Starbucks has stopped using the AI tool it rolled out across its North American stores last September to count inventory. Reuters reported on May 21, 2026, based on an internal newsletter, that the company’s “Automated Counting” feature had been retired after about nine months. The tool, built by Seattle-based NomadGo, used tablet cameras and LiDAR to scan shelves of milks, syrups and other beverage components and produce counts automatically.
The problem, according to Reuters’ reporting as relayed by The Next Web and Gizmodo, was that it frequently miscounted and mislabelled items — particularly products that look alike, such as oat milk and dairy milk. NomadGo’s marketing had claimed 99% accuracy.
/ 01What happened
The tool had been in development for several years and was expanded nationally after Brian Niccol became chief executive in September 2024, as part of his “Back to Starbucks” turnaround. Inventory was a real pain point: per the company, fewer than a third of deliveries to its distribution centres arrived on time and in full in early 2024, and shortages were hurting sales.
The internal memo was blunt: “Starting today, Automated Counting will be retired,” adding that beverage components and milk “will now be counted the same way you count other inventory categories in your coffeehouse.” Starbucks told Reuters the move reflected “a decision to standardise how inventory is counted across coffeehouses.” One employee’s note, quoted by The Next Web, summed it up: “The thought behind it was great, but the execution was proving difficult.”
NomadGo told Reuters it is “continuously learning from customer and user feedback” to improve its products. Starbucks continues to invest in other technology, including AI tools for sequencing orders and assisting baristas.
/ 02Why it matters beyond coffee
This is a small, specific failure, and that is what makes it useful. It was not a moonshot. It was a well-defined task — count the bottles on a shelf — with a measurable outcome and a vendor promising near-perfect accuracy. It still failed in production, and the reason is familiar from many enterprise AI projects: a system that is good at the general skill (seeing objects) without a reliable grip on the specific facts of the business (which product is which, in this store, on this shelf).
Telling oat milk from dairy milk is not a vision problem in the abstract. It is a product-identity problem. The two cartons may look nearly identical to a camera, but they are different stock-keeping units with different suppliers, par levels and reorder rules. A system that guesses at identity from appearance alone will make confident errors, and in inventory those errors compound: a miscount drives a wrong order, which drives a stock-out or waste.
The second lesson is about how errors surface. A count that is wrong but plausible is hard for a busy barista to catch, and Reuters had already reported miscounts in February, three months before the tool was pulled. When a system produces a number with no visible basis, the only check is a human doing the job again — at which point the automation has saved nothing.
/ 03Designing for checkable outputs
A sturdier pattern pairs perception with an explicit model of the domain. The store’s product catalogue, planogram, delivery records and sales are all structured facts that can live in a knowledge graph. A vision system’s job then becomes proposing observations (“12 cartons at this shelf position”) that are reconciled against what the graph says should be there: what was delivered, what was sold, what the shelf is assigned to hold. Disagreements are flagged with their evidence rather than silently written into the count.
- Identity comes from the catalogue, not the pixels. Look-alike products are distinguished by position, barcode or delivery record, and the system knows which ones are easily confused.
- Every count carries its basis. A reviewer can see whether a number came from a scan, a reconciliation or a manual entry.
- Anomalies are explained. A count far from what deliveries and sales predict is surfaced as a question, not accepted as a fact.
None of this would have made the camera better. It would have made the system’s mistakes visible and bounded, which is often the difference between a tool that staff trust and one that gets retired.
Most enterprise AI does not fail at the skill. It fails at knowing which facts are true for this business.
/ 04What to take from it
For teams running AI pilots, three practical checks follow. Test on the confusable cases, not the average case: the gap between a 99% marketing claim and a production failure usually lives in the look-alikes and edge conditions. Decide in advance how a user will verify an output, and build that into the interface. And keep the business’s own facts in a structure the AI system has to respect, so that the model is an input to a decision rather than the decision itself.
Starbucks’ experience is a reminder that the hard part of enterprise AI is rarely the model. It is the context the model is missing.
This was a catalogue failure dressed up as a vision failure. The camera could see cartons; nothing in the system knew that oat and dairy are separate SKUs with their own shelf slots, suppliers and reorder rules.
Put the product catalogue, planogram, delivery manifests and point-of-sale data in one graph and the vision model’s output becomes a proposed observation, not a count. Each observation links to a shelf position, the SKU assigned to that position, and the stock level that deliveries minus sales predict. A scan reporting dairy where the planogram assigns oat, or a figure far from the expected level, is held for review with the conflicting records beside it. A district manager asking why an oat milk order doubled can trace it to the scan, the reconciliation and the delivery record behind it. Known look-alike pairs live in the ontology as explicit relations, and store-operations staff can maintain them without writing code, which is where confusable cases actually get fixed.
How we’d model it
Shelf observation—proposes→Stock countShelf position—assignedTo→SKUSKU—confusableWith→SKUStock count—reconciledWith→Delivery record
- Starbucks pulls its AI inventory tool nine months in, after it kept confusing the milks — The Next Web thenextweb.com ↗
- Starbucks Abandons Borked AI Inventory Tool That Couldn't Count: Report — Gizmodo gizmodo.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.