Starbucks’ AI Inventory Experiment Failed Fast — Here’s What Went Wrong

Starbucks attempted one of the largest retail AI inventory deployments in the food‑service industry—and shut it down less than a year later.

In September 2025, the company rolled out an AI system called “Automated Counting,” developed with startup NomadGo, across more than 11,000 North American stores. The goal was simple: replace manual inventory checks with a tablet-based computer‑vision system that could scan shelves and automatically identify and count products such as milk, syrups, and other beverage ingredients.

But by May 2026, Starbucks abandoned the program entirely after stores reported persistent counting and identification errors that undermined day‑to‑day operations.

What the Automated Counting system was supposed to do

The system used tablet cameras combined with computer vision and spatial sensing to scan refrigerators, shelves, and backroom storage areas. Store employees would sweep the device across items, and the software would automatically recognize products and tally quantities in real time.

The technology promised several operational benefits:

• Faster inventory checks than manual counting
• Improved visibility into low-stock ingredients
• Fewer product shortages affecting menu availability

NomadGo’s system processed images directly on mobile devices and overlaid product identification data to help employees validate counts quickly.

In theory, that automation would free workers from routine counting tasks while giving Starbucks better data about inventory levels across thousands of locations.

Why the system failed in real stores

Once deployed at scale, the AI system struggled with basic reliability.

According to reporting citing internal company communications, the tool frequently miscounted items, mislabeled products, or missed inventory entirely.

One recurring problem involved similar-looking products, especially milk varieties such as dairy, oat, or other alternatives. The system sometimes confused these items or failed to detect them correctly.

For Starbucks stores—where milk availability directly determines whether drinks can be prepared—those mistakes created operational friction. Instead of saving time, inaccurate counts forced workers to double‑check inventory manually.

When inventory systems are unreliable, they can cascade into larger problems:

• Incorrect restocking decisions
• Ingredient shortages
• Menu items temporarily unavailable

At that point, the automation no longer delivered its intended benefit.

Why Starbucks returned to manual counting

In May 2026 Starbucks informed stores that the Automated Counting program would be retired and that beverage components and milk would return to the same manual counting process used for other inventory categories.

The change was tied to a broader operational reset inside the company.

Under CEO Brian Niccol, Starbucks launched a strategy often described internally as “Back to Starbucks,” aimed at fixing persistent product shortages and improving store execution.

As part of that effort, the company emphasized:

• standardized inventory practices
• more consistent store processes
• daily replenishment of key ingredients

Returning to manual counts was seen as the more dependable option for ensuring stores had the ingredients needed to keep menu items available.

A cautionary example for enterprise AI deployments

The shutdown of Starbucks’ inventory AI highlights a broader challenge in enterprise automation: technology that works in controlled demos can struggle in messy real‑world environments.

Retail stores present difficult conditions for computer vision systems, including:

• cluttered shelves and refrigerators
• changing packaging designs
• similar-looking products
• inconsistent lighting or storage layouts

Even small recognition errors can become operational problems when multiplied across thousands of locations.

In this case, Starbucks concluded that the system’s accuracy issues outweighed its efficiency gains, leading the company to scrap the tool after just nine months of full deployment.

The experiment still demonstrates how aggressively large retailers are testing AI in physical operations—but also how quickly those pilots can be reversed when reliability falls short of the realities of daily store work.