How a Retailer Used Computer Vision to Reduce Shelf-Stock Errors

The shelf problem was costing more than lost sales

A mid-sized British retailer with 180 supermarkets had an inventory problem that conventional audits could not explain. Its central systems showed average availability above 96 per cent, yet weekly store walks found that roughly one in 12 promoted products was missing from the shelf during trading hours. In many cases the stock was technically present: cases remained in the back room, products had been placed in the wrong bay, or a shelf label showed a line that had been discontinued. Customers saw an empty space; the replenishment system saw no reason to act.

The mismatch carried a measurable cost. Internal analysis estimated that shelf-stock errors were suppressing sales by £18 million a year, with the sharpest losses in chilled food, soft drinks and household essentials. Manual gap scans helped, but they were inconsistent and labour-intensive. A full scan of a large store took two employees nearly three hours, and results depended on when the walk happened. Checks conducted before the evening rush routinely missed the gaps that appeared afterwards. The retailer wanted more frequent evidence without asking staff to spend their shifts photographing shelves.

Management set a deliberately narrow objective: identify empty facings, misplaced products and incorrect shelf labels quickly enough for store teams to intervene. The project was not framed as autonomous retail, customer analytics or loss prevention. That boundary mattered. It kept the technical design focused, gave employees a comprehensible use case and reduced the risk that cameras installed for stock accuracy would be perceived as a covert surveillance network.

A limited pilot established what vision could see

The first deployment covered six stores and 14 high-velocity aisles in each location. Fixed edge cameras captured shelf images at scheduled intervals, generally every 20 minutes, with additional captures after known replenishment windows. The system divided each bay into digital shelf positions and compared visible products, gaps and labels with the store’s planogram. Rather than attempt to recognise every item in the estate, the pilot concentrated on 3,200 stock-keeping units whose packaging and shelf locations were relatively stable.

The computer-vision models performed three tasks. Object detection located products and shelf edges; classification estimated which product occupied each facing; and optical character recognition read price and promotional labels. Processing occurred on a small server in each store, so most raw video did not leave the premises. The system retained event images showing a suspected problem, not continuous footage. Faces and bodies appearing in those frames were blurred before upload to the central review platform.

Early results exposed the limits of a laboratory benchmark. The model achieved 94 per cent accuracy on clear, front-facing packaged goods, but fell below 80 per cent on reflective pouches, products stacked behind cardboard trays and seasonal packaging. A red pasta sauce label could be confused with another flavour; a dark gap might actually be a deep shelf holding stock at the back. These errors were useful. They showed that the operational challenge was not simply recognition, but deciding when the evidence was strong enough to interrupt a colleague.

Human review turned alerts into usable work

The retailer resisted the temptation to send every model output directly to handheld devices. During the pilot, suspected exceptions entered a review queue staffed by an existing offshore retail-support team. Reviewers saw the current image, the expected planogram position and the previous two captures. They could confirm an empty facing, mark a likely misplacement, reject the alert or escalate an ambiguous case. High-confidence gaps on fast-selling products were routed automatically, while uncertain classifications required review.

This hybrid design changed the economics of accuracy. The raw model initially generated about 420 alerts per store each day, far more than a team could reasonably action. Confidence thresholds, temporal checks and human review reduced that figure to around 55 tasks, of which 44 were typically valid. The system also waited for the same condition to appear in two consecutive captures unless the item was designated critical. That rule delayed some interventions by 20 minutes, but eliminated many alerts caused by customers briefly blocking a shelf or moving products while choosing.

Store employees received tasks phrased as operational instructions rather than algorithmic judgements: “Check four empty facings in soft drinks, bay 7” or “Product may be in the wrong position beside the £2.50 label.” Staff could close a task as replenished, no stock available, planogram incorrect or false alert. Those responses became labelled data for retraining and, just as importantly, revealed process failures. In one store, repeated alerts for bottled water traced back to an overnight delivery being parked in the wrong back-room zone, not poor forecasting.

Inventory integration made the images actionable

Computer vision alone could identify an apparent gap, but it could not explain whether stock existed or what action made sense. The retailer therefore integrated confirmed shelf events with perpetual inventory, delivery schedules, sales velocity and replenishment records. If the shelf was empty and the system showed 24 units on hand, the task was prioritised as a likely back-room or counting error. If both shelf and inventory were empty, the event went to the ordering workflow rather than a floor colleague.

