Skip navigation
10 Items or Less
artificial intelligence monsitj/iStock/Thinkstock

Can artificial intelligence solve e-commerce inefficiency?

Instacart’s deep learning; how not to speed checkout

This post is part of the 10 Items or Less blog. 

E-commerce technologies and brick-and-mortar supermarkets can’t stop running down the aisle together, but from my seat in the pews, it appears a marriage borne more out of necessity than true compatibility.

Take, for example, the fundamental inefficiencies inherent in generating fast and precise shopping trips — a necessity of any store-based e-commerce effort with any prayer of profitability — with the fact that the setting for them was designed in most cases to facilitate leisurely trips and unplanned impulse buys. That’s a challenge for fast-growing click-and-collect operations at places like Kroger and Walmart, and for concierge shoppers like Instacart and Shipt, that use the local grocery store as a warehouse.

Job Portraits

Jimmy Stanley, Instacart

For Jeremy Stanley, VP of data science at Instacart, one way to a happy marriage is artificial intelligence. In a recent interview, he described to me how the San Francisco-based company has used machine learning to direct its shoppers to take the most efficient routes through stores while shopping on behalf of its customers.

Instacart provides shopping and home delivery in a variety of stores — some with the ability to track and locate the location of every item on the shelf, others with inventories that might include a truckload of watermelons in the parking lot. Or as Stanley puts it, “some are like warehouses, some are like haunted houses.” Oftentimes, staple items like milk are deliberately placed in the rear of the store. That, of course, can encourage traditional shoppers to wheel their carts past a variety of profit-making items but is an enemy of those for whom shopping time is money.

“In a grocery store all of the yogurt is presented in front of the shopper in a single frame — there’s 1,000 different flavors and varieties. If you want something specific it can be very hard to find it. If you were to design a store for picking speed, you wouldn’t have that,” he said. “You’d have on the shelf a particular yogurt, sitting next to a soda, sitting next to a cereal, so if you wanted to get a particular yogurt it would be the only one there to grab.”

Given that backdrop, Stanley led a process at Instacart that over the past several months tested various means to suggest how its employees shop for items on a list most efficiently, ranging from an alphabetical list used as a base; to a route based on clusters of items mapped in a store (“the traveling salesman route”); and finally to an artificial intelligence approach that used data from the company’s most efficient shoppers — oftentimes, store stockers who “moonlight” taking items off the shelves for Instacart — to predict a sequence of picks that would be the most efficient. 

Explaining exactly how Stanley’s team crushed its data is a subject for scientists, but Stanley said it was not unlike the way other complex questions are being tackled today. “Increasingly, people are using AI to solve hard problems like translation. How do you translate an English sentence into a French sentence? Well, it turns out that algorithm we’re using to try and figure out ‘how do you sort 50 items to mirror how a human would pick them in a store?’ is very similar to ‘how do you translate a sentence?’” 

The results need little translation. Instacart found shoppers utilizing the “traveling salesman” approach to be twice as fast as those using a random, alphabetized list — but that the AI approach was twice as fast as the traveling salesman route (Stanley declined to specify a per-item pick time, saying the data was confidential). Using AI, Instacart is able to “guess” the next item a shopper will pick 60% of the time.

“It’s not 95%,” he confessed, “but there’s room for variance and error. When we look at the overall sequence it mimics what the shopper does very closely, and usually only reverses a few items per trip.”

About a month ago, Instacart rolled out AI-sequenced shopping lists as the default order sorter for all its shoppers. Stanley noted that its shoppers aren’t required to follow the suggested order — some, he noted, like to scroll from the bottom — but he said the technology is winning their trust, and trust is one part of a good marriage.

Bad Idea

The Coles supermarket chain in Australia this week dialed back a campaign designed to call attention to open check stands to meet an expected Easter shopping rush. It’s not that the service initiative didn’t work, but promoting it by asking its checkers to hold up signs saying “I’m Free” led to some staff members complaining they were subsequently harassed by shoppers checking them out, the Sidney Morning Herald reported.

Hide comments


  • Allowed HTML tags: <em> <strong> <blockquote> <br> <p>

Plain text

  • No HTML tags allowed.
  • Web page addresses and e-mail addresses turn into links automatically.
  • Lines and paragraphs break automatically.