The Gap That Earns the Read
Walmart has deployed AI more broadly than almost any company on earth. It runs AI across catalog management, supply chain routing, equipment maintenance, fashion production, and associate tooling — at a scale that dwarfs most enterprise AI programs by an order of magnitude. And its most significant AI bet, a proprietary retail large language model called Wallaby, still hasn’t shipped.
That gap is the story. The operational wins are real, verified, and significant. The strategic crown jewel is still in internal testing. Understanding both — what Walmart has actually built and what remains unproven — tells you something important about what enterprise AI at scale actually looks like when a serious organisation attempts it.
Infrastructure First, Always
There is a logic to the sequence of Walmart’s AI deployment that becomes clear in retrospect. The company spent the first major wave embedding AI where no customer would ever see it: supply chain routing, refrigeration monitoring, product catalog enrichment. These are not glamorous applications. They are, however, exactly the right place to start.
Operational AI wins don’t require customer trust. They can be validated quietly, measured precisely, and iterated without reputational exposure. When Walmart deployed digital twin technology across store infrastructure, the results were concrete: a 30% reduction in emergency maintenance alerts, a 19% drop in refrigeration maintenance spend, and 842 equipment failures caught proactively across 20 stores in six months. A refrigeration technician whose emergency call volume has dropped by nearly a third isn’t being replaced — not yet — but their job has changed materially. The work that used to come in as a crisis now arrives as a scheduled intervention. That’s a different kind of work, and it requires a different kind of readiness.
Route optimisation followed the same logic. The AI-driven delivery routing eliminated 30 million delivery miles and avoided 94 million pounds of CO2 emissions. For the delivery driver, that means a route that is now algorithmically optimised to a degree no human dispatcher could match — tighter windows, fewer inefficiencies, and less room for the informal judgment calls that drivers have always made. The work is more efficient and more prescribed. Walmart was direct enough about the value of this capability that in March 2024 it commercialised the technology through Walmart Commerce Technologies, selling it as a SaaS product to other retailers. The AI investment stopped being purely an operational cost and became the foundation of a platform business — the same pivot IBM made when it began selling the tools it had built internally.
On the catalog side, GenAI is now responsible for improving 850 million product data points, a number Doug McMillon cited in the August 2024 earnings call. Fashion production timelines have been cut by 18 weeks using Trend-to-Product AI. These are not incremental gains. They represent a fundamental restructuring of how Walmart manages the information layer of its retail operation.
The Honest Accounting Problem
In 2024–25, Walmart reported 26% EPS growth, and AI was named as a contributor. The attribution here deserves scrutiny, because the honest answer is that no one has cleanly separated the AI signal from the noise.
Walmart simultaneously ran a major e-commerce expansion, a private label investment cycle, and a broad supply chain modernisation program. When a company this large improves performance across multiple simultaneous initiatives, any single-cause narrative should make you suspicious. The AI contribution is real — the operational metrics support that — but the financial narrative is directional, not quantified. Attributing a specific percentage of earnings growth to AI deployment is methodology that no serious analyst has actually demonstrated. What the data supports is: AI is contributing. What it does not support is a clean number attached to that contribution.
This matters because the enterprise AI investment case is frequently built on exactly these kinds of aggregated claims. Walmart’s story is stronger than most, but even here the honest answer is “we believe it’s working” rather than “we can prove it’s working at this ROI.”
What McMillon Said Out Loud
In September 2025, Doug McMillon told CNBC that AI will change “literally every job” at Walmart. Most CEOs won’t say that. They’ll say AI will augment workers, or create new roles, or improve productivity. McMillon said something more honest, and it deserves to be taken seriously.
In May 2025, Walmart cut 1,500 corporate jobs. In July 2025, hundreds more followed. The company launched AI tools for its 1.5 million associates in June 2025, and in partnership with Google, offered 1.6 million workers free eight-hour AI training. The sequence matters: tools deployed, training offered, headcount reduced. That is the actual shape of how workforce AI transitions happen in large enterprises, regardless of how they’re framed in press releases.
The associate AI tools are real, but the question of what those 1.5 million workers are actually being asked to do differently is harder to answer from the outside. What does an AI assistant change for a stocker managing inventory on a night shift? Or a customer service associate whose system now surfaces product information before they’ve finished the customer’s sentence? The work changes before the job title does, and the gap between those two things is where the real workforce story lives. McMillon was honest about the destination. The path between here and there is what most companies, including Walmart, have not yet figured out how to manage well.
The Wallaby Question
At the centre of Walmart’s long-term AI strategy is Wallaby, a proprietary large language model trained on decades of Walmart transaction data, product catalog records, and supply chain information. The theory is sound: a domain-specific model trained on retail-native data should outperform general-purpose LLMs on retail tasks — personalisation, demand forecasting, associate queries, supplier negotiations. General-purpose models are trained on the internet. Wallaby would be trained on Walmart.
As of late 2024, Wallaby was still in internal testing. That’s a long runway for a model that was positioned as a strategic priority.
The stakes are asymmetric. If Wallaby ships and performs, Walmart has a proprietary AI capability that no competitor can license, replicate, or buy. It becomes the intelligence layer for a company with 1.6 million US employees, 11,500 global stores, and transaction data going back decades. That’s a defensible advantage of a kind that Amazon’s AWS-native ML infrastructure currently holds and Walmart currently lacks.
If Wallaby doesn’t ship, or underperforms, Walmart is left running its most sensitive competitive workloads — demand forecasting, personalisation, supplier intelligence — through OpenAI, Google, or Anthropic APIs. That’s a significant strategic liability. It means Walmart’s most valuable data is being processed through infrastructure it doesn’t control, by vendors who also serve its competitors, under terms that can change.
The Wallaby bet is partly a response to Amazon. Amazon has 15 years of machine learning infrastructure built into its fulfillment and recommendation systems, plus AWS as a commercialisation vehicle that funds continued model development. Walmart is not a fast follower here — its operational AI program is serious and genuinely competitive — but Amazon’s head start in training models on its own transaction data at scale is real. Target, for comparison, has had notable AI deployment failures and has proceeded with more caution, which puts Walmart in the middle: more aggressive than Target, not yet as embedded as Amazon.
The Position
Walmart is running the right playbook for enterprise AI at scale. The sequencing — operational infrastructure first, customer-facing second, proprietary model as the long-term bet — is coherent and defensible. The workforce communication has been more honest than most. The commercial pivot through Walmart Commerce Technologies changes the ROI calculation in ways that most AI investment analyses don’t account for.
The operational wins are real and replicable. Any enterprise with serious data infrastructure can learn from the digital twin and route optimisation work. The methodology is visible and the results are measurable.
The Wallaby bet is high-stakes and unproven. The workforce story is more honest than most, but honesty about displacement is not the same as managing it well. And the financial narrative — while directional — is not the clean AI-drives-earnings story that gets told in earnings calls.
Whether Walmart is running this playbook well enough to stay ahead of Amazon is the open question. The gap between what’s already deployed and what hasn’t shipped yet is where that answer will be written.