Where AI helps most in debt collection

Debt collectors are already using AI to call borrowers. Wired reported in May 2026 that AI agents handle millions of collections calls a month for select firms.¹ AI is doing two jobs here: helping to decide which borrowers to call, and making the calls themselves.

Of these two jobs, helping to decide whom to call arguably offers the bigger gain today. Most collections models now predict who is likely to pay. But the question that truly improves collection efficiency is whether a call will make a borrower more likely to pay, and that is where AI models can add value. In a Dutch field experiment, software built to answer that second question outperformed human judgement, while a study of AI agents making the calls found they recovered less than call centre staff did.

Many of these borrowers are in financial difficulty, and the FCA holds the firm responsible for how it treats them. However a firm uses AI in collections, it needs clear guidance and close oversight.

Segmentation models: how firms choose which accounts to chase

Firms were sorting accounts by risk long before AI. In its 2016 review of how lenders handle early arrears (TR16/10), the Financial Conduct Authority (FCA) found most firms already set their contact strategy by risk segment.² They put most effort into the riskiest accounts and made fewer attempts on customers expected to catch up without help (self-cure).

AI segmentation models refine this approach. Instead of a few broad risk bands, they use payment history, past contact and other account data to estimate each customer’s own chance of paying. But they still answer the same question the risk bands did, which is who is likely to pay. What matters for deciding whom to call is whether a call changes the outcome. Some customers would cure without a call, so ringing them wastes effort. Others are the opposite and would pay only if someone rang them. To tell these two groups apart and improve collection efficiency, a firm needs a model that estimates how much a call would change each customer’s chance of paying. AI makes this practical by learning from each customer’s own history.

Evidence from the field

We looked at two studies that test AI in collections.

The first study, by Zhou and Wang, took place at a Dutch collection agency in 2020. 7,839 delinquent borrowers were split randomly into two groups. For one group, human collection officers chose whom to call each day. For the other, software made that choice by estimating how much a call that day would raise each borrower’s chance of repaying. The same team of human agents called both groups, without knowing which list they were working from.

When the software chose whom to call, 53% of borrowers repaid, compared with 43% when people chose.³ The software’s group also needed fewer calls per borrower. The same agents called both groups, so the difference came from the software making better choices about whom to call.

The second study, by Choi and co-authors, tested AI agents making the calls at a large online consumer lender in China, on cases that entered collection between April 2021 and December 2022. Borrowers could easily tell the AI voice was not human, and the lender recovered less when AI agents handled the calls than when human agents took them over.

The lender also ran a randomised test on larger debts. Half the accounts were handled by people from the start of collection. The other half were handled by AI agents for the first four days and then passed to people. After a month, the group that started with AI had recovered slightly less, and the gap was much the same 11 months later.⁴ Even counting the salaries saved, the AI agents were less cost-effective than people at that lender.⁴

These results come from AI voices built in 2021 and 2022. Today’s voices already sound more natural and are improving quickly, so we expect AI agents to handle a growing share of collections calls in future.

Putting AI to work in collections

Drawing on Gini’s own experience, the most effective use of AI in collections today is helping to decide whom to call, where the empirical evidence is also the strongest. As noted above, we expect AI voice agents to catch up. The published evidence does not yet show them making collections more efficient, and little has been published on how the newest voices perform in practice. As they improve, it could make sense to use them for the calls themselves, provided firms set clear rules to protect the customers they speak to. Until then, human agents should keep making the calls, with AI helping to decide which calls to make.

A firm can start with the data it already holds, such as account and contact history. A model that ranks borrowers by the difference a call would make can be built from that data and introduced cautiously, for example by letting it choose a random share of each day’s calls, or by running it in shadow beside the current strategy and comparing its choices.

A more advanced step is to bring what borrowers say on calls into the model, since conversations reveal intent and circumstances that account data cannot. The Dutch study tested this directly. Its human agents typed a short note after every call, and when the AI model could read those notes alongside the account data, it got better at predicting who would repay, even for borrowers it had never seen.

The most useful notes were about the things an experienced collector already listens for. These included signs that the borrower was in financial hardship, a statement that they intended to repay, and talk about what would happen if they did not pay. However, some of these cues point to customers who may need support, and a model focused only on recovery may give them less attention than they need.

Reading conversation notes therefore brings a risk as well as a gain. In the Dutch study, the software that decided whom to call became less likely to call borrowers whose notes recorded hardship.³ Calling them less may improve recovery rates, but it leaves the most vulnerable customers with the least contact. The FCA’s Consumer Duty guidance (FG22/5) expects firms to justify any outcome that is consistently worse for one group of customers.⁵ Hardship and vulnerability cues should therefore be used to route those customers to a support team.

Call recordings could supply the same signal without relying on each human agent to decide which details to write down. Each call can be converted into a small, fixed set of fields, such as "promised to pay" or "reported job loss", which can be audited and checked against the recording. Keeping to these fields matters. A large language model can take in and produce almost any text, so left to write its own summary it may add things the borrower never said. Much of the work is limiting what it records, so the output is repeatable and someone can be held accountable for it.

However, moving from typed notes to recordings brings its own risk and legal requirements. The risk is transcription error. Each call has to be converted to text by a speech-to-text model, and telephone audio is hard to transcribe. The developers of Whisper large-v2, an OpenAI speech-to-text model, reported that it got 13.8% and 17.6% of words wrong on two sets of recorded telephone conversations.⁶ A 2020 US study of five widely used commercial speech recognition systems⁷ found their error rates differed between speakers, so any system needs careful testing across accents and dialects before its output reaches a model.

Data protection law adds further requirements. Using recordings to predict repayment is a new use of data that was originally collected for reasons such as staff training and quality checks. Before any work starts, a data protection officer should confirm the firm has a legal basis under UK GDPR to use them in this way (Article 6).⁸ Furthermore, recordings can contain health information, such as a borrower mentioning an illness, which UK GDPR allows firms to process only under specific circumstances (Article 9).⁹ The firm needs one of those conditions in place before using them.

Governing AI in collections

Data protection is one of several checks a firm needs before it lets AI help decide whom to call. AI can make those choices more accurate, but the firm stays accountable for every one of them, so it needs clear rules for what the model may do and oversight of the results it produces. Before the model goes live, those rules should set out which decisions the model makes alone, how often a borrower may be contacted, and how hardship and vulnerability cues reach the support team. Once it is running, a named person in the firm should be responsible for its decisions, its results should be checked regularly, including whether any group of customers is consistently worse off, and its decisions should be recorded so they can be reviewed and challenged.

AI voice agents, when a firm is ready for them, need the same care before they speak to customers. A firm should test them for bias, check how customers are treated on the call, and make it easy for any customer to reach a human agent when they need one.

With clear rules and oversight in place, AI can make a collections firm more effective and provide more opportunities to look after struggling customers well.

Frequently asked questions