Starbucks pulls AI inventory tool after nine months of errors
Computer vision couldn't tell oat milk from whole milk cartons
Starbucks officially ended its 'Automated Counting' AI inventory program across all North American stores after only nine months. The system, built with computer vision firm NomadGo, was designed to use on-device 3D spatial intelligence, computer vision, augmented reality, and LiDAR sensors to provide real-time stock visibility. CEO Brian Niccol had hoped the tool would free up baristas from administrative work, but real-world deployment revealed fundamental flaws. The computer vision model struggled with spatial awareness: it overcounted items, overlooked stock, and mislabeled products. Most notably, it could not differentiate between whole, oat, and almond milk cartons. A promotional video even showed the app missing a syrup bottle entirely.
Workers complained the system required them to wave and angle tablets in specific ways to trigger sensors, making the process slower than manual inventory checks. Starbucks has since deleted blog posts praising the tool and returned to manual inventory counts. Despite this failure, the company remains committed to its 'Back to Starbucks' transformation plan, which includes a Smart Queue system to prioritize orders across in-store, mobile, and drive-through channels. Financially, Starbucks reported a 9% revenue increase in Q2 to $9.5 billion and 7.1% rise in comparable store sales, suggesting the AI misstep hasn't hurt overall performance.
- Starbucks and NomadGo's AI inventory system failed to differentiate milk types (whole, oat, almond) and missed syrup bottles in tests
- Workers had to physically manipulate tablets to trigger sensors, making the process slower than manual entry
- Company reverted to manual checks but continues other tech like Smart Queue for order prioritization
Why It Matters
Highlights the gap between AI hype and real-world reliability in complex physical environments like retail stockrooms.