You may not realize it, but for many years, AI has been successfully used in supply chain planning, particularly in demand forecasting and demand planning.
The public launch of ChatGPT in 2022 created a frenzy of interest in AI. But there’s still a lot of confusion over what AI is and how it can help users of supply chain planning software.
There are a few reasons why you may be hesitant to use AI in your supply chain planning (SCP):
- A growing general skepticism because of a perception that the value companies are getting from some generative AI (Gen AI) solutions doesn’t match the investment and hype.
- A number of research studies saying that the majority of AI projects fail.
- Fear that AI requires a huge investment, hiring a staff of PhDs, and incurring a lot of risk, leading to an impression that it’s only for huge corporations with large R&D budgets.
- SCP vendor hype, which has only added to the confusion.
This article will address these concerns and clarify New Horizon’s viewpoint on AI in SCP.
A few quick definitions will help set the stage.
1. AI Isn’t Just One Thing (and It’s Much More Than ChatGPT)
AI is an umbrella term for a whole range of distinct technologies.
They all perform tasks that previously required human intelligence, such as learning, reasoning, problem solving, perception, and decision making. Many of the enabling technologies have been around for years and have been proven in business applications.
The below diagram provides a high level view of the relationships among the dominant AI approaches common in supply chain planning today. Note this is a simplification that leaves out important nuances and many other forms of AI, but it is sufficient for the purposes of this article.
Below are descriptions of each approach shown in the diagram:
Machine Learning — A set of algorithms designed to learn patterns and trends from historical data. The purpose of machine learning is to predict future outcomes and generalize beyond historical data without being explicitly programmed. In SCP, machine learning is used in demand forecasting, intelligent decision-making, and much more.
Deep Learning — A type of machine learning that uses artificial neural networks, which are layers of inputs, outputs, and hidden layers to filter data, process it, and make predictions. Deep learning technologies can take advantage of diverse data types simultaneously, including for example structured data (like sales history and inventory levels), unstructured data (social media sentiment, news articles, weather reports), along with sensor data (IoT signals from warehouses and transportation telemetry).
Generative AI — A set of AI approaches including deep learning that can create new content —like audio, video, code, simulations, and text—based on patterns learned from large datasets.
Our article, AI for Demand Forecasting: A Primer on Smarter Predictions, discusses some of these technologies for demand forecasting.
Why there’s so much confusion around AI
Much of the recent skepticism around AI relates to the new wave of Gen AI technologies based on large language models such as ChatGPT.
But AI is much bigger than these newer technologies, and we can’t generalize about AI based on the performance of any one AI technology.
Another source of confusion is the AI Effect, which states that the definition of AI is constantly shifting. Once an AI technology is proven, it’s often dismissed as not really AI but brute force calculation. As a result, AI is constantly being redefined to refer to unproven, cutting-edge technologies (which by definition aren’t proven and warrant skepticism).
2. AI Has a Long, Proven Track Record in SCP
Machine learning has been successfully applied to various areas of supply chain planning over the last 20 years.
Forecasting has been particularly successful because it uses a form of machine learning called supervised learning. Supervised learning is a process in which a model learns to predict an output (e.g., demand) from input data by being shown many examples in which the correct answer is known.
Because a machine learning demand forecasting algorithm can continuously learn by comparing its forecasts with actual demand, it can fine tune the forecasting model and avoid going off the rails or hallucinating the way ChatGPT can sometimes.
Machine learning approaches have evolved and now often incorporate artificial neural networks (ANN) and deep learning technology. For instance, New Horizon uses a number of ANN forecasting models in our Demand Planning application.
We have successfully deployed different AI technologies in other areas as well, such as in detecting erroneous data feeds, optimizing safety stock levels, doing root cause analysis of supply chain issues and recommending resolution actions.
While many vendors have raced to market with ChatGPT-like add-ons for their SCP solutions, the jury is still out as to whether Gen AI is useful in such applications.
New Horizon constantly evaluates new AI technologies. Just recently, we launched the New Horizon AI Agent. It uses generative AI and other AI technologies to allow users to get answers to natural language queries based on supply chain data in the New Horizon application. We’re working with early adopters now to evaluate the usefulness of this technology and improve the functionality based on user feedback.
