The Memo Said One Line.Eleven Thousand Stores Went Back to Counting by Hand.

The Memo Said One Line. Starbucks Pulled NomadGo From 11,000 Stores. — Aethreallegence Field Notes

On Monday, May 19, 2026, Starbucks sent an internal newsletter to every location in North America. Reuters obtained it and confirmed the contents with two employees. The key line was seven words.

“Starting today, Automated Counting will be retired. Beverage components and milk will now be counted the same way you count other inventory categories in your coffeehouse.”

Starbucks internal newsletter · May 19, 2026 · Reuters

That was it. Nine months of deployment across more than 11,000 company-operated locations in the United States and Canada. The tool that CEO Brian Niccol had described as a centerpiece of his “Back to Starbucks” turnaround. The system that Starbucks’ own Chief Technology Officer said would free baristas to focus on “crafting high-quality beverages and connecting with customers.” Gone in a single sentence.

The announcement from September 2025 has since been deleted from the Starbucks website. The promotional video is gone too.

This is the most prominent enterprise AI rollback in retail so far in 2026. And what it reveals is not just a product failure. It is a category argument.

What was promised

NomadGo is a Redmond, Washington-based startup. Its Automated Counting system combined LiDAR sensors and tablet cameras with augmented reality to generate inventory counts. The concept was straightforward: a barista picks up a tablet, sweeps it across a shelf of syrups and milk varieties, and the AI generates the count overlaid in AR. Fast. Validated. Done.

At launch in September 2025, the numbers were compelling.

99 percent accuracy claimed publicly.

Up to eight times faster than counting by hand.

A “unique synthesis of on-device 3D spatial intelligence, computer vision, and augmented reality.”

Described by NomadGo’s CEO as transforming inventory from “manual, tedious, and inaccurate” to “automated, intelligent, and fun.”

Starbucks’ CTO endorsed it publicly. Starbucks posted about it. NomadGo posted about it. The financial logic of the deployment was not crazy. Counting milk and syrups is a real, recurring labor cost. If you can make it faster and more accurate, that is a genuine operational improvement.

What followed was a different story.

What actually happened on the store floor

Reuters first surfaced the problems in February 2026 — five months into the deployment. Employees and managers across multiple locations described the system frequently miscounting and mislabeling items. It confused similar milk types. It missed items on shelves during scan sessions. It failed, in at least one documented case, to recognize a peppermint syrup bottle in a Starbucks promotional video that Starbucks itself had uploaded to demonstrate the tool.

That video has since been deleted.

A Starbucks shift supervisor of nine years named Carl Addison told Fortune the experience clearly: “It started off not particularly accurate and got less accurate over time.” He also noted something that deserves particular attention from any operator evaluating AI inventory tools: the system required stores to rearrange their back-of-house storage to accommodate the scanning workflow. It did not adapt to the store. The store had to adapt to it.

“If the system counted too much of a product, it wouldn’t send enough of a product a store was running low on. If the system counted too little, it wouldn’t ship enough of a needed product.”

Carl Addison, Starbucks shift supervisor of 9 years · Fortune, May 2026

That is not a technology inconvenience. That is an inventory truth failure with direct downstream consequences. The system’s counts fed replenishment. When the counts were wrong, replenishment was wrong. When replenishment was wrong, product was either over-ordered or missing. The gap between what the system believed and what was physically true caused operational damage that compounded over time.

That is exactly what Truth Decay does.

The timeline tells the story

NomadGo at Starbucks · Verified Timeline

Reuters · Fortune · BusinessWire · May 2026

Sep 2024

Brian Niccol becomes Starbucks CEO. Inherits NomadGo deployment project, decides to push it forward as part of “Back to Starbucks” turnaround.

Sep 2025

NomadGo deployed to 11,000-plus company-operated North American locations. 99% accuracy claimed. CTO says it will free partners to focus on customers. NomadGo CEO calls it “automated, intelligent, and fun.”

Feb 2026

Reuters reports employee complaints: frequent miscounts, mislabeled items, confusion between similar milk types, missed items during scans. Peppermint syrup unrecognized in Starbucks’ own promo video. Starbucks tells Reuters the tool has “improved product availability.”

May 19, 2026

Internal newsletter retires Automated Counting across all North American locations. Staff return to manual counting immediately. Starbucks frames it as “standardization,” not failure. NomadGo says it is “continuously learning.”

Post-retirement

Starbucks deletes its September 2025 NomadGo announcement and promotional video from its website. Neither Starbucks nor NomadGo publicly acknowledges the accuracy failures Reuters documented.

The timeline contains a detail that should not be overlooked. In February 2026, when Reuters first reported the problems, Starbucks’ response was that the tool had improved product availability. Three months later, the same company issued a memo retiring it entirely and telling staff to go back to counting by hand immediately. Those two statements cannot both be true in the way Starbucks intended them.

That is the credibility cost. Not the tool breaking. Tools break. The cost is the distance between what was said at launch and what was happening in the store the entire time.

The operational failure that matters most

The accuracy problems are visible and documented. But there is a structural failure underneath them that is more important for operators to understand, because it would have been a problem even if NomadGo’s accuracy had been exactly what was claimed.

Every scan NomadGo produced required a human to verify it before the business could trust it. That was not a workaround. It was the operating model. When the count looked wrong, someone checked. When it looked right, someone still had to confirm it was right before acting on it, because the error rate had removed the basis for automatic trust.

