Amazon killed cashierless checkout in supermarkets and rolled it out to 360 other stores. Both calls were right.

In the same stretch of time, Amazon pulled its Just Walk Out cashierless system out of its US grocery stores and expanded it into more than 360 other locations. Read quickly, that looks like a company contradicting itself. Read closely, it is the clearest lesson AI retail has produced: the technology works where the basket is small and breaks where it is big, and the variable that decides which is the use case, not the capability. Same cameras, same models, opposite calls, both correct.

The retreat that got the headlines

Amazon removed Just Walk Out from its large-format US Fresh grocery stores in late 2024, replacing it with Dash Carts that customers scan items into as they shop, across a fleet of 44 US Fresh stores, 27 of which had run the cashierless system. CNN, fairly, called the cashierless push a flop in full-size grocery. The Dash Cart that replaced it has since had its own complexity reeled in until it is essentially a self-checkout on wheels. And by January 2026 Amazon was closing Go and Fresh locations and converting some to Whole Foods. The supermarket bet, as sold, did not pay.

The expansion nobody put on the front page

Now the part that complicates the obituary. In the same period, Just Walk Out scaled to more than 360 third-party locations worldwide and processed 36.7 million items across 17.7 million shopping sessions in a single year, with stadiums and arenas, now more than 80 of them, an early and fast-growing deployment area. Amazon’s own read explains the split: customers value the technology on mission-driven trips with relatively few items, while on a big weekly grocery run they prefer the Dash Cart. The same system the company pulled from supermarkets it is licensing aggressively to everyone else.

Why the big basket broke it

The failure was economic, not magical. A cashierless store charges you by tracking every item you pick up, and the more items in the trip, the more chances to get one wrong and the more expensive the correction. That correction was human: the “autonomous” store leaned on more than 1,000 reviewers, primarily in India, labeling footage and checking transactions to make the AI work. On a three-item stadium beer run, that overhead is trivial. On a sixty-item weekly shop, it compounds until the math stops working. The technology did not get worse between the stadium and the supermarket. The basket got bigger.

The wider retail signal points the same way. Walmart, Costco, and Dollar General have all been rethinking self-checkout as it drives merchandise losses from mistakes and theft. Automating the checkout does not have a universal return. It has a return that depends entirely on the context you drop it into.

The altitude shift

Follow a single mis-scan to the income statement. In a small store it is one flagged transaction a reviewer clears in seconds, lost in the rounding. Multiply it across a full grocery cart, across thousands of carts a day, across dozens of stores, and the review labor and the correction cost stop being rounding and start being the line that decides whether the format is profitable. The capability looked identical in both settings. The unit economics did not, and unit economics, not a demo, is what closed the supermarkets and opened the stadiums.

The rule for retail automation

Stop asking whether cashierless, or any retail AI, “works.” It is the wrong question, and it produces both the breathless launch and the gleeful obituary, often about the same technology. Ask instead what use case it works for. Match the automation to the basket: frictionless capture for small, fast, mission-driven trips, and a scan-as-you-go cart, or a human, for the big weekly shop. Amazon did not fail at cashierless and then succeed at it. It learned, in public and at expense, exactly which trips it fits. That map is the asset. The technology was never the question.