AI debt collectors fare worse than humans

While loan servicers are eyeing AI to make collection calls cheaper and faster, a recent study suggests non-human debt collectors may struggle getting borrowers to cough up cash.

The collections industry is still slow to adopt generative, large language AI models like ChatGPT, Tim Collins , member of the Association of Credit and Collection Professionals , told Auto Finance News .

However, he said, loan servicers are interested in future use of generative AI debt collectors, which call borrowers, inform them of their due payments and ask them to pay.

“Everybody is looking to the AI side — the voice AI especially — because they associate voice with debt collection,” Collins said.

Lenders, retailers test bots Major lenders and retailers have also signaled their interest in AI-powered collections. At the Non-Prime Auto Financing Conference on June 4, SAFCO Chief Executive George Fussell predicted that outsourced collections agents could be replaced with AI in the future.

Bridgecrest , the servicing partner for retailer Carvana , is also testing an AI chatbot to encourage payment activity through text and analyze customers’ financial situations.

Chatting with an AI agent could offer a safer space for customers to talk out why their payments are late, Alan Kibler , associate director of strategy and analytics at Carvana, said last month at Auto Finance Summit East 2025.

“If you think about the situation that customer is in, they probably don’t want to talk to someone about it,” he said. “It’s a much more comfortable conversation to have, potentially, with an empathetic AI agent that also knows all your options.”

Some lenders working with electronic payment processor REPAY use AI debt collectors for initial calls with borrowers. So far, human agents are taking 15% to 20% fewer initial collections calls due to AI, providing them with more time to work on complex cases, REPAY Enterprise Account Executive Lesley Hastings told AFN.

AI can easily optimize initial collections calls, Maggie Pugesek, partner at C&M Coaching , which trains auto dealers and lenders how to handle collections, sales and service calls , told AFN .

“A lot of times your collectors are basically cashiers. ‘Yeah, Hi, I’m calling to make a payment.’ Nice and simple stuff — we don’t really need collectors to take those calls,” she said.

Occasional human “check-ins” with borrowers who are mainly talking to AI is still useful for verifying information and maintaining a relationship, Pugesek added.

Can AI twist arms? While they free up time, AI debt collectors’ ability to get borrowers to repay their debts may be weaker than humans, according to a National Bureau of Economic Research working study published April 14.

The study, which took data from a finance company in China that was using both AI and human agents from December 2022 to December 2023, found:

AI callers collected 9% less money from repayments than human callers during the first 30 days; AI callers collected 5% less money from repayments after a year; Borrowers were less likely to make repayment promises to AI; and Borrowers were less likely to keep promises they made. Borrowers don’t feel as obligated to pay when they are talking with an AI agent, Dong Huang , a finance doctoral student at Yale University who co-authored the study, told AFN .

“Those promises are just made with some emotionless machines instead of humans that are watching them and judging them,” he said.

The cost of paying human callers is about even with employing AI for smaller-scale debts, Huang said. However, for larger debts, “AI still underperformed humans even after considering the labor costs,” he said.

Initial contacts with AI can also have long-lasting damages on borrowers’ repayments, according to the study.

The case for humans For some debts, a human caller would take over a case after an AI caller was on it for more than five days. Human callers were able to improve the AI’s underperformance over time, according to the study.

However, borrowers who never interacted with AI performed slightly better than borrowers who initially had contact with AI. Non-AI borrowers continued to perform slightly better for up to a year, according to the study.

That finding suggests that even as generative AI continues to improve, knowing a debt collector is not a human will hurt repayments, Huang said.

“As long as the borrower knows that they are talking to some AI, it will make a difference,” he said.

Customers need to trust their collectors, which is often better achieved through human interaction, C&M Coaching’s Pugesek said.

“When customers feel comfortable with your collector, they’re more likely to reach out to you,” she said. “The computer isn’t going to be able to establish that relationship and build that rapport.”