What NomadGo replaced

+One manual inventory cycle. Staff count shelves. Staff record the count. Done.

What NomadGo created

—Staff conduct a scan with the tablet.

—Staff examine the AI-generated count for errors.

—Staff manually verify items the system miscounted or missed.

—Staff re-enter corrections into the system.

—Two cycles where one cycle existed before.

A system that requires workers to verify every output delivers no net efficiency gain. It does not reduce inventory labor. It changes its form and — in cases where the verification takes longer than the original count would have — it increases it.

This is not a criticism unique to NomadGo. It is a structural property of any checkpoint-based AI system deployed in an environment where trust in the output has not been established. The gap between what the demo promised and what the store floor delivered was the gap between a controlled condition and an operating one.

Why the store floor broke what the demo held together

Starbucks stores are not test environments. They are high-traffic, high-movement operations running under time pressure, with visually similar products stored in close proximity, constantly handled by multiple staff across multiple shifts, in lighting that changes across the day.

The core failure mode — confusion between similar milk types and syrups — is not an edge case in a working café. It is a central operating condition. Half-and-half and heavy cream look similar. Two-percent and whole milk share the same jug form. Syrups of different flavors share the same bottle shape. These are the exact products the system was supposed to count. And they were the exact products it consistently failed on.

The training data did not match the production environment.

The lighting on a promotional video shelf does not match mid-shift in a working store.

A product in its designated position does not behave the same as that product after it has been moved, restocked, and handled across three shifts.

The same shelf photographed in a controlled pilot does not perform the same as that shelf in 11,000 locations across two countries with different layouts, different lighting, and different operational rhythms.

This is the gap between controlled demo performance and real-world operating conditions. It is not new. It shows up consistently when physical-world AI tools scale beyond the environment in which they were trained.

More importantly for operators: none of these conditions are unusual. They are the normal state of a working retail environment. Any AI inventory tool that cannot perform reliably under normal conditions has not solved the inventory problem. It has moved it.

The checkpoint architecture and what it cannot see

There is a deeper reason the NomadGo model faced these challenges that is worth naming clearly, because it is architectural rather than executional. NomadGo is a checkpoint system. A staff member initiates a scan. The AI processes that scan. The count is recorded. The scan is done.

Between scans, the system is blind.

Milk gets used and restocked between scans.

Syrups get moved between scans.

Products are touched, shifted, and reorganized between scans.

Deliveries arrive between scans.

Returns land in incorrect locations between scans.

At the moment of each scan, the system captures what is visible. Then the environment keeps moving. The next scan starts from a physical reality that has changed since the last one. If the accuracy of each scan is less than claimed, the cumulative drift between the system’s records and physical reality grows with every cycle.

Carl Addison’s description of the system getting “less accurate over time” is consistent with this dynamic. It is not that the AI degraded. It is that the gap between snapshots and physical truth accumulated.

The Architecture Argument

“Cameras see the store at a moment. They do not maintain the truth of the store over time.”

What the market now has to answer

The NomadGo retirement from Starbucks does not mean AI cannot help with inventory. It can and it will. The category is real. The market need is genuine. Counting syrups and milk is a real problem worth solving.

What the retirement makes harder to ignore is this: making counting faster is not the same thing as maintaining inventory truth. Those are two different problems. They require two different architectures. And confusing one for the other — or assuming that solving one solves the other — is the error that produced this outcome at scale.

For operators evaluating any AI inventory tool, the questions that the NomadGo case now makes unavoidable are not about speed or accuracy claims at the moment of a scan. They are about what happens between scans.

The Questions Operators Now Cannot Avoid

01Does the system observe what happens between counts, or only at the moment of the count?

02When the system and physical reality disagree, does the system require human verification of every output, or only when confidence genuinely breaks?

03Does the system get better from its corrections, or does it repeat the same errors across thousands of locations?

04Can it distinguish between visually similar products reliably, in real operating conditions, not in a demo environment?

05Does the system treat inventory as a continuous living state, or as a series of periodic snapshots?

06What is the inventory confidence of the record between the last count and the next one?

The last question is the one the industry has not had language for. NomadGo’s deployment made it visible. An inventory record that was accurate at the last scan may no longer be reliable by the time the next decision is made from it. The gap between the two is where inventory truth decays.

Eleven thousand stores just learned that lesson at full scale.

Closing thought

The memo that ended NomadGo at Starbucks was seven words. The operational story behind those seven words is far longer: an AI system promised to make inventory accurate, deployed to more than 11,000 locations, confused milk types, missed items on shelves, required staff to verify every output, and left stores less able to trust their inventory records than they were before the tool arrived.

Starbucks called it standardization. NomadGo called it a learning process. Neither statement acknowledges the real lesson.

Faster counting at a checkpoint does not stop inventory from drifting after the checkpoint closes.

A more accurate scan does not maintain truth between scans.

A verification burden transferred to staff is not inventory accuracy — it is a new form of the same manual labor.

And a system that gets less accurate over time in a real operating environment has not solved the inventory problem at all.

The real inventory problem is not that counting is slow. It is that inventory truth decays continuously — between counts, between shifts, between scans — and a checkpoint can only confirm what was true at the moment someone looked.

The question the market has to answer now is what kind of system actually changes that.