A 99% validated computer vision PoC in sandbox can fail in production at the scale of 11,000 stores: Starbucks' withdrawal of the 'Automated Counting' tool at the end of May 2026 demonstrates this. The chain removed the 'Automated Counting' tool developed by NomadGo during the week of May 21, 2026, after nine months of deployment across more than 11,000 North American stores to inventory milk and syrups. They returned to full manual counting. The solution was designed by NomadGo (Redmond, Washington), led by David Greschler. According to a BusinessWire statement on September 3, 2025, the developer claimed a speed up to eight times faster than manual counting and an accuracy of 99%, self-reported figures with no independent validation known to date. The withdrawal is documented by the analysis 'AI Autopsy' by Bigeye published on May 23, 2026.
From a claimed 99% accuracy to a withdrawal in nine months - across 11,000 stores.
Figure from the official NomadGo statement (Sept. 2025): an accuracy measured in controlled conditions, not validated independently before large-scale deployment.
A Key Tool in Niccol's Recovery Plan, Following a Previous Amazon Example
The 'Automated Counting' was part of the operational recovery program led by Brian Niccol, who became CEO of Starbucks in September 2024, aimed at reducing stockouts on high-turnover items. The launch statement of September 3, 2025, recalled that the technology had been tested for 'several years' in sandbox before its general deployment, a pilot duration that did not prevent failure in production. The scenario mirrors the trajectory of Just Walk Out at Amazon Fresh, which Ars Technica reported in April 2024 that the technology was not as autonomous as announced and relied on a workforce of about a thousand people in India to watch video recordings to validate transactions (free translation) - a revelation published the same year Amazon stopped the device. ActuIA documented as early as 2023 the penetration of AI in retail, a field where the gaps between claimed accuracy and field performance become observable at scale. For any organization undergoing a computer vision PoC in retail, the case raises a direct operational question: do the sandbox validation conditions replicate the variables that led to failure in deployment at 11,000 stores?
A Mobile Architecture and a Known Failure Class Beyond 1,000 References
The NomadGo solution combined computer vision, LiDAR sensors, three-dimensional spatial processing, and augmented reality overlay, all operated from a smartphone or tablet assigned to store personnel, known as a mobile scan architecture. Reported failures included confusion between types of milk and counting errors on visually dense items. According to the TechnoLynx analysis published in April 2026, which describes the phenomenon as a composite failure class, computer vision models that pass precision tests on 500 references fail in production beyond 1,000, not due to a single cause but across four simultaneous failure axes (free translation). The thesis suggests that accuracy validated in a controlled environment (limited reference range, stable lighting, standardized layout) degrades by accumulation when the real environment imposes proliferation of references, lighting variations, and non-standardized packaging. The Bigeye analysis of May 23, 2026 summarizes the documented gap in its title, tracing how Starbucks' AI inventory tool went from a 99% claimed accuracy to withdrawal in nine months (free translation).
The 1,000 SKU Threshold: A Failure Class, Not an Isolated Bug
According to TechnoLynx (April 2026), computer vision models validated on less than 500 references fail in production beyond 1,000 product classes - not due to a single cause, but under the effect of four simultaneous failure axes: SKU proliferation, lighting variability, heterogeneous layouts, non-standardized packaging. This composite failure mechanism is documented independently of the Starbucks case.
Focal Systems at Morrisons: A Large-Scale Retail CV Deployment on Fixed Camera Architecture
In the same market, the publisher Focal Systems claims at the British retailer Morrisons a deployment presented, in its corporate statement, as one of the largest in computer vision in large-scale food retail. According to this same statement, the deployment delivered 'more than 2% improvement in product availability in stores across the network, and up to 4% in the best-performing stores' (free translation) - a self-reported figure by the publisher, without third-party replication published. The project received the 'Digital Transformation Project of the Year 2024' award. The architecture mobilized by Focal Systems relies on fixed cameras in the aisles, where NomadGo operated via mobile devices handed to staff: the architectural variable constitutes an observable structural difference between the two systems. Before any large-scale deployment of a computer vision system in retail, three variables need to be tested in real conditions: the validated reference catalog (sandbox versus production, with specific measurement beyond the 1,000 SKU threshold), the capture architecture (mobile scan assigned to staff or fixed cameras in the aisle), and accuracy measurement documented by an independent third party. NomadGo published none of the three results before its rollback to 11,000 stores.