Cashierless Checkout: Growing in Venues, Retreating from Supermarkets, Rarely Checked by Outsiders

The use case

On 27 January 2026 Amazon announced that it would close all of its Amazon Go and Amazon Fresh stores, the places where it had built its checkout-free system, while stating that the system now operates in over 360 third-party locations across five countries (operator-reported)¹. The rest of this paper explains how both statements can be true.

The system works as follows. Cameras and shelf sensors feed machine-learning models that work out which shopper took which item, and the shopper's payment card is charged on exit. Amazon's version, Just Walk Out, combines computer vision, sensor fusion and deep-learning models; it assigns each shopper a temporary numeric code for the visit and, according to Amazon, does not use or collect biometric information². Other suppliers include Trigo, whose technology Tesco uses, and AiFi, which Aldi used³. Grabango, a United States supplier, is covered below⁴.

Scope. Only the machine-learning step counts as AI here. Card-tap entry gates, self-checkout kiosks, scan-and-go apps and handsets, and radio-frequency tags, which Amazon has added for clothing and fan gear, are older technologies and are not credited with any result below². Amazon's Dash Cart uses the same vision technology inside a shopping cart and appears only where it explains a decision². Amazon One palm payment is a separate biometric service and is out of scope². Amazon opened its first Go store in Seattle in 2018, outside the three-year window used for recent evidence⁵.

Where it is running

The sites fall into two groups: Amazon's own grocery stores, which are closing, and smaller or faster-turnover formats, which Amazon describes as growing.

The third-party sites are mostly venues. Amazon lists travel retailers, sports stadiums, entertainment venues, conference centres, theme parks, convenience stores, hospitals and college campuses². Amazon says Lumen Field in Seattle has nine stores serving more than 68,000 fans and that Delaware North has opened more than a dozen at United States stadiums (April 2024)². In January 2026 Amazon added that more than 40 North American fulfilment centres use the system in staff breakrooms (operator-reported)¹. A trade report in September 2026 relayed an Amazon manager's LinkedIn claim of more than 400 stores globally (vendor-reported, search extract)⁸. The two counts use different words (locations and stores), and neither was independently checked.

Geography. The evidence is almost entirely from the United States and the United Kingdom. Canada and Australia appear only in Amazon's 2024 location count², and the five countries in its 2026 count are not named¹. Deployments in continental Europe and Asia surfaced only in vendor material and older trade items and were not verified, so no conclusion is drawn about them.

What the evidence shows

No independent measurement of accuracy, theft reduction or cost per store was found, and the stadium and hospital figures all come from Amazon or its venue customers. The evidence is ordered from strongest to weakest.

Shopper preference (peer-reviewed experiment, 2024). A Journal of Retailing study tested how the features of unstaffed stores affect store patronage with a conjoint experiment. Consumers preferred staffed stores on every dimension except checkout; checking in with a credit card, limited access to remote staff and being unable to verify the basket before paying were significant barriers. Community-based and rural locations suited unstaffed stores better than anonymous traffic hubs (abstract; sample and design details not read)⁹. The study covers unstaffed stores in general, not only AI-based ones. Its finding on traffic hubs sits uneasily with the venue deployments, and no equivalent experiment on venue stores was found.

Shopper behaviour (working paper, not peer-reviewed). A working paper compares campus stores at one university across two academic years before and after Just Walk Out was introduced. It reports that demand shifted towards the periods before classes and that sales of sensitive personal items rose, including a 164% increase in condom sales (abstract; search extract; the excerpt read does not mention a control group)¹⁰.

Retailer decisions (independent trade reporting). Aldi ended its checkout-free trial in Greenwich in March 2026 and kept the site as a conventional store with self-service tills. Aldi said the trial ended as planned, and some shoppers had complained about a £10 pre-authorisation taken at the gate⁷. Sainsbury's removed the technology from its Holborn Circus store and now uses a kiosk till and two small self-checkouts³. Aldi's UK chief executive said in 2023 that self-checkouts were the right next step rather than more Shop & Go stores (as reported by The Grocer)³. The Grocer also reported a survey it commissioned in which 13% of shoppers had used checkout-free technology, against 26% who had self-scanned with an app and 40% with an in-store handset (sample size and date not shown)³.

Operator-reported results. All of the following figures are operator-reported. In April 2024 Amazon relayed that Lumen Field recorded an 85% increase in transactions and a 112% increase in sales per game at its first Just Walk Out store, that Delaware North's stores serve 20% to 30% more customers than traditional concession stands, and that Market Express at ExCeL London can serve 300% more customers on its busiest days and raised annual revenue by 56%². In January 2026 Amazon said that waiting times in a hospital cafeteria fell from 25 minutes to 3 and that fans at Scotiabank Arena can get concessions in 30 seconds¹. None of these states a control group or separates the effect of the technology from that of opening a new, differently designed outlet. The mechanism is plausible: in venues, queues form in short bursts and the till is the bottleneck, so removing it adds capacity. That explains why the largest gains are reported where demand peaks, and it says little about a supermarket with steady traffic.

Human review. A Business Insider report relaying The Information (date not shown) said, citing an unnamed source, that about 700 of every 1,000 Just Walk Out sales in 2022 needed review by a team in India, against an internal target of 50 (search extract; disputed by Amazon)¹¹. Amazon replied that its associates annotate real and synthetic shopping data to improve the models and do not watch live video to generate receipts². Both statements can be true, because human labelling is normal in machine-learning systems. What decides whether a store is genuinely cashierless is the share of visits that need a person, and no published rate was found.

