How Amazon Go built Just Walk Out (and how much of it was people)
Amazon Go let the customer grab a product and walk out without a checkout. The system behind it is impressive and was sold as autonomous. It was not, not in the sense the headline suggested.
Amazon opened the first Go to the public in 2018, in Seattle, with Just Walk Out technology. Ceiling cameras and shelf sensors identified who took what, and the charge happened on its own to the customer’s account. No line, no register.
What the system does for the operation
It removes the checkout. The customer enters, grabs, leaves. For the operator, that attacks two problems at once: the line, which makes people abandon the basket, and the cost of the register.
The promise is strong because the line is real. Anyone who runs a store knows the peak-hour line costs abandoned baskets.
What the public record says, with the honest correction
This is where the story needs discipline. In 2024, reporting (The Information, then Business Insider) revealed that about a thousand people in India took part in reviewing Just Walk Out transactions. The reading that went viral was “the autonomous checkout is just people watching video”.
That reading overstates it. Amazon pushed back, and was partly right: the reviewers trained and validated the model, they did not type each purchase by hand. But the point that remains is the one that matters. The system relied on heavy human review for the hard cases, and that carries a cost the camera hides.
The proof is Amazon’s own decision. In 2024, the company pulled Just Walk Out from its larger Amazon Fresh stores and kept the format in small stores, while licensing the technology to third parties. It removed it from where the math did not close.
The non-obvious part
The cost was never the camera. It was the combination of three things: the dense per-store sensor install, the review team for the cases the model cannot resolve alone, and the economics of the grocery basket.
A supermarket has a large basket, loose items, lookalike packaging, fruit with no barcode. Each of those is a hard case. In a convenience store, the basket is small and predictable, and the math closes. In a supermarket, the number of hard cases explodes, and each hard case calls a human.
The real cost
For a normal chain, copying all of Just Walk Out is the wrong budget decision. The per-store install is expensive, the review team is continuous, and the gain only shows in small format with high margin.
The expensive part is not what shows in the demo. It is what keeps the demo working every day.
The version that fits your chain
You do not need the invisible checkout. The version at 5 to 10 percent does not chase frictionless checkout for the whole basket.
Take the specific problem where computer vision already pays: measuring OSA at the shelf without a manual audit, or flagging the shrinkage point at self-checkout. Narrow scope, clear ground truth, a human in the loop only where the error is expensive.
The same camera that fails to charge a whole basket succeeds at saying “shelf 12 has been empty for two hours”. One problem is the frontier of the technology; the other is Tuesday’s work.
Think about the system your operation already considered buying because “Amazon does it”. Can you write, in one line, the operational result it has to deliver? If you cannot, the project does not exist yet, the envy of the demo does.
Tell me the result your operation wants, the specific number: OSA at 96% with no audit team, or self-checkout loss below X. In one hour I will send back a sketch of the architecture sized for your case, with a build or buy marker on each piece, and what is technology frontier versus what is known work. It is the kind of thing we deliver. If the shape confirms, the two-week Diagnóstico becomes a signed spec, and that document becomes the contract for the Implementação.