Artificial intelligence is rapidly transforming consumer financial services, and consumer debt collection is no exception. AI already is being used in a variety of first-party and third-party consumer debt collection activities, ranging from account scoring and communication strategies to payment-plan optimization and compliance monitoring.

In our Consumer Finance Monitor podcast show being released today, our special guest John McNamara, Chief Growth Officer at Avtal (a fintech and software as a service (SAS) company that provides an AI-powered, white-labeled digital engagement platform to help third-party consumer debt collection agencies automate communication and process self-service payments) and  a former CFPB senior official who played a significant role in developing Regulation F promulgated under the Federal Debt Collection Practices Act explained that the debt collection industry needs to distinguish genuine AI applications from the marketing hype surrounding the technology. He also emphasized that the use of AI must be accompanied by careful attention to data governance, explainability, consumer protection, and human oversight.

Our show is hosted by Alan Kaplinsky, founder and former leader for 25 years and now Senior Counsel of our Consumer Financial Services Group.

Where AI is being used in debt collection

According to McNamara, several AI applications are already in production.

One is account scoring and segmentation. AI can analyze large data sets to help determine how accounts should be segmented and how collection strategies should differ among consumers.

Another is determining when and through which channel to communicate with a consumer. AI can analyze interactions and other available information to help identify the timing and communication channel most likely to result in productive engagement.

AI also is being used for payment-plan, settlement, and other optimization. McNamara views these applications as among the less risky uses of AI because they generally involve analyzing information to determine how best to engage a consumer rather than having AI interact directly with the consumer.

Perhaps the most interesting development, in his view, is the ability to combine pre-charge-off and post-charge-off data to identify patterns that may help a collector determine how to engage a consumer about resolving an outstanding debt.

AI also has potentially important back-office applications. One that McNamara highlighted is compliance monitoring. Historically, collection agencies could monitor only a small percentage of their consumer interactions. AI can potentially analyze virtually all calls, emails, and text interactions, including tone, sentiment, and other characteristics, making it possible to identify problematic practices much earlier.

Consumer-facing AI presents greater risks

The more AI moves from analyzing information to directly interacting with consumers, the greater the risks become.

Examples include AI chatbots, outbound voice bots, and “agent assist” systems that provide real-time guidance to collectors during conversations with consumers. Skip-tracing applications also can raise concerns if AI is used to exploit information about consumers in ways that may be viewed as unfair or deceptive.

McNamara expressed particular caution about outbound AI voice bots. He noted that the Supreme Court’s decision in Facebook v. Duguid has implications for calls using artificial or prerecorded voice technology, including the continuing risk of litigation under the Telephone Consumer Protection Act.

Generative AI also presents challenges when it is used to communicate directly with consumers. Collection communications must comply with applicable disclosure and substantive requirements, and AI-generated communications can create additional risks if the system does not understand the legal significance of what a consumer says.

For example, an AI system needs to recognize that an unusual or colorful statement by a consumer may nevertheless constitute a refusal to pay. That may be easy for an experienced human collector to understand but considerably more difficult for an automated system.

AI can also improve consumer outcomes

The risks should not obscure the potential benefits of AI for consumers.

One potentially significant benefit is the ability to give consumers more control over how and when they interact with collectors. McNamara noted that consumers increasingly may prefer asynchronous communications through email or text rather than intrusive telephone calls.

AI and data analytics can help identify consumer preferences regarding communication channels and timing. Instead of repeatedly calling a consumer in an effort to make contact, a collector potentially can communicate through the channel the consumer is most likely to use.

AI also can help identify patterns suggesting that a particular communication is producing a negative reaction. Collectors can analyze consumer responses, including tone and other characteristics of interactions, and use that information to determine whether subsequent communications are improving or damaging the relationship.

Another potential benefit is giving consumers greater visibility into their options for resolving a debt. Rather than approaching collection as a binary choice—pay the entire amount or do nothing—digital engagement can present available payment plans, settlements, and other options in a manner that allows consumers to consider them at their own pace.