What it takes: cost, integration, governance

Cost. No public price per store was found. Indirect evidence points one way: C-Store Dive reported in October 2024 that most convenience retailers that piloted automated checkout chose less expensive self-checkout kiosks for wider rollout, and it described the technology as pricey⁴. Amazon says it is building edge computing and new recognition algorithms to improve latency and accuracy in large-format stores, which suggests the current versions fall short there². Amazon's January 2026 statement that its stores lacked "the right economic model needed for large-scale expansion" concerns the stores as businesses, not the checkout technology alone¹.

Integration. The system needs cameras and sensors across the sales floor, a way to identify the shopper at entry and a payment link. Entry is by card, app or palm scan²³; Aldi's pre-authorisation shows that the entry step can itself deter shoppers⁷. Sainsbury's Holborn Circus was the first store Amazon retrofitted with the technology, and Sainsbury's was Amazon's first non-US customer³. Tesco's four stores keep a checkout option beside the cameras³, consistent with the experiment's finding that limited access to staff deters shoppers⁹.

Governance. Privacy law is the main rule set found. New York City's Local Law 3 of 2021 requires retail stores, food and drink establishments and places of entertainment that collect, retain, convert, store or share biometric identifier information to post a clear sign at customer entrances. It defines the term as a physiological or biological characteristic used to identify or help identify a person, including "any other identifying characteristic"¹². An aggrieved customer may sue after giving 30 days' written notice and a chance to cure, with damages of $500 for each signage violation and $500 or $5,000 for each unlawful sale of such data¹². In March 2023 a customer filed a proposed class action, Perez v. Amazon.com, Inc. (S.D.N.Y. No. 1:23-cv-02251), alleging that Amazon Go stores in New York collected biometric information, including the size and shape of customers' bodies, without proper notice (allegations; search extract)¹³. Amazon states that Just Walk Out does not use or collect biometric information². No ruling or settlement was found, so whether body-shape tracking falls within the law is unresolved in the sources read.

Supplier risk. Grabango, which had piloted its system with several United States chains, ceased operations in October 2024 after failing to secure funding, and the effect on its retail partners was unclear at the time⁴. A retailer that chooses a smaller supplier takes on the risk that the supplier does not survive.

Where it falls short

This section answers the challenge question: what makes the use case look less successful than it is usually presented?

Format fit. Amazon's own account is that Just Walk Out suits quick trips for a few items, and that in larger grocery stores customers have so far preferred the Dash Cart because they want a running tally of spending (operator-reported)². Amazon removed Just Walk Out from its United States Fresh grocery stores in April 2024 and moved them to Dash Carts (AP report, search extract)¹⁴². In January 2026 it decided to close all Amazon-branded grocery stores¹. The mechanism is visible in the experiment: shoppers valued verifying the basket before paying⁹, which a receipt sent after exit does not offer¹⁴.

Where the results come from. Reported gains cluster in venues with bursts of demand and in newly opened outlets, so they do not transfer to steady-traffic grocery. Amazon says businesses value reduced theft but gives no theft figure².

Persistence. Of the four United Kingdom checkout-free retail experiments in the sources read, Aldi's has ended⁷, Sainsbury's has had the technology removed³ and Amazon has proposed closing all 19 of its UK Fresh stores less than five years after the first opened⁶. Only Tesco's four hybrid stores continued at the time of reporting³. Amazon had planned 260 UK Fresh stores by 2025 but never opened more than about 20 (Retail Week report; search extract)¹⁵.

Reach. No source examined whether benefits reach shoppers without smartphones or bank cards. Aldi's gate pre-authorisation drew complaints from shoppers who spent less than £10⁷.

People and process. Annotators and reviewers sit behind the model², and the hybrid format Tesco kept leaves a checkout in the store³. Gains credited to AI therefore include the contribution of that labour.

Bad news searched for. Searches for regulator action and for independent audits of Just Walk Out accuracy found none. The withdrawals, the supplier closure and the privacy lawsuit are described above, and no court outcome was found.

Phase: In early use — why

The hypothesis, In early use, is kept. Machine-learning checkout is confirmed, because Amazon describes computer vision and deep-learning models², and Amazon reports over 360 commercial third-party locations in five countries in January 2026¹. That count is operator-reported and could not be checked, and the sites are concentrated in venues, workplaces and small stores. Grocery, the format that would make the use wide, is where Amazon, Aldi and Sainsbury's retreated¹⁷³. A phase of In wide use is therefore not supported, and Pilot with results undersells the commercial third-party sites. The phase should be reviewed if an independent count or accuracy audit appears.

What to watch next

• Whole Foods Market Dash Cart. Amazon says the newest version will be in dozens of stores by the end of 2026, which tests whether vision inside a cart, rather than in the ceiling, becomes the grocery format².
• Fulfilment-centre breakrooms. Amazon plans more sites in 2026¹.
• Perez v. Amazon.com, Inc. A ruling or settlement in S.D.N.Y. No. 1:23-cv-02251¹³.
• New York City Int. No. 217 of 2024. It would amend the biometric law to cover public accommodations; its status was not established¹⁶.
• Published review rate or independent count. Any Amazon figure for the share of visits needing human review, or an independent location count that confirms or revises the 360-plus figure¹.
• Tesco GetGo. Any announcement that changes the four stores' hybrid format³.

Questions to ask before adopting

• What share of shopping visits in a comparable store needed human review in the last 90 days, and who measured it?
• Do our shoppers make quick trips for a few items or weekly shops, and does our format match where operators report gains?
• How will shoppers verify their basket before they leave, and what staffed fallback do we keep?
• What is the all-in cost per store against self-checkout, scan-and-go and smart carts, and what happens if the supplier fails?
• Which biometric and privacy rules apply where we operate, and does the supplier's description of what is collected hold up under the legal definition?