At Sam’s Club, an AI algorithm tells workers how many Key lime pies to prepare based on factors like weather and what other items might be out of stock. At Home Depot, an algorithm tells workers which items to restock first, based on demand. And Chipotle is developing an algorithm designed to predict tortilla chip demand, and is creating a robot named Chippy that could ultimately make them.
The algorithms have a number of goals, including making employees’ jobs easier, saving on costs and getting more done in a shorter amount of time. But in each case, companies are confronting workers who have their own expertise and routines with AI recommendations that aren’t always 100% accurate. And it is causing friction.
“Our members have told us: listen, we feel like guinea pigs,” said Ivana Saula, research director for the International Association of Machinists and Aerospace Workers, or IAM. She said workers have the opportunity to complain and give feedback and watch the tools improve, but still have to deal with poor early iterations.
As AI expands in scope and capability, especially thanks to rapid developments in generative AI, the role that advanced tech could play in many jobs across a swath of industries is only expected to increase. A recent study by researchers at the University of Pennsylvania and OpenAI found that around 80% of the U.S. workforce could have at least 10% of their work tasks affected by new AI capabilities.
That means the stakes are high for how accepting employees will be of new tools and exactly how businesses go about deploying them.
Getting feedback is a critical part of the development process, said Pete Rowe, vice president of merchandising & AI labs at Sam’s Club. He said that as the algorithms improve in accuracy and functionality, workers build up more and more confidence in them. But, he added, it takes time.
About five years ago, the warehouse-club chain, owned by Walmart, started rolling out algorithms that predicted daily, or sometimes more frequent, demand for more than 100 freshly-prepared items at an individual club level.
The algorithms would use factors like weather, seasonality, local events, promotions and what other items were out of stock to tell associates how many croissants, rotisserie chickens or Key lime pies they should prepare to meet demand and minimize waste. But the company quickly found that associates often weren’t following the AI recommendations.
“They didn’t trust it yet. And I understand it, because they were in the situation where—if you don’t produce enough rotisserie chicken, my AI engineers are not the ones dealing with angry members,” Rowe said.
Part of the friction around building trust comes when the algorithm is supplanting an existing worker’s expertise. For example, earlier this year, Home Depot rolled out an algorithm designed to tell store associates what order to restock items in, said Paul Antony, senior vice president of technology. But sometimes a manager would want associates to do things in a different order than the AI did, creating a dilemma for workers, he said.
“I think part of it is they feel like they have expertise, right—you feel like you know what you’re doing. And in the past, our systems may or may not have done the best job,” Antony said.
“My expectation is, the very experienced people are going to feel like they know better than the tool,” said Chipotle Chief Customer and Technology Officer Curt Garner who is working on an algorithm that predicts how many chips to prepare in a given restaurant. However, he said, less experienced or newer employees will be more likely to follow the tool closely, and therefore it still create benefits for the company.
In other cases, the barrier is a lack of convenience. For example, in 2012, Sephora rolled out a tool that let store workers take photos of customers’ skin tones and then used an AI algorithm to match them with the appropriately-colored foundations from Sephora’s more than 8,000 offerings. But at the time, it was a clunky piece of hardware permanently attached to a corner of the store, said Sephora Pro Artist and Manager of Client Experience Programs Shawn Lumaban. After Sephora relaunched the tool as a smartphone attachment in 2021, usage increased, he said.
But oftentimes the core of the problem is when end-user employees aren’t consulted early on enough, IAM’s Saula said.
“Consultation means I’m part of the process from the beginning. Not when you’ve already taken three steps into the process,” she said “That’s where some of the bad blood is created, and why workers tend to distrust employers and distrust the technology that they’re using.”
Buy-in is still achievable, even after an initial lack of trust, Rowe at Sam’s Club said. Over the last five years, it has taken time, patience and education to build that, he said. The company collected tons of feedback from associates, continued to refine the accuracy of the tool, and in some cases, tweaked the way certain information was presented, as well as continuing to teach workers about the algorithm’s capabilities.
Rowe said it was also important to ensure the associates felt they had ownership over their workflow, which is in part why they still maintain the freedom to make changes to the algorithm’s recommendations. But, he said, today they are making changes much less frequently than they used to, and the company has ultimately made excellent progress in terms of achieving that trust.
“We work with our associates to show that the machine is what we want them to consider and use, but we don’t penalize them and we don’t discourage them saying: never override,” he said. “The point is: let’s earn the trust of our associates that the machine’s making their life easier.”
