Starbucks has quietly retired its AI-powered inventory management system after just nine months in the wild, marking a rare public failure for enterprise AI deployment at scale. The NomadGo-provided tool, which was supposed to automatically track beverage components like milk and syrups across Starbucks locations, officially bit the dust following reports from Reuters in February—and the coffee chain confirmed the discontinuation to Fortune this week. This isn't some experimental pilot either; Starbucks announced the system with fanfare last September as part of its broader operational overhaul under CEO Brian Niccol's turnaround strategy.

The Accuracy Problem Nobody Wanted to Talk About

Sources close to the matter told Reuters that the NomadGo app consistently miscounted or mislabeled inventory items, regularly failing to identify bottles sitting right there on store shelves. We're not talking about edge cases here—this was a systemic issue affecting daily operations across multiple locations. When your AI system can't correctly count milk cartons in a refrigerator, you have a fundamental problem with the core value proposition. The tool's inability to accurately detect presence of products meant stores were constantly working with bad data, creating downstream chaos in the supply chain ordering process. Carl Addison, a Starbucks shift supervisor based in Shoreline, Washington who's spent nine years with the company, gave Fortune an earful about what this looked like on the ground. "The automated counting app required stores to rearrange back-of-house storage, which was a time-intensive process," he explained—and that's before you even get to the accuracy issues that made employees' workflows actively harder. Addison's assessment was brutal: "It started off not particularly accurate and got less accurate over time." That's not a bug you patch; that's foundational model failure at inference.

When Over-Counting Hurts as Bad as Under-Counting

Here's where it gets interesting for anyone building AI systems that touch physical inventory. The NomadGo tool's failures created a bidirectional problem: if the system counted too much of a product, it wouldn't trigger enough orders for items the store was actually running low on. Conversely, under-counting meant over-ordering and potential waste. Neither direction is acceptable when you're trying to optimize a supply chain that spans thousands of locations. Starbucks partners (employees) were essentially working around an AI system that was supposed to work for them—a classic case of technology creating friction instead of eliminating it.

Inside the 'Back to Starbucks' AI Playbook

Despite this setback, Starbucks hasn't exactly gone cold turkey on artificial intelligence under Niccol. The company's "back to Starbucks" turnaround plan includes other AI deployments like Green Dot Assist—an app running on store iPads that serves up recipe cards and ingredient substitution suggestions while also troubleshooting equipment issues. There's also Smart Queue, a tool designed to sequence orders more efficiently. These initiatives suggest Starbucks is still very much in the AI game, just perhaps with a clearer sense of where computer vision meets reality.

What Baristas Actually Said

Starbucks did provide Fortune with some positive barista feedback about the automated counting tool, including one comment stating it "improved inventory processes and the interface to view inventories." But not everyone was convinced. One particularly pointed response cut through the corporate PR speak: "Thanks for discontinuing Automatic Counting! The thought behind it was great, but the execution was proving difficult." That's the kind of honest feedback that should make any AI vendor sweat during their next enterprise pitch.

Key Takeaways

  • NomadGo's computer vision system couldn't reliably detect inventory items on shelves across Starbucks locations
  • The tool degraded in accuracy over time rather than improving—a red flag for production ML systems
  • Supply chain downstream effects meant both over-ordering and under-ordering problems depending on the direction of miscount
  • Employees had to rearrange back-of-house storage to accommodate a system that was supposed to work for them
  • Starbucks' broader AI strategy (Green Dot Assist, Smart Queue) remains intact despite this failure