AI in Retail: Beyond The Hype and Into Reality for Store Operations.

Understanding AI.

The term “AI” is often thrown around, but what does it really mean? Here’s a quick rundown of the 3 broad layers of AI that people are talking about:

 AI is essentially technology that mimics human intelligence. In a broad sense, all computing can be considered AI as it calculates and processes information, much like the human mind. In fact the word “computer” comes originally from a professional role whereby a person used to compute numbers.

Machine Learning is a subset of AI where a computer optimizes itself to arrive at the correct answer, effectively learning from its experiences rather than relying on set rules.

Deep Learning, a subset of machine learning, uses specific neural network architectures that mimic the human brain. This is the tech used by language models like ChatGPT.

Understanding these definitions and differentiating between them can help us evaluate the true potential and limitations of AI.

The practical application of AI.

Imagine wanting to identify whether there’s a problem in a specific category of store sales that you need to take action or improve. Using standard computing, we could program a relationship between current sales and last year’s sales, flagging it as actionable when the relationship crosses a certain threshold.

Basing this on just last year’s sales and current sales looks pretty simple (and most store managers will be familiar with this LY comp analysis), but where the broadest definition of AI computation starts to get more accurate than the human mind is when we put in multiple data streams (“features” in AI speak). Instead of two inputs, we could also optimize including; last week sales, sales of a relevant other category, inventory patterns, weather conditions etc.

In the context of Machine Learning, this problem can be approached differently. Machine learning algorithms can use historical labeled data as either actioned or not, the machine itself will optimize for the relationship between the different inputs, providing a more accurate relationship by adjusting the bias and weighting towards the different inputs (shown by the line identified on Chart 2). As more and more items are actioned, the model becomes more and more accurate. This is where AI simplifies complexity and enables more efficient decision-making.

Deep learning, with its neural networks, further enhances this capacity by dealing with an extensive range of features and weightings. For example, ChatGPT uses 175 billion weights to determine the most suitable next word based on the input string of words. This is great for very very complex problems, but can be very expensive to train and compute.

The System Behind AI.

AI needs to be part of a larger system to function optimally. Our system, Quorso, serves as an example. We designed Quorso as an Expert System to function alongside retail managers, pinpointing valuable opportunities for action and learning from the results to continually improve productivity.

Here are six essential components of an effective AI system:

1. Objective: What are you trying to optimize for? In our case this is constantly providing the most actionable information that will drive the most productive outcomes.
2. Data: The quality and relevance of the data used in training the AI model are crucial. It can only be optimized based on what it’s been provided. For us we take all the relevant sales, inventory, product, labor etc. data that a manager may need to use to optimize their store.
3. Algorithms – There are two core types of algorithms used:
   
   Classification algorithms: which classify and group outputs together. In our case actionable or not actionable, but equally could be used to determine dog or cat in an image model.
   Regression algorithms: which are used to predict future outcomes based on past data. We use this to understand how an action is likely to perform and therefore what value it is likely to bring.
4. Learning Models – There are three types of learning that AI models have:
   Supervised Learning: Supervised learning is a method where AI algorithms learn from labeled training data and this data guides the model to predict outcomes.
   Unsupervised Learning:Unsupervised learning identifies patterns in datasets with no pre-existing labels and with minimal human supervision.
   Reinforcement Learning: Learns by performing certain actions and observing the results or rewards from those actions by a human agent.

Here at Quorso we use supervised and reinforcement learning to train our model. The requirements in retail to have some control over changing strategy mean from an operational perspective unsupervised learning is more of a black box.

1. Computation: The sheer computational power possible today for AI is massive. Advances in computation have been a significant factor (arguably the most important one). The types of calculations we create today in Quorso were just not possible in a cost or time efficient manner even 5 years ago.
2. User Experience: Interactions with the AI tool are critical for two reasons: they ensure that the tool is beneficial and gives value to people, and they can ensure they can give feedback for reinforcement learning.
   
   

   AI Use Cases in Retail Operations.

Finally, we arrive at the practical applications of AI in retail operations. Here are four notable use cases:

1. Generative AI: The area everyone is talking about. Generative AI uses deep learning to build models that either generate text, generate images or generate other forms of content. They are focused on generating content based on learnings from all content out there on the web. Their use case is fantastic for broad applications of content but less so for specific and accurate use of content/data (as can be seen from hallucinations and inaccuracies).

2. Expert Systems: Like us here at Quorso are designed to sit alongside a professional and scale their intelligence and knowledge to achieve a certain outcome. A non-retail example would be how radiologists use AI to better predict cancer from scans vs the naked eye. Our focus is on helping Field leaders make the most productive decisions in their stores.

3. Predictive Models: look to make more accurate assessments of future economic or behavioral patterns vs human instinct. Therefore they are deeply embedded into demand forecasting, customer behavior trends, labor forecasting.

4. Computer Vision and Robotics: understand their external environment, label it and make adjustments associated with it. We are seeing this with anything from inventory cameras to warehouse robotics.

Hopefully this is a good primer for how AI can and is being used in a retail operations environment but if you would like to learn more, please send us a message at contact@quorso.com.