AI collections in banking is one of the easiest business cases to build, starting with early delinquency. In the Turing Bank example for auto loans, the goal is to raise response from 7% to 14% and collectability from 4% to 7%, recovering USD 3,600,000 extra over the benchmark.
We know artificial intelligence offers major opportunities and benefits for every industry, and in particular for the financial sector. The volume and speed of information is almost overwhelming, as the big consulting firms say, at every event and across social media. What usually proves hard, however, is identifying how to start in a practical, realistic way: the process of building a business case and testing a hypothesis on how this technology can improve an existing process that has specific current results and is part of the business. We already covered why so many smart chatbots still frustrate banking customers ; the next step is choosing a case with measurable impact. In this blog post we'll walk through a recurring case in the banking industry that we find fairly simple to build, and the one many of our clients choose to begin with: managing early-stage delinquency portfolios with generative AI . Current limitations vs. the new technology To build our hypothesis, we can first analyze the value generative AI can add, and then use that as the foundation for the business case. In managing a delinquency portfolio, we can highlight what an AI collections agent brings: Negotiation capability: Holding a natural back-and-forth conversation with the customer where, using information pre-approved by the bank, the conversational agent "works" within a negotiation threshold. Handling queries off the happy path: Answering questions that come up outside the ideal negotiation flow when the customer has doubts or concerns about security, product details, complaints, and so on. Scalability and efficiency: Handling thousands of conversations in parallel, at any time, at a marginal cost compared with other contact channels. Inclusion: Adapting to language and to the different formats on WhatsApp (audio messages, spelling errors, local languages or dialects) without trouble. The starting point The next step is to define where we are today, so we have a clear reference for what we want to improve and so the hypothesis becomes measurable. In this case, we'll use the example of a fictional bank, Turing Bank, with which we'll run a proof of concept for auto loans of our early-stage delinquency portfolio management agent, deployed on WhatsApp. We always recommend running a proof of concept as the initial step, one we can use as a reference to test against our hypothesis and that also serves as a reference for scaling to the rest of the use case. To define the current benchmark, we need to define the KPIs or key metrics of the business case. For example, in collections management we'll use: Response rate : the percentage of customers who reply to the contact. Collectability rate : the percentage of customers who bring their delinquency current. Amount recovered from the portfolio : the monetary value we manage to regularize from the delinquency portfolio, which directly affects the loan-loss provision in the P L. Turing Bank's benchmark For Turing Bank, we'll set the following numbers for auto loans as the current benchmark: Total customers in delinquency: 8,000 Average ticket: USD 15,000 Total provision: USD 120,000,000 Response rate: 7% Collectability rate: 4% Amount recovered: USD 4,800,000 These numbers must reflect the current reality of the business, since we'll use them as the reference to validate our improvement hypothesis or not. The goal to reach With the benchmark defined, we can set the objectives and quantify them to estimate the potential benefit of the business case. For Turing Bank, we'll set these objectives: Response rate: 14% (100% performance improvement vs. the current benchmark) Collectability rate: 7% (75% performance improvement vs. the current benchmark) Amount recovered from the portfolio: USD 8,400,000 (USD +3,600,000 on top of the current benchmark) These values will act as the judge of the new solution we'll test, and then the agent's real performance will give us results that we'll position against this measurement, in order to draw our conclusions. Controlled rollout Although it goes beyond building the business case, it's also worth mentioning how we suggest testing this type of implementation. Beyond selecting a specific product, in this case auto loans, we recommend working with a controlled population , avoiding any disruption to operations and preserving the health of collections. In previous implementations, we usually see tests of 2 to 3 months with a sample of 20% to 30% of the total portfolio. This runs in parallel with current methods to get a clear comparison. With a platform designed for your bank , this kind of test fits into operations without friction. If you made it this far, thanks for reading! If you'd like to dig deeper into this use case or another in the banking sector, write to us and we'll set up a demo.
Why start with early-stage delinquency collections for a generative AI business case in banking? It's a contained, measurable process with clear KPIs (response rate, collectability rate, and amount recovered). That lets you build the hypothesis quickly, test it against a current benchmark, and scale to the rest of the use case once it's validated.
Which KPIs are used to measure a generative AI collections agent? Three metrics: response rate (customers who reply to the contact), collectability rate (customers who bring their delinquency current), and amount recovered from the portfolio, which directly impacts the loan-loss provision in the P&L.
How is the rollout recommended to be tested? With a controlled population: tests of 2 to 3 months over 20% to 30% of the portfolio, run in parallel with current methods for a clear comparison without disrupting operations.
What is AI collections in banking? It is the management of delinquent portfolios through generative AI conversational agents that contact customers on channels like WhatsApp, negotiate within thresholds approved by the bank, answer questions outside the ideal flow, and handle thousands of conversations in parallel, at any hour and at a marginal cost compared with other contact channels.
What results can a bank expect from AI collections in early delinquency? It depends on the starting benchmark. In the Turing Bank model case for auto loans, the goal is to double the response rate from 7% to 14%, raise collectability from 4% to 7%, and recover an extra USD 3,600,000 over the benchmark, validated with a POC of 2 to 3 months.