The dream of any retail merchandiser is to have a replacement item on the way to the shelf every time that product crosses the checkout counter. For Star Trek fans, it’s like the computer that can automatically process, formulate or fabricate anything the user wants, from a three-course meal to a replacement pair of the high-top leather boots Spock seems to prefer.

Sivakumar Lakshmanan

Unfortunately, we still have several decades before replicators and teletransportation is real. Also, and perhaps more important, consumers are a volatile group. Drug retailers are looking for artificial intelligence techniques to improve accuracy and reduce the wastage that plagues the industry. They are seeking truer demand forecasts to make it more likely that they will have the right product on the shelf for the customer, at the right time. Of course, if your company knew exactly what the demand was going to be, you would have little to no wastage. But in the real world, with its myriad of supply chain considerations and vagaries of consumer tastes and decisions, the perfect insight into demand has many shortcomings.

The reality is that no AI model is going to totally imitate an individual’s brain, let alone the collective human psyche. But it’s still critical for drug retailers to apply AI to forecast the true consumption demand for products so they can maximize revenue while reducing the costs associated with markdowns and returned goods. Even companies considered the best at forecasting demand can never be sure of the revenue opportunities they are missing from strong growth markets and high-growth categories.

Through the application of machine learning and AI, drug retailers can take advantage of growth opportunities and still cut their unsalable goods costs by as much as one-fourth. This translates into millions of dollars in savings and additional revenue annually.

Traditional demand forecasting methods rely primarily on historical data, but if that data influences the forecast too heavily, a retailer can miss out on strong growth opportunities. For example, a drug retailer sells 100 bottles of aspirin, 10 per week at this store, 25 at that store and so on. They also know how many bottles are likely to sell on a Saturday as opposed to a Wednesday. So, they stock the shelves based on traditional patterns. They know how demand varies by season or in response to promotions. This is the typical approach, and it generally works well.

The problem with relying on historical data, and your understanding of the various external factors, to project demand is that you end up operating in a paradigm that assumes that what happened in the past will continue to happen in the future. But even if that were the case, when you start thinking about a growth category or a growth market, your demand models would still fall apart.

With AI, drug retailers can do better at ensuring they have the right product, at the right time, in the right place. This improvement translates into more revenue, better shopper engagement and the ability to leverage growth opportunities. A key point to keep in mind is that your demand forecasting does not have to be perfect; it just needs to be better than your competitors, making it a strategic differentiator for your business.

Using the same aspirin example, if an AI model is provided with the trends regarding the larger consumption of over-the-counter painkiller drugs, promotional response of another drug that drives demand for aspirin, and local events such as a parade passing right in front of the store — the drug retailer can come up with a hyperlocalized strategy for aspirin for a specific set of stores for a specific period of time.

How do AI and ML help determine better forecasts? The approach is similar to the way a computer learns to play chess and soon becomes hard to beat by humans. It involves constantly exploring and exploiting the options that give the best returns. Not every move will be a positive one, but over time, the trend will be toward more moves in the direction of better returns. The learning process involves constant projections of how much more, or less, ibuprofen or bandages or vitamins should be placed into retail stores, and when, and taking lessons from the results. The ever-smarter system will tell you, essentially, “Here is where you have the best growth opportunity and least chances of return for this product.”

AI models can effectively model interactions between multiple external and internal factors to provide a forecast that is responsive to market consumption shifts. For example, improved forecast accuracy can be attained through reinforcement learning, an ML technique that helps determine which actions will lead to the greatest rewards. Combining the potential of reinforcement learning with human decision making delivers far more effective forecasting models, for both direct-to-store and centralized distribution business frameworks.

At the core of reinforcement learning is something that statisticians call the “multi-arm bandit” algorithm. It refers to the analysis of a group of multiple slot machines (one-armed bandits) in a casino, and the testing and learning involved in determining which of those machines will pay off better than others in the long run. Once determined, a gambler can then focus his or her time and money on those machines for a better return.

Multi-arm bandit algorithms are designed to identify the most profitable policies in the face of statistical uncertainties. In contrast to traditional A/B testing, which simply can’t scale to massively complex environments, multi-arm bandit algorithms will do so through a series of explorations of uncertainties and exploitations of those instances when there are positive outcomes. Learning more with each round of explorations and exploitations, the system will eventually make more advantageous forecasts.

Another challenge drug retailers need to solve for is slow-moving/low-turn items. AI can be equally effective in dealing with low-turn items like walking canes and some pharmacy products that are inherently hard to forecast. Firstly, AI-based segmentation techniques can classify forecastability of the product before we invest money into solving the problem through forecasting. Perhaps an alternative min-max approach to replenishment is a better choice.

If forecasting is deemed appropriate for a slow-mover, AI can help improve accuracy by dynamically identifying the right level at which to forecast. Using ML, AI can build a model over several products, locations and external attributes, then intelligently disaggregate them back to the store/SKU level, resulting in a more accurate forecast than standard replenishment methods.

Among the many reasons drug store retailers need to embrace AI-driven demand forecasting, here are a few critical ones to consider:

• Financial Gain — Generally, a few percentage point improvement in forecast accuracy translates into multiple millions of dollars in ROI. The goal is 100% on-shelf product availability, because a more productive inventory maximizes revenue and margin potential. And accuracy is self-replicating; as you ratchet up the precision of the forecast, you drive even more intelligence into the planning process across functions. This enables better decisions and helps companies transform away from siloed, stand-alone forecasting capabilities.

• Variability Predictability — Some sales patterns are more variable due to factors such as seasonality or new competitors, while others are fairly stable. In situations that are more variable, it is essential to distill the signal and separate what is important from the noise. That is done by breaking the total demand signal down into actionable parts, which enables companies to tweak those individual parts. By using better prediction of the true impact of different levers and scenarios, a business can bolster ­performance.

• Elasticity — Understanding price elasticity item by item is paramount to demand forecasting. By producing a set of demand drivers — such as price reductions, features, displays, product launches, competitor information, etc. — a company can determine which strategy will deliver the best result.

• Productivity — A forecasting platform provides the ability to address a business problem in a repeatable way. This adds immediate value because you don’t have to reinvent the wheel each time you forecast. Low- or no-touch planning is a reality with AI infusion.

• Single View — Distilling all the elements into a single, unified demand signal has tremendous value across the enterprise. It drives intelligence to each function of the organization, making it clearer what each area needs to do in response to the signal as well as keeping each function informed on the impact other areas are making to the signal. No forecast or signal may ever be perfect, but by distilling the signal, it eliminates all the problems caused by multiple forecasts and multiple signals.

In the end, it is the difference between simply looking back and basing decisions on hindsight, or looking forward and leveraging a vastly more sophisticated capability.

Sivakumar (Siva) Lakshmanan is executive vice president of forecast and supply chain analytics at antuit.ai. He previously was a leader for supply chain analytics at IBM Global Consulting Group. He can be reached at [email protected].