At Mercari Group, our mission is to “circulate all forms of value to unleash the potential in all people.” To achieve this mission, Merpay has continued to develop a credit model that utilizes AI technology. In doing so, we increase the potential to deliver value to people who, for one reason or another, have been ineligible for credit.

For this edition of Mercan, we spoke with three employees at length about the unique strength of Merpay’s AI credit as well as debt collection that leverages AI. Sitting for the interview were Iori Yoshida (@yotta), who is involved in debt collection and management operations at Merpay, Yasukuni Asuke (@asuke-yasukuni), who is in charge of development related to debt collection, and Takuya Wada (@phi), who works on developing and operating credit score models.

A multifaceted approach to analysis that uses AI Credit will allow us to deliver suitable credit to more people

──To start off, can I ask each of you to explain what your role is in debt collection at Merpay?

@yotta: At the moment I work as the manager of two teams involved in debt collection, Collection Ops, and Special Debt Management. My team’s main missions are maximizing debt collection, managing risk when dealing with customers, and building safe and stable operations for debt management.

@asuke-yasukuni: I’m an engineer on one of the company’s Credit Backend teams, where I work on projects related to the development of Merpay debt collection. I also build and revise systems that the Collection Ops and Special Debt Management teams use. In addition to my work on debt collection, I also work on a team that leverages AI technologies for work optimization.

@phi: I’m on the Credit Modeling Team. I work as a machine Learning (ML) engineer in charge of developing and operating credit score models for our credit services including Merpay deferred payment.

──It’s my understanding that Merpay was the first certified comprehensive credit purchase intermediary that leverages AI credit to be certified by the Ministry of Economy, Trade and Industry (METI). Could you explain what a certified comprehensive credit purchase intermediary is?

@phi: As METI puts it, certification is a part of a framework that makes it possible for companies to perform credit checks based on such things as a user’s sizeable payment history by using methods developed independently. Certified companies use AI to analyze a user’s sizeable data set, also referred to as big data, which is acquired in the course of providing services to a user, and to perform credit checks. It’s a little long-winded to describe it that way, so we just call it AI credit internally.

@phi: Credit limits for credit cards are conventionally set based on the results of a payable amount assessment, which looks at such things as a person’s annual salary, debt, and living expenses. However, companies that have the certification that Merpay now holds can use assessment logic that has been developed independently to determine a customer’s credit line without performing a traditional payable amount assessment.

──How do our users benefit from Merpay being able to determine their credit line independently?

@phi: There are a number of benefits, but one big difference is that we expand the pool of users who can receive credit among the segment of people who have already been turned down by other companies.

Mercari is used widely by a diverse segment of the population, including housewives and students. Conventional screening methods that look at salary as a criteria don’t favor applicants like these and do not allow them to get credit. Even if they pass the screening, they usually only receive a very small amount of credit.

But trust can’t be measured merely by how much money a person makes. If an applicant is a Mercari user, we look at the items that they list and the profit they gain from their sales as income. By using AI credit, we’re able to use the big data that Mercari Group has accumulated across all of our services, allowing us to screen people from angles that had not been available until now. In this way, people like the housewives and students I mentioned earlier can also receive credit appropriate to their actual circumstances.

I think that AI credit is effective as a means of achieving Merpay’s mission of “building trust for a seamless society” as well as Mercari’s group mission of “circulate all forms of value to unleash the potential in all people.”

Using AI to estimate how long collection will take, and tailoring our response to each and every user

──When you grant a loan to a user based on their credit, at some point you have to collect. How do you leverage AI when collecting on a loan?

@yotta: For collections, we’re focusing on developing a collection difficulty model that estimates how long it will take for users who have defaulted to settle their loans. Changing how we approach this process based on the values calculated using this model reduces the workload on the operations side. Most importantly, we expect it to cut down on the stress our users experience from collection tactics like making debt collection calls when in fact they had every intention of repaying their loan.

Take a user assessed to have a low collection difficulty value—in other words, a user who is highly likely to repay their loan even if we don’t send them any reminders. If they are late repaying, we can choose not to call them for a certain period of time without there being a significant impact to overall collection rates. On the other hand, this also allows us to contact users assessed to have a high collection difficulty value early to confirm their situation and discuss payment with them.

──How accurate is the collection difficulty model?

@yotta: Looking at our actual operations, I think we can say that the model is very accurate. For example, where operators contacted users assessed as having a high model value early, it resulted in an improvement in overall collection rates.

@phi: I think that what makes it possible for the predicted collection rate to be so accurate is the diverse big data that Mercari Group holds. Like I mentioned before, at Merpay we screen users from a variety of angles other than just their salaries when granting them credit. This is why, for users who default, we are able to use a detailed scoring system to determine which users will be able to make payments by themselves, pay if they are reminded, or be unable to pay for one reason or another.

@asuke-yasukuni: Our collection difficulty model is working well at the moment, but it also has an issue; there’s currently no mechanism in place to provide more detailed feedback on debt collection. If we were to build a mechanism like that, it would be the job of the team I’m on, so I would like us to work harder on feedback going forward. Once a feedback mechanism is put in place, I think it will go on to further improve the accuracy of the collection difficulty model.

Optimizing debt collection using AI to further increase quality

──So other than optimizing collection work, are there any other ways that we’re using AI to improve our work and the user experience?

