BETA
This is a BETA experience. You may opt-out by clicking here

More From Forbes

Edit Story

Companies Improve Their Supply Chains With Artificial Intelligence

Following
This article is more than 2 years old.

Many large enterprises use one form or another of a supply chain application to help manage their supply chains. Supply chain vendors have been touting their investments in artificial intelligence (AI) for the last several years. In the course of updating our annual research on the supply chain planning market, I talked to executives across the industry. Alex Pradhan, Product Strategy Leader John Galt Solutions, told me that “all planning vendors have bold marketing around AI.” But the trick is to find suppliers with “field-proven AI/ML algorithms” that “have been delivered at scale.”

Further, while artificial intelligence helps solve certain types of problems, Jay Muelhoefer - the chief marketing officer at Kinaxis pointed out - optimization and heuristics work better for other types of planning problems. This article, which is focused on the different types of artificial intelligence used and the types of problems they are solving, is aimed at helping practitioners cut through the hype.

Let’s start with a definition: 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 artificial intelligence (AI). AI can refer to several different types of math. But, in the supply chain realm, machine learning (ML) is where most of the activity surrounding artificial intelligence has been focused.

Garbage In, Garbage Out

It is also worth pointing out, that based on this definition, not all forms of machine learning are particularly complicated. Planning applications don’t work well if the master data they rely on is not accurate; this is known as the “garbage in, garbage out” problem. Artificial intelligence is beginning to be used to update the data. Lead times, for example, are a critical form of master data for planning purposes. Having an agent detect how long it takes to ship from a supplier site to a manufacturing facility, and then doing a running calculation on how the average lead time is changing, is trivial math. The agent technology is much more complicated than the math. Relying on humans to update this data has not worked at all well; people just don’t want to do it.

But sometimes fixing the bad data problem is complicated. In process industries the supply chain models used for optimization are much more complex than those used in other industries. The processing units in an oil refinery, for example, operate at high temperature and high pressure. These constraints need to be understood. So, models for heavy process industries often include first principle parameters. First principles reflect physical laws such as mass balance, energy balance, heat transfer relations, and reaction kinetics. The first principles are important to understand yields, as well as the energy requirements for running the equipment.

AspenTech has developed in a process simulator which is tuned with real plant operating data. During development, the models automatically perform thousands of permutations and perturbations of the first principles model to create a large data set to which AI algorithms applied. The AspenTech models combine the classic first principles approach with the modern pure data-driven approach. Starting with a first principles model, according to AspenTech, improves accuracy significantly. They tell me, the model with either a first principles or pure data plus AI, the model accuracy would be in the 90-97% range. But hybrid models that combine first principles, data-driven models, and AI, they have 99+% accuracy.

Demand Planning Models Really Can Learn

A supply chain planning model learns when the planning application takes an output, like a forecast, observes the accuracy of the output, and then updates its own model so that better outputs will occur in the future.

When you look at machine learning this way, artificial intelligence for supply chain planning is nothing new. Machine learning has been used to improve demand forecasting since the early 2000s. But machine learning for demand forecasting is much better than it used to be. There are far more forecasts being made in far more planning horizons and at a greater degree of specificity today than 20 years ago. For example, forecasting how much of a particular product will be sold in a particular store is far more intensive than forecasting how many products in a product family will be sold in a region. This explosion in the number of forecasts would not be possible without the latest generation of machine learning. There were only a few SCP suppliers with mature capabilities in this area a few years ago. Since then, virtually every supplier I talked to in the process of updating this year’s Supply Chain Planning Market Analysis Study has said they are investing in this area.

One example of the value of machine learning in demand planning comes from Mahindra & Mahindra. Aniruddh Srivastava, Head of Demand and Supply Planning at Mahindra & Mahindra, said at Blue Yonder’s Icon user conference that artificial intelligence and machine learning algorithms are the cornerstone of their strategy. Through their partnership with Blue Yonder, Mahindra & Mahindra was able to increase forecast accuracy by 10%. A better forecast leads to carrying less inventory while maintaining or even improving service levels. The improvement in forecasting contributed to an increase in service levels by 10% while reducing inventory investment by 20%.

But that was pre-COVID. But after the pandemic hit their safety stock was increased by 30%. “Post-COVID it was not about savings,” Mr. Srivastava explained. “The game changed to a global competition for the same set of raw materials.” This division makes automotive spare parts, so the competition was to secure semiconductor chips.

During the pandemic, forecasting accuracy was terrible. Forecasting is based on the presumption that history repeats itself. As an E2open forecasting benchmarking report pointed out, “for companies trying to predict demand in March of 2020 as the world was descending into lockdown and everything was being turned upside down, what happened in March of 2019 had little to no relevance.”

But if there was any silver lining it was that companies that made use of planning systems that combined demand sensing – the use of multiple, real-time signals (like sales in a particular store or shipments from a retailer’s warehouses to their stores) – and machine learning, had significantly less error. And the companies that used these solutions, saw their forecasts improve much more quickly than traditional solutions.

In making demand forecasts, one can look at product history. An alternative is to look at customer behavior surrounding how clusters of customers buy these products. QAD Dynasys is one of several suppliers investing here.

One thing that is difficult to forecast are new product introductions. The way this forecasting is done is through the use of attributes. If you are looking at a purse, attributes would include the material it is made of, size, color, and other things as well. To the extent that one product is like another, it may be easier to forecast. But which attributes matter? Infor is using machine learning looking at attributes and past launches to make this determination. Solvoyo and Lily AI are using another form of AI, image recognition, to tackle this problem. Getting merchandisers to enter the attributes has not worked well. Merchandisers see this as an unimportant, dull task and they just don’t take the time to do this properly.

