I have been using the Sales and Operations Planning process, or S&OP, to manage supply chains in various companies and industries for many years, and I remember some of the challenges and issues I faced when managing the supply chain of an international cosmetics retailer in Mexico. Now, I advise clients who want to support these kinds of processes with artificial intelligence platforms, but I still hear the same concerns about three significant challenges in the area of forecasting. Along with Kamil Folkert, Chief Revenue Officer for Occubee, we will introduce those inconveniences in managing the supply chain on a daily basis and explain how they can be turned into opportunities by using AI-based tools.
First challenge: I used ABC product classification based on volume sales for a cosmetics retailer with more than 6,000 active SKUs. Products A were fairly easy to forecast using Excel and basic regression formulas, and we had a high accuracy rate. The problem was the B and C products — some would sell very slowly, and we had to keep them in stock. How can the AI and ML models improve the quality of these forecasts?
Suggested solution: ML models are set up to automatically analyze huge data sets and draw conclusions for each individual product category. Of course, manual product classifications based on an expert approach are very common, and not only in the cosmetics industry. The wider the range, the more difficult it is to distribute products evenly between shops. The problem with classifications is that they are based on a limited number of determinants and a short time horizon. Therefore, it is inherently difficult to advance a product to a higher grade after it has once been classified as B or C. A huge advantage of AI-based forecasting is that products are not dependent on these classifications, as all items can be forecast using different models, with individually assigned hyperparameters. The algorithms automatically analyze all SKUs, which is impossible for the employee in charge of assortment planning — this one focuses on accurate forecasts for the most profitable products. Of course, expert-based classifications can still make sense in another area of the business, such as reporting.
Second challenge: The French cosmetic brand has been around for more than 50 years; however, this retailer was fairly new to Mexico at that time — it has been three years since it started operating in the country, with three to five stores opening each year, new brands appearing every three to four months and the assortment becoming increasingly complex. How can it overcome the challenge of a lack of historical data for new products and brands?
Suggested solution: Every data-driven decision-making process requires a set of historical data to generate a valuable recommendation. There are some ways of factoring new products into the baseline. This can be based on manually defined reference product associations, analysis of the entire product category to detect demand trends or simply ensuring minimum shelf capacity. All of these strategies will require the replenishment of a different amount of product, but this will result in product availability on the shelf, which is mandatory to initiate data collection for a new product. Let’s also consider the next three stages of moving from manual to automated forecasting. With naïve methods, such as moving average, we can react to demand as soon as a new product is launched. The next step involves time series-based algorithms, a method that can capture trends based on sales data collected since the product launch. The final step comprises ML-based algorithms. The more sophisticated the sales characteristics, the more detailed the data we need. However, in the early days of a product, it makes sense to use naïve methods and human support.
Third challenge: We sold our own brand in our shops, which was our bestseller, helped position the company and was an asset we had to take special care of. However, due to the high cost of shipping, we had to stop flying freight and ship it by sea, so lead times became long. Can AI somehow help manage lead times?
Suggested solution: AI by itself will not reduce lead times, but it can help in two ways. Firstly, it provides accurate forecasts through historical data, which are key to ordering enough goods in advance. Secondly, it helps to automate the ordering process for most products, freeing up staff time to devote attention to the most problematic SKUs. The combination of both factors — ML-based processes and the input of experts who can supplement forecasts with additional domain knowledge — determines the success of AI implementation.
AI is being embraced in the supply chain industry to improve efficiency, accuracy and decision-making. AI algorithms analyze historical data and market trends to create more accurate demand forecasts, leading to improved inventory management and reduced waste. It can also support processes for introducing new products and dealing with long lead times. AI in the supply chain helps to operate more efficiently and optimize product availability, leading to customer satisfaction as well as business competitiveness and profitability.