@asuke-yasukuni: Well, we’re developing a repayment consultant chatbot. When users who have defaulted on their loans consult with us about repayment, I imagine they’re under immense psychological strain. So, in order to direct users to the right information more easily, we thought it would be better to use an AI that was able to handle inquiries of this nature. We felt confident that we could use retrieval-augmented generation (RAG) technology to answer questions specifically about Merpay, but we also hypothesized that AI would be able to analyze emotions from text-based inquiries. We came up with and implemented a mock model that would read the emotions of users based on the text in their inquiries and even be able to determine an appropriately empathetic way of handling the situation and responding in kind.

This mock model was actually able to analyze our user’s emotions especially well and adapt flexibly as it handled interactions. It even won top prize at the “Gugen Conference,” an inhouse competition at Mercari built around using AI technology. However, there remain a variety of risks (i.e., legal, compliance, and reputational risks) that we still need to address, so we have yet to implement this model in our production environment.

@phi: In the field of debt collection, if an AI model gives incorrect information or suggests an inappropriate response, there’s a risk that this could cause harm to the user. That said, we are still working on development. We’re now carefully examining to what extent we can use AI while exploring how to minimize risk.

@yotta: We saw a range of potential in the repayment consultant chatbot. The nature of the negotiations we have with users is really diverse and complex. You could say that, despite dealing with such highly complicated communication, the technology showed potential for efficient debt collection while also providing a better user experience. It depends on the user, but I think that when it comes to consultations about repaying a debt, many users would choose to use the app over receiving a phone call.

──Is there anything else you would like to share about ways of using AI that you are looking into in order to improve the user experience?

@phi: If we can see ahead of time when a user will call us, I think we can expect improved work efficiency. Currently, our operators are constantly on standby without knowing when they’ll receive a call from a user. If we can use big data, AI might be able to predict which users will call us and when. Once that’s a reality, it will be easier for managers to draft shifts and operators will also be able to prepare and get themselves in the right frame of mind to take calls.

@yotta: I would like to put more energy into creating personas for our users and to provide more meticulous feedback on improving credit and our product. Fine-tuning a persona is extremely difficult. However, I think that if we leverage the big data that Mercari holds, such as a user’s transaction history, communication history, and payment patterns, our goal will be achievable.

@asuke-yasukuni: Something occurred to me as I was listening to what @yotta said just now; I think it would be a good idea if we could use AI to present to operators an optimized persona created based on user information.

When a user makes an inquiry, I think they expect the operator to have enough information about their situation to be able to handle it smoothly. Especially where Mercari is concerned, users enter a variety of information on our platform, so that we have the information about their transactions and payments. And yet, when a user sends us a query, if the operator asks the user to explain everything from scratch and the transaction information they have entered on Mercari to date is not considered at all, it’s kind of disappointing.

Ideally, when we receive a query, the AI model should be able to adeptly summarize the data on Mercari and instantly display the information needed to deal with the user’s current inquiry. Ultimately, based on this information, it would determine how to resolve the query and present the operator with advice on how to handle it appropriately. If we can make this work, I think more of our users will feel that they have had a good experience.

However, the nature of debt collection reminders and regular customer support differs greatly, so we have to be careful.

──Could you talk more about the differences between the two types of inquiries?

@phi: For regular inquiries, we talk about “communication,” whereas for reminders, we talk about “negotiation.” With reminders, it’s not a matter of simply responding to a user’s requests. We can’t say that we were successful if we don’t manage to collect on a debt.

@yotta: Yes, but I don’t think hyper-focusing on efficient debt collection is quite right either. Users can have a variety of reasons for defaulting on their debt. For instance, are temporary money problems what caused them to be late with their payment, or are they facing long-term financial troubles due to something like an illness? We have to think about the most efficient way for a user to repay their debt while at the same time empathizing with their situation. I think there’s potential for us to use AI for this. We determine the user’s situation from big data, compare it with similar cases we’ve seen in the past, and then perform analysis to determine what kind of approach would most likely allow the user to pay their debt. I think that using AI in this workflow will allow us to insert this element into our operations seamlessly.

It’s of course important to use AI technology to improve the efficiency of debt collection. However, I think one perspective that we must never forget is that optimizing our work is necessary in order to increase the quality of the operations that contribute to the user experience.

A smooth and comfortable experience regardless of whether the contact point was an AI agent or a person

──To wrap things up, I’d like to know what you think Mercari Group should aim for in order to promote AI usage at the company.

@yotta: As a company that is aggressively promoting AI usage, I think we want to be “the first penguin.” I would like us to take the examples we have of effective AI usage within the company and proactively share information about them externally.

In order to accomplish that, we need to use AI even more proactively within the company. For example, we’d like to hone our AI agents to the point where users who contact us with inquiries have such a good experience that they’re uncertain as to whether they were speaking with an AI agent or a person. Ideally, they’ll say that they weren’t sure which they were dealing with, but that the process was very smooth.

Examples like this will also spread to other companies, improving the quality of customer service not just in Japan, but around the world, which I think will bring us closer to a society that accomplishes our mission to “circulate all forms of value to unleash the potential in all people.”

@phi: Yes, that’s it exactly! Great answer. (laughs)