As Retail AI Faces A Reality Check, Winning Operators Build For Precision Over Scale

AI in retail has arrived primarily where it's easiest to deploy, in customer-facing personalization, app-based recommendations, and marketing automation. Those are visible, demonstrable, and relatively low-risk. However, the operations that actually determine whether a retailer makes money, like inventory accuracy, supply chain coordination, store-level decisioning, and loyalty activation, run on back-end systems that are far more complex, fragmented, and less ready for the large language models that vendors have been selling into boardrooms for the past three years.

Howard Lake is the publisher of Retail Slop, a newsletter covering global retail operations, technology, and market intelligence. He's spent over two decades analyzing retail infrastructure, store operations, and technology adoption across European and international markets, with recent work focusing on the gap between AI's marketed potential and its operational reality in retail. Given the industry's structured, repetitive, and security-sensitive workflows, Lake believes small language models may be more suitable than the general-purpose LLMs that have dominated the conversation.

"A lot of retailers bought into the hype. The capabilities of AI have been oversold, and we're now seeing people take a step back," Lake says. He points to recent high-profile rollbacks and withdrawals as evidence that the industry's initial enthusiasm is giving way to a harder set of questions about what actually works.

Legacy systems block the path forward

The structural challenge facing most retailers when it comes to AI adoption is whether their existing infrastructure can support it. Large retail operations are built on layers of legacy systems, and getting those systems to communicate with each other, let alone with a new AI layer, is an enormous undertaking. Lake notes that decades of M&A activity have created a tech debt that generic AI simply cannot digest. "Carrefour in Brazil has the license for Sam's Club, so they're inheriting Walmart legacy systems. We've seen this recently with ASDA in the UK, too. The business was acquired from Walmart four or five years ago and they still haven't finished untangling the Walmart IT legacy system. They can't really do anything on the back end with AI until they've dealt with that."

Large models carry risks retail can't absorb

Beyond integration challenges, Lake raises a concern that most vendor conversations underplay: security. Retail operations handle vast volumes of customer loyalty data, purchasing behavior, and supplier information. Large language models, by their architecture, introduce attack surfaces that many retailers haven't fully mapped. "Because it's such a new technology, we're still only uncovering how vulnerable it can make systems," he says. He references recent discoveries of exploitable flaws in systems previously considered secure, arguing that the risk calculus for a supermarket chain is fundamentally different from that of a defense contractor with dedicated cybersecurity teams.

The argument extends to model utility, as well. Store operations are highly structured, with repetitive processes, predictable workflows, and outcomes that depend on precision rather than creative reasoning. A model trained on billions of parameters carries information the store doesn't need, Lake explains, and that excess breadth increases the likelihood of errors. "There's no need for a model to operate on multi-billion parameter scales just to run a store. If you only train it on the thing it needs to know, it's far more likely to give you the outcomes you want."

Small models, built for the store

Lake's central thesis is that small language models, tightly scoped and hermetically sealed within individual store or regional networks, are a better architectural fit for retail than general-purpose LLMs. "The problem with large language models is whether they’re actually suited for the kind of work retailers want them to do at all," he says. With his approach, each store could run its own contained model, trained on its specific operational data, with outputs feeding into a centralized model at headquarters that aggregates intelligence across the network. He points to Mistral's Forge model and its recent partnership with Tesco as an example of this direction taking shape. "The Forge model seems very much targeted at very specific use cases where the data is highly protected," Lake says, noting that it's the same profile that defines most retail back-end work. "I see that being more transferable to the retail environment."

On the infrastructure side, Schwarz Group, the parent company of Lidl, represents the most aggressive move toward sovereign, self-contained AI architecture. The company is building a dedicated AI campus at its Heilbronn headquarters, has launched its own European cloud platform called StackIt, and has previously connected all its European stores into a single centralized server network. "They will be way out ahead. The whole idea of digital and data sovereignty and reducing the areas where you can get attacked will become much more important," Lake asserts. 

Talent and vendor gaps persist

Even with the right architecture, retail faces a talent problem. Most AI projects at major retailers are being run by IT teams who were already in place, not by specialists trained in prompting, business-specific model design, or creative application of AI to retail workflows. Lake sees this as a structural bottleneck that will persist until vendors with deep retail expertise create purpose-built AI solutions that integrate directly into existing retail systems. "I see a massive opportunity for any supplier in the AI space that understands the nuts and bolts of the business. When CTOs start saying, 'This is the thing we've been waiting for,' that's when we'll see a real leap. But I think we're a ways off."

He's candid about the gap between what technology leaders inside retail organizations are experiencing and what C-suites are being told. "I'm sure there are people in tech teams across retailers who are saying, 'It can do some of it, but it's not doing exactly what we want.' Somewhere out there is somebody listening to those tech teams, going away to build it, and coming back in six months with the thing that actually works."

Precision over scale

In Lake's view, the next phase of retail AI will be defined by brands that deploy the most precise, secure, and operationally integrated systems within the constraints of how stores actually run. Rather than attempting to achieve the broadest capabilities, the vendors who win will be the ones who deliver solutions that connect to inventory, supply chain, and store-level workflows out of the box, built by people who understand the business deeply enough to know what a store actually needs. "The search is currently on for the optimal AI solution for retail," Lake says. ""Retail knows what it wants. It just hasn't appeared yet."