AI for Retail

CHAPTER 21AI for Checkout

Current Roles, Processes, and Inefficiencies

I could write an entire book on this subject in lieu of this chapter, but I will do all I can to summarize my thoughts as best as possible.

From day one at Focal, we created two product teams, one to apply AI to inventory and another to apply AI to checkout. For checkout, we tried a host of solutions. We tried AI on shopping carts, AI‐powered shopping cart bays, AI on the conveyor belt, Amazon Go format, you name it, we tried it. We spent nearly $10m in R&D over four years trying to beat the economics and accuracy of a self‐checkout machine, and despite our AI prowess, we just couldn't win, so in 2019, we canceled all AI checkout product development.

The role of checkout could be summarized as the following task: to accurately assign the products a customer wishes to buy to that customer and verify that they have purchased them before letting them leave the store.

Whittling down to specific success criteria, in the ideal case, our checkout method would have the following features:

• Be low cost per transaction (or depreciation cost if there exists a large up‐front cost);
• Work for all products (frozen, big, small, produce, weighted, etc.);
• Be very fast and convenient for the shopper (should not be confusing to use, should not delay the shopper, etc.)—for this, we introduce a term that we want to minimize called total checkout latency, which is the time from when a customer decides to check out to the time ...