The News: Starbucks has ended the AI program it used to automate inventory counts across its North American stores, roughly nine months after rolling it out.
The tool, built by Seattle-based NomadGo, used LiDAR sensors and tablet cameras to tally shelf stock of milks, syrups, and other beverage ingredients. Reuters reported in February that the system frequently miscounted and mislabeled products, confusing visually similar milk varieties or failing to register them at all.
In a statement, Starbucks said it is moving to "standardize how inventory is counted across coffeehouses as we continue to focus on consistency and execution at scale," and is working toward more frequent, daily store replenishment.
The tool had been championed by CEO Brian Niccol as part of his push to fix the product shortages he has blamed for hurting sales.
What changed: Starbucks pulled a computer-vision counting layer out of more than 11,000 North American stores and is reverting to a standardized manual count, paired with more frequent replenishment to keep shelves accurate.
The machine's core job was to tell near-identical SKUs apart on a back-room shelf, separating whole milk from oat from 2%. That is precisely where it broke.
Rather than retrain the model on better data, Starbucks is changing the process beneath it: define one counting method for every store, then replenish often enough to keep the numbers true.
The pattern: The more revealing comparison is the company doing the opposite.
Walmart is scaling the same underlying technology, computer-vision inventory mapping, and credits it with holding inventory growth to 2.6% year over year, roughly half its sales-growth rate. Walmart now runs automation in 23 of its 42 regional DCs, and 60% of US stores are fed by automated facilities.
Two retailers adopted the same class of tool and landed in opposite places. What separated them was not the AI but the operating model underneath it.
Reporter's notes: This was a sequencing failure rather than an AI failure.
Walmart layered computer vision onto a standardized, instrumented network it had spent years building, so the model had a clean, consistent process to read. Starbucks bolted the same class of tool onto thousands of stores that each counted inventory a little differently, then asked it to resolve an ambiguity, which milk is which, that the stores themselves had never standardized.
Starbucks' own statement points to the lesson: the fix is to "standardize how inventory is counted," which is process first.
For any multi-unit operator weighing a computer-vision rollout, the order matters. If your stores cannot count the same way by hand, a camera will not save you. It will only be wrong faster, and at scale.
Room for disagreement: A smart ops leader could argue Starbucks gave up too soon.
Nine months is barely past the scar tissue of a pilot, computer-vision models improve with labeled data and retraining, and NomadGo's system works in other settings.
By that reading, the real failure was change management and store-labor adoption, not the technology, and pulling the tool resets a learning curve Starbucks has already paid for. There is also a cost tail: daily replenishment plus manual counts across more than 11,000 stores is a labor line the automated count was meant to remove.
