Enabling Smarter Decisions in Retail with Artificial Intelligence

IN Artificial Intelligence — 04 October, 2017

A previous blog post has explored the importance of high-quality data and how to build an AI strategy. But is all this necessary? Why should companies invest in becoming a digital enterprise if they're already successful?

While businesses may have thrived in the past without digitalization, it is also true that businesses have always strived for optimization. The introduction of power-looms in the 18th century is widely regarded as the first industrial revolution, introducing a disruptive technology which changed how a whole industry sector worked. In the past decades, most industrial production processes have been supported or automated by robots and companies have used data in analytics projects and business intelligence to get detailed insights into the performance of the enterprise.

blog_smarter_decisions_with_AI.png

AI – turning companies into a predictive enterprise

But the real game changer is the introduction of AI-based systems. Until now, people have been at the center of production and decision-making, even if they were heavily assisted by machines. Industrial robots may be able to build a car, but they follow strict programming and are not capable of redesigning the steering wheel. Business Intelligence doesn’t make decisions by itself, but rather focuses on human decision-making by pulling all information together and presenting it in the way that most suits the task at hand. AI systems are based on data – lots of it and high quality – and are able to make their own decisions, such as the best price at any given time for a specific product sold online or at a particular store or how many items to order during the next replenishment window. Facebook recently reported that the bots that were set to negotiate with each other also started to develop their own language without human input as a more effective way to achieve their goal. Suddenly, humans were not only no longer directly involved and they didn’t understand the language being used in the negotiations.

What kind of task is ideal for AI systems?

Aren’t human experts better at making decisions then AI systems? The answer to this questions is – as always – it depends. Strategic decisions and occasional decisions are best done by human experts at the moment – and that will likely remain this way for the foreseeable future. For example:

  • What is the key vision of our enterprise? What are the core values?
  • Should the enterprise expand into a new vertical or region?
  • Where should a new distribution center be built?

Unless your business makes these decisions all the time, there is little historic data to go by. In many cases, further constraints have to be taken into account which are based on assumptions with large uncertainties. Artificial Intelligence is still a long way from being able to make these kind of decisions. 

But when it comes to the various operational decisions that companies must make every single day, AI is a capable partner: 

  • How many items of a particular product should be ordered for a specific store on a particular day?
  • What price should be set for each product within the next hour or day?
  • What incentive should I give a particular customer?
  • Is this particular customer browsing on the online store or are they looking for something in particular and would engagement help?

In many companies, these decisions are still made by people – everyday and all day long. However, many businesses, particularly retailers, would rather have the human experts focus on the high level decisions rather than the mundane day-to-day work.

A leading fashion retailer said: 

"Our fashion experts are very good at designing a customer experience for each individual, store-in-store, adapting to the size and anticipated audience at each location. However, we also have to deal with never-out-of-stock articles such as socks, undergarment, etc., which takes between 30 percent to 50 percent of our time. Can’t this be automated away?"

Or in the case of an electronics online shop:

"We want our human experts to focus on the experience of buying the latest flagship smartphones. However, they also have to make sure that all the cables, cases, screen protectors, chargers and hundreds of other things are ready to be shipped from our warehouse – can’t this just work in the background?"

The sheer number of these kinds of decisions that need to be made in a very short amount of time make them ideal for automation through a sophisticated AI system. Human experts may be very good at deliberating difficult decisions, but find it difficult to make high-quality operational decisions all day long. People simply cannot focus for many hours and treat each decision as if it was the only one we had to do during the day – and then make a decision in a few seconds and move on to the next one. The complexity of the situation also makes it very challenging to come to a good decision – people already struggle if a decision is influenced by two or more factors which may also be correlated. Real world decisions, however, are driven by tens if not hundreds of correlated and competing factors – far too many to be understood by even the best experts, even if they had the time to think about each operational decision as long as they would like to. Furthermore, people don’t have an intuitive understanding of statistics and find it difficult to incorporate large variances into deliberations. Even when considering fluctuating patterns – such as the demand for a product in a supermarket – people tend to think that the demand narrowly varies around the mean demand, maybe with large spikes caused by promotions. However, if a given product has a large variance, we don’t find it “natural” that the actually observed values can be very low on one day, just to be very high on the next.

