Starbucks Drops AI Inventory Tool After It Keeps Confusing Milks

Starbucks is scrapping its AI-powered inventory system across North America, pulling the plug on a high-profile technology initiative that was supposed to solve persistent stock shortages but instead struggled to tell oat milk from dairy. The chain is returning to manual counting, marking another high-profile example of an enterprise AI pilot that failed to deliver in the real world.

According to an internal newsletter reviewed by Reuters and confirmed by the company, the Automated Counting tool is being retired immediately. “Starting today, Automated Counting will be retired,” the Monday memo stated. “Beverage components and milk will now be counted the same way you count other inventory categories in your coffeehouse.” Translation: employees will go back to counting by hand.

The system, developed by Seattle-based NomadGo, relied on tablet-mounted cameras and LiDAR to scan shelves of syrups, milks, and other beverage components, generating automatic inventory counts that were supposed to replace manual stock-takes for select categories. It had been in development for years and was rolled out nationwide after CEO Brian Niccol took over in September 2024 as part of his “Back to Starbucks” turnaround strategy.

The tool’s fatal flaw, as Reuters reported in February and internal company materials confirmed, was its inability to reliably distinguish between similar-looking liquids. The app frequently miscounted or mislabeled items, especially products like oat milk and dairy that appear nearly identical. A promotional video Starbucks itself released showed the system failing to register a bottle of peppermint syrup sitting on the shelf, a glitch that now looks far more damning in hindsight.

Starbucks framed the decision as a standardization effort rather than an admission of failure. In a statement to Reuters, the company said the move came from “a decision to standardise how inventory is counted across coffeehouses as we continue to focus on consistency and execution at scale.” The chain is shifting toward more frequent daily replenishments and ongoing supply chain improvements. An internal note shared by the company quoted an employee thanking the team for ending the program: “The thought behind it was great, but the execution was proving difficult.”

The stakes were high because inventory management was supposed to be the easy fix. Over five years, four different Starbucks CEOs have blamed lost sales on the company’s inability to keep stores reliably stocked. In early 2024, fewer than a third of deliveries to Starbucks distribution centers arrived on time and in full, by the company’s own admission. Automated Counting was designed to provide the live, store-level visibility that had been missing, and it was one of Niccol’s headline operational fixes.

This failure arrives at a moment when the broader track record for enterprise AI is looking far less impressive than the initial hype. MIT’s NANDA initiative found last year that 95% of enterprise generative-AI pilots delivered no measurable impact on the P&L, despite roughly $30 to $40 billion in spending, with only 5% reaching production. While Starbucks’ tool was not generative AI, the shape of its failure is familiar: a deeply integrated, store-level workflow proved far harder to automate reliably than the demo suggested.

The financial context makes the decision open to interpretation. Starbucks posted its strongest quarterly sales growth in two and a half years last month, and the stock is up 24% so far in 2026. But operating margins in its core North American market have fallen to 9.9%, down from 18% two years earlier. Niccol has continued investing in other technology bets, including AI tools to sequence orders and assist baristas during peak hours. NomadGo told Reuters it is “continuously learning from customer and user feedback” to improve its products.

The next challenge is whether daily replenishments and manual counts can succeed where the algorithm failed: keeping peppermint syrup on the shelf.