How retail giant Walmart is managing its AI costs

Last year, at Walmart Global Tech’s flagship technology and retail forum, it launched Wibey, a super-agent for the company’s software teams across regions to build intelligent, context-aware systems. It also unveiled Sparky, an agentic AI shopping assistant; and Marty, a super-agent designed for suppliers, sellers, and advertisers’ ecosystem. 

At the company’s latest earnings call in August, John Furner, President, CEO & Director, Walmart Inc., while speaking on AI adoption,  shared that the usage of Sparky by customers was up 70% from last year,  and among those who use it, Walmart saw shopping spend 40% more per order . “We continue to take a people-led, tech-powered approach. We’re using AI to make our work easier and to help our associates grow and be at their very best. We believe AI will improve nearly every part of our business by making shopping better, and our associates work easier,” Furner said . 

Walmart’s approach towards harnessing tech innovation rather than building entirely new capabilities market by market—focussing on building once and improving continuously, scaling those globally and making them faster and more efficient—allows customers in more markets to benefit from innovations developed anywhere across the company. Of the super-agents launched last year, Balu Chaturvedula, Senior Vice President & Country Head, Walmart Global Tech India, says the most progress  has been in Sparky and Wibey. 

At a time when tokenmaxxing is giving way to prudence, for enterprises, AI usage is focussed more on who needs what AI tools and outcomes from its use. At Walmart’s scale, being EDLP (Every Day Low Prices) and EDLC (Every Day Low Cost) are core to its operations. “ Cost is the most important dimension for us. We don’t have the luxury of overspending, so every time we look at something, we have to choose the most viable and cost-relative option,” says Chaturvedula. 

AI Innovation and Balancing Costs 

At the adoption curve of super-agent Wibey, the primary challenge was adoption among developers. While developers could initially write code in Wibey, taking that entire code back to their traditional development environment limited the gains. The IDC teams then redesigned the developer experience so users could simply describe what they wanted and complete the entire development workflow within the platform. Once usage increased, attention shifted to cost management. “We came up with some guardrails: every leader has a budget, and you need to manage your budget. How do you manage the budget?” Chaturvedula asked.

“What we’ve seen so far is that even with the guardrails, there are a number of people asking for exceptions, and there are good use-cases. Some have become power users. It’s a spectrum of people who want to experiment more and do a lot more with the existing tools,” he explained. 

On the business side, a year ago when the company started using frontier LLMs to enhance content quality—for instance, to enhance listings for users—the cost seemed non-viable. “We have half a dozen such use-cases. We have to do duplicate detection for listings. We also have to do variants. We also have to do resold items. There are multiple variations, and for each one of them the cost can grow exponentially. We’re talking about 850 million data points on a daily basis,” said Chaturvedula who also looks after the catalogues for Walmart globally. 

With the learnings Walmart is looking at a combination of open-weight models, frontier LLMs, and building its own small language models for organisational AI needs. “We’re no longer in the red zone. We’re able to tell the rest of the organisation how using these model orchestration layers, you can control cost and build software in a more scalable and economical way,” Chaturvedula said. 

The Future of Enterprise AI: Beyond Models to Interoperability 

Enterprise AI conversations rapidly evolving beyond the LLM race and models. “Six months ago, the way everybody described LLMs, the number of parameters was very important, 8 billion parameters, 12 billion,” said Chaturvedula, pointing at discussions that once centred on model size and parameter counts, but is no longer a test to convince enterprises on how good a language model is.  

As the focus is no longer just on the cost of training models, but enterprises looking to balance both training and inference costs, he also sees industry moving towards universal interoperability between models. “Eventually you’ll get to a stage where there is inter-connectivity, or interoperability, across any model. That standardisation will soon happen,” he says. Drawing parallels with the evolution of cloud computing, he sees AI also following a similar pattern where seamless integration becomes essential for enterprise adoption at scale.  

Globally, Big Tech firms, frontier LLM companies, and tech infrastructure firms account for trillions of dollars of bets that investors have placed on AI. However, the current AI adoption trend suggests that no organisation is likely to put all eggs in one basket or with any one company. As market pressures start building, for AI infrastructure companies and model providers the need to generate value will provide more impetus. “They still have to figure out how to stay relevant, and to stay relevant in the organisation and find new revenue models—interconnectedness is the only way, according to me. I have a feeling that not many companies are talking about this publicly, but my gut tells me this is the next big phenomenon for LLMs and the whole model orchestration,” said Chaturvedula.