Making AI Really Work for Retail

Like every industry, retail is experiencing an artificial intelligence boom — and like other industries that operate with such small profit margins, any competitive edge is not only welcome but necessary to survive.

AI offers benefits in five key areas for retailers: personalization, visual search, virtual reality (VR) and augmented reality (AR), predictive analytics, and chatbots. The first three offer improvements in the form of improved customer experience which can increase sales and average cart value; the latter two are more about reducing costs, which can be clearly seen in your bottom line.

While some of these functions are more beneficial for specific retailers than others, all can provide the competitive advantages and improved margins you’re looking for while reinforcing your customer-first, future-thinking brand identity.

Personalization is perhaps the most important use of AI in retail right now. Effective search and product recommendation algorithms may seem simple, but if you’ve ever experienced not being able to find a specific item, such as the perfect peach A-line dress for an event, you know how complicated they really are. On the other hand, when your search turns up the dress of your dreams, you also experience how powerful they can be as you then add a new pair of shoes, earrings and even a handbag to your cart. Search and product recommendation algorithms are essential for increasing cart value by enhancing the customer experience, making it easy to find what the shopper is looking for and suggesting additional, related items they may find interesting.

Even the retailers with the largest physical presences in the world, like Walmart and Target, understand the importance of visual search in tandem with personalization. Most consumers now expect an excellent online experience — and it comes with benefits: early research has suggested personalization alone may improve AOV by 6 percent to 10 percent.

The complication comes in the data that the model is trained on. If a model doesn’t have high-quality labeled data, it can’t find our example dress in the first place, let alone pivot to recommended items. Moreover, a model needs an extensive database to match the experience of shopping in-store. In Sama’s work with Walmart, the team covered over 2.5 million items — upping retail item coverage from 91 percent to 98 percent — and Walmart has reaped the benefits in AOV increases.

VR/AR can provide what online stores once lacked: the ability to try before you buy, all without leaving your couch. Cart abandonment currently stands around 70 percent across online retail. Helping the shopper visualize themselves with the product, whether it’s a new tube of lipstick or a couch, may reduce that rate. However, retailers will only realize the benefits if models are trained to display items accurately. If the couch they’re looking for is displayed with completely wrong dimensions or scale, then that visitor may land in that 70 percent cart abandonment rate that retailers are trying to avoid. Alternatively, if the dimensions are wrong but the shopper does order it, then retailers could face issues like poor customer experience (losing future sales) and return costs (which get expensive quickly for large pieces of furniture, for example).

Chatbots can be a double-edged sword and are best used with caution. They can offer help 24/7, but they have a reputation for frustrating customers by providing prepared answers that don’t completely answer their questions. Advancements in generative AI have offered improvements, but chatbots can still hallucinate and create new policies that must be honored, like this example from Air Canada. To prevent this, retailers should find partners that follow a human-in-the-loop (HITL)-based approach to model evaluation. Giving a model direct feedback from humans is an extremely effective way of improving and retraining a model. In addition, red teaming can play a vital role in preventing hallucinations. This process involves experts deliberately trying to break a model’s safeguards to make it give incorrect or completely false information. Knowing about these vulnerabilities before deployment gives developers the opportunity to fix them before they become a real problem.

Predictive analytics will be one of the fastest-growing retail technologies in the coming years. COVID-19 proved that getting supply chains and inventories wrong can be very expensive, as Target found out in 2022. Retailers can use analytics to optimize inventory allocation across stores, warehouses and fulfillment centers as well as offer real-time re-allocation recommendations to reduce stock-outs, improve sell-through rates and avoid unnecessary markdowns.

These use cases are all well defined and documented, and we’re entering a period where these models need to be enhanced and refined to meet their full potential. Retailers now need deeper visibility and understanding of how their models are performing and why they may be failing. Obviously, identifying these issues in the initial development process is best for budgets. But when that doesn’t happen, it’s best for a retailer to find a partner that can identify problems at any stage of a model’s life cycle, including deployment and retraining.

AI is moving out of the experimental phase in retail, and the pressure to rush ahead and implement solutions quickly is real. However, with sales more dependent than ever on good customer experiences, it’s better to take your time, deploy solutions wisely, and find a partner that understands the need for a high-performing model and what that entails. If you blindly follow the trends or go with the lowest bidder and hope for the best, you’re likely to miss out on the best AI has to offer.

Lisa Avvocato is the vice president of global marketing at Sama, a provider of data annotation solutions that power the AI models of the future.