The game-changer of large-scale retail: Artificial Intelligence

The large-scale retail industry has never been simple:  Plagued by heavy processes, challenging logistics that involve numerous third parties, personnel that remain focused on siloed tasks to ensure day-to-day operations, the industry’s well-established practices have become difficult to challenge. Innovation appears to be a daunting task despite a few advancements that have seen the light of day, both in the aisles and back in the warehouse (ex:  real-time price checking, real-time add-to-cart and checkout as you shop, digital price tags, or even contactless payment connected to loyalty profiles).

In the warehouse, the industry has seen its fair share of investments around robotics, with the aim to automate logistic processes. Indoor and Auto Positioning Systems are such innovations that have greatly eased pick-and-lift operations on the floor, by using robots to reduce time to load and unload. Such innovation offers a reduced need for human workers and minimises errors and security risks (workplace health and hazards), whilst offering the possibility to have full governance and monitoring throughout the process.

Beyond the help of physical machinery and robotics, a much finer innovation has disrupted (and continues to disrupt) traditional large-scale retail.

Let’s take a closer look at the impacts that Artificial Intelligence (AI) may have in the large-scale retail industry – both from the Point of Sales (POS) standpoint and the warehouse one.

Point of sales and retail store

Inventory operations for large-scale retail (think Walmart, Costco, Carrefour, or even any very large surface supermarket) are known to be incredibly complex and hampered by processes with numerous manual operations that require constant human supervision and oversight. Details become critical and could easily mean the difference between fully replenished shelves and long-standing backorders. Errors can be costly, be it as a result of entering wrong quantities in the system or by failing to remove a product lot that’s had a crucial recall.

Traditional retail software platforms have been equipped with features that alert when a product must be re-ordered – yet such alerts are based on fixed lower-bound values, and not on predictive analysis.

Enter AI and it’s now possible to ensure that the right item is ordered on time and in the right quantity. In contrast with traditional systems that alert for stock replenishment according to pre-set static thresholds, AI introduces self-trained algorithms that dynamically adjust the threshold based on multiple parameters. These parameters determine the right time and quantity required for order adjustments. In a nutshell, AI goes far beyond analysing mere sales trends – it factors in other elements, from estimated supplier inventory stocks and competitor pricing to geopolitical news that could have cascading impacts on the supply chain.

Accuracy

“Dynamically predictive” and the ability to easily factor in so many variables (many of which can be considered “unconventional” for traditional forecasting methods), is impressive… But just how accurate can AI get when it comes to inventory management? The answer is simple: AI solutions are as accurate as to how well they’ve been modelled, how much they’ve been trained, and the extent of their maintenance. Training datasets (real-life data) get fed into the AI model: Whether you’re feeding data to merely train the AI model, or you’re running the AI model for actual production decision-making in your business, all data fed into the AI model typically doubles as training data.

Transparency and Scalability

And while accuracy may sound like the end-all and be-all that there is to predictive analysis, there are other factors to consider when it comes to AI. Senselessly feeding structured and unstructured data into an AI model may get you close to accurate results, but the element of Transparency is essential for your AI model to be trustworthy. “Transparent AI would allow you to judge why (and how) your AI model is making a decision (or not making a decision) for your data.” (https://towardsdatascience.com/ai-beyond-accuracy-transparency-and-scalability-d44b9f70f7d8)

Beyond that, achieving that additional 1% of accuracy may come at a cost… The trade-off to reach further accuracy against the additional time required to train the AI model in addition to the computational overhead required to attain greater accuracy should not be overlooked: AI models need to be scalable.

So back to our Point of Sales and Retails Stores, assuming that we have an accurate, transparent, and scalable AI model in place – we can therefore consider the AI model to be Trustworthy. A trustworthy AI model would crunch the data and send an order proposition that would simply require review and approval, before hitting “Submit”. This is in stark contrast to the traditional way of running inventory management – which has often been carried out during closed hours in order to control stocks and the ERP, by matching the period’s orders against sales.

Warehouses

Some ten years ago, Amazon extended its heavy investments into warehouse automation, by acquiring Kiva Systems for $775 million and to create a fleet of more than 200.000 robots today – many of which are powered by AI. These 200.000 mobile robots work alongside hundreds of thousands of human workers in Amazon’s warehouses.

