Poojitha Jayadevan
Staff Writer

How Walmart’s Indian IT team used AI to predict customer preferences

Feb 20, 2022
Artificial IntelligenceRetail Industry

Walmart Global Tech India developed applications to improve customer satisfaction and employee efficiency during the pandemic.

Online orders skyrocketed at Walmart, the largest retailer in the US, when the pandemic hit, making more work for in-store employees. At the same time, demand for certain products led to frequent stock outages.

While Walmart’s ordering app allowed customers to indicate their preferred substitutes for out-of-stock products, customers usually skipped this step. This forced the Walmart employees who pick and pack items on behalf of the customer to make the decision themselves.

As a result, dissatisfied customers returned one in ten substitute items, leaving Walmart to refund the full amount of the product and pick up the cost of restocking.

To reduce the number of returns and the accompanying losses, and to improve customer experience, the company’s innovation hub, Walmart Global Tech India (WGTI), rolled out an AI system to learn customers’ preferences. It uses data to predict consumer behaviour, preferences, and needs.

“The AI-driven system learns individual preferences of every customer over a period of time and gives the pickers hints to what the customer likes if a particular item is not available,” says Rohit Kaila, WGTI’s vice president of US tech.

Further adding to the workload of in-store employees assembling online orders for delivery, a few thousand Walmart stores added the option of curb-side pickup to reduce customers’ exposure during the pandemic.

“Earlier, the supply chain part was optimized only for people coming into the stores. Now you have a lot more and different kinds of goods that come in, so a lot of the supply chain aspects had to be designed accordingly. You need to build a much stronger workflow system,” says Kaila.

That prompted WGTI to develop the Me@Walmart app, launched in June 2021, to help employees — known at the company as associates — to manage their work schedules and operate more efficiently. It includes push-to-talk communications to help employees stay in touch around the store, and a way to quickly check the availability of an item in inventory to respond to customers’ questions.

Most importantly for the online commerce operations, it also offers in-store pickers sophisticated routing and batching algorithms to maximize the number of orders served per trip and thereby serve more customers.

Pick-path optimization

A picker always has multiple orders to pick up, and earlier used their instincts to figure out how to collect all items in less time. The pick-path optimization feature of Me@Walmart helps employees fulfil orders while visiting fewer aisles by grouping similar orders in one pick walk.

Kaila explains: “Think of it as an in-store Google map. It bunches orders together and creates a path for an associate to pick up. And while you are doing it, if things are out of stock, you can go back on the system and report.”

WGTI used multiple approaches to innovation to come up with the right solutions: Kaila describes pick-path optimization as a “sideways” approach.

“There was an organization that was doing a lot in the supply chain, like optimizing in terms of how the trucks drive.  We decided to do the same thing on a micro-scale for our storage data. So that’s a sideways innovation. There’s also bottom-up innovation that happens, and top-down which is looking at trends.”

To cope with the additional development work, WGTI hired UI/UX engineers, and data scientists to work on core algorithms and build new ones. Analysts and ML engineers were also needed.

Kaila explains their roles: “We use a lot of existing data science technologies and build our algorithms where a lot of analytics is involved. So, there are a lot of analysts that we hire because input for any data science is analysis. ML engineers may not build algorithms but are building platforms on which multiple algorithms can be trained at the same time.”

The company also needed to more back-end engineers, cloud engineers, and experts in database technologies.

In addition to hiring, WGTI also trained up its existing employees, particularly those who possessed the basic technology skillsets necessary, but were from a different domain. WGTI also worked with universities to offer internship programs for students, who were also a crucial part in developing the solutions.

Less is more

One of the biggest challenges Kaila faced in training the models to accurately predict customer preferences was a result of the pandemic itself, as it caused customers to significantly change their behaviour.

“Traditionally it was always expected what people will buy during holidays, summer or winter. But with the pandemic, people’s consumption patterns changed drastically. To figure out those consumption patterns and enable the entire supply chain mechanism and substitution system was a daunting task,” Kaila says.

A large data set is used to train the AI system: a combination of past purchases, returns, and cross learnings from other customers in the region.

Usually, more data is better when it comes to training AI models, but for WGTI the most relevant data to train the AI algorithm is the most recent data. “Rarely would we use data beyond 90 days or 120 days. During the pandemic suddenly everybody’s ordering lots of milk, and lots of toilet paper and sanitizers. But if you go into the history beyond 120 days it’s nothing like this, so the recent aspect is extremely important,” Kaila says.

Initially, a small set of customers were given the substitute products suggested by the AI model, and their responses studied. If a customer is dissatisfied with the substitution, the amount is refunded, and it becomes valuable data for the AI solution for further learning about the customer’s preference.

Development of the system started in 2020, and since it was introduced customer acceptance of substitutions has improved: only 2% of the AI-suggested substitutes are returned, compared to 10% before, saving staff time and cost.

While the adaptations that WGTI has made to Walmart’s systems have been a success, there is much more work to be done. “Everybody misjudged the enormity of the problem COVID threw in front of all of us. I think if we go back in time, not just us, but as an industry, we need to look at it from a different perspective so that we could have been more stable with our supply chain systems and have much better alternatives for the customer,” Kaila concludes.