Amazon Just Walk Out

What They Said

Amazon launched Just Walk Out in 2018 with a promise that sounded like science fiction made real. Customers would walk into a store, pick up what they wanted, and leave. No checkout. No cashier. No app to scan. Cameras and shelf sensors, powered by computer vision and machine learning, would track every item picked up and charge the customer’s account automatically.

The technology was the centerpiece of Amazon Fresh’s expansion strategy and its Amazon Go convenience stores. Executives positioned it as the future of physical retail — a labor model that would let Amazon scale stores without the cost or friction of traditional cashier operations. Third-party retailers were pitched the same vision: license Just Walk Out, eliminate the checkout line, redefine the store.

What Actually Happened

In April 2024, The Information reported that Amazon was pulling Just Walk Out from its US Amazon Fresh stores. The reason was not customer rejection. The reason was that the technology had never actually worked the way Amazon described.

Behind the cameras sat roughly 1,000 workers in India per store, manually reviewing video footage to confirm what customers had picked up. The “AI” was a labeling pipeline. As of mid-2022, Just Walk Out required human reviewers to validate around 700 of every 1,000 transactions. Receipts that customers expected within minutes sometimes took hours to arrive while the offshore team caught up.

Amazon disputed the 1,000-per-store figure but did not dispute the core mechanic. The company replaced Just Walk Out in Fresh stores with Dash Cart, a smart shopping cart that scans items as the customer adds them — a far less ambitious but functioning system. Amazon Go stores remain, but the dream of licensing Just Walk Out as a third-party retail platform has effectively collapsed.

The reporting exposed an open secret inside Amazon. The economics of the system never worked. Each store generated tens of thousands of video segments per day requiring human review, and the cost of that review never declined fast enough to justify the absent cashier.

The Root Cause

Amazon shipped a research project as a product. The computer vision models behind Just Walk Out were genuinely impressive, but they weren’t accurate enough to operate without a human-in-the-loop. Rather than acknowledge that gap, Amazon staffed the gap with offshore labor and marketed the system as autonomous.

The deeper failure was incentive design. Internal teams were rewarded for store openings, not for the unit economics of each store. The human review pipeline was treated as a temporary scaffolding that the model would eventually outgrow. It never did. Five years in, the scaffolding had become the building.

The Pattern to Watch For

Any AI product whose accuracy depends on a hidden review layer is not an AI product yet. It is a labor arbitrage with a marketing wrapper. Ask your vendor what percentage of transactions, decisions, or outputs require human validation today, and what the trajectory has been over the last 18 months. If the percentage is flat or rising, the model isn’t learning fast enough to ever be autonomous.

What You Should Steal

Build a “review rate” metric into every AI deployment from day one. Track the percentage of model outputs that require human correction or validation, and report it monthly to the executive sponsor. If the rate isn’t dropping by a measurable amount each quarter, you don’t have an AI system on a path to autonomy. You have a labor pipeline with a confidence problem. Decide which one you’re actually running before you scale.