AI vs. Human Debt Collectors: Why Borrowers Still Respond Better to a Real Voice

A large-scale study by Yale, Tsinghua, and Shanghai Jiaotong finds AI is significantly less effective than humans in debt collection, especially in eliciting moral commitment and long-term repayment. Even brief AI contact causes lasting damage to borrower cooperation.

In a world increasingly dominated by artificial intelligence, a critical question looms: can machines replace humans in jobs that require emotional intelligence, moral persuasion, and interpersonal finesse? A new study from Yale School of Management, Tsinghua University’s School of Economics and Management, and Shanghai Jiaotong University tackles this question head-on by analyzing AI’s effectiveness in one of the most psychologically demanding domains debt collection. Drawing from over 22 million loan cases at a major online consumer finance firm in China, the researchers find that AI still lags far behind human agents when it comes to persuading people to perform difficult, personally costly actions like repaying delinquent loans.

The Performance Gap Between Humans and Machines

The study implemented both natural and randomized experiments to assess the effectiveness of AI versus human callers. In a regression discontinuity design, loans just below 300 yuan were assigned to AI callers, while those just above were handled by humans. Over time, this provided a clean comparison. The findings were stark: by day 30, the net present value (NPV) of repayments collected by AI was 9 percentage points lower. Even after a full year, the gap was still around 5 points. A separate randomized trial gave some borrowers five days of AI calls before handing them off to humans. These borrowers paid back 1% less than those who were always contacted by humans, a small figure, but one that proved persistent and costly over millions of accounts. Just four days of AI contact appeared to impair a company’s long-term ability to collect.

Why AI Struggles With Moral Persuasion

The researchers dug deeper to understand the roots of AI’s underperformance. They found that humans were 21 percentage points more likely to secure a verbal promise to repay. And crucially, borrowers who made such promises to humans were far more likely to fulfill them. A promise made to an AI simply didn’t feel as binding. The moral and psychological weight of a human interaction, it turns out, still matters. Borrowers likely perceive less guilt or social consequence when reneging on a commitment to a machine. Additionally, AI conversations were shorter by an average of 31 seconds and less varied, suggesting AI was less capable of adapting to the emotional or informational needs of the borrower. Importantly, these differences persisted even when adjusting for call timing, frequency, and borrower profiles.

Even Advanced AI Still Misses the Mark

The company periodically upgraded its AI software during the study period, particularly improving speech recognition and language understanding. These enhancements led to modest gains in early repayment rates. For instance, one upgrade boosted NPV collections by 3 percentage points within five days. However, this benefit was offset by a new problem: easier cases were resolved by the improved AI, leaving more difficult ones for human collectors. These tougher calls yielded lower returns, resulting in no net improvement in overall collection outcomes. Instead of AI complementing human labor, the study found signs of displacement AI handling more straightforward tasks but not making human workers more effective. Even more telling, the long-term alienation caused by AI contact did not diminish with newer versions. Borrowers contacted by AI remained more likely to miss each of their next twelve monthly payments, regardless of software updates.

The Human Voice Still Holds Power

From a cost-efficiency perspective, AI appeared to be more competitive on small loans, where human effort yields diminishing returns. When accounting for direct labor costs, salaries and phone time, AI narrowed the performance gap on smaller cases. However, for larger debts, human callers remained decisively more effective. Notably, the researchers only considered direct costs, excluding both the hidden overheads of managing human labor and the substantial investment required to develop and maintain AI systems. What truly set the results apart was how damaging AI contact proved in relational terms. Borrowers who first dealt with AI were consistently more likely to miss future payments, while borrowers contacted by less effective human agents, such as weekend callers or newer employees, typically returned to normal repayment patterns. The enduring damage from AI was unique.

The research delivers a sobering counterpoint to the often-hyped promise of AI automation. While machines may rival or even surpass humans in analytical, predictive, or repetitive tasks, they still lack the capacity for emotional intelligence and moral resonance. Debt collection, a profession rooted in social cues, guilt, trust, and negotiation, remains firmly in human territory. As AI continues to evolve, there’s no doubt its reach will expand. But if the goal is to convince someone to make a sacrifice, such as paying up when they’d rather not, the human voice continues to carry power no algorithm can yet match.