A fridge door reflection doubled the milk count. A rubbish bin got logged as syrup. Somewhere between the pitch deck and the stockroom floor, Starbucks' shiny new AI inventory tool ran headfirst into reality and lost.
Nine months after rolling it out nationwide, the coffee giant quietly pulled the plug. The gap between what was promised and what actually happened in stores tells you everything about where this went wrong.
The Ten-Minute Promise
The tool, called Automated Counting, was built with startup NomadGo and used iPad Pros running computer vision to scan backroom shelves. It went live across all 11,300 company-operated stores in North America on a full national bet.
In controlled testing, the system reportedly hit 99% accuracy, an easy sell to a company that size. The promise: cut an hour-long manual chore down to ten minutes and free up staff to actually talk to customers.
Where 99% Accuracy Meets a Glare on a Steel Door
The situation in the stockroom was nothing like that of a laboratory. Baristas reported that the shiny fridge doors were confusing the cameras, causing them to double-count milk. Additionally, trash cans and syrup bottles were frequently mixed up.
Patchy in-store Wi-Fi made things worse, wiping scans mid-count and forcing staff to restart multiple times a shift. Seasonal cups and limited-time packaging created another hurdle, requiring up to six weeks of retraining every time Starbucks changed its lineup—which, for a chain reliant on constant limited editions, was a continuous bottleneck.
Why Legacy Infrastructure Stays
Much of the trouble traced back to Starbucks' backend, which runs on a legacy IBM AS/400 system dating to the 1990s. Wiring cutting-edge computer vision into three-decade-old architecture is like plugging a Tesla charger into a rotary phone switchboard.
The cameras were never just competing with milk cartons or trash bins; they were up against legacy infrastructure nobody had budgeted to replace. In the end, the old infrastructure won.
The Sudden End of a Startup Contract
The end came without warning for NomadGo. The startup was blindsided when Starbucks called off the partnership with no prior signal of a strategy shift. Within days, NomadGo laid off a large portion of its small team, including the engineers who built the Starbucks integration.
Six weeks later, baristas were told to strip tracking codes off the shelves and pick the clipboards back up.
Why Starbucks' AI Roadmap Remains Fragile
Starbucks framed the pivot around familiar corporate logic: technology should enhance human connection, and abandoning an underperforming tool is simply good management.
That explanation misses the core problem. Automated Counting didn't fail because computer vision is flawed but because Starbucks tried syncing 2020s spatial AI with 1990s IBM plumbing across thousands of chaotic, low-bandwidth backrooms.
Yet the coffee giant isn't slowing its automation push. Development continues on an AI ordering assistant and a ChatGPT tool designed to suggest drinks based on a customer's mood or outfit.
Deploying flashy consumer-facing AI while backrooms revert to clipboards addresses symptoms rather than underlying technical debt. Until the legacy foundation is rebuilt, every new AI layer risks the exact same fate: impressive in the boardroom, but useless on the shop floor.
