Starbucks pulls the plug on its inventory-counting AI: the real lesson isn't the milk mix-up
99% accuracy promised, 8x faster than human counts. Nine months later, Starbucks retires its NomadGo AI. The real story isn't the algorithm.
99% accuracy promised by the vendor, 8x faster than baristas. Nine months after launch, Starbucks pulled the AI that counted inventory across more than 11,000 stores in North America. The internal memo landed on May 21, 2026, in tidy corporate prose: "Automated Counting will be retired."
The headline version going around: the AI confused oat milk with almond milk, end of story. True, short, and missing the point.
The Niccol bet and the precision threshold
Brian Niccol took over in September 2024. He arrived with a turnaround mandate at a chain where fewer than a third of trucks were delivering complete, on-time orders at 95%. The "Automated Counting" tool built by Seattle startup NomadGo was supposed to fix that: LiDAR-equipped tablet, shelf scan, automatic counts of syrup bottles and milk cartons, deliveries triggered by real stock levels.
Niccol greenlit the national rollout. By fall 2025, more than 11,000 locations were running it.
The trouble lives in the number. 99% in a lab demo doesn't replay on the floor. Shelves are messy, packaging changes, bottles hide behind each other. Oat milk and almond milk come in the same carton format with similar color codes. Two seconds for the human eye, a nightmare for the vision model.
The field confirmed it. Carl Addison, a Shoreline (Washington) shift supervisor with nine years on the job, told Fortune: "It started off not particularly accurate and got less accurate over time." Starbucks' own launch promo carried the tell: the video shows the system counting a row of syrup bottles and skipping a peppermint bottle parked in the middle.
At 95% accuracy in a store, the barista has to recount every shelf to verify what the machine saw. The work gets done twice.
The cost no one wrote into the business case
The tool also required a back-office reorg to work properly: bottles arranged in a specific order, corners cleared, references spaced out. The machine dictates the layout. To save a few minutes a day, you hand baristas a physical protocol, even though their in-store time was exactly the variable Niccol was trying to protect.
Because next to all this, Niccol is running a plan called "Back to Starbucks": ceramic mugs for sit-down customers, handwritten notes on cups, smoother lines through more staffing. The story is the barista returning to customer relationships. The NomadGo AI took human time to replace it badly. The retirement looks less like an isolated failure and more like a course correction.
Comparable sales: +6.2% in the most recent quarter, across all regions including China. The turnaround is holding. Without this AI.
The operational threshold, the real variable
Starbucks joins a growing list: McDonald's scrapped its IBM voice-ordering drive-thru solution in 2024, Google had to disable parts of AI Overview after it suggested putting glue in pizza. Same pattern every time: a tool with very respectable average precision that whiffs in the specific case where the customer is right there and the cost of a bad output is immediate.
An AI's operational threshold is never its average precision. It's its precision in the worst relevant case. Below that, it doesn't remove human work. It shifts it, and changes who pays the cost.