Machine Learning and the Sustainability Feedback Loop

One real trend ARC has seen this year, is the increasing investment supply chain planning suppliers are making to improve the ability of SCP to help companies reach their sustainability goals. Cyrus Hadavi, the CEO of Adexa, provides a good explanation for how SCP solutions can calculate the carbon footprint associated with a plan. “The way this works is that every element in the supply chain is given a carbon index, absolute or relative. That is every machine, factory, DC, mode of transportation, supplier, product, material, etc. These indices then become attributes of these objects. Every time we plan and use any of these elements, the system can project the total carbon footprint of the projected plan. Additionally, we have embedded self-correcting algorithms - using ML - in building the model of the digital twin. For example, we learn the energy efficiency of the resource in the month of June vs. December.” In addition to carbon emissions, these attributes can be used for other forms environmental and governance goals as well.

So, a plan can be produced that predicts the emissions. After a plan is executed, the actual emissions that occurred can be measured, and it is possible to see how close the plan came to what occurred. Just as a demand planning solution compares the forecast to what actually sold and uses machine learning to improve the machines forecasting capabilities, a similar feedback loop can exist with sustainability.

AI is Also Used for Supply Planning

Artificial intelligence can also be used in supply and factory planning. But on the supply planning side it is not about using machine learning to select the right algorithms to improve the plans. When supply plans don’t pan out it is less about the model than it is about a data quality issue or an unexpected occurrence. An example of an input issue would be, “We thought it took 20 minutes to set up this machine to make product C, but it really takes 60 minutes when product A was made right before product C.” An example of an unexpected occurrence would be a critical piece of machinery breaking down.

Machine learning is being used to predict machine breakdowns. But very few vendors are taking those alerts and automatically feeding them into their manufacturing planning solutions. AspenTech has probably done the most in this area. AspenTech, for example, is using predictive analytic inputs on when key machinery in a refinery will break down to allow alternative production schedules to be generated in a more autonomous manner. AspenTech’s advantage is that they have both asset management (a solution that can use machine learning for the predictive maintenance alert) and the supply chain planning models those alerts can feed.

Natural Language Processing is Also Being Used

A less commonly used form of AI in supply chain applications is natural language processing (NLP). Google’s Alexa uses NLP to understand a person’s command and then play the music they want. There is a desire to use NLP to allow planners to tell a planning system what to do so they can focus more of their time on higher priority problems.

But Coupa and Oracle are also leveraging natural language processing for supplier risk assessment. Humans don’t speak with a clarity that machines can understand. A company can go bankrupt, and a machine could be programmed to understand that. But on social media someone might say that a company is about to go “belly up.” Machines don’t understand this type of “unstructured” data. NLP helps to make sense of these kinds of data. Oracle’s DataFox is accessing databases with important company information, but it also has web crawlers examining huge numbers of online news sites as well as social media to discover negative news about a company. That news could be an impending bankruptcy, unhappy customers, key executives leaving the company, or many other things. These “events” are turned into supplier scores, and if significant the score goes flagged in the Oracle procurement system. Now Oracle is working to connect these scores to the planning systems. At Coupa supplier risk is also flagged for single source or capacity constraints. This can then be leveraged by their supply chain design solution to improve risk mitigation.

AI Drives Autonomous Decision Making

Companies need to take machine learning driven demand-side predictions – which are particularly good at granular short-term forecasts - and adjust production accordingly. The closer in time a plan’s creation comes to the actual execution of an order, the more a planning system becomes an execution system. The idea is for supply planning application to digest a short horizon demand signals into meaningful plans by using machine learning to suggest courses of action for planners. These suggestions are based upon the way planners had previously solved the same kind of demand/supply disruption. However, this kind of AI does not work out of the box. The system observes planners’ actions over time and then learns to make the pertinent suggestions. QAD and Noodle.ai are among several suppliers working in this area.

In the last couple of years, RELEX Solutions has developed new capabilities for autonomous capacity balancing. In short, the AI algorithms can pull orders forward (for products with longer shelf-life) to level out the flow of goods into distribution centers and stores as well as to adhere to time-dependent capacity limits. Johanna Småros, Co-founder & Chief Marketing Officer at RELEX, points out “the current difficulties in finding staff have really raised awareness of the value of being able to plan ahead to ensure availability and efficient use of human resources as well as to plan around this availability when it becomes a bottleneck in the supply chain.”

Blue Yonder, in turn has developed a machine learning powered Dynamic Segmentation solution that automatically groups customers with similar fulfillment or procurement needs based on data changes, and then develops distinct supply chain operations to meet those specific requirements. This enables planners to provide differentiated service levels based on customer value and business parameters

While this article stemmed from research ARC is doing on the supply chain planning market, and most AI investments have been focused on planning applications, it is worth pointing out that AI investments are increasing in the supply chain execution realm as well. Companies like Oracle, Manhattan Associates, Koerber, and Blue Yonder, are all increasing the R&D in AI in their supply chain execution systems. A transportation system that applies machine learning to predict how long it will take a truck to make a delivery is one example of this. A warehouse management system that can digest a prediction of what ecommerce customers are apt to buy, and then drop the right work orders at the right time to the warehouse floor, is another example of this. 

Final Thoughts

To sum it up, Madhav Durbha – the vice president of supply chain strategy at Coupa Software – said that artificial intelligence is becoming much more widely adopted “due to progress occurring on several fronts at the same time. These include the development of new machine learning algorithms, “computing power, big data analytics, and acceptance by industry leaders.”

But remember, AI only fixes supply chains to a degree; this is not like waving a magic wand and seeing your supply chain problems suddenly vanish. Nevertheless, AI really is improving planning, and it is increasingly being used to improve order execution as well.

Check out my website