Data governance may be the most important issue

McNamara identified data governance as one of the most important considerations for companies deploying AI.

AI systems generally perform better when they have access to more data. But that does not mean a company should provide an AI vendor with every piece of information in its possession.

As McNamara put it, “with more data sharing is more risk.” Companies need to understand what data an AI system is using, where the data came from, why the company is permitted to possess and use it, and whether additional disclosures or other legal requirements may apply.

This issue can be particularly important in debt collection, where agencies may have decades of accumulated consumer information. The fact that information is available does not necessarily mean that it is appropriate or necessary to feed that information into an AI system.

McNamara suggested a straightforward framework: start with the use case, identify the risks associated with it, and then determine what data is actually necessary to accomplish the objective.

In other words, companies should resist the temptation to “dump everything” into an AI system simply because a vendor says that more data will produce a better result. The incremental benefit of additional data should be weighed against the additional legal, compliance, privacy, and reputational risks.

The “black box” problem

Another important issue is explainability.

McNamara drew a parallel to the use of AI and alternative data in lending. A company cannot simply say that an AI system produced a particular result and leave it at that. Companies need to understand enough about how their systems operate to explain and defend important decisions.

The same principle applies in debt collection. Companies should understand what an AI system is doing, what information it is relying upon, and why the system is producing particular recommendations or outputs.

This does not mean that every AI system must be completely transparent in every respect. Rather, companies should understand the boundary between what an AI system can do and what the company can explain and defend. If a particular application requires a degree of explainability that the system cannot provide, the company may need to limit the data or functionality being used.

Hallucinations and the importance of a human in the loop

Generative AI’s potential to produce inaccurate information, often referred to as “hallucinations”, is another obvious concern when AI interacts with consumers.

McNamara’s principal safeguard is simple: keep a human in the loop.

This principle is particularly important in debt collection because an apparently minor error in an automated communication could have legal consequences if it results in inaccurate information being provided to a consumer or causes a collector to take an inappropriate action.

Vendor management remains critical

Companies also need to understand how third-party AI vendors operate.

A creditor or collection agency should not assume that a vendor’s AI solution is automatically compliant merely because the vendor describes it as “AI-powered.” The organization deploying the technology remains responsible for understanding the use case, the data being supplied to the system, and the resulting risks.

McNamara cautioned against overly prescriptive requirements imposed by creditors on collection agencies. For example, restrictions on the use of email and text messaging can sometimes have the unintended consequence of forcing collectors to rely more heavily on telephone calls, which may be more intrusive and generate more complaints.

The better approach, in his view, is for creditors and agencies to understand what the technology is intended to accomplish and to establish appropriate principles and guardrails rather than imposing requirements without understanding the operational consequences.

AI may become infrastructure rather than a separate technology

McNamara’s long-term prediction is that AI eventually will become less remarkable precisely because it will become ubiquitous.

He compared AI’s development to earlier technologies such as ATMs and cloud computing. Technologies that once generated considerable concern eventually became ordinary infrastructure.

The same thing may happen with AI in debt collection. Rather than thinking about AI as a separate category of technology, companies may increasingly use AI as an embedded component of ordinary collection operations.

That could produce a significant shift toward a more consumer self-service model of debt collection, in which consumers can interact with collectors digitally, on their own schedules, and with greater visibility into available options.

At the same time, McNamara does not expect humans to disappear from debt collection. He anticipates that human agents will continue to be needed for escalated and complicated situations, even if the industry has less need for large numbers of agents handling routine interactions.

Conclusion

AI presents debt collectors with substantial opportunities to improve both operational efficiency and consumer engagement. But the technology also raises familiar, and some new, legal and compliance questions.

The central lesson from McNamara’s discussion is that responsible deployment should begin with the use case rather than the technology. Companies should ask what they are trying to accomplish, what data is actually necessary, what legal authority they have to use that data, whether the system’s output can be understood and defended, and where human oversight is required.

If those questions are addressed carefully, AI may ultimately help move debt collection away from repeated, intrusive attempts to reach consumers and toward a more personalized, digital, and consumer-directed process.