Self-checkout has always had an exception problem.
The technology works efficiently when the transaction follows the happy path: item presented, barcode read, item paid for. Retailers lose the efficiency the moment something ambiguous happens. A product is not scanned. The wrong barcode is read. The bagging area disagrees. A colleague has to leave another task, inspect the transaction and decide whether the shopper made a mistake or something more deliberate occurred.
Morrisons is now trying to automate part of that exception loop.
The grocer has signed a long-term partnership with Everseen and plans to deploy the company’s Evercheck computer-vision system to 200 UK stores in the initial phase. The platform analyses self-checkout transactions in real time and identifies potential missed or mis-scanned items. Instead of immediately calling a colleague, the checkout can prompt the shopper to review the incident and correct it.
That workflow is the important part of the deployment.
According to descriptions from Everseen, Morrisons and coverage of the rollout, the system can show the customer a playback of the relevant moment and give them another opportunity to scan correctly before escalating to staff. The technology is transaction-focused; Everseen and Morrisons say it does not use facial recognition or profile the shopper.
This is a narrower use of computer vision than the industry’s more futuristic “intelligent store” pitches. It may also have a clearer return on investment.
Self-checkout creates a structural tension for grocery. Retailers deploy it to increase throughput and reduce the labour required per transaction, but the format can also introduce loss and a new class of customer-service interruption. An ECR Retail Loss study published in 2026, summarised by industry coverage of the Morrisons deployment, found materially higher loss at stores operating self-checkout than at stores without it. The exact causes vary—deliberate theft, accidental errors, product recognition issues and process design all matter—but the operating problem is real.
The crude response is to add more friction: more interventions, more verification and more staff watching screens. That undermines the original point of self-service.
Evercheck is an attempt to make the exception cheaper.
If the system is sufficiently accurate, a large share of minor errors can be resolved by the shopper without waiting for a colleague. Staff attention is then reserved for the smaller set of cases that remain unresolved. In systems terms, the AI is not replacing the checkout. It is prioritising human intervention.
That distinction should shape how the rollout is evaluated.
Detection rate alone is not the right metric. A system that identifies every ambiguous movement but generates constant false alerts could make the checkout experience worse. Morrisons should care about the share of prompts that shoppers resolve themselves, the change in staff interventions per thousand transactions, transaction time, customer satisfaction and actual loss reduction.
Privacy is another reason the architecture matters. Computer vision at checkout can easily trigger concerns about surveillance, especially when retailers talk loosely about “AI cameras.” Evercheck’s transaction-centric design draws a useful boundary: analyse the act of scanning rather than identify the person doing it. That does not remove every privacy question, but it is a substantially different proposition from facial recognition or persistent shopper tracking.
The deployment also fits Morrisons’ broader push to attach technology investment to operating outcomes. Chief executive Rami Baitiéh said earlier this summer that data and AI had materially contributed to a wider £940 million savings programme over three years. That figure should not be read as “AI saved £940 million”; Morrisons itself describes technology as one contributor inside a broader cost programme. But it shows the standard management is applying to new tools: they are expected to change the economics of the business.
Evercheck will now face that test at scale.
Two hundred stores is large enough to expose the awkward edge cases that disappear in pilots. Different layouts, lighting, checkout hardware, baskets, packaging and customer behaviour all complicate computer vision. The software also has to integrate with Morrisons’ existing checkout estate rather than assume a greenfield store.
If the rollout works, its success may look boring. Fewer unnecessary calls for help. Fewer genuine miss-scans. Shorter queues. More colleague time spent elsewhere.
That is a useful correction to the way retail AI is often discussed. The most valuable vision system in a supermarket may not be the one that recognises every shopper or maps every shelf.
It may be the one that knows when not to bother an employee.
The design choice is also a response to a broader credibility problem around self-checkout AI. Retailers have often framed computer vision as loss prevention, which immediately puts the shopper under suspicion. A system that first treats a mismatch as an error to be corrected creates a different social contract. The camera is still watching the transaction, but the interface assumes recoverability before escalation.
That can matter operationally because not every miss-scan is theft. Multi-buy confusion, awkward packaging, barcodes on fresh goods and simple distraction all create errors. If the system cannot distinguish intent—and it generally should not claim to—it can at least distinguish between an event that was corrected and one that remains unresolved.
The absence of facial recognition is therefore not just a privacy talking point. It keeps the technical scope aligned with the task. Evercheck needs to understand objects and transaction sequence; it does not need to know a shopper’s name. That narrower design can reduce both data risk and unnecessary model complexity.
Morrisons will still need to prove that the computer-vision layer does not become a new failure mode. Store lighting changes. Cameras are bumped. Packaging changes. Customers stack items or move them unpredictably. Software updates interact with legacy checkout systems. At 200 stores, support and observability become as important as model accuracy.
This is where the retailer’s operating model matters. A central dashboard may show detection rates, but store teams experience the system as a queue that either moves or does not. If an alert cannot be explained quickly to a customer, the store pays for the AI twice: once in technology cost and again in damaged service.
For Everseen, the Morrisons contract is also a chance to demonstrate that its global deployment claims translate into a large UK grocer with an existing estate. The company says its technology is live across more than 150,000 checkouts and 10,000 stores globally. Those are vendor-supplied figures, not independent performance measures. The Morrisons rollout should generate a more useful test: measured before-and-after outcomes in a named retailer.
The most revealing number would be the escalation rate. If the vision layer can resolve routine anomalies with the shopper and dramatically reduce colleague interventions while also lowering loss, the business case is strong. If it simply creates more prompts, the AI has automated suspicion rather than checkout.