area involves more than an investment in a modern user interface. Anaplan provides financial, logistics, supply-chain, workforce, and other connected planning models. These models now can consume predictive intelligence from machine learning and translate them into time phased plans and decision support. Changes in one model flow through and diverse planners in different departments can see how changes in one area affect margins, revenue, transportation costs, capital investments, and many other areas across the business.
Bigger, better solves
There are always new problems to solve. Omnichannel is the best current example of that. Manhattan Associate’s Scott Fenwick, director of product strategy, points out that when a new flow is supported, like order online but pick-up-in store, inventory allocation decisions need to change. But picking up that shift in the demand signal can be difficult. They are using machine learning to help solve this true demand problem.
JDA has exciting product development centered on the idea that demand management solutions generally approach forecasts as if they were always a normal distribution from which the center point, or mean, is taken as the forecast. But not all distributions are normal. They are seeking a probabilistic approach where the costs of lost sales are balanced against the costs of replenishment and waste. The analysis uses machine learning technologies at the item/store level to demand. Influencers include weather, competitor promotions, store traffic by day of week, and potentially many other things.
The more constraints an optimization solution can model, the better the optimization. Over time, most suppliers’ models have gotten bigger, and the in-process memory technology used to solve the problems have gotten much quicker. River Logic, for example, has added labor/shift availability and regular/over time components into their ability to solve capacity planning problems. Further, what used to be distinct models for demand, inventory optimization, and replenishment, have been condensed into one model for several suppliers.
The self-healing supply chain
The outputs of supply-chain planning systems depend upon key supply-chain parameters. One of the most important set are lead times. The longer the lead times, the greater the variability around a lead time, the more safety stock companies must invest in. With its self-healing supply chain, Kinaxis is investing in machine learning to help keep these parameters up to date. In many cases, lead times are taken out of an enterprise system, and never verified. Even if they were right to begin with, they are never updated. Kinaxis is investing in comparing the historical lead time against what more recent enterprise data is showing. But they seek to use not just internal data but also mapping the historical lead times against third-party data like weather data and examine how lead times have been affected by various types of weather.
In supply planning, a self-healing supply chain would seek to determine when a key production machine might go down, and then using planning to proactively deal with that situation. AspenTech is probably the closest to a productized solution in this area.
There are also views and workflows centered on exception situations. Exception resolution comes with inline analytics and the ability to drill right into the problem. Oracle, for example, is approaching exception situations by including a business matrix that includes a variety of metrics like demand at risk, fill rate, and costs. When a planner comes into the solution, the top problems are identified visually. For example, a factory planner may see the demand at risk in a certain region based on throughput issues at the company’s factory in Mexico. For exception situations that occur with some regularity, the planner can see options that are specific to the problem at hand. The planner could say to himself, “For demand at risk, these are the options and here is the cost to resolve that issue.” The planner can accept the proposed resolution in the user interface. In the future, several suppliers foresee using Siri-type technology that will allow a planner to verbally accept a resolution.
Alternatively, if the resolutions do not appear straightforward, the planner can jump into the application and run more scenarios to generate other options. Often SCP solutions include a social media platform that allows the problem to be shared in real time with the right team members needed to resolve the situation.
How much exception situations can be automated is a subject of debate. Exceptions can be so numerous and diverse that large scale exception automation may be problematic. Some suppliers of supply-chain planning are looking to “solve” this problem by using pattern recognition to see how planners resolved similar problems in the past and then suggest a similar resolution the next time the problem arises. The more at-bats a system has, the faster it gets smarter. But a particular exception may occur only a couple of times a year. And the resolution that was used in June may be completely different from the one in December. It may take hundreds, but more likely thousands, of at-bats before the machine can begin to unravel why a planner did one thing in one situation and a different thing in the next.
Oracle is approaching this issue by developing a planning advisor who will sit in the background and observe how a planner is doing his or her work. The advisor might ask, “In the past when this problem arose, you had the factory pay overtime to meet the order commitment. Are you sure you want to go forward with short shipments?” If the planner does want to proceed with a new alternative, then the system asks him to provide reasons. This context may help either people or systems more quickly build a hierarchy of exception-resolution rules.
But other vendors believe the solution to this is a built-in robotic process automation to build exception rules. “Without the need for programming, superusers do a lot of stuff that looks like programming,” says Johanna Smaros, cofounder and chief marketing officer at RELEX Solutions. “We expected to get 100 of this SKU today; we only got 50. That could trigger allocation rules.” The rule might make the decision based on what customer needs the inventory the most, or prioritize some store formats over others, or some other decision criterion could be used. AERA Technology has the same approach.
Many suppliers’ demand planning engines, while not fully automated, have become much more automated. Many demand planning solutions will recommend when the algorithm used to forecast a stock keeping unit may need to be changed, based on changing demand patterns, and then suggest the algorithm that should be used.
Replenishment and demand forecasts are tightly linked, and semi-automated planning is also becoming more common around replenishment. Logility, for example, briefed me on one of their retail customers that has automated 80 percent of standard replenishment to stores of their apparel products at the size/color level. Freeing up time for planners to engage on promotions and other forms of event planning. Planners are now able to move away from routine activities and focus their efforts on more strategic initiatives. The result is more precision on replenishment orders and priorities combined with greater employee satisfaction.
Fully automated planning
Fully automated planning is the most controversial form of intelligence. E2open’s demand-planning system is perhaps the best example of fully automated planning. But E2open’s competitors are quick to point out it is a black box solution. Humans cannot analyze the output and understand how the system generated the forecasts it produced. There are big, sophisticated companies that believe in fully automated planning. Procter & Gamble, for example, has publicly spoken about using the E2open solution.
Other providers of SCP solutions point to some problems this kind of planning system. Harish Iyer, vice president of industry-solutions marketing at Kinaxis, asserts, “Someone needs to be accountable.” If something goes wrong, and users have just accepted a plan from a black box, how do you hold the planner accountable? “And the further up the hierarchy you go, the more unacceptable this is.” If a public firm misses its numbers in the current quarter, “can you imagine a CEO telling Wall Street [he doesn’t] know why?”
Currently, fully automated planning is not viable for supply planning for a variety of reasons.
In conclusion, supply-chain planning systems have become more intelligent in numerous ways, some of which include machine learning, though many do not. Machine learning has great potential, but customers are looking for productized solutions, not science projects. In some areas of SCP, machine learning is proven. In other areas, we are still in the science-project phase.