Reasons the Starbucks AI inventory system did not succeed at full scale.

      On April 3, Starbucks informed NomadGo that it was discontinuing the AI inventory tool that the Redmond-based startup had created for the company. Just days later, the 30-person firm significantly reduced its workforce, as reported by GeekWire.

      These layoffs included the technical team managing the Starbucks account, and Starbucks did not notify its baristas for an additional six weeks. “When leadership and strategy change, there’s nothing you can do,” said David Greschler, NomadGo's CEO, describing the decision as a total shock.

      This information emerged from a Fast Company investigation published on Monday, which included insights from numerous Starbucks employees as well as the startup itself. We initially reported on the tool's cancellation in May, but have lacked the vendor's perspective until now. The baristas became aware of the situation on May 18 when they received a memo instructing them to remove the QR tracking codes from the backroom shelves and revert to manual counting.

      The issues encountered were mostly straightforward and practical. A shift supervisor near Seattle pointed an iPad at a steel fridge, which caused the camera to capture a reflection, miscounting five cartons of oat milk as ten. In other instances, the app misidentified milk types, mixed up syrups, and inaccurately counted a bin as food.

      Conversely, a store manager in Graham, Texas, faced a different challenge: inconsistent Wi-Fi disrupted her inventory count midway, while her managers deemed manual counts unacceptable. Consequently, the store was left without a usable count.

      The issue was not the model itself. NomadGo's computer vision achieved 99% accuracy in controlled tests, with a launch promise of counts up to eight times faster than manual methods. Greschler offered a more intriguing explanation than mere accuracy failure, noting that computer vision has difficulty adapting to continually changing inventory. He mentioned that seasonal cups and limited-time packaging could require six weeks of retraining, and that developers often discovered new items only after they appeared on the shelves.

      Another longstanding constraint was Starbucks' backend system, which operates on a legacy IBM AS/400 platform from the 1990s, making the transfer of real-time data from stores challenging. Insiders estimated that the program might have cost over $10 million over several years.

      This was not merely a trial run. The Starbucks AI inventory tool, named Automated Counting, was implemented across all 11,300 company-operated cafés in North America by the end of September 2025, but it was discontinued by May 18. Failures of enterprise AI are typically labeled as pilots that never fully scaled; however, this instance scaled first, which resulted in substantial costs when it ultimately failed.

      The MIT NANDA initiative discovered that 95% of enterprise generative AI pilots failed to produce any measurable profit impact, and most enterprise AI expenditures remain confined to labs. Similar trends were noted by the UK’s Office for National Statistics, which indicated that while AI adoption is expanding, it is not becoming more entrenched within organizations. In contrast, Starbucks opted for comprehensive implementation all at once.

      Before deciding to terminate the tool, Starbucks publicly defended it. After Reuters reported inaccuracies in February, the company claimed that the tool had enhanced product availability, according to CNBC. Roughly eight weeks later, it instructed NomadGo to cease production.

      Greschler attributed the decision to a change in leadership and strategy, though he did not specify which change affected the decision. The timeline is documented: Deb Hall Lefevre, then Starbucks' chief technology officer, lauded the tool’s rollout in NomadGo’s launch announcement on September 3, 2025, yet resigned on September 29, the same month the implementation concluded. Starbucks subsequently hired a permanent replacement from Amazon in December. The tool managed to outlast its executive sponsor by about six months.

      As for Starbucks' current stance, a company spokesperson stated, “We use technology to support human connection, not to replace it,” while highlighting a $500 million investment in increasing staff numbers at its coffeehouses. “When it fell short, we listened to feedback and changed course.”

      This deserves careful consideration; discontinuing an implemented tool across 11,300 locations is significantly more challenging than allowing a pilot program to fade away. Many companies continue to pay for software their staff no longer trust. Starbucks has maintained its other AI projects, including Green Dot Assist for baristas and a ChatGPT integration for customers.

      What remains unaddressed in the statement is the impact on the supplier. Tools that create work instead of reducing it are a known failure pattern, and researchers have labeled it accordingly. Additionally, there is less discussion surrounding who bears the financial burden when a major client decides to alter course. The Starbucks AI inventory tool's failure affected all 11,300 sites, and the cost was ultimately borne by a company of just 30 employees.