Asda’s recent technology history offers a useful illustration of the difference between systems working and retail working. After separating from Walmart, the British grocer undertook a vast technology program involving more than 2,500 applications and services across finance, stores, logistics and ecommerce. By late 2025, the final systems cutover had disrupted the flow of stock between depots and stores, contributing to inconsistent availability online and in stores. The systems were later stabilized, but the episode exposed a distinction that extends well beyond Asda: a technology program can succeed on its own terms while the retail operation around it struggles.
Asda was not an AI failure. It was a reminder that technology only creates value when the operation around it can absorb the change. That lesson matters more as artificial intelligence moves deeper into retail. Much of the industry’s technology agenda has been built around improving the intelligence of the system — better forecasts, recommendations, automation and, increasingly, more capable AI models. Yet a technically superior output has little commercial value if it reaches the wrong person, arrives too late or cannot be acted upon.
Retailers are becoming better at producing signals. The harder problem is turning those signals into better decisions.
The distance between insight and action
Consider a system that identifies an empty shelf with impressive accuracy. Detection is useful, but it is not the business outcome. Someone still has to receive the alert, determine whether stock exists elsewhere in the store, decide whether replenishment should take priority over other work and move the product before the lost sale occurs.
The problem may be detected in one system, caused in another and solved in a third. The empty shelf may first appear as a computer-vision alert. The cause may be an inaccurate forecast, a delayed delivery, insufficient labor or a replenishment failure. The remedy may depend on an associate receiving the right task at the right time, with enough information and authority to act.
The same logic applies elsewhere. A better demand forecast matters when it changes an order quantity. A labor recommendation matters when it changes a shift plan. An alert matters when it changes replenishment quickly enough to improve availability. The relevant chain is longer than the technology itself: signal to decision, decision to action, action to measurable retail outcome.
A model can therefore be accurate without being useful. An automation can function without improving the business. A recommendation can be correct and still fail commercially because the surrounding workflow is too slow, too fragmented or too rigid to respond.
For decision-oriented retail technology, value is created when a digital output changes operating behavior.
Adoption has moved faster than proof
The gap is particularly visible in AI because adoption has accelerated more quickly than measurable returns. In McKinsey’s 2026 European Grocery CEO Survey, 47 percent of respondents ranked AI and automation among their three leading priorities. Yet 70 percent said they had seen no measurable EBIT impact from AI so far, or that it remained too early to determine one.
Those figures do not amount to a verdict on AI. They suggest that technical capability is only one part of the economic problem. Retailers can generate more predictions, recommendations and alerts than they could a few years ago, but the organization around those systems has to absorb them.
KPMG’s 2026 research on consumer and retail technology describes a similar shift, with companies moving beyond isolated pilots toward scaled deployment while fragmented technology stacks and weak governance remain barriers to sustained returns.
The bottleneck is increasingly not whether the system can produce a useful answer. It is whether the business can use that answer repeatedly, quickly and at scale.
The execution layer
Retail operations do not conform neatly to technology boundaries. Inventory, labor, logistics, pricing and store execution may all contribute to the same customer outcome, while the systems supporting them operate separately.
That fragmentation matters only because it interrupts execution. A retailer may have an accurate forecasting model, a capable inventory platform and effective store-tasking software and still fail to improve availability if the signal does not travel cleanly from one part of the business to another. Inconsistent data, disconnected systems and unclear governance are not the central story on their own; they are the reasons a useful recommendation may never become a useful action.
This is why the execution layer deserves more attention in discussions about retail AI. The retailer has to know who owns the recommendation, what authority that person has, how quickly the decision must be made, what happens when local knowledge contradicts the model and how the resulting action will be measured. These may sound like operating-model questions, but they determine whether the technology ever produces a return.
A useful AI system is not simply a model attached to data. It is a model embedded in a decision process.
Start with the decision
Retail technology projects naturally produce technical measures. Deployment progress, uptime, latency and model accuracy all matter, but they describe whether the technology is functioning. They do not establish whether the retailer is improving.
The stronger starting point is the operating decision that needs to change.
If the objective is higher on-shelf availability, the project should define how a signal becomes a replenishment action before debating the sophistication of the model. If the objective is lower waste, the important link is between the forecast and the ordering decision. If the objective is higher picking productivity, the technology has to work inside the picker’s actual workflow.
This reverses a familiar sequence. Instead of deploying a capability and then attaching a business case to it, the retailer begins with the decision, the action and the metric that should move.
That also changes the meaning of a pilot. A pilot can show that a model is accurate or that automation is technically feasible. It says much less about whether the surrounding operating model will work across hundreds or thousands of stores, where local processes, maintenance, staff turnover, data quality and exceptions are normal conditions rather than edge cases.
At scale, the workflow is part of the technology.
AI will be judged by what changes
The next phase of retail AI may therefore look less dramatic than the first. Some of the most valuable work will concern integration, decision rights, workflow design and the removal of manual bottlenecks rather than another visible AI feature. These are not the elements that make the strongest product demonstration, but they are where much of the economics is decided.
Retailers will continue to experiment with assistants, forecasting tools, recommendation systems and other forms of AI. Some will prove durable and others will not. The more important capability will be the ability to convert machine intelligence into operating behavior consistently.
The unit of value in retail AI is not the prediction. It is the decision that changes because of it.
A more capable model can improve the signal, but it cannot complete the chain on its own. The retailer still has to decide who receives the output, what authority that person or system has, what action follows and which commercial metric should move if the technology is working.
A retailer can buy the same model as its competitor. Building the operating discipline to turn useful signals into action is much harder to replicate.
For years, retailers have asked what new technology can do.
The harder question is what the business is prepared to do with it.