At Walmart, artificial intelligence (AI) and machine learning are everywhere.
You won’t see it as you walk down the aisles of a Walmart store. You won’t feel it when you pick up a Walmart package from your stoop. And you won’t notice it when you search Walmart’s website for everything from paper towels to toys.
But today, AI and ML is embedded throughout the Walmart organization – from supply chain management and shopping to search.
As the world’s largest retailer, it’s no surprise that the Bentonville, Arkansas-based retail leader has invested in cutting-edge AI for years: In 2017, for example, VentureBeat highlighted Walmart’s massive rise in inventory thanks to AI, while we have covered Walmart’s AI efforts on everything from express delivery to grocery delivery robots over the past half-decade.
And over the past six years, Walmart has gone from a handful of in-house data scientists to hundreds, according to Srini Venkatesan EVP, U.S. Omni Tech at Walmart. These data scientists serve on teams related to supply chain forecasting, optimization and labor/demand planning; search and personalization; as well as emerging technologies. “We are really spending a lot on internal development because we feel this is our competitive secret sauce,” he told VentureBeat.
Venkatesan, who runs all technology teams that enable Walmart’s global marketplace, omni supply chain and stores, said that Walmart is “evolving from being an automator of retail to becoming an enabler of retail – that is where AI and ML are very relevant for us.” What he means by “automating” versus “enabling,” he added, is that the company has moved from simply using technology to make Walmart’s tools and processes more efficient to taking a step back. “We look at the overall end-to-end picture to enable improvements throughout the entire shopping journey and across the organization,” he said.
Which, of course, leads directly to Walmart’s biggest goal: figuring out what the customer wants and providing it. “Walmart has always been about what the customer wants,” he said. “The customer is number one.”
AI suffuses Walmart’s entire supply chain
To give customers what they want in an era of global supply chain woes, Walmart has emphasized AI on the supply chain front: Last week, the company announced it will open four next-generation fulfillment centers (FCs) over the next three years, with the first debuting this summer in Joliet, Illinois.
These FCs will be the first of their kind for Walmart, using robotics and machine learning to speed up fulfillment. The company claims that combined with its traditional fulfillment centers, Walmart will now be able to reach 95% of the U.S. population with next- or two-day shipping.
In addition, this past Monday Walmart announced it will bring Symbotic’s next-generation robotics and AI technology to all 42 of its regional distribution centers over the next eight years – it’s already used in 20 – as the retailer works to modernize its supply chain network. The technology should help Walmart increase its inventory accuracy and boost its warehouses’ capacity to receive and ship products to stores, the company said in a statement.
Symbotic, which went public this week and enjoys a significant investment from Walmart, said its AI-powered software and robotics system – including its Symbots, which are fully autonomous vehicles that leverage machine learning, vision and algorithms – addresses some of the biggest challenges of Walmart’s complex supply chain.
“When you look at things like the accuracy and reduced errors and reduced scrap, there’s just incredible savings from a working capital perspective, inventory management perspective and the overall labor piece,” said Symbotic CEO Michael Loparco. “So I think there are powerful cost drivers – but I think the biggest catalyst for Walmart is changing consumer demand and the need for market pull through.”
Walmart’s evolved supply chain
Walmart’s supply chain efforts using AI have evolved over the past few years, said Venkatesan, moving from simply predicting sales demand – how much will sell that is already in the stores – to predicting consumer demand in terms of what the customer will actually want to buy, by analyzing data across channels, from Google searches to Tik Tok social feeds.
During the pandemic, however, what was a tricky demand problem to solve also became a thorny supply problem.
“We learned that we needed to understand what was not going to be in stock and what we should substitute it with,” he said. “So we invested a lot of effort into AI and ML for substitution logic.” Deep learning AI considers hundreds of variables — size, type, brand, price, aggregate shopper data, individual preference and current inventory, among others — in real-time to determine the best next-available item.
AI powers Walmart’s search and personalization
Historically, much of the activity around Walmart’s search and personalization was around automated decision-making, said Jan Pedersen, VP of search and personalization, U.S. Omni Tech, at Walmart. But more recently, computer vision AI model performance has become much better than it used to be because of deep learning, he explained. “You can use these things in production and get results,” he said.
