Amazon announced that it would phase out the Just Walk Out checkout model at its Fresh format stores. It will be replaced by smart shopping carts.
Amazon currently operates 44 Fresh concept stores with an average floor space of 2300 – 4200 square meters. Autonomous Just Walk Out checkout was used in 27 of them. Just Walk Out is still in use at 20 smaller (170-900 sq. m. size) Amazon Go retail locations.
Just Walk Out debuted in 2016. Touted as one of the biggest innovations in retail, the concept failed to go mainstream. In nearly a decade it was only deployed at a handful of stores, while the biggest technological hurdles for practical uses couldn’t be addressed.
In 2022 sources noted that 1000 autonomous checkouts required 700 human interventions, when the shopping video footage needed to be manually reviewed and validated by an Amazon employee. The company estimated that 1000 autonomous transactions would only require 50 human interventions.
Clearly the autonomous retail revolution failed to materialize. So what’s next for autonomous store technology?
“Using Computer Vision to identify moving objects in a real-time setting – at a store or, for instance, in traffic – is extremely difficult. There are multiple unpredictable scenarios that can throwa the technology off. Even experienced security staff cannot notice all fraud and shoplifting scenarios while monitoring the shop floor. The positioning of store security cameras might not be ideal to cover all the shopping area. There are variations in lighting. Objects can be blocked by other shoppers. This presents many technological challenges for machine learning. It is notable that even a tech behemoth such as Amazon could not solve these issues. When in 6 years from deployment you still require human interventions for 70% of transactions, the development is clearly not going according to plan”, says CEO of retail technology solutions provider Neto Baltic Rokas Budvilaitis.
According to R. Budvilaitis, similar solutions might be perfect in a lab environment. Yet they break down when faced with the unpredictability of the real world.
“In retail checkout is a mission critical process. The solution cannot work well most of the time. It needs to be precise and accurate all of the time”, added R. Budvilaitis.
Neto Baltic has experienced the challenges of AI in retail training firsthand. Together with a technology company Agmis, Neto Baltic develops Crime Predictor – an image recognition fraud prevention app for self checkouts. The solution identifies products scanned at the checkout counter. It recognizes when bulk items are mislabeled in the picklist menu (e.g. cucumbers turn into avocados), barcode switching, no-scan and no-pay scenarios.
“While Crime predictor uses a similar method to identify scanned products, its operations are confined to the checkout space. This space is more akin to a laboratory, with less unpredictability. Despite that, a lot of time and effort went into the development that the solution could prevent 98% of checkout fraud scenarios”, said R. Budvilaitis.
Crime Predictor debuted in 2020. It is currently in use at major retail locations in the Baltic states as well as other European markets.
Neto Baltic also offers AI solutions for shelf stock level monitoring. When stock of a particular product – e.g. milk – is running low, the app automatically notifies a store employee to restock the shelf. A similar solution is also used to ensure planogram compliance.
“We are also testing an AI solution for high-risk merchandise theft prevention. When a bottle of an expensive alcoholic drink is picked from the shelf, the solution can track the shopper across the store to the checkout – ensuring that the bottle is paid for. The solution is fully compliant GDPR, as it does not identify the shopper using biometrics or face recognition, just assigns a tracking tag for a higher-risk item”, says R. Budvilaitis.
Yet – while these AI solutions have already proven in preventing losses or ensuring more efficient management of stores – they still operate as auxiliary supportive tools to augment other store security measures.
“Their operations do not affect mission-critical retail processes. What we can learn from the Amazon example, AI is still not ready for mission-critical retail primetime”.