Amazon has transformed e-commerce through machine learning and personalization, but its acquisition of Whole Foods brought a new opportunity to apply machine learning to the physical retail world. Amazon’s experiment with cashier-less, fully automated tests the limits of machine learning technology, physical retail business models, and customer privacy concerns.
Researchers have been studying how to computationally identify people by their faces since the 1950s, and advanced machine learning techniques have successfully been applied to facial recognition beginning in the early 1990s. Amazon is now working to apply this to the grocery store industry. With its Amazon Go stores, consumers can simply walk in, take what they need, and walk out, without any sort of physical checkout. An array of ceiling-mounted cameras and sensors collect data on people and movements through every corner of the store, while sensors on shelves precisely measure the weight of inventory remaining. Putting together all of this data, Amazon applies computer vision to identify specific individuals as they enter the store, track their movements as they pick up items, create a virtual cart of items they are purchasing, and complete the transaction as they exit the store. Those entering the store scan a barcode identifying them as Amazon members, the transaction completes to the method of payment on file, and the customer is emailed a receipt.
Typically, FMCG retailers (fast-moving consumer goods, industry lingo for grocery stores) have traditionally not brought much tech or machine learning into the store experience. While Wal-mart has reduced labor costs with self-checkout kiosks and Nordstrom has experimented with WiFi tracking to create customer in-store “heat maps”, the Amazon Go concept is fundamentally different in that it requires a large investment of sensors throughout the entire store to collect vast amounts of data. Amazon connects a specific person to their journey through the store, identifying specific items that were picked up, considered for purchase, and ultimately bought. Ultimately, machine learning is important for Amazon in this product development because this data creates a strategic moat against experienced brick-and-mortar competitors. While the FMCG industry has traditionally had very low margins, this advantage could create a more lucrative business for Amazon and provide a rationale for its Whole Foods acquisition and expansion into the industry.
In the short term, Amazon has tested the concept with employees outside its offices in Seattle. After months of learning about and refining the user experience, Amazon is now slated to open several more stores, initially with locations in tech-savvy, urban hubs in San Francisco and New York, with aspirations to open up to 3,000 stores nationwide. These stores are currently small to medium-sized stores, about the size of a typical 7-11, and designed to suit those in cities looking for a quick snack on the go. While these are great short-term steps, in the medium term Amazon needs to really analyze if this is a profitable business model. The incremental costs of installing a fully outfitted cashier-less store relative to a normal 7-11 store are large, with some estimates at $1M or more. It’s unclear what the breakeven point would be and if this type of store really delivers a superior customer experience that would meaningfully attract more customers than the traditional store format. Additionally, the store still needs staff to stock the shelves, and presumably would still need to be attended with staff whenever it’s open. A typical convenience store may have only 1 employee working at a time, at perhaps $10-12 per hour. Supposing the cost to pay an employee with benefits were $15 per hour, the total outlay to staff a convenience store for 24 hours every day of the year would be $131,000. If Amazon were able to eliminate one employee in the convenience store model, the $1M cashier-less investment would take 8 years to pay back. This suggests then that Amazon sees real value in the customer data and other reasons to pursue the machine learning approach, not just in labor savings.
There are a few other steps Amazon should take as it develops this technology. Consumers are increasingly savvy the level of data collected about them by companies, and Amazon should actively address this concern by explaining in simple terms what type of data it collects in order to operate these stores and how it is used.
Some questions remain outstanding: will these stores be economically viable at a mass scale? Second, are these stores a superior experience that customers will seek out or are they simply a novelty? Finally, to what extent will consumers’ privacy concerns limit the pace of adoption, given recent heightened attention? While these questions will need to be solved, one thing is for certain — Amazon will continue to test new ways of applying data and machine learning to change retail.
“Stealing From a Cashierless Store (Without You, or the Cameras, Knowing It),” September 13, 2018, The New York Times, https://www.nytimes.com/2018/09/13/technology/standard-market-retail-automation-behavioral-data.html, accessed November 2018.
“Inside Amazon Go, a Store of the Future,” January 21, 2018, The New York Times, https://www.nytimes.com/2018/01/21/technology/inside-amazon-go-a-store-of-the-future.html, accessed November 2018.
“Amazon Will Consider Opening Up to 3,000 Cashierless Stores by 2021,” September 19, 2018, Bloomberg, https://www.bloomberg.com/news/articles/2018-09-19/amazon-is-said-to-plan-up-to-3-000-cashierless-stores-by-2021, accessed November 2018.
“Stepping Into An Amazon Store Helps it Get Inside Your Head,” October 23, 2018, Wired, https://www.wired.com/story/stepping-into-amazon-store-helps-get-inside-your-head/, accessed November 2018.
“Amazon’s Latest Cashier-less Go Store Opens in San Francisco Today,” October 23, 2018, The Verge, https://www.theverge.com/2018/10/23/18010022/amazon-go-cashier-less-store-san-francisco-location-opens, accessed November 2018.
“Understanding The AI Behind Amazon Go,” February 27, 2018, Analytics India Magazine, https://www.analyticsindiamag.com/understanding-ai-behind-amazon-go/, accessed November 2018.
“Amazon Go, deep learning, and a better retail experience,” January 12, 2017, VentureBeat, https://venturebeat.com/2017/01/12/amazon-go-deep-learning-and-a-better-retail-experience/, accessed November 2018.
“WTF is computer vision?” November 13, 2016, TechCrunch, https://techcrunch.com/2016/11/13/wtf-is-computer-vision/, accessed November 2018.
Turk, Matthew and Alex Pentland. “Eigenfaces for Recognition.” Journal of Cognitive Neuroscience, Volume 3, Number 1 (1991). http://www.cs.ucsb.edu/~mturk/Papers/jcn.pdf, Accessed November 2018.
Word count: 794