Prediction is at the core of running an efficient supply chain. Inventory management, in particular the prevalent challenge of stockouts for sustained periods, has made intelligent prediction more important than ever in enabling our vastly complex supply chains to operate in real time and deliver on customers’ needs. However, the current state of the art in machine learning relies on past patterns and correlations to make predictions of the future — which makes it prone to fail amid shifts in data distribution.

The starting point for putting this right is shaking off the misconception that machine learning is synonymous with Artificial Intelligence; the real AI revolution only starts when machines can learn like scientists – looking at causal factors, as well as data, and making reasoned judgements for true intelligence in decision making. Today's learning machines have superhuman prediction ability but aren't particularly good at causal reasoning, even when fed ever greater amounts of historical data to meet their appetite for finding and exploiting statistical regularities.

It is largely true that supply chains face an inevitable transformation by AI. However, AI’s focus so far on data has missed the fundamental problem: the world changes very fast, and basing actions purely on past data can lead to sub-optimal decisions. However complex available methodologies may be, they are only able to infer correlations.  For outcome-focused decision making, machine learning needs fusing with, for example, “domain” expertise from humans to make more sophisticated ML algorithms.

The need to understand the cause and effect of possible actions in order to affect desired outcomes has long been understood in fields such as economics and medicine, yet only recently has it begun to emerge in industry, let alone in the supply chain. When the causal drivers of demand or supply in the world change, even sophisticated curve-fitting models can make worse decisions than the toss of a coin. It’s not just a question of a data scientist retraining a model to reflect external changes; the model is still left working with static — albeit revised — data.

Progressing from predictive to prescriptive solutions for supply chain problems requires more than data scientists; it needs a holistic approach to building systems. At an extremely granular level, this involves taking data about time frames and products, shipping distances and times of manufacture, and using it in knowledge and context about prevailing external factors. At present, decisions are being optimized largely at a “macro” level.  It's not unusual for a major business to use only three machine learning models to address 10 million possible permutations of time frame and location for shipping. However, by building in causality, it’s possible to reduce waste, help the environment and improve profitability.

The next generation of AI will deliver KPI optimization platforms that look at a business as a whole, including understanding causal elements, in order to make critical business decisions. Be warned: the technology won’t be downloadable from the open-source community for bending into shape by a few data scientists; it will involve “thinking” at a much higher level of automation.

Causal AI enables visibility of the entire supply chain in order to quickly understand and take action to mitigate delay. Next-level AI isn’t about being satisfied that predictions are right. It asks: are we making the right decisions? Can we infer the impact of our decisions? Do we know the root cause of our outcomes?  After all, KPIs and ROI are outcomes to decisions, which require a causal element. Supply chains are affected by myriad external factors: relationship management; regulatory environment; operational risk; expert judgement; budget constraints; business context. Improving On Time, In Full (OTIF) service levels to a significant degree will be achieved more speedily with a full consideration of causal factors, as well as using casual AI to assess what-if scenario and optimization planning.

So how might it work in practice? Through causal discovery and inference algorithms, millions of data features are defined and connected — not just statistical, but also causal relationships. Supply chain knowledge is embedded, allowing subject matter experts to inject more accuracy to a causal graph. Since supply chains are generally very complex, this domain element is vital. The combination of top-down human expertise and bottom-up data discovery is very powerful.

Specific intelligence on materials and shipment, supply and demand, production and purchase orders, or sales activity and orders are all areas from which it may be possible to identify factors conspiring to cause delay and friction in the chain. When a combined approach is adopted to achieve next-level causal AI, it may be deployed in numerous use cases, for example, a process of root cause ID to remedy sales order process delays.  This root cause may be used within a decision application that is able to provide actionable recommendations for business users and domain experts. In the case of such delays, it might consider the impact of capacity optimization in identifying the top five centers in a network for processing an order. These may be actioned programmatically, with outcomes to be tracked, reducing time and cost.

AI has for some time been proposed as a transformative development for the supply chain. McKinsey estimates that organizations globally stand to gain between $1.4 and $2 trillion in revenues by using AI in manufacturing and supply chains. Nonetheless, in reality a level of machine learning capable of truly optimal decision making is only now emerging.  Using cause and effect, a new category of intelligent machines that reason as humans do will become a revolutionary tool for solving real world challenges. On a three-to-five-year view, the future for AI in the supply chain is very bright indeed.

Jerry Stephens is GM of Supply Chain Management at causaLens