Demand planning systems are often fixed systems. Retailers rely on a rule-based approach to help them forecast, replenish and manage stock. While rules do work, they tend to work best for rather consistent, predictable product categories and known trends. However, the COVID pandemic taught us that purchasing patterns can shift and become unpredictable. Think of shipping costs, labor shortages or waves of panic buying. This left a surge of retailers looking for artificial intelligence coupled with data and analytics, from online supermarkets to pharmacies [1].
A spaghetti bowl of rules
A rule in a demand planning system may state if there's an overstock in store A, first use up this amount and then refill from your DC. Within this exclusively rule-based system, a rule could have a minimum volume that acts as a threshold. As soon as you are down to X amount, you need to order Z units. As a retailer, you might only want four items but the minimum order quantity from the supplier is six. You therefore order six items just to have them on time. Two too many.
Many software as service (SaaS) solutions try to help retailers navigate these inconveniences. In general, while most SaaS solutions offer a lot of experience with the product, they are usually built in .NET and use logic.
Retailers think SaaS will simplify their work. Instead, they are left with a spaghetti bowl of rules and logic. Unraveling these connections and getting a clear overview is often easiest with the support of Artificial Intelligence.
A robot that discovers the concept of doing the dishes
Artificial Intelligence is frequently shrouded in a bit of mystery, which is why pragmatic explanations are important. When asking people about AI, their first point of reference is usually a movie. There, artificial intelligence is often portrayed as an evil robot of sorts. The reality, however, is completely different.
Sticking to the robot example, imagine you have a friendly robot living in your house. You program it to serve dishes and the robot becomes smarter over time. However, after breakfast, lunch and dinner the robot runs out of plates. What now? Is it just a simple serving robot or does it discover the concept of washing plates? When the robot starts figuring things you didn’t program, that's the big difference!
Did you end up with a spaghetti bowl of rules? Rule-based systems are like an excel sheet with if-this-then-that relations. If A happens, then do B.
In contrast, an AI algorithm makes relationships: If A and B happened then this influences C but also affects the relationship of C on D.
AI offers complex prediction support
In a rule-based approach, a retailer sets the rules and reviews them once a month. This means things are just optimized for the time being. After a week, the retailer compares the prediction to the actual sales they made. As soon as predicted sales don’t match actual sales, the rule-based model must be revised again…at the end of the day, the system lacks adaptability.
A solely rule-based system is not learning over time. AI enables retailers to react to disruption proactively.
A product forecasting example
Gaining insights and becoming smarter over time is a huge advantage. Looking at product forecasting, we can use running shoes as an example. Say the new version of a certain model comes out. You have some experience with new editions; you know roughly how the market reacts. There is a previous logic to base your mapping on.
But what happens when your brand releases a completely new shoe model? This is unknown territory. You can use artificial intelligence to compare scenarios; you can predict what normally happens when introducing a new product, gauging if the volume looks about the same and what the price range is.
Taking it a step further, making predictions for a completely new brand is more challenging. The retailer needs to start gathering and analyzing data to see the current processes. Unfortunately, present-day processes and data aren’t good enough. Completely new relationships need to be predicted and a vast number of factors must be considered. That's why the supply chain professional needs artificial intelligence to start storing and recognizing things. The software sees both logic and patterns, while it's constantly optimizing itself. Ultimately, this means that the software has the basics covered and can direct the retailer’s attention to where it is needed most.
When is it worth it to invest in AI powered demand planning?
First, when conducting complex planning isn’t feasible manually anymore. Second, when you exceed a certain volume of what you work with and have in stock. Small businesses may therefore not have the need. Roughly speaking, the business case is the clearest for companies with an annual revenue of > 40 million euros.
Whether you make this switch to AI-driven demand planning also depends on your current demand planning quality. If forecasting is completely manual and laborsome, an uplift thanks to an advanced solution will be more noticeable. Depending on how much a product typically costs on average, you can estimate how many products are necessary for a combination of revenue, the number of stock-keeping units (SKUs) and the complexity.
A company must have the size, tech skills and number of products right when they implement a more sophisticated demand planning solution.
For example, IG&H works with a sports e-commerce retailer that meets all these qualifications. This pure player benefited from a data-driven way of working and achieved:
How long does it take for the algorithms to learn?
The well-known adage of ‘practice makes perfect’ applies to Machine Learning and AI as well. Though full perfection isn’t an option, the technology benefits from doing ‘exercises’. In short: How long the algorithm needs to learn depends on the volume. Especially machine learning models are a convenient forecasting option thanks to the sheer volume of data they can handle [2]. When talking about variables, the algorithm’s learning improvement is not specifically linear. The more volume you have, the easier it is to achieve a statistically significant outcome.