The combined data produced a more useful measure than nominal availability: shelf availability by trading hour. A store might report 97 per cent inventory availability across a day while falling to 90 per cent on key lines between 5pm and 7pm. Managers could now see that distinction and adjust replenishment rounds. In the pilot, two stores moved their main ambient top-up from 3pm to 4.30pm after images showed predictable gaps emerging during the commuter peak. Sales of the affected lines rose 3.1 per cent without changing order quantities.

Integration also uncovered bad master data. About 9 per cent of persistent vision exceptions were caused by planograms that did not match the physical store, often after a local manager expanded a popular range or a promotional display ended early. The retailer created a weekly reconciliation process rather than forcing stores to conform blindly. This prevented the system from becoming an automated source of nuisance tasks and improved the planogram database used by merchandising, procurement and store operations.

Privacy was designed into the operating model

Cameras in shops create a legitimate concern: infrastructure introduced for shelf management can easily be repurposed to monitor customers or employees. The retailer addressed that risk before the pilot went live. Camera angles were restricted to shelving wherever possible, audio capture was disabled, and edge software discarded footage that did not contain a stock event. Retained images were deleted after 30 days unless attached to an active technical investigation. The company documented these controls in a data protection impact assessment.

It also imposed a purpose limitation that was organisational, not merely technical. The stock system could not perform facial recognition, estimate age or gender, track individuals across aisles, or feed disciplinary and loss-prevention processes. Access logs were reviewed by the privacy team, and any proposed change required approval from a governance panel including operations, technology, legal and an employee representative. Clear signs explained that cameras were being used to monitor shelf availability, while an internal briefing showed staff exactly what images and task data managers could see.

The most important decision was to consult employees before installation. Colleagues worried less about customers being identified than about productivity scoring: would the system measure how quickly a named person filled a gap? The retailer responded by assigning tasks to departments, not individuals, and reporting completion at store level. That concession reduced the granularity of performance management, but it helped secure adoption. It also reinforced a sound operational principle: a gap may result from supply, layout or system data, not from the conduct of the nearest employee.

The rollout improved accuracy, but not uniformly

After 16 weeks, the six pilot stores recorded a 28 per cent reduction in verified shelf gaps across monitored aisles. The proportion of alerts resolved within 45 minutes rose from 41 per cent in the first fortnight to 73 per cent by the end of the trial. On-shelf availability improved by 2.4 percentage points, while sales in monitored categories increased by 1.7 per cent against matched control stores. The business case projected an 18-month payback if similar gains held across 120 larger locations.

The results varied considerably. Stores with disciplined back-room routines captured more value because employees could find stock quickly. In cramped urban branches, an alert often confirmed a problem that staff had no inventory or space to solve. Fresh produce remained difficult: loose items, variable colours and frequent customer handling produced more classification errors than boxed grocery. The retailer excluded produce from the next rollout and added chilled ready meals, where packaging and planograms were more consistent.

There were also real costs. Each store required cameras, an edge server, installation outside opening hours and network support. Model maintenance increased when suppliers redesigned packaging, while human review remained necessary for roughly 35 per cent of accepted events. The retailer judged those costs against avoided manual scans and recovered sales, but did not claim that vision removed labour. It shifted labour from repetitive inspection towards replenishment and exception handling, while creating new work in review, data quality and system governance.

The durable lesson is disciplined automation

The programme succeeded because the retailer treated computer vision as one component of an operating system, not as an all-seeing replacement for store judgement. Cameras supplied timely evidence; human reviewers controlled ambiguity; inventory systems established context; and employees completed or challenged the task. Each layer corrected a weakness in the others. The model could see a gap but not its cause, the inventory file could show stock but not its location, and a colleague could resolve the issue but could not continuously inspect every bay.

Scaling depended on resisting feature creep. Executives considered using the same cameras for queue measurement and suspected-theft alerts, but kept those proposals outside the stock programme and subject to separate assessment. That restraint protected trust and preserved clear technical requirements. It also prevented a commercially useful tool from inheriting the legal, ethical and accuracy risks of unrelated surveillance applications.

For other retailers, the practical sequence is clear: start with categories where packaging and layouts are stable, measure valid interventions rather than model accuracy alone, and integrate alerts with inventory before expanding coverage. Set retention, access and purpose rules before cameras are installed, then publish them in language employees and customers can understand. The most valuable outcome is not a perfect digital replica of every shelf. It is a reliable flow of evidence that helps stores correct the right problems quickly, without turning ordinary shopping and work into subjects of continuous personal monitoring.