AI in Retail

It is no secret that physical retail is under pressure. Consumers are rapidly moving from physical shops to online stores, leaving physical shops to compete with online businesses that can offer products at any time, anywhere, at lower margins. Physical retailers need to optimise their processes and offer customers a more appealing experience than they can get online. At DSL, we believe that data can play a major role in the success of physical retailers. In this article, we look at ways to collect data in physical shops and present several AI use cases that use this data to create value across different departments of a business.

Customer data

It sounds counterintuitive. Hundreds of customers visit a physical shop every day; staff see who is browsing, hear what people are interested in and interact directly with customers when they make a purchase.

Online store staff rarely interact directly with customers. Yet it is much easier for an online store to collect data about them. Customers in an online environment can be tracked from the moment they click on an advert until after they leave the online store. Every click on a product, category page or shopping basket, for example, can be recorded. This gives online stores numerous data science applications for building a detailed picture of their customers and approaching them with relevant offers to provide a personalised experience.

The challenge in physical retail is that staff often have a good understanding of the customers at their branch, but this knowledge is qualitative and cannot be converted into usable data. As a result, it cannot be shared systematically with other branches, even though decisions about promotions, product ranges and operations are made across the organisation. Point-of-sale systems also collect a great deal of transactional data, but the AI possibilities are limited because purchases cannot be linked to an individual. Fortunately, there are ways to collect quantitative customer data at scale and link it to purchases in physical shops too.

Most retail businesses have already started using loyalty programmes. By encouraging customers to scan their loyalty card or app when they make a purchase, you can link each purchase to a customer. This shows you how often customers return and what kinds of products the same person buys. It also makes AI applications such as recommender systems and churn prediction possible. If you encourage customers to share some demographic information, you can link their behaviour to their characteristics. You can then segment customers and reach out to them in a targeted, personalised way. Add an app to the loyalty programme, and you can track every interaction a customer has with it.

There are also ways to collect data on customers who do not make a purchase. These range from simple entrance sensors that count customers as they enter to more advanced methods, such as cameras that record the routes customers take or recognise a customer’s age and gender.

Although we distinguish physical retail from online retail in this article, most physical retailers also have an online shop. Data collected in the online shop can be used to optimise the physical store.

So physical retailers can also collect large amounts of data. In the rest of this article, we discuss some examples of how we would use it.

Product recommendations

To maximise revenue, retailers commonly use cross-selling to sell more to customers. It also improves the customer experience when products that are often bought together are not too far apart. AI can give you insight into which products are often bought together (joint purchases) and which additional products customers are interested in after buying a particular product.

Even when you cannot link a customer’s separate purchases, you can still gain valuable insights from sales data. Market Basket Analysis identifies relationships between different products across large volumes of past transactions. Statistical models reveal which products are often bought together and calculate the likelihood that one product will be purchased alongside another.

These insights can help you decide which products to place together to encourage cross-selling. Point-of-sale systems can also be set up to alert a cashier when a customer buys just one product from a frequently purchased combination, showing which other product might interest them. This allows the cashier to make a targeted recommendation during the sale. Such models can also help determine which products to put on offer: promoting the right products may boost sales of other products linked to them by the market basket model.

A loyalty programme links a customer’s separate purchases. A customer profile can also be linked to these purchases. A recommender system can then provide better product recommendations. While Market Basket Analysis looks only at products bought within a single transaction, a recommender system considers products the customer has bought before or that similar customers have bought.

There are two categories of recommender systems: collaborative filtering & content based. Collaborative filtering uses a customer’s purchases to find similar customers. If the customer has not yet bought a product that those similar customers purchased, that product is recommended. This method therefore does not require a linked customer profile.

Content-based methods use additional information about the customer (such as age and gender) or the product (such as price and category). A model is then built to identify relationships between customers and products (for example: women aged 25 to 45 often buy higher-priced items in the home category). A person with certain characteristics can then be recommended a particular product. This method requires customers to provide additional information about themselves. Unlike collaborative filtering, however, it can give customers a relevant product recommendation even if they have not made a purchase before.

A product recommendation can be made at the checkout when a customer’s loyalty card is scanned. If an email address is available, customers can also be reached when they are not in the store. A relevant product recommendation may then encourage them to visit.

Analysis with computer vision

As described earlier, collecting data about store visitors is a challenge. A target audience is often defined across the organisation based on ‘gut feeling’ or through inefficient methods such as surveys. Staff have a general sense of customers’ routes through the store and their behaviour, but these are not measured systematically. We have already mentioned that this data can be collected using cameras. With computer vision, camera footage can be converted into quantitative data.

Cameras with facial recognition make it possible to estimate the gender and age of each store visitor. They can also record whether someone enters alone or with others. This allows you to build a clear picture of the target audience for each branch. You can also measure when different customer segments shop and how long they stay. Does a branch in a village attract a different kind of visitor in the morning from a branch in the city? Are lunchtime visitors mainly people aged 25 to 45 who stay briefly and buy little? Do families make up most visitors on Saturdays? With insights into who visits each branch at different times, you can design promotions for specific branches and customer segments.

As well as estimating age and gender, cameras can track the routes shoppers take through a store. What route does a customer take to reach the product they buy? How do shoppers who end up buying nothing move through the store? Is the whole store being used, or do shoppers rarely venture beyond the first few aisles? Insights into these routes can help you optimise store layouts. Because the data comes in continuously, you can measure the effect of a change immediately. You can also experiment with new store designs through A/B testing, for example by giving store A a new layout while leaving store B unchanged.

Computer vision can also record what shoppers do. How do customers interact with products in the store? How often is a product picked up and put back, and how does that compare with its sales? Are products at a particular shelf height viewed and picked up more often? This information helps you make data-driven decisions about where to place products on shelves. Here too, you can experiment with different shelf layouts to make an informed decision.

Of course, you must comply with privacy legislation when using computer vision with cameras. Personal information must not be stored without consent, and the last thing you want as a retailer is for people to avoid your store because of privacy concerns. By storing only the estimated age and gender data, rather than the recorded face, you prevent the data from being traced back to an individual and avoid storing privacy-sensitive data.

Demand forecasting with machine learning

Accurate demand forecasting is important for every retail business. An overestimate ties up more capital in stock and leads to higher storage costs than necessary. An underestimate means turning customers away because products are out of stock. A good demand forecast also helps you set an effective budget. Making an accurate forecast is not easy, however. Many changing factors — from shifts in the weather, competition and availability (online/offline) to store type and new product reviews — affect whether customers visit your store and make a purchase.

Businesses have long forecast demand using traditional statistical methods, such as time series analysis. These methods often work well for businesses operating in a stable market and selling products with a long lifecycle. One drawback is that they rely solely on historical sales data and assume that what happened in the first year will happen again in the second. They are less suited to dynamic markets and cannot forecast demand for products that have not been sold before.

Predictive models using machine learning are based on statistical models, but can draw on additional internal and external data sources (such as weather data, economic indicators and social media reviews). They use complex algorithms that can identify more than just linear relationships. Instead of producing a one-off forecast at the start of the year, machine learning models can be retrained automatically and adapt to changing circumstances.

Conclusion

So it is possible to collect large amounts of customer data in physical retail too. In this article, we have presented only some of the ways you can use AI with this data. We are convinced that applying AI in physical retail can create significant value too! Get in touch to discuss the possibilities for your organisation!