Walmart: when shopping starts somewhere else

In January 2024, Walmart offered customers a different way to shop for a football watch party. Instead of searching separately for snacks, decorations and tableware, they could ask for help with the occasion. The search service combined a language model Walmart described as open source, adapted using its retail data, with OpenAI models supplied through Microsoft Azure.¹

Walmart wanted to act as a shopping partner, helping customers work out what they needed and arrange how to get it. It called this “adaptive retail”: connecting stores and digital services around customers’ needs. Technology chief Suresh Kumar described developing shared technology once and deploying it across Walmart U.S., Sam’s Club and its international businesses.²

Finding the right dress

Helping customers choose suitable products depended on reliable information about what Walmart sold. Walmart explained the problem from the shopper’s perspective: someone searching for a red cocktail dress might be shown a blue maxi dress. Across its extensive product range, keeping billions of catalogue data points accurate by hand was impractical.³

In August 2024, Walmart said generative AI had created or improved more than 850 million pieces of product information. Executives described better discovery across its product range and clearer packaging images that helped employees identify items when picking orders.⁴

The work needed to be accurate and affordable at that scale. Balu Chaturvedula, who leads Walmart Global Tech in India, described catalogue workloads for which frontier-model costs were impractical. Walmart combined models, including smaller ones trained for particular tasks, to make the work more economical.⁵

Its technical account explains how extraction and checking are separated. One model identifies product details in descriptions and images; another checks the results. Human quality assurance specialists assess those details against defined specifications, using judgement where descriptions or measurements require interpretation. Their assessments provide training examples for the checking model. The models are tested on separate examples that were not used in training.³

Larger models also help improve the smaller ones. Where a smaller model struggles to identify particular product details, Walmart uses a larger model that performs reliably on those details to generate additional training examples. This helps the smaller model learn the task without requiring the larger model to process every item.³

From search to a shopping assistant

Sparky brought more of the shopping conversation into Walmart’s app. Launched in June 2025, it helped customers compare products, ask questions and get recommendations for occasions. Its usefulness depended on more than a model’s answer: recommendations had to connect with Walmart’s products, inventory, delivery capabilities and customer preferences.⁶

Customers subsequently began using Sparky for more routine shopping. Walmart reported growing use for food and household essentials as it added meal planning, replenishment and personalisation.⁷

Walmart doubled down on its strategy by adding a new senior role to accelerate AI adoption across the business. Daniel Danker joined from Instacart in August 2025, bringing experience in product development and online grocery. Reporting directly to then-CEO Doug McMillon, his remit combined accelerating AI adoption with leading product management and design across Walmart.⁸

By August 2026, Walmart said customers using Sparky spent 40% more per order than those who did not.⁷

Taking the business into someone else’s interface

Walmart wants to remain part of customers’ shopping as AI services open new ways to choose and buy. It is investing in its own shopping tools, creating connections for outside agents and working with partners such as ChatGPT and Gemini to help shape how those channels develop.

One opportunity is to reach people who have not yet decided where to shop. Walmart illustrated this with someone asking Gemini about camping equipment and discovering relevant products from its range. Turning that discovery into a purchase depends on making the shopping experience work for the customer.⁹

The first ChatGPT checkout did not do that well. From November 2025, customers could buy a limited selection of Walmart products through Instant Checkout without visiting Walmart’s website. But they had to buy items individually, separately from baskets they were already assembling with Walmart. Danker told WIRED that shoppers completed purchases at one-third the rate recorded when ChatGPT sent them through to Walmart.¹⁰

Danker described a different approach. ChatGPT or Gemini would first surface a Walmart product; when the customer selected it, Sparky would take over the Walmart shopping experience. Customers could use their accounts, combine purchases with existing baskets and retain membership benefits. Google announced plans for those features in Gemini. OpenAI launched a Walmart-run experience inside ChatGPT in March 2026, acknowledging that its initial checkout approach had lacked the flexibility merchants needed.¹¹⁹

Sparky gives Walmart a way to bring its customer relationships and shopping capabilities into another company’s service. Danker saw an opportunity to win business beyond customers’ usual purchases. Someone who buys groceries there each week might discover its electronics, beauty products or home furnishings through an AI conversation. He also described a commercial test for the platform partnerships: were they bringing additional customers and shopping occasions, or serving business Walmart would already have received through another route?¹¹

Keeping a role in recommendations matters to advertising too. Suppliers pay Walmart to help their products reach customers. Danker argued that Sparky could continue selecting and ordering product recommendations inside another company’s assistant. He said their relevance remained essential: unsuitable promotions would fail to generate the sales advertisers wanted.¹¹

The platforms have commercial ambitions of their own. OpenAI began testing ads in ChatGPT in February 2026; Google has been testing sponsored offers and new advertising formats in AI Mode in Search. Walmart’s partners could help it win sales while competing for some of the advertising spending it seeks from suppliers.¹² Google’s Universal Cart, launched in October across Search and Gemini in the US, also brings products from multiple retailers into a shared shopping experience.¹³

