Starbucks has withdrawn its AI-powered “Automated Counting” inventory system from stores across North America, roughly nine months after rolling it out, following persistent errors in counting milk and other stock, Reuters reported on May 22, 2026.
The system was built with NomadGo, a Seattle-area startup, and used tablets fitted with LiDAR depth sensors and computer vision to scan shelves and back rooms and tally what was on hand. The goal was to replace manual stock counts, reduce the shortages that leave customers unable to order items on the menu, and free up baristas’ time.
Why the tool was pulled
According to the reporting, the technology struggled with a task that sounds simple but is hard for a camera: telling apart similar-looking products such as different types of milk, and producing accurate counts in cluttered store environments. Mislabeled items and inaccurate tallies disrupted store operations rather than smoothing them, and Starbucks has told stores to return to manual inventory counts as the standard procedure.
The project had been treated as part of chief executive Brian Niccol’s turnaround plan, which emphasizes faster service and better availability of core items. Starbucks said it would instead focus on improving daily replenishment so that items listed on the menu are actually in stock when customers order them.
A setback for the supplier
GeekWire reported that the decision caught NomadGo off guard. Starbucks had been the startup’s most prominent customer, and the deployment across thousands of stores had been held up as a showcase for computer vision in retail operations.
The episode illustrates a broader pattern in enterprise AI: pilots that look promising in controlled settings can fail when exposed to the variety and mess of real stores. For now, counting cartons of oat milk at Starbucks remains a job for people.
