AI-based systems do not always make better decisions than humans. Find out why, and how to make sure your system is getting the best of both worlds.
Man vs. Machine
When it comes to Assortment Optimization, AI (artificial intelligence) and machine learning play a key part behind the scenes. While a human perspective helps, there are two important factors that make AI very important: scale and automation. For example, a large retailer with 1,000 physical stores and 10,000 products has as many as 10 million assortment decisions to determine which stores should carry which SKUs, what quantities, and when. Even though a significant portion of product offering is chainwide, that still leaves a formidable number of micro-decisions to be made. This amount of decision-making for humans is inefficient and not very profitable.
Consequently, relying on a team of time-constrained analysts would necessarily lead to adopting a one-size-fits-all approach such as the A-B-C store tiering logic. This results in sub-optimal assortment decisions and leaves money on the table. Conversely, an intelligent system can process millions of decisions much more quickly and free up precious analysts’ time, which can be invested in the small sets of decisions where an automated system may lack the knowledge or sufficient data to make accurate recommendations.
While we are on the man vs. machine topic, let’s also dispel the notion that an AI-based system always makes a better decision than a human being. Any assortment decision represents a “bet” (a resource allocation like inventory) under uncertainty (demand) whose goodness is revealed only after the demand manifests itself, as in a sale.
Humans make these bets and automated systems, AI-based or otherwise, also make these bets. In fact, typically thousands or even millions of these bets are made daily in retail chains. Say for example we are comparing two systems, X and Y. It may seem that AO system X will be better than system Y if X’s outcome on every single bet is better or more profitable than Y’s. However, this is quite unlikely to happen. What we will likely observe is that in the aggregate X produces more wins with a larger payoff than Y does. A well-designed AO system employing the right AI will be more profitable on most – not necessarily all – assortment decisions, when compared to a human-driven one.
On top of that, AI helps assortment optimization to improve over time. In fact, AI systems by definition become smarter over time as they ingest more and more diverse data. The ability to improve over time derives from the intrinsic ability of AI systems to re-calibrate themselves to achieve better predictive power without having to be re-programmed every time as they are fed more data. That is how AO systems increase efficiency and profitability without requiring added human intervention.
That is one of the fundamental defining features of AI. Examples abound: from the ability to recognize what product attributes are relevant to demand prediction, to refining the substitution estimation among products with some degree of demand transference, to the ability to recognize specific instances of demand as spurious and discard them instead of incorporating them in future predictions. All of these functionalities and more are refined over time as the system observes more instances and processes more historical data.
The Main Role of AI in AO
AI is used in AO to arrive at the optimal decisions, but not necessarily making the optimal decision itself. Specifically, AI generates the information that is then fed into the decision-making part of the process, the latter being mostly AI-free.
The best analogy known to everyone is autonomous driving. In self-driving cars, the machine uses a variety of AI techniques (mostly supervised classification) to interpret the signals collected by the hardware (radar, lidar, infrared, laser, etc.), map it to a dictionary of known objects and situations, and eventually understand the surrounding environment. It needs to recognize simple things such as whether a stoplight is green or red, as well as more subtle nuances, such as the current state of a multi-lane roundabout, or the most likely actions that an approaching vehicle is going to take.
Once it paints a picture of the world around and a prediction of its most likely future states, the AI layers hand it off to a decision layer that uses a set of pre-determined rules to achieve a given objective, such as going from point A to B within target safety and time boundaries. The decision layer employs little or no AI at all. In fact, we wouldn’t want a “learning” process to figure out foundational truths such as green=go and red=stop, or what the different curb colors represent (red, green, yellow, white, blue, etc.). It would be a very costly learning process indeed!
Similarly, AI and machine learning in AO does not ultimately make optimal decisions. Instead, it is used to read, interpret and predict the environment. For example, it can:
- forecast demand by attribute even for new products or a combination of stores and products not offered before
- understand the effect of demand transference among substitutable products
- predict the effect of new product introduction on demand of existing products
Once this wealth of information is generated, it is used by an optimization layer to maximize one or more financial objectives, such as profit or revenue. So, AI enables better decisions but does not make those decisions directly.
Is Machine Learning Always Necessary in Assortment Optimization Solutions?
Many AO solutions on the market do not use machine learning but appear to work well. However, even the steam engine worked pretty well and was widely accepted, until the advent of the internal combustion engine. Yet, of the hundreds of car companies that existed in the early 1900s, only a handful survive today, which proves that even a better technology still needs good execution to bring benefit to both the provider and the consumer of that technology. This is why Machine Learning does in fact matter for assortment optimization. It is the future of the industry and the solution.
What Questions Retailers Should Ask Before Choosing an AO Solution?
With so many providers pushing AI-powered solutions today, retailers should ask specific questions and look for some depth to answers beyond the generic, “AI figures things out,” or “AI recognizes patterns.” AI and machine learning are a collection of specific techniques each doing a different function (prediction, grouping of like items, classification, etc.). Retailers should ask which techniques are used for what purposes and beware of vague answers disguised as intellectual property protection. Think of it this way: A 10-year old laptop and a brand new laptop marketed today are both said to have a “very powerful processor.” That’s a vague answer that doesn’t really tell you how powerful each laptop is, or not.
Also, several AO solutions from different vendors may be equally and legitimately good, but one may work better than others for a specific retailer given that retailer’s demand characteristics, assortment process, and data quality and availability. Therefore, time permitting, we recommend running a blind validation bake-off, consisting of the following:
- feed two or more competing vendors 2 years of complete historical data starting 3 years back; these two years represent the training period
- share assortment data only for the most recent historical year: what stores carried what SKUs, but not demand data; this one year represents the validation period
- ask the competing vendor to produce demand estimates for the validation period
By comparing the vendors’ predictions to the actual observed demand values, we can assess which solution produces the most accurate estimates. While forecast accuracy does not necessarily translate into better decisions per se, it is a quantifiable and very important aspect of any AO systems: being able to accurately estimate demand levels for combinations of SKUs and stores which were not previously observed or offered.
This blind validation approach requires resources: time, IT personnel, analytics manpower to run the comparison, etc. It does pay off many times over given that an AO system is a huge investment, not only monetarily, to acquire it and run it, but also in terms of missed opportunities and lost profit due to not finding and using optimal assortment decisions.
Retailers can make smart choices when choosing an AO solution. It needs to be an informed decision, and there are best practices that help make the decision easier based on how it works with your existing inventory systems, your current human resources, and even your budget.
Stefano Alberti | Senior Vice President, Analytics
Stefano Alberti joined 4R Systems in 2002 and brings expertise in forecasting, statistical data analysis, algorithm development and price optimization. Prior to joining 4R Systems, he was a Lead Analyst with Manugistics, where he…
Read more about Stefano here.