How a checkout-free shop works out what you took

You pick up two drinks, put one back, and leave a shop without presenting either at a till. For the charge to be right, the system must work out what you kept, connect it to your shopping visit and apply the retailer’s prices. Amazon’s Just Walk Out supports checking a payment method at entry or at exit, with the basket assembled through the shopping process.[1]

That makes checkout-free retail a useful example of AI interpreting events. Recognising a bottle in one picture answers only a small part of the question. The final basket depends on a sequence: what moved, who moved it, whether it returned and which payment account belongs to the visit.

The basket has to survive an ordinary change of mind

Amazon’s earlier technical account describes a temporary numerical identity for each visit, allowing observations from different cameras to stay associated with the shopper. Several people can also share one payment account. A family’s purchases therefore involve both individual movements and a common basket. Tracking a visit serves a different purpose from identifying a person by name.[2]

The physical evidence varies with the item. Cameras can follow a visible package, while a small product may be harder to distinguish from above. The same account describes shelf weight sensors supporting products such as chewing gum. A weight change helps establish that something moved; the visual evidence helps connect that movement with an item and a shopper.[2]

In September 2024, Amazon described a foundation model combining multiple camera views, selected shelf sensors, a store map and product catalogue images. Encoders turn these inputs into numerical representations that the model can use together. Training also covers activities such as detecting, tracking and counting, rather than rewarding only a final receipt.[3]

The newer video model can go back for another look

A June 2026 technical account describes a camera-only Large Video Model. For straightforward events it predicts the time, location, product and quantity directly. When the view is blocked or the sequence is ambiguous, it can invoke learned capabilities such as searching earlier footage, cropping, tracking and counting. The receipt follows that investigation.[4]

The distinction is useful: a system can ask where the missing evidence is within its recorded observations. A later view might clarify what an earlier view could not show. Amazon describes this as an improvement over its camera-and-weight-sensor system, but gives no numerical accuracy benchmark or complete deployment schedule in that account.[4]

A folded shirt creates a different recognition problem

Clothing can fold, stretch, disappear inside a bundle or be worn out of the shop. Amazon introduced an RFID alternative in 2023: each item carries a unique radio-frequency identification tag that readers detect at the exit. The system can identify the tagged garment without reconstructing its appearance in every position.[5]

Reading a radio signal still requires interpretation. A reader may detect a tag near the lane that the shopper is not buying. In its August 2026 update, Amazon says a retrained deep-learning model filters these stray reads, distinguishing carried merchandise from nearby shoppers’ items or products outside the lane. AI helps assign the detected signals to the right transaction.[6]

That release also adds an itemised on-screen basket and total including tax before exit. If the list looks wrong, the shopper can request an associate in the lane. It is a useful place to put correction: while the person and the goods are still together, before an uncertain basket becomes a payment dispute.[6]

Recognising the goods is part of running the shop

The basket must connect to the retailer’s operating systems. In August 2025, the Seattle Seahawks described a two-way integration between Just Walk Out and MyVenue at Lumen Field. It synchronised menu and sales information while supporting loyalty and customer functions. Identifying the item and knowing the applicable price or benefit are separate jobs that need to meet in the same transaction.[7]

Employees remain part of the arrangement. Amazon’s FAQ describes staff assisting customers and replenishing goods, and says the retailer handles customer service and returns. For a shopper, keeping the receipt is still useful. For an operator, the work shifts towards stocking, assistance and resolving exceptions around the automated basket.[9]

The commercial setting matters too. Amazon’s 2026 announcement closing its Amazon Go and Amazon Fresh physical stores said it had not established the customer experience and economics needed for large-scale expansion. The same account described Just Walk Out continuing in more than 360 third-party locations across five countries. Those are different businesses: operating a grocery chain and supplying a checkout system.[8]

For a venue with a short interval to serve hungry spectators, removing the scanning stage addresses a clear problem. It does not by itself determine the range, prices, layout or economics of the shop. The useful achievement is narrower and concrete: turning a messy sequence of ordinary human actions into a basket that the shopper and retailer can recognise.

Report an error or suggest a correction