Learning by Doing: What Netflix teaches us about AI inventory

Some years ago, a small company called Netflix had a very good idea. Back when they started their business in 1997, it was a DVD rental company. To become more user-friendly, Netflix began renting DVDs through a mail service in April 1998, which was a tricky task as the business model was completely new. Beyond that of course this wasn’t the only problem Netflix faced, because at that time only 2% of Americans owned a DVD player. You may remember that the home entertainment market still belonged to the VHS.

To put it simply, Netflix’s good and far-sighted idea was to make it easier for customers to access their desired movies by delivering them to their home. One year later, Netflix decided to switch to a subscription system: Customers paid a monthly fee that allowed them to order a variety of DVDs online. This was the next step towards their now very successful on-demand service. Later they took their idea of being available any time to the extreme by turning themselves into a streaming service and the rest is history. In the area of inventory management, however, we can all still learn something from Netflix.

What you really need to do

They threw out all their DVDs and since then Netflix never ever had to deal with inventory management or transport logistics again! Problem solved; blog post closed.

Just look how easy supply chain management can be!
…

All right, that’s not what I meant, of course.

You are surely aware of many other industries where the expectation of getting a product or service on demand, any time or at least as fast as possible, has become the new norm. For a long time, Netflix was perfectly able to provide exactly this service to their customers. Amazon adding same-day-delivery to e-commerce might be another famous example. People, even in B2B markets, have become accustomed to this kind of strict expectations in terms of delivery speed and on-time delivery. In most cases, you must not get your delivery wrong!

Generally speaking, a company that relies on on-demand services can benefit from high customer satisfaction and relatively low stocks, as production is relatively timely. At best, production is supplied with the right material just in time. On the other hand, on-demand business requires very complex and flexible processes. The requirements for sufficient resources are considerably higher than for conventional production processes – which usually leads to higher production costs. If you can’t become a video-on-demand provider yourself, you either must stock up or – as Netflix did – turn to modern technology to help you out.

That’s why an intelligent inventory management solution will remain an essential part of your supply chain management. You’ll always want to accurately forecast what your customers are going to demand and timely order all the goods you will need to meet this demand. Otherwise bad things will happen to your business.

Let’s take a look at Netflix again:

Learning with algorithms

Today, artificial intelligence is divided into knowledge-driven and data-driven technologies. The former, which includes, for example, tools for optimized planning and decision-making, already have a long history of success. The latter, on the other hand, are still comparatively young and are classically referred to as machine learning, in which algorithms can learn from vast amounts of data by observing them. Although there are many good reasons to deploy both types of AI in an hybrid approach, here we will focus on the data-driven aspects of AI.

One of the best-kept secrets to Netflix’s success has always been their self-learning algorithms which recommend movies and series to the subscribers that they will most likely enjoy. This is how it works: Every second of video content people watch on Netflix has meticulously been tagged to tell the machine many different aspects about the content. Which actor is currently appearing? Is the mood creepy or funny? Is the movie set in space, in a city or in the woods? This data is combined with user behaviour data like what has the subscriber been looking before the current video. What did they watch a year ago? How much time do they spend with what and what feedback do they give?

With machine learning they find out which of these factors must be weighted in order to accurately forecast what a subscriber might want to watch next. Building on this, subscribers can be categorized into several categories of more than 2,000 taste groups, on which the recommendations are based. On top of that, the forecast is not only being used for recommendations but also for pro-actively creating content that fits personal preferences.
Okay, I ask once again: What can we learn from Netflix here?

AI in inventory management

The idea of providing the right articles at the right time is essential to inventory management. While optimization tools are probably most important, machine learning can help you to create a better basis for decision-making. For example, which material or part must be purchased and when it needs to be ordered to make sure it doesn’t arrive too early or too late for production or assembly.

If you want to order an item, you will usually investigate your ERP system to see what the replenishment lead time is. It’s a fixed value and belongs to the classical master data. For example, it says an item will take ten days to arrive. Logically, you or your planning tool will then take ten days into account when ordering the item. In reality, because of everyday operational disturbances and the fact that life is crazy, the order will not always arrive in ten days, but sometimes take eight, nine, eleven or twelve days. This planning blur had to be dealt with until now.

Like Netflix, we can use machine learning to learn from the past and improve our forecasts. By examining historical sales data, as well as many other forms of data, from a bird’s eye view for recurring patterns and correlations, machine learning algorithms can uncover and evaluate causal factors for specific developments. Maybe they find that the ordered item usually would take ten days to arrive if it was ordered on Monday, but twelve days if ordered on Friday. This of course, would be very easy to identify, but in practice they use innumerable data sets to find much more complex patterns. Even if a pattern is easy to understand, an inventory manager usually has to overlook tens of thousands of different articles, each with its own upstream supply chain, contractual provisions and conditions as well as other framework conditions. They couldn’t even handle all the easy patterns. With machine learning, the previously fixed master data would now be flexible and thus a more realistic basis for decision-making.

Just as Netflix uses historical data to forecast future user behaviour, inventory managers can also use this self-learning technology to more accurately assess how future ordering processes will play out. They thus make an important contribution to making their companies more agile and economical.

Are you already using machine learning to improve your supply chain management?