What you’re likely getting wrong about AI in supply chain planning

As global supply chains struggle to recover from the pandemic, new environmental crises and the war in Ukraine create further disruptions. And we’re kidding ourselves if we think that at some point things will settle into a ‘new normal’.

In fact, Gartner says we need to get used to the idea of a ‘no normal’, which it describes as “an environment of uncertainty and ambiguity that requires continuous flexibility, innovation, and investment (or reinvestment) in data and analytics strategy.”

In light of all this, it may seem that using AI to plan supply chains is a no-brainer. Find the right smart planning tool and take all the guesswork out of the process, right?

Wrong!

As we’ll explore in this article, applying AI to supply chain planning is a smart decision for sure, but it’s not a quick fix. There’s considerable groundwork to lay.

Like so many aspects of supply chain technology, AI gets hyped up in the media and, consequently, is often misunderstood. In this article we’ll demystify AI’s role in supply chain planning–what it excels at, and its limitations.

Spoiler alert: the power of AI automation is amplified in supply chain planning when combined with the knowledge and ingenuity of human planning teams.

AI is not the same as ML

First let’s make a distinction between two correlated, but often mixed-up, parts of computer science — AI and ML. AI is a broad term used to describe intelligent pre-programmed software, which is designed to simulate human thinking and behaviors. When we talk about automation systems that are engineered to replace human capacity, we’re usually describing some form of AI. Machine learning is a subset of AI that enables a machine to automatically learn from past data, so it can produce more accurate results.

Not all AI systems operate on the same principles. Some are opaque systems engineered to carry out a specific function with no human intervention — for example, the new ‘driverless pods’ being trialed at airports and other controlled settings. More flexible AI systems expose some or all of their algorithms and may also allow for human overrides. Think of the self-driving car that lets the human driver take over in heavily pedestrianized areas, narrow streets, and other complex situations.

How AI-driven supply chain planning systems differ from traditional ones

AI-driven supply chain planning systems are built using algorithms that have been coded to optimize this process to meet a set objective. Our software, for example, specifically works to optimize inventory levels against a target service level. Much like the self-driving car, it also allows planners to override the system to handle exceptions, like store closures during a pandemic.

AI-driven planning systems generate what’s known as probabilistic forecasts. Instead of giving one number, probabilistic forecasts give a range of possible numbers centered on the one that’s most likely. From there, planners can intervene using their own knowledge or integrate machine learning to refine forecasts further.

In the case of supply chain planning, ML is often applied to improve demand forecasts that AI-driven systems generate. It has shown to be particularly useful in identifying seasonality patterns, promotions planning, new product introductions and other processes that are repeated and gain intelligence over time.

Lennox Residential Heating and Cooling, which has a highly complex, multi-echelon supply chain applied machine learning to reliably model highly variable seasonal demand patterns. It sifts through hundreds of thousands of SKU-Locations to identify “clusters” of those with similar seasonality profiles. These enhanced seasonality clusters substantially increase peak period forecast accuracy. As a result, it was able to boost service levels by 16%.

Because machines learn from past experiences, there needs to be enough past data for ML to be able to make the right decisions. That means ML is not appropriate for planning in intermittent, exceptional, and changing circumstances — in other words, crises. For that you need a smart and controllable AI system that can allow human or machine overrides.

Key takeaway: Despite ML’s considerable prominence in supply chain literature, it is not a ‘magic wand’. You can’t just sit back and let ML run. You need to invest time and effort to build and maintain ML engines.

The foundation for AI: robust data and demand modeling

Before you can expect to benefit from AI automation, you need to invest the time in building a robust data model of your distribution network. The data model specifies which data you want to capture, where it’s hosted, and how it relates back to different parts of your business. The goal here is to ensure clarity and consistency to all the people and systems involved in supply chain planning.

After you’ve built your data model, then you can start modeling your demand. Your demand model should take into account as many different internal and external factors that shape demand as possible. Internal factors include planned new product introductions, promotions, and price reductions. [link to blog]. External factors might include social media feeds, weather forecasts, and competitive information.

Key takeaway: data and demand modeling is not something that you can do in the midst of a crisis. In most cases it takes around three months to get right.

Franke — managing extreme planning complexity during Covid

Franke is a global $2 billion kitchen products company operating in 42 countries. It manages roughly 300,000 finished goods SKUs using our digital planning system for 140 sales areas in a five-tier distribution network. That translates to 1.7 million SKU_Location combinations that are forecasted in each of Franke’s bi-monthly planning cycles.

Fortunately, Franke had built its data and demand models and had its ‘normal’ planning system and processes up and running smoothly before Covid hit. At the global level, only two people managed the process centrally using our planning system to generate baseline probabilistic forecasts. Crucially however, a team of roughly 100 people across Franke collaborate on demand planning and forecasting, adding their knowledge of customers and local market dynamics to adjust the baseline.

At the start of the pandemic, when planners mostly had to contend with supply delays coming from China, the situation was pretty manageable. The team was able to isolate all suppliers of ‘at risk’ materials and manage customers’ expectations.

When Covid spread to Europe however, the situation really escalated. Each country had different government policies and demographics, which meant the crisis affected each in different ways. Its planning process was too complicated for the new climate. At the same time, many people involved with planning were taking time off or furloughed, while others were working around the clock. There was simply no existing strategy to deal with a crisis of that dimension.

Fortunately Franke was using an automated planning system that allowed for manual intervention. This meant it was able to modify its approach to manage through the crisis successfully. Among the steps Franke took:

• Focused on obtaining high-priority ‘A’ products to avoid stock outs, requesting immediate delivery and using air shipments
• Temporarily moving to a simplified weekly demand planning cycle; reallocated all planners from non-essential tasks to crisis management
• In the new weekly cycle, planners continuously cleansed the probabilistic forecasts to reflect sales order postponements and purchase order changes; updated planned forecast deviations daily.
• Closing the calendar on days where market closures made it impossible to sell. This ensured that these non-sales days didn’t impact any future forecasts.

Franke’s full presentation on their Covid strategy is available here.

By deploying AI automation and human resources effectively, Franke limited forecast degradation surprisingly well. It only lost 15% accuracy during the peak of the Covid crisis in an environment where total variations were around 80%. It also managed to avoid the bullwhip effect, even in markets like the UK, which had the most variability due to market closures.

An even brighter future

The past decades have seen huge progress in AI-augmented supply chain planning. However, I believe the best is yet to come. Today it’s mainly large global companies like Amazon and Google that enough have collected enough data to use machine learning for exception planning.

However there is some momentum building behind the scenes to eventually move to a new paradigm — open cloud platforms for supply chain. The idea here is that multiple organizations collaborate to share data, systems, and knowledge. Theoretically there is the potential in this model to gather enough data to apply machine learning to significantly improve resilience during crises. But it’s some way off.

In the meantime, I hope this has helped clear up some of the misconceptions out there about AI in supply chain planning. It won’t offer an instant cure for all supply chains’ current ills. However, for companies like Lennox and Franke with highly complex distribution networks facing ongoing change and disruption, it might be just what the doctor ordered.