Linking wages to performance increases the use of AI in decision making

Artificial intelligence has improved dramatically over the past few decades, changing the way many people, including corporate managers, do business.

However, the use of algorithms in management decision-making is not universal and there are several factors that encourage greater use of AI. According to a new study co-led by Cornell researchers, how managers pay and how artificial intelligence can be framed.

Contrary to highly cited research over 30 years ago, incentive wage structures lead to greater reliance on AI in decision-making than flat, fixed compensation. Also, when AI is described as combining both data and human expertise, people are more likely to use it than if they were framed as strictly algorithmic advice.

Martin Wiensparger, assistant professor of accounting at Samuel Curtis Johnson School of Management at Cornell SC Johnson College of Business, is co-author of “Reliance on Incentives, Framing, and Algorithm Advice: An Experimental Study.” Management Science. His co-author is a graduate of the Vienna University of Economics and Business.

Two studies in the 1980s and 90s feel that incentivized decision makers (paid based on performance or paid in a “tournament” setting where the top performers of the group get paid, less in algorithmic advice and instead are forced to “earn” rewards through their own efforts. In other words, the backstroke of incentives due to a phenomenon known as “algorithm aversion.”

“They are very well-known studies on the paradoxical effects of incentives and decision-making, and when we have very different types of decision support and algorithms, I thought it was probably worth reconsidering this question,” Wiernsperger said. “We wanted to see if this backfire was still held.”

He and co-authors Philip Grünwald and Georg Lintner were all doctoral students in Vienna when they attended seminars in 2021 and were assigned to work on the project together. They designed an experiment to test how AI reward structure and framing affect decision-making at a particular task. Estimates of night rates for Airbnb apartments. Other co-authors, Professors Ben Greiner and Professor Thomas Lindner, met students at the seminar and proposed to turn the project into a research paper.

For their research, the researchers recruited approximately 1,500 participants from three large universities in Austria. Subjects were randomly assigned to one of nine experimental conditions. Performance payments; or tournament payments, and within each of those terms: no AI advice. AI advice; or human advice.

Strictly artificial intelligence framing via algorithms or algorithms with the involvement of human experts was a key factor in whether decision makers trust AI under certain conditions.

Participants were presented with a list of 10 Airbnbs from their Vienna apartments. This was extracted from a dataset of approximately 12,000 apartment lists where the algorithm was trained, and was given all relevant list information except prices. They were tasked with estimating the night rates for all 10 people. In the AI-Ai-Advice group, participants were required to estimate prices based on the list information they received.

However, both AI groups performed two estimates. The first group is only list information, and the second is list information and algorithm advice. It determined the “weight of advice.” How much does this affect the impact that advice (in this case, algorithmic help) has on people's decisions?

Researchers found that individuals compensated based on either performance or tournament incentives rely more on AI than those who received fixed payments.

Also, those who used algorithmic advice performed a task that better estimates the nighttime rate than those who did not receive AI support, regardless of how it was framed.

“In general,” Wiernsperger said, “When it comes to AI use, managers or decision makers have a positive effect based on performance, and no negative effect.”

These results make sense for companies looking to implement AI in their decision-making, researchers said. If you want people to use AI tools, how to motivate them and talk about technology is both.