Walmart’s own Agent Gateway opens another route. Announced in October as an invitation-only beta, it connects outside agents with Sparky to shop on customers’ behalf. Its published terms restrict agents’ reuse of Walmart’s proprietary data, including for model training, advertising targeting or resale. Walmart is opening access while setting conditions intended to preserve its role and information.¹⁴

Those protections do not ensure that a customer reaches Sparky. An outside assistant may recommend another retailer. Even when it brings a purchase to Walmart, it may already have helped settle choices that Walmart would previously have helped the customer make. Walmart is therefore working to earn a place in the shopping journey as well as retain its role once the customer arrives. Its investments and partnerships give it ways to pursue both; they do not settle how much business or influence it will retain.¹¹

The work behind the promise

A successful shopping conversation still leaves an order to fulfil. AI was also helping Walmart’s store managers prepare the day’s work and employees get started.

In June 2025, the company reported that managers estimated a reduction in overnight-shift planning time from 90 minutes to 30.¹⁵ A manager later told Business Insider that employees could see their assignments without waiting for someone to give them instructions.¹⁶

But receiving an assignment sooner did not mean its time allowance covered everything the job required. Other workers interviewed by Business Insider described estimates that left out checking expiry dates, removing damaged goods and cleaning spills. Those tasks still had to be done before the shelf was ready for customers.¹⁶

Brooks Forrest, Walmart’s vice-president for associate tools, said stores retained discretion over how they used the tools. “We do not punish people for not following the tech guidance,” he told Business Insider. He described store visits and feedback sessions as part of improving the tools.¹⁶ Greg Cathey, its transformation and innovation chief, had also described testing with employees and obtaining operational approval before wider rollout.¹⁷

Employee feedback can identify work the system has missed. It can also question how much work is reasonable in the time available. Correcting an estimate and reconsidering an expectation require different judgements. Walmart’s accounts of time saved and employees’ accounts of work left out both matter when assessing what the tools achieve.

The results have more than one audience

How to assess the effects on employees has also become a shareholder question. United for Respect, a retail-worker advocacy organisation, led a proposal co-filed by the Congregation of St. Joseph in Ohio, seeking additional reporting on AI and automation’s workforce effects.

Walmart’s board argued that its existing disclosures, skills investment and oversight by technology and workforce committees were sufficient, and recommended voting against the proposal. Shareholders followed the board’s recommendation and rejected it on 4 June 2026.¹⁸

Walmart’s investment serves the business it operates today while preparing it for a market that may work differently. More reliable product information and economical model choices have immediate uses. Developing Sparky and connections to outside agents gives Walmart ways to carry its strengths into new shopping channels, while its partnerships help it participate in shaping them. The checkout experience prompted a change in how Walmart and OpenAI worked together.

Walmart is trying to protect and develop its future before the direction of shopping becomes clear. Its technology gives it more ways to compete, but customers still have to choose it, partners have ambitions of their own and employees have a stake in how the work changes. The investment creates room to act; the position it secures remains something Walmart has to earn.

Connected reading

Open models. Competing ambitions. examines what startups and established businesses develop from available models, and where the resulting value accrues. Some assembly required explains models, adaptation, software and services.

What the company knows considers organisational knowledge and stakeholder interests.

The alphabet soup of agentic commerce examines customer access, authority and value as agents enter commerce. Its companion map identifies the initiatives and participants. The position is taken and Adoption is the value thesis explore how investment changes an organisation’s opportunities and its ability to capture value.

Notes

Footnotes

1. Walmart Global Tech, Walmart’s Generative AI search puts more time back in customers’ hands, 30 January 2024; Microsoft, announcement of Walmart’s generative-AI capabilities, 9 January 2024. Walmart describes a fine-tuned open-source model; Microsoft identifies Azure OpenAI Service. These accounts do not name the base model, its licence or the OpenAI model versions. ↩

2. Walmart, CES announcement, 9 January 2024; plan for scaling AI and other retail technologies, 9 October 2024. McMillon and Kumar articulated the adaptive-retail direction; Kumar described shared technology across Walmart’s businesses. ↩

3. Binwei Yang, Walmart Global Tech, How Walmart uses LLMs to manage its massive product catalogs, 20 May 2025. Describes extraction, quality checking, human evaluation and training examples generated by larger models. It does not provide a complete named model inventory. ↩ ↩² ↩³

4. Walmart, second-quarter FY2025 earnings call, 15 August 2024, pp. 4 and 10. These results concern the wider catalogue programme, rather than an evaluation of the particular configuration described the following May. ↩

5. Rukmini Rao, Fortune India, How retail giant Walmart is managing its AI costs, 21 October 2026. Interview with Balu Chaturvedula about the economics of different model choices. It does not quantify the resulting increase in processing capacity or sales. ↩