How do cognitive biases hurt retailers?

From the perspective of behavioral science, it is well known that people are subject to cognitive biases, which are very hard to overcome and determine our thinking and decision-making on a fundamental level. For example, we frequently overestimate how well we make decisions (overconfidence effect) and tend to subconsciously replace questions and decisions we encounter with easier to answer questions (substitution). For example, the question “Are there enough items on the shelf available for sale? might be substituted by “Are there any gaps on the shelf?” Both questions seem the same, but the first is aimed at anticipating future demand while the latter is a yes/no question about what is already available. Since these effects happen subconsciously, human experts are not even aware that they didn’t address the original question or decision to start with. Business Insider has a list of popular cognitive biases, an even longer one can be found on Wikipedia.

An academic study on cognitive biases in retail performed by G. Bolton and E. Katok focuses on the replenishment of “ultra-fresh” food, i.e. food that perishes at the end of the business day and all excess inventory has to be disposed of. Hence, if too little is ordered before the store opens in the morning, sales will be lost and the store will experience stock-out. If too much is ordered, excessive waste costs have to be taken into account. This setting is known academically as the “Newsvendor problem” (from the time when the daily newspaper was sold by newsboys around the streets). Crucially, the optimal solution can be derived analytically. Hence, at least in principle, all participants in a study should know which amount should be ordered to maximize the (expected) profit of the simulated store.

However, although the optimal solution was known, none of the participants placed the optimal order. The study analysed whether extended experience (by providing time to experiment with the order quantity before actions were recorded) or performance reviews (by giving feedback about the payoffs of options taken and not taken) would improve the ordering decision of the participants. However, it turned out that neither improves the performance of the participants. The third part of the study, however, did find a significant performance improvement: Instead of being allowed to order at each period (e.g. a single business day in the case of ultra-fresh food), the participants were forced to place standing orders for 10 periods (e.g. days) in advance. This forced the participants to carefully choose the next standing order as it cannot be amended for the next 10 periods. Suddenly, participants were no longer able to react to fluctuations they observed in the previous period or the one before. This is also known as “demand chasing” where experts – even with the best intentions in mind – try to compensate for the values recently observed and fall from one extreme to the next due to the volatility of the underlying statistical demand distribution.

But placing standing orders is just a crude way of automating the replenishment – and using a sophisticated AI system to do so results in much higher gains. As the above and other studies recruited mainly MBA students as participants, one might wonder how experienced procurement managers perform when compared to students. Bolton et al. address this in their study “Managers and Students as Newsvendors” that compares the performance of freshmen (no formal training in operations management), graduate students (at least one course in operations management) and managers (at least one year of relevant experience in the field). The authors of the study found that students and procurement managers alike show the same biases and experienced managers are no better at utilizing relevant information or experience they gained during the study.

Introducing a sophisticated AI-based procurement would alleviate the negative consequences that arise from our human nature. 

The impact of introducing AI systems on the workforce

Automating decisions does not reduce thte workforce at all – but it enables them to work smarter and make better decisions.  However, that does not mean that people will make the same decisions as they did in the past. As AI-based systems become more powerful and sophisticated, business owners and the senior leadership team need to consider carefully which decisions in a digital enterprise are best made by people – and which are best left of an automated AI system. It is important to keep in mind that introducing a sophisticated AI system in a company does not only imply significant changes the way decisions are made – the work of many individuals will change as well and the senior leadership team has a special responsibility to guide the company and its employees through this change and empower them with additional training and understanding how they work.

Dr. Ulrich Kerzel Dr. Ulrich Kerzel

earned his PhD under Professor Dr Feindt at the US Fermi National Laboratory and at that time made a considerable contribution to core technology of NeuroBayes. He continued this work as a Research Fellow at CERN before he came to Blue Yonder as a Principal Data Scientist.