In September 2025, Starbucks put an AI inventory-counting tool into more than 11,000 North American stores. The vendor had reported 99% accuracy and counting roughly eight times faster than doing it by hand. Nine months later, in May 2026, an internal memo retired the tool, and baristas went back to counting shelves manually.
Most of the coverage has focused on the symptoms. The system confused similar milk types. It missed items. It miscounted visually complex products. Staff ended up counting twice — once for the system, then again to check it — which is precisely the opposite of what the deployment was meant to achieve.
I don’t think this was a reckless bet, and I don’t think it is a story about one vendor. It was the predictable outcome of a validation standard that the whole industry still accepts. The interesting question isn’t why it failed. It’s why we keep being surprised when this happens.
A 99% accuracy score describes the test, not the system
The conditions that produce a 99% accuracy figure in a pilot are never the conditions of a live deployment. Pilots run on controlled lighting, a curated SKU range, standardized packaging, predictable layouts, and a handful of stores where staff know they are part of something being watched. Strip all of that away across thousands of locations and the model encounters a fundamentally different problem from the one it was trained and measured on.
The failure modes also don’t arrive politely, one at a time. SKU proliferation, lighting variability, heterogeneous store layouts, and non-standardized packaging land simultaneously, and they compound. A model validated on a few hundred product types can degrade sharply once it has to separate several thousand — and the degradation shows up first exactly where it hurts most: on near-identical items that differ by a flavor band, a size marking, or nothing visible at all from the angle the camera had. That is a structural problem, not a tuning problem. No amount of post-launch parameter adjustment repairs a system that was never asked the harder question.
There is a second issue worth naming plainly. The 99% figure was the vendor’s own, and it was not independently validated in a live store. A number produced by the party selling the system, under conditions the party selling the system selected, is a marketing input. It is not an engineering one.
Architecture matters more than the accuracy number
Once you accept that, the question changes. You stop asking how accurate a system is and start asking how it reaches an answer because that is what determines whether the number survives contact with 11,000 stores.
Start with how the image is captured. Mobile scanning inherits human variability, angle, distance, lighting, operator consistency and it inherits all of it at exactly the moment scale makes that variability uncontrollable. A fixed camera sees the same shelf from the same position every hour. That consistency is what makes model training reliable and production performance predictable.
One of the biggest mistakes in retail AI is assuming that product recognition is a single-model problem. It isn’t. Our system doesn’t just try to match a picture of a product to a catalog picture and call it done. Every detected item gets converted into a numeric fingerprint, essentially a signature of what it looks like which we compare against candidates using a similarity score rather than a single yes/no guess.
Every image contains multiple sources of information, and the product itself is only one of them. There is also the shelf position, the price tag, neighboring products, promotions, planogram information, stock on hand, and historical observations from previous scans. Each source provides evidence, and each has a different level of reliability depending on the situation.
Rather than asking one model to make the entire decision, production systems should combine these independent signals into a single prediction. If the product image is partially occluded, the price tag may carry more weight. If the price tag is unreadable, neighboring products or shelf location become the stronger indicators. All of those weighted votes combine into one confidence score, and the system only commits once that score clears the bar. That’s the difference between a single-shot guess and a system built to reach an answer through corroborating evidence the way a person glances at the tag when the packaging alone leaves them unsure.
This architecture is also why fixed-camera systems have an inherent advantage over mobile scanning. Every image is captured from the same viewpoint, at the same distance, under similar conditions. That consistency dramatically reduces variation, which lets models learn more stable patterns and makes performance far more predictable across thousands of stores.
Built for the shelf, not the sandbox
Architecture explains how a system decides. Training breadth explains what it already knows on the day you switch it on and that is the second place pilots quietly mislead you. A model that covers 80–90% of national-brand assortment out of the box handles SKU scaling very differently from one trained only on a pilot catalog.
Production performance starts long before the first camera is installed.
Around 80–90% of products in most grocery retailers are national brands. These products appear across thousands of stores worldwide, which makes it possible to build a strong foundation model before deployment. That foundation allows a retailer to achieve meaningful recognition on day one instead of starting from scratch with a small pilot dataset.
The remaining challenge lies in what makes each retailer unique: private-label products, proprietary price tags, store fixtures, promotional materials, and local assortment differences. Those elements are incorporated through targeted training and continuous data collection after deployment.
Just as importantly, production systems need mechanisms to learn continuously. Packaging changes, seasonal products arrive, and new SKUs are introduced every week. Without an efficient process to identify missing products and continuously improve the recognition pipeline, accuracy inevitably degrades over time. Production AI is not a one-time deployment — it is an evolving system.