As a result, there are several areas where Walmart uses AI technologies and natural language processing in search and personalization, he explained. There are English language queries – understanding what people mean when they request a product type, understanding what parts of the query are important.
Understanding the quality of the image is also key, he added. “Maybe even doing attribute extraction, so knowing that it’s a red shirt because it’s red in the picture is important.” Finally, there is machine translation. “We don’t have to do manual translation of anything, so that’s a big boost,” he said.
Search is an expanding frontier
Some queries, however, are much easier than others, he pointed out. “You may have a query that’s repeated many times and people give you a very strong signal about what it means, or you might have a query where if you look at it the intent is very obvious what the user is interested in, but if you attack it from a standard approach, you won’t really get good results.”
An example of this just recently, he explained, was ‘avocados from Mexico.’ “The reason that’s interesting is that most avocados don’t tell you that they’re from Mexico.” On the other hand, he explained, the query itself is very obvious — it’s clear what the user wants. “So we put that in the bucket of semantic queries where you have to really be on top of that, understand that the avocado part is important or infer in general from other things that you know about an item that is likely to be important.”
Finally, Pederson discussed Walmart’s efforts related to multilingual queries, which enables Spanish-speaking customers to find specific items on the Walmart.com site and in the app.
“One of the interesting things about search experiences in general is that people can type in whatever they like, because it’s an empty box,” he said. To service Spanish-searching customers, Walmart uses language detection using AI. “You detect that this query is likely to be in Spanish and then you do machine translation to translate the query into English,” he said. “Then, when we get the results, we return the results in English. The next level is to do machine translation of the content of the product description so we can translate the titles.”
AI-powered fitting room tech
Computer vision also powers one of Walmart’s most recent AI-driven offerings: dynamic virtual fitting room technology from Zeekit, which Walmart acquired last year. It allows customers to shop for clothes online and see how an item will actually look on them.
Walmart’s “Choose My Model” experience, which launched in March, offers customers a choice of 50 models between 5’2” – 6’0” in height and sizes XS – XXXL. Customers can choose the model who best represents their height, body shape and skin tone.
“Based on the millions of images we have from the catalog, we analyze all the different points of articulation on the models and use that to create the dress simulation,” said Desiree Gosby, VP, new businesses and emerging tech at Walmart. “It’s about breaking down everything from whether it’s supposed to fit loosely or not, where the waist is supposed to fit, how the length should adjust depending on your height.”
Currently, Gosby’s team is working on an experience using Zeekit’s technology where customers can actually upload their own photos. “It’s actually a harder problem for AI and computer vision,” she said. “And customers have to make sure they are taking a good picture that they feel good about.”
Conversational AI in the mix
Walmart also recently launched, after several months of testing, its conversational AI technology called Text to Shop. Customers can text or say what they need and Text to Shop will add it to their cart. If they need an item they’ve never purchased before, Walmart will provide product recommendations
“This is really about how we make it easier for the customer to express what it is that they want or that they need from us,” said Gosby. “It’s basically a digital assistance platform that leverages voice and text chat – we work with across the company including customer care and we power the Walmart shopping assistants on Google and in Siri.”
Text to Shop is the result of a lot of investment in natural language understanding, she added. “We’re leveraging GPT-3 underneath the hood and then really leveraging our data to create natural language understanding that is natural.”
But, she admits, “making this simple is actually really hard – being able to understand if you say things like add chocolate milk and pizza to my cart that you really mean chocolate milk as opposed to chocolate versus milk.”
Overall, these technologies are about giving customers confidence to make purchases, said Gosby. “Are we actually saving them time? Do we decrease return rates for apparel?” Everything Walmart does has to be about removing friction for the customer in some way, she said: “We don’t do technology for technology’s sake.”
Walmart’s AI is tightly focused on the customer
When asked about the future of AI at Walmart, Venkatesan circled back to focusing on the customer. “Our prediction of the future has always been what the customer wants – we observe the customer very carefully,” he said. “We can understand how customer trends are going and we will then adapt ourselves to it, because it’s very tough to predict exactly where it will go.”
Walmart will continue refining, he added. “I think there are a lot more improvements to be done,” he said. “It will be a constant evolution or upgrading of what we do continuously, because it’s only going to get more complex as the customer demands change.”
VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Learn more about membership.