Starbucks terminated its AI-powered inventory counting system across all North American stores this week, nine months after deploying it as a centerpiece of CEO Brian Niccol's "Back to Starbucks" turnaround — the most prominent enterprise AI rollback in retail so far in 2026. An internal newsletter reviewed by Reuters and confirmed by two employees stated plainly on Monday, May 19: "Starting today, Automated Counting will be retired. Beverage components and milk will now be counted the same way you count other inventory categories in your coffeehouse."
The decision reversed a deployment that had spanned more than 11,000 company-operated locations across the United States and Canada. It also exposed a gap that enterprise buyers of physical-world AI tools cannot afford to miss: a system that requires workers to verify every output delivers no net efficiency gain — it doubles the task.
NomadGo's 99% Accuracy Claim Did Not Survive the Store Floor
Developed by Redmond, Washington-based startup NomadGo, the Automated Counting system used LiDAR sensors and tablet cameras to tally syrups, milk varieties, and other beverage components stored on store shelves. Employees would scan a shelving unit with a tablet; the AI would generate the count and overlay it using augmented reality.
At launch in September 2025, NomadGo claimed its system counted inventory up to eight times faster than manual methods, with 99% accuracy. Starbucks CTO Deb Hall Lefevre described the technology in the official announcement, saying it would allow workers to "spend more time focusing on what matters: crafting high-quality beverages and connecting with customers." That announcement has since been deleted from the Starbucks website.
The floor reality was different. Reuters reported in February 2026 that employees and managers across multiple locations described the system frequently miscounting and mislabeling items — confusing similar milk types, missing items entirely during scan sessions, and in at least one case failing to recognize a peppermint syrup bottle in a promotional video Starbucks itself had uploaded to showcase the tool. By the time Monday's newsletter arrived, that video and the original September 2025 blog post had both been removed from the company's site.
Double-Counting Became the New Standard
The operational loop the errors created is the clearest measure of the deployment's failure. When a scan produced an inaccurate count, workers had to verify the output manually and re-enter corrections. That is not a time saving — it is a time tax. Baristas were conducting two inventory cycles where one had been designed to eliminate the other.
This pattern is one of the more common failure modes for computer vision tools deployed in uncontrolled environments. Datature's 2026 Enterprise Vision AI Adoption Report identifies distribution shift — the gap between training data and production conditions — as the second most common cause of computer vision deployment failure, noting that "a retail shelf recognition model trained in one store chain fails when deployed to a chain with different shelving heights." Lighting changes, shelf reorganizations, and repositioned products occur routinely in a working café. Once the error rate rises above the threshold at which workers trust the output, the tool has lost its only value: accuracy they don't have to re-verify.
In February, Starbucks Said It Was Working. In May, It Was Gone.
The timeline of Starbucks' own statements is significant for enterprise buyers. In February 2026, when Reuters first reported worker complaints about miscounts, Starbucks told Reuters the tool had improved product availability in its stores. Three months later, the same company issued a memo retiring it and instructed staff to return immediately to manual counting.
Starbucks framed the May 2026 decision as a choice, not a concession — telling Reuters it was "a decision to standardize how inventory is counted across coffeehouses as we continue to focus on consistency and execution at scale." NomadGo said it is "continuously learning from customer and user feedback" as it works to improve its products. Neither statement acknowledged the accuracy failures Reuters had documented five months earlier.
The financial backdrop adds context. Starbucks posted its strongest quarterly same-store sales growth in two and a half years in its most recent report, and the stock is up 24% in 2026. But North American operating margins have fallen to 9.9%, down from 18% two years earlier before Niccol took the helm. Morningstar analysts, according to Reuters reporting, had written as recently as April 2026 that AI inventory initiatives could support long-term improvements in restaurant-level margins by reducing labor hours and waste. That assessment now requires revision.
Starbucks Still Spending on AI — Just Not This Kind
The retirement of Automated Counting does not signal a full technology retreat. Niccol is rolling out Green Dot Assist, a generative AI chatbot built on Microsoft's Azure OpenAI platform, initially piloted at 35 locations in June 2025 and now in broader rollout across the United States and Canada as part of fiscal 2026 operations. The tool helps baristas look up recipes, troubleshoot equipment, and identify shift cover — tasks where an imperfect AI suggestion can be reviewed and overridden before it causes a downstream problem.
That distinction matters. A language-model assistant that suggests the wrong ingredient substitution can be corrected before the drink is made. An inventory count that is off by two units of oat milk feeds directly into restocking decisions, supply orders, and the product availability numbers that Niccol has cited as a key measure of the turnaround's progress. The tolerance for error is categorically different.
Niccol has also recruited logistics executives to address what current and former employees described to Reuters as a supply chain that is fragmented and hampered by outdated systems. Whether manual counting can solve what the AI could not remains an open operational question.
What Enterprise AI Buyers Can Take From This
The Starbucks case is now a reference point for any enterprise evaluating physical-world computer vision. Several dynamics that contributed to this failure are not unique to Starbucks.
Computer vision tools in operational settings require near-perfect accuracy before workers can trust them. Below that threshold, workers verify every output — and if workers are verifying every output, the software has not reduced labor. It has redistributed it while adding cognitive overhead. The efficiency case collapses entirely.
Retail and food-service floors are not controlled environments. Lighting shifts, shelves are reorganized, products are moved. Computer vision systems trained in more uniform conditions can degrade significantly when deployed at scale into high-variation production settings. Nine months is a short window to expect production-grade performance from a novel deployment — but it is also the window enterprise IT teams typically get before executive patience runs out.
NomadGo's claim of 99% accuracy and 8x speed was not independently verified before deployment across more than 11,000 stores. Volume of scans is not the same as accuracy of scans. AI that cannot work reliably in the conditions it was deployed into is not a productivity tool — it is another task for employees to manage.
Originally published on Tech Times