What “production-ready” actually looks like
If I could change one thing about how retailers evaluate computer vision vendors, it would be to replace “how accurate is it” with three harder questions before any chain-wide rollout.
- How was accuracy measured, and was it validated in the store?
Ask who ran the test, across how many SKUs, in how many stores, over how long, and whether anyone outside the vendor verified the result. A pilot number and a production number are different claims and should be presented as such.
- What happens to model performance as SKU count grows beyond pilot range?
A vendor that has operated at scale can show you that curve. A vendor that hasn’t will change the subject. This single question separates most of the field.
- How does the system handle the variables your stores introduce — your packaging, your lighting, your layouts, your private label?
And what is the concrete process for teaching it the things it will still get wrong in month three? If there isn’t one, you are buying a snapshot, not a system.
None of these are trick questions. Any vendor operating at genuine production scale should welcome them.
The standard the industry should hold
One characteristic separates production AI from a great demo: it knows when it doesn’t know.
Every recognition system will encounter products it has never seen, packaging redesigns, damaged labels, or ambiguous images. The goal isn’t to force a prediction every time — it’s to accurately measure confidence. High-confidence recognitions can be automated, while low-confidence cases should trigger a different flow. Those uncertain cases then become new training data, allowing the system to continuously improve instead of gradually degrading as products evolve.
The Starbucks story shouldn’t discourage retailers from investing in computer vision. It should encourage them to ask better questions.
Accuracy measured in a controlled pilot tells us very little about how a system will perform across thousands of stores, millions of products, and constantly changing retail environments.
The next generation of retail AI will not be defined by who builds the most accurate model in a laboratory. It will be defined by who builds the most resilient production system, one that continuously adapts as stores, products, and customer behavior evolve.
That’s the standard retailers should expect before committing to enterprise-scale deployments.
In the next post I’ll go a level deeper into how that actually works in our stack — how an image becomes a feature vector, how the weighted votes resolve two near-identical SKUs, and why the business layer matters more than most people expect.
Frequently asked questions
Why did Starbucks stop using its AI inventory system?
Starbucks retired the NomadGo computer vision inventory tool in May 2026, nine months after deploying it to more than 11,000 North American stores. Store staff reported that the system confused similar products such as milk types, missed items, and miscounted visually complex shelves — so employees had to recount manually, eliminating the time savings the tool was meant to deliver.
Why do computer vision systems that score 99% in a pilot fail in production?
Because a pilot accuracy score describes the test conditions, not the system. Pilots use controlled lighting, a limited SKU range, standardized packaging and predictable layouts. In production, SKU proliferation, lighting variability, inconsistent store layouts and non-standard packaging occur simultaneously and compound, so a model meets a materially harder problem than the one it was measured on.
How many SKUs can a retail computer vision model handle reliably?
There is no fixed ceiling, but the risk is in the gap between validation range and production range. Models validated on a few hundred product types commonly degrade once production pushes them past a few thousand, because near-identical SKUs — same brand, different flavor or size — become the dominant source of error. Ask any vendor for the performance curve as SKU count grows, not a single headline number.
Are fixed cameras or mobile scanning better for retail inventory accuracy?
Fixed cameras are more reliable at scale. Mobile scanning inherits human variability in angle, distance, lighting and operator consistency at exactly the point where scale makes that variability uncontrollable. Fixed cameras capture the same shelf from the same position every hour, which reduces variation, produces more stable training data, and makes production performance far more predictable across thousands of stores.
What questions should a retailer ask an AI vendor before a chain-wide rollout?
Three: how accuracy was measured and whether it was validated in a live store by someone other than the vendor; what happens to model performance as SKU count grows beyond pilot range; and how the system handles the variables specific to your stores — your packaging, lighting, layouts and private label — including the process for correcting errors after go-live.
What does “production-ready” mean for computer vision, retail AI?
A production-ready system knows when it doesn’t know. It measures confidence accurately, automates high-confidence decisions, routes low-confidence cases to a different flow, and feeds those uncertain cases back as training data. That makes it improve over time rather than drift as packaging, assortment and store conditions change.
Does the Starbucks failure mean retailers should avoid computer vision?
No. It means retailers should change how they evaluate it. Shelf computer vision continues to deliver measurable results in grocery — including reduced manual scanning cost and improved on-shelf availability. The lesson is to evaluate architecture, training breadth and confidence handling rather than a single self-reported accuracy figure.
