The hardest part of retail AI may soon be the part nobody puts in the demo.
A model can recommend a markdown, predict a stockout, adjust a picking route or identify a pattern in store traffic in milliseconds. But a retailer only gets value when that decision becomes an accurate action in the physical store. Whether it’s a price change or a sensor reading, it has to arrive promptly for it to be useful.
There is already a striking gap between AI ambition and execution. Deloitte’s 2026 survey of 200 retail and consumer products executives found that 75% consider AI a top strategic priority, yet only 16.5% can quantify the return they are getting from it. Clearly, retailers believe in AI. The harder problem is turning isolated capabilities into something that works consistently across the business.
Latest Videos FromTechRadar
For physical retailers, the challenge does not end with models, data or software. AI-driven decisions have to pass through stores containing thousands of devices, existing networks and years of installed technology. In complex environments, the building itself can affect connectivity. As the number of connected endpoints grows, so does the importance of the wireless foundation carrying information across the store.
For CIOs and CTOs, that means the next constraint on AI may be something that has been underappreciated: whether the store estate underneath it can execute reliably at scale.
A store is not a cloud workload
Physical retail introduces conditions that software teams cannot abstract away. Stores have walls, metal shelving, refrigeration, back rooms, loading areas and legacy systems installed. They also have trading hours. Infrastructure changes that look simple on an architecture diagram can become expensive or disruptive when repeated across hundreds of sites.
Take inventory, for example. A forecasting model can improve a replenishment decision, but store-level execution may depend on shelf sensors, handheld devices, picking tools and accurate product information all exchanging data.
Computer vision, digital displays and environmental monitoring add further endpoints and a wider mix of data, from routine device updates and sensor readings to visual content and real-time operational data. These workloads place very different demands on the network.
Some require high bandwidth or low latency. Others involve large populations of battery-powered devices sending regular updates across difficult environments. Individually, none is necessarily a network problem. Collectively, they can become one.
If a pricing decision cannot reach every store quickly, a replenishment signal arrives too late or each new application requires another costly infrastructure layer, the retailer loses part of the value the technology was supposed to create.
One response is to add another gateway, controller or application layer for the next project. That can work for a pilot. Repeated often enough, though, every new device gains a separate integration path, security owner, management console and failure mode.
In difficult store environments, adding project-specific infrastructure to overcome coverage gaps can also change deployment economics. The technology works, but operating it becomes progressively harder.
The scaling problem appears after the pilot works
Retail technology teams are used to asking whether a new application meets its business case. AI adds another question: what happens to the infrastructure if the application succeeds?
A pilot may involve one store and a limited number of endpoints. Scale changes the economics. Coverage has to work throughout the building, updates must remain reliable during normal trading and security teams need a consistent way to onboard and manage devices. Operations teams also need clear ownership when something fails.
Connectivity should therefore be treated as an architectural choice, not a specification buried inside each procurement. The objective is not to replace every existing network or create one for every application, but to match workloads to the wireless layer best suited to their requirements while keeping the estate manageable.
Current designs may be adequate in some stores. In large or difficult formats, or where suitable retail IoT endpoints are growing quickly, an additional purpose-built wireless layer may be justified. The decision should be deliberate rather than inherited one project at a time.
Different workloads need different wireless foundations
The future-ready store is unlikely to depend on one network doing everything. It will more likely use a combination of connectivity layers, each supporting the workloads it suits best. Now, retailers should consider whether enterprise Wi-Fi should remain the de facto transport layer for every new category of connected device.
As electronic shelf label estates expand and retailers introduce more sensors and retail IoT applications, suitable workloads may be better served by a purpose-built wireless layer alongside the enterprise network.
This should not mean an isolated network for each use case, but a scalable, reusable connectivity layer able to support large device populations over time. Security belongs in the same conversation. Every endpoint needs consistent management, from trusted onboarding and credentials through to decommissioning.
Open standards remain important because they provide a more consistent foundation for future devices, applications and platforms. But openness alone is not the architecture. The decision is which workloads belong on which connectivity layer; standards help those layers remain manageable and adaptable.
The most successful retail AI programs will be judged by more than the quality of their predictions. As connected estates grow the test will be whether organizations can turn those predictions into dependable action across real stores.
Before approving the next AI pilot, retailers should ask what will their estate need if it succeeds, which workloads will grow, and should they all continue to be carried in the same way?
If success means more devices, richer data, more frequent updates or deployment across a much larger estate, the wireless architecture supporting those workloads has to be a central part of the AI strategy.
Retailers are not investing in AI tools or connectivity for their own sake. They want to protect margin, improve availability and make stores easier to operate. AI capability and AI scalability are different problems.
The objective should be an architecture in which new workloads can use the wireless layer best suited to them without forcing the estate to be rebuilt each time. The retailers that design for both will be in a much stronger position to turn ambition into reliable, repeatable execution.
We’ve featured the best AI chatbot for business.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit
