Starbucks pulls AI inventory tool after real-world accuracy gap

Starbucks has quietly shut down Automated Counting, an AI-powered inventory system it had deployed to roughly 11,000 stores across North America just months earlier, after the tool failed to perform reliably on store floors.

The system, built by supply-chain vision startup NomadGo, used on-device computer vision and augmented reality so baristas could scan shelves and stockrooms with a tablet to generate real-time counts of ingredients and supplies. Starbucks rolled it out nationwide in September 2025, positioning it as a fix for one of the chain's most persistent operational headaches: knowing what's actually in stock behind the counter.

In lab benchmarks, NomadGo's technology reportedly hit 99% accuracy. That number did not hold up once the tool met actual store conditions.

Baristas found the system regularly miscounted items, mislabeled products, and struggled to tell similar milk types apart on camera. Employees ended up manually correcting the AI's counts, undercutting the labor savings and speed the tool was supposed to deliver.

Why it mattered to Starbucks' turnaround

The stakes went beyond a single software rollout. Product availability sits at the center of CEO Brian Niccol's "Back to Starbucks" turnaround plan, which has aimed to simplify operations and restore consistency after years of menu complexity and service slowdowns. An inventory tool meant to keep shelves stocked and speed up restocking decisions instead added a new layer of manual verification work for store staff.

The gap between controlled benchmark performance and messy real-world deployment is a familiar failure mode for computer vision systems, particularly ones asked to distinguish visually similar products, like different milk cartons, under variable lighting and clutter typical of a working cafe. A 99% accuracy figure measured in testing conditions does not necessarily transfer to a busy store during a morning rush.

Starbucks has not detailed what, if anything, will replace Automated Counting, and NomadGo has not issued a public statement responding to the reported shutdown. For a company that leaned on the tool as a visible piece of its operational fix, the retreat suggests that AI-driven inventory management at scale remains harder to execute than the initial pilot numbers implied.