Why Starbucks’ AI Inventory Tool Failed at Full Scale
A Fast Company investigation reconstructs how a national AI rollout collapsed, highlighting the challenges of implementing enterprise-level artificial intelligence.
The Story Unfolds
Starbucks told NomadGo—a 30-person startup behind the AI tool—on April 3rd that they were scrapping the inventory management system. Within days, the company laid off a significant portion of its technical team, including those responsible for managing the Starbucks account, as reported by GeekWire.
"There’s nothing you can do when leadership and strategy change," shared David Greschler, NomadGo’s CEO, expressing his surprise at the decision.
The Technical Glitches
The issues weren’t merely software-related; they were grounded in practical, everyday challenges:
- Scanning Issues: A shift supervisor near Seattle encountered a problem where the iPad camera captured a reflection, leading to incorrect counts of oat milk cartons.
- Typing Confusion: The AI app occasionally mistook different types of milk and syrups.
- Connectivity Problems: A store manager in Graham, Texas, faced an issue with patchy Wi-Fi that wiped her count partway through, leaving the store without a usable inventory count.
Despite promising 99% accuracy in controlled tests, Greschler attributes these failures to the dynamic nature of Starbucks’ inventory—seasonal cups and limited-time packaging required frequent retraining of the computer vision model.
The Backend Challenge
An older system, an IBM AS/400 from the 1990s, used by Starbucks for its backend operations, posed another significant challenge. This legacy system made it difficult to move real-time store data, hindering the AI tool’s effectiveness.
A Massive Investment with No Pilot
Estimated to have cost over $10 million over several years, the AI inventory tool was implemented in all 11,300 company-operated cafés across North America by September 2025. However, it was discontinued just six months later, on May 18th.
This widespread rollout differs from typical enterprise AI failures described as pilot projects that never scaled. Starbucks went all-in, and the outcome was costly.
The Bigger Picture
According to MIT’s NANDA initiative, 95% of enterprise generative AI pilots fail to deliver measurable profit impact, while the UK’s Office for National Statistics reveals that AI adoption often widens rather than deepens. Starbucks’ approach stood out in its rapid and comprehensive implementation, marking a significant lesson in the challenges of implementing AI at scale.