3. AI Isn’t the Answer to Every Problem
The key to AI success is using the right technology for the right problem. AI isn’t always the solution.
A number of software vendors talk about an “AI-first” strategy: they use AI wherever it can be used, ahead of any other technology. This echoes the mobile-first strategy employed by application vendors to build their apps with a mobile focus from the beginning, rather than adapting desktop apps for mobile devices.
We take this approach ourselves, but with the caveat that we only use AI if it is the best approach.
AI often isn’t the right solution, and a number of AI-first vendors have taken the approach too far, producing software that underperforms traditional technologies.
In demand forecasting, AI is best suited to massive amounts of data where the AI can outperform traditional statistical forecasting models in figuring out what is driving demand.
With much smaller data sets, proven traditional models often yield more accurate forecasts.
At New Horizon, we blend tradition and AI models to get the best of both worlds.
Our forecasting engine uses over 20 different forecasting models, some traditional (such as ARIMA and Holt-Winters), some more sophisticated models like Facebook Prophet, and some cutting edge AI models that use artificial neural networks.
We use AI to automatically choose, for each product-location combination, the best model to maximize predictive accuracy. This approach yields the best of both worlds.
4. Transparency and Explainability of AI Models is Key
One reason for AI skepticism is a lack of transparency as to how it works and how it comes up with particular answers.
Case in point — ChatGPT hallucinations. ChatGPT now provides sources but when users dig in, those sources often don’t seem to have anything to do with ChatGPT’s answers.
There’s a similar issue with some AI technologies used for SCP — users are less likely to trust black box technologies which provide no explanation for their answers.
When it comes to demand forecasting, one simple first step to provide transparency and explainability is “due to” analysis.
As one example, an AI model may examine lots of historical demand and causal data to come up with a forecast.
But it’s not enough to get a total demand forecast number. Instead, it’s better to be able to see the components of demand — the baseline portion that’s driven by long term market trends and seasonality, with incremental components driven by factors such as weather and promotions shown separately.
This lets users see how the forecast was derived and determine whether it passes the sanity test.
Another example is forecast model choice. It’s not enough to come up with a forecast without providing any insights as to what’s behind the forecast.
To address this issue, New Horizon allows users to see which forecasting model our technology chooses for a given forecast and gives them the ability to override the AI recommendation and choose their own model.
This is appropriate if the user knows something about changing market conditions that the software doesn’t.
We continue to invest heavily in developing explainability features for our AI technology.
5. Purpose-Built Apps Often Outperform Custom Projects
The truth is, you don’t need a big company R&D department staffed with PhDs to develop custom AI models.
That’s fine for huge corporations with unique challenges that can justify the investment.
But for most midsize companies, a better approach is to use supply chain applications with proven AI technology embedded within it. This eliminates the risk of expensive R&D science projects that may yield little benefit.
Take demand forecasting, for example. A custom R&D effort staffed by PhDs may come up with a better forecast, but typically such efforts are applied only to a small subset of forecasts because this approach is not scalable across every product, location, and time horizon.
When proven AI technology is embedded in supply chain applications, companies can benefit from an AI-based forecast that is almost as good as that from a custom system, without having to go outside their operational supply chain systems.
These solutions can be used to create a forecast for every single product and location, every time they generate a new forecast.
Remember, “perfect is the enemy of good.”
A custom AI approach may yield a perfect answer for a small subset of your business but a proven, standardized technology embedded in your SCP system can provide 80% of the benefits across your entire business.
Rethink AI in Your Organization’s Supply Chain Planning
AI represents a huge opportunity to help companies improve supply chain performance. But it must be used in a thoughtful way. That means:
- Embedding proven AI technology in SCP applications. This makes AI benefits available for all planners, not just PhDs in the research department.
- Using the right technology for the right problem, which may be a combination of AI and traditional techniques.
- Providing transparency and explainability for all forecasts and recommendations.
To Learn More
Interested in learning how New Horizon’s AI-powered supply chain planning solutions can help you achieve your organization’s supply chain goals? Talk to one of our experts.