Why Did Starbucks Give Up on Computer Vision Inventory Management?
language
eng
date
Jun 9, 2026
slug
starbucks-computer-vision-inventory-failure
author
status
Public
tags
Computer Vision
Machine Vision
Inventory Management
Edge AI
Retail Tech
Starbucks
AI Failure Case
summary
Starbucks pulled NomadGo’s computer-vision-based inventory management system just nine months after rollout. This case highlights the limits of marketing claims like 99% accuracy, the constraints of FOV and on-device AI, and the difficulty of simply adding vision technology onto existing store operations.
type
Post
thumbnail
category
Computer VIsion
updatedAt
Jun 9, 2026 08:31 PM
While browsing computer vision news, I came across an unexpected name: Starbucks. A coffee chain may not seem closely related to computer vision, but Starbucks had actually introduced a computer-vision-based inventory management system for store inventory—and then removed it just nine months later.
As part of Starbucks CEO Brian Niccol’s “Back to Starbucks” strategy, the company rolled out a computer vision inventory management system in September 2025 to reduce stockouts of key ingredients such as milk and to improve supply-chain visibility. Nine months later, it decided to pull the system. I am not sure whether it is appropriate to name the vendor in a failure case—in Korea, at least, it feels like that might be sensitive—but the system was provided by NomadGo (https://www.nomad-go.com/). NomadGo’s approach uses a tablet equipped with a camera and LiDAR to capture images and recognize product inventory with on-device AI. The company promotes the system as being able to work even when a store goes offline, while also providing fast recognition and processing speed.
The core message repeated across many NomadGo pages is “8–10x faster work” and “99% accuracy.” At first glance, those numbers may not seem problematic, but there are several traps hidden inside them. Are both the speed and the accuracy really good enough? From an accuracy perspective, 99% is a very vague number in computer vision or industrial machine vision. If it means missing one item out of every 100 scanned, or producing one failure or wrong result every 100 runs, the error rate is far too high. Even barcode or OCR recognition typically needs to claim accuracy in the 99.xxx% range. NomadGo may have had no choice but to market a more realistic-looking number. Accuracy could have been affected by issues such as:
- The precision limits of the camera and LiDAR
- Issues in the 3D spatial vision algorithm
- Model constraints or insufficient precision caused by using on-device AI
In a store, this level of accuracy is directly tied to inventory data, so errors immediately become costs. Misrecognition or missed recognition can lead to over-ordering, under-ordering, or store-operation disruptions—and those consequences are not always easy to detect from the numbers alone.
The speed claim also reveals its limits in real-world use. Starbucks seems to have aimed to improve labor efficiency by using AI to handle repetitive work. Since the system was deployed at large scale across more than 11,000 company-operated stores in North America, it must have looked impressive in the initial demo. Starbucks also released a partially public YouTube video showing how NomadGo’s product was being used, although the video may disappear at some point.
Looking at the actual usage video, the camera’s FOV (Field of View) appears quite narrow. In practice, an employee has to point the tablet camera at items almost one by one, at a speed that is not far from manually counting them. That makes me question whether the productivity improvement really reaches the 8–10x speedup claimed by NomadGo.
According to the original article linked above, Starbucks found that the computer-vision-based inventory system frequently made errors in real operations, such as confusing similar milk products or failing to recognize some items. The attempt itself was interesting, but the result was ultimately a failure for real store operations. Was the technology the problem, or was the target market simply a poor fit? For example, in a shoe store, a 1% inventory error might not create a large cost—though that is not guaranteed either. But in stores that handle fresh products with short shelf lives, such as dairy, errors are directly tied to cost. Personally, I think attempts to improve an existing operating system merely by attaching camera sensors and adding a computer vision layer are extremely difficult and likely to produce disappointing results. The store layout, shooting path based on warehouse rack height, camera placement, and sensor position all need to be considered from the moment the store is designed. Beyond the hardware layer, the inventory management system itself probably needs to be integrated under a broader plan to operate with an AI-based inventory management system from end to end. Only that kind of integrated strategy seems likely to survive in practice.
Related posts
Computer Vision in the World Cup: From Goal-Line Decisions to Semi-Automated Offside
Jul 6, 2026
World Cup officiating technology is moving beyond simple video review and turning stadiums into real-time data spaces. This post explains how goal-line technology, semi-automated offside, connected-ball data, and optical tracking systems help improve both fairness and the flow of the game.
Between InGaAs and Si: SWIR Sensor Supply-Chain Risks and the Practical Alternative of NIR CMOS
Jun 20, 2026
As supply-chain risks grow around InP and InGaAs-based materials, not every SWIR application necessarily requires an InGaAs sensor. Below 1,100 nm, silicon-based BSI CMOS can be a practical alternative in terms of cost and supply stability. This article looks at the technical boundary between InGaAs and silicon.
The Speed of Chinese Manufacturing Seen at Shanghai Vision China and Semicon China 2026
Jun 3, 2026
A recap of visiting Shanghai Vision China and Semicon China 2026. The shift toward edge devices by Chinese companies, the overwhelming scale of Semicon China, and the vertical integration of materials, components, and equipment made the speed—and the urgency—of Chinese manufacturing feel very real.