The 2020 MHI Annual Industry Report reports that only 12% of warehouses are currently using AI technology. Despite the low adoption, more than half of the respondents believe that AI has the power to disrupt the industry and to create a competitive advantage for their business. Beyond AI’s robotic applications, this could extend to machine learning, natural language processing, or even augmented reality.

Machine learning uses algorithms to “learn from experience” and make practical decisions. Using data gathered from sensors, from ERP systems, or even from external sources, it detects patterns and suggests actions such as anticipated replenishments of nearly out-of-stock items, shorter routes, and better inventory positioning.

Wearable technology paired with AI can also have important impacts in the warehouse.: smart wristbands and proximity sensors can help employees follow specific paths (for security reasons). Smart glasses could be used to scan items or to optimise picking operations. Speech recognition uses AI and natural language processing so that workers can operate hands-free and more safely. Also, cameras placed in different areas of the warehouse could automate visual tracking of workers, forklifts and inventory items to have a detailed overview of floor operations.

From the perspective of operations management in the warehouse, the possibility of having specific reports on how critical processes are carried out is gold. Such information can support decisions that are based on timing, efficiency, reliability of workers, or internal routes. Managers will use dynamic dashboards to have a quick and good overview of the data. Insights and customised panels will summarise the needed information and allow management to share different aspects of the data with stakeholders.

The world of large-scale retail has seen an explosion in its adoption of AI and machine learning through the infinite possibilities and applications that it has to offer.

Some practical applications

Technology and innovation should not be an obstacle; their purpose is to enable and facilitate access to opportunities. Here, we present some examples of AI and predictive applications that we helped to develop.

One such example is a customer churn forecast application which we helped to develop for a global leader in the food and beverages sector. The customer churn forecast application was developed for a specific product category. A “churner” is a customer who has stopped purchasing our product for a certain time. The objective of churn analysis is to identify potential churners before they effectively count as churn. The objective of the churn forecast application was: 1) To produce a churn score that would identify potential churners (based on purchasing patterns and individual characteristics), 2) To use the score to take actions such as targeted market campaigns to avoid the churn (with also in mind the customer lifetime value).

Our first step was to define a churner. For our project we classified Customers as: Active (last order placed in less than X months); Passive (last order placed between X and Y months); Inactive (no orders in the past Y months). The objective was to forecast the inactive customers. We followed the standard steps for this kind of project: Business understanding, Data understanding, Data preparation, Modelling, Evaluation, Deployment, Model management.

This is a typical classification problem, the usual approach would be to use methods like Decision Trees, Random Forest, Logistic Regression, Gradient boosted tree, and a specific development environment (R, Spark, Python, SPPS, SAS). For this project, we opted to use one of the out-of-the-box, user-friendly tools instead. The tool provided an automated analytics engine based on Structural Risk Minimisation. The tool helped in some of the standard initial data preparation as well (scaling, outliers). At the end of the Evaluation phase, the results were promising, and we were also able to identify some of the indicators of churn.

We proceeded on to the next phase: Industrialisation (or deployment) of the model. This phase is crucial and needed to be accurately planned and executed as the production environment is often different from the development environment, and the most sophisticated models can fail if they’re not correctly deployed and managed. For our project, we deployed the model on SAP’s latest technologies, which at the time, were not flexible enough to develop and test, but which were sufficiently robust to support the production model.

We equally supported the design and development of related applications for this client, including sales forecast, distributor sales forecast based on buy-in and buy-out and order change forecast.

From the onset and throughout the project, we worked in close harmony with the client’s stakeholders – keeping them informed and involved at every step of the way. As a result, the user community demonstrated much ease and confidence to use the models that got deployed.

Processes and people: technology is an opportunity, not a limit.

While the ethics of implementing AI remains a grey area for many, the mere idea of intelligent machines can also be quite intimidating as well. The reality is that we often focus too much on technology rather than on the people and processes on which these very innovations depend. As for any innovation, simple or complex, it is fundamental to have a clear vision, to set goals, and to have a clean design of the end-to-end processes that will make use of technology.

At Nembrini Consulting, our approach calls for meticulous attention to people and processes. Technology is secondary. Our experience allows us to understand the real needs of our clients and to propose optimal solutions that are built around people, processes and technology.