6. Walmart, The Future of Shopping Is Agentic. Meet Sparky., 6 June 2025; Goldman Sachs conference transcript, 9 September 2025, pp. 9–10. The launch distinguished available shopping assistance from proposed reordering and other capabilities. ↩

7. Walmart, Q1 FY27 earnings call, 21 May 2026, p. 14; Q2 FY27 earnings call, 20 August 2026, pp. 5–8. The spending comparison is between users and nonusers, not the same customers before and after adoption. No controlled comparison or Sparky-specific profitability calculation is supplied. ↩ ↩²

8. Walmart’s 2026 annual report, executive-officer biographies; Danker biography; Bloomberg, appointment reporting, 23 July 2025. Danker joined in August as Executive Vice President, AI Acceleration, Product and Design. McMillon’s explanation of the new role, accelerating AI adoption while leading product management and design, was reported in Drug Store News, 23 July 2025. This was an additional senior role, not a replacement for Suresh Kumar. ↩

9. Walmart, Walmart and Google Turn AI Discovery Into Effortless Shopping Experiences, 11 January 2026, supplies the illustrative camping example and plans for linked accounts, combined baskets and membership benefits in Gemini. OpenAI, Powering Product Discovery in ChatGPT, 24 March 2026, announces the Walmart experience on the web, with mobile access forthcoming. The announcements establish plans and a change in approach, not an improvement in subsequent conversion. ↩ ↩²

10. Paresh Dave, WIRED interview with Daniel Danker, 18 March 2026. The routes offered different product selections; no controlled comparison was disclosed. ↩

11. Walmart, Morgan Stanley conference transcript, 4 March 2026, pp. 4–9, on additional shopping occasions, the handover to Sparky, selection among retailers and advertising; ICR conference transcript, 13 January 2026, pp. 10–12, on additional business, customer information and platform economics. Danker’s account distinguishes initial discovery in the outside assistant from Walmart’s subsequent shopping experience. No specific fee or revenue-sharing rate is disclosed. The possibility of losing a purchase or influence before the handover is a commercial inference, not evidence of a contractual right to remove Walmart or Sparky. ↩ ↩² ↩³ ↩⁴

12. OpenAI, Ads in ChatGPT, checked 29 October 2026, dates the start of US testing to 9 February 2026 for eligible Free and Go users; it describes labelled placements separate from ChatGPT’s answers. Google, AI Mode advertising announcement, 20 May 2026, describes tests of new AI Mode advertising formats and an expanded Direct Offers pilot. These Google announcements concern Search, not an advertising launch in the standalone Gemini app. Potential competition for suppliers’ advertising budgets is an inference; these sources do not establish lost Walmart revenue or the advertising provisions of its negotiated partnerships. ↩

13. Google, Universal Cart announcement, 15 October 2026. Announces a US rollout across Search and Gemini, including products from Walmart and other retailers, with Google Pay checkout or transfer to a merchant’s cart. Retailers remain sellers of record and provide confirmations and customer support. This establishes a platform shopping capability, not evidence that Google can bypass Sparky under its agreement with Walmart. ↩

14. Walmart, Agent Gateway announcement, 21 October 2026; Gateway access page; Agentic Access Terms, version 21 October, checked 29 October. Access was an invitation-only early beta. Section 2 explicitly defines the Gateway as enabling agent-to-agent shopping with Sparky; section 6 restricts reuse of Walmart proprietary data. The Gateway addendum sets customer-authorisation and transaction requirements. Interaction with Sparky does not necessarily require a customer-visible Sparky conversation. The public terms do not disclose separately negotiated arrangements with individual AI platforms. ↩

15. Walmart, announcement of employee AI tools, 24 June 2025. The planning-time reduction is an early estimate from team leads and store managers, initially for overnight stocking. It is not a measurement of total labour productivity. ↩

16. Dominick Reuter, Business Insider, Walmart store workers have a new responsibility: correcting AI errors, 18 August 2026. The report includes a company-arranged tour, worker interviews and Walmart’s response. It does not establish how widespread the reported experiences are or trace a particular complaint into changed staffing or task timings. ↩ ↩² ↩³

17. Greg Cathey, Omni Talk interview at NRF, 13 January 2025, particularly 08:36–13:39. The published transcript describes employee feedback and staged operational testing. This is management’s account of its approach, not an independent evaluation of each subsequent tool. ↩

18. Walmart, 2026 Proxy Statement, pp. 100 and 108–109; Committee on Workers’ Capital, 2026 voting tracker, identifying the co-filer; Walmart’s voting results, recording rejection on 4 June 2026. The vote preceded Business Insider’s August reporting. The proposal concerned additional reporting; its rejection was not a vote to prohibit workforce assessment or employee consultation. ↩