Back in 2016, before AI was the hot topic it is today, Amazon asked what was, at the time, a question that very few could understand, let alone answer: ‘What if we could weave the most advanced machine learning, computer vision, and AI into the very fabric of a store so that no one had to wait in line?’ In 2024, we have the answer: existing technology can’t do that yet – at least not in a way that makes sense for a company notorious for its laser focus on seeking profit.
Amazon named their pioneering project ‘Just Walk Out’—a network wherein customers could walk into their grocery stores and leave, without ever having to go through the tedious checkout process. The grocery store would instead track their purchases through a combination of in-store surveillance that established what goods were purchased, and a phone scan at the door to link these up with the individual who made them. As can be expected, this necessitated that each Amazon Fresh location be equipped with hundreds of cameras as well as a fixed stocking system and item positioning just to ensure an AI could take over the role of a cashier. These changes were carried out in 20 Amazon ‘Go’ stores, two Whole Foods (now owned by Amazon) locations, and 40 Amazon Fresh grocery stores. To branch out with the new business model, the company even convinced several third parties to sign up for the new service: 30 stores operated by other companies in sports stadiums, one university, and 12 airports.
Last year, six years after the initiative was launched, performance reports surmised that it had been in simple terms, a failure. The AI element was just not ‘learning’ at a speed that would make the switch to AI from human cashiers cost-effective. One probable cause for these high costs would have been the over 1000 Indians that were manually revising transactions, and labelling images from video for Just Walk Out’s machine learning model to train on. These workers were also tasked with reviewing problem transactions to prevent disputes from escalating. 6 years of reviewing and machine learning later, however, the AI model still needs to be reviewed manually at least 70% of the time. The target had been to only review transactions about 2.5% of the time. Reportedly, Amazon consistently missed out on its goals to cut down on human intervention.
All this culminated with a formal announcement that ‘Just Walk Out’ was no more, which was accompanied by a statement on behalf of the company:
“We’ve invested a lot of time redesigning a number of our Amazon Fresh stores over the last year, offering a better overall shopping experience with more value, convenience, and selection – and so far we’ve seen positive results, with higher customer shopping satisfaction scores and increased purchasing. We’ve also heard from customers that while they enjoyed the benefit of skipping the checkout line with Just Walk Out, they also wanted the ability to easily find nearby products and deals, view their receipts as they shop, and know how much their money saved while shopping throughout the store.”
Despite the promise of AI-powered automation cutting down on the need for human labour in production processes, any machine-learning model requires a vast amount of human labour to gather data – human labour that companies like to source as cheaply as possible. Cheap labour is easier to source in developing countries, hence the need for human workers. This is where the Indian workers come into play. In addition to ‘teaching’ labelling to object recognition, these workers also have to constantly rate AI responses, which are then fed back into the machine for it to improve; much like students are graded in schools.
A multi-billion dollar company like Amazon however always lands on its feet, which is why it’s phasing out the ‘Just Walk Out’initiative in favour of a more effective model. Enter the ‘Dash Cart’, which scans each item as it’s dropped in, which is far less intensive than tracking them through cameras and sensors mounted all over the store. It will however still allow people to skip going through checkout. This will be rolled out at over 40 locations, based on the testing the Dash Cart went through in a few Fresh and Whole Foods stores. Dash Cart will also be accompanied by both self-checkout kiosks and assisted checkout to ensure that everyone has the best experience. Assisted checkout also allows human management to step in whenever something questionable is happening: most of America’s retailers that installed self-checkout options are rethinking their choice due to the significant losses incurred due to customer error as well as outright theft. For now, only Amazon’s ‘Go’ convenience stores will hang on to ‘Just Walk’. It will also continue to be licensed out to third parties.
It looks like AI machines are not quite there to match the human eye and reasoning in engaging with complex physical environments – to the disappointment of AI enthusiasts. If it’s any consolation, there is little doubt that it will get there very soon.
(Theruni Liyanage)