Artificial intelligence has been a part of the credit landscape for a while now, but generative AI promises to fully change the game. From the ubiquitous chatbots to enhanced credit scoring to personalized loyalty programs, AI is trained on every aspect of the credit industry.
In From Hype to Impact: How AI Is Transforming Credit, a new report from Javelin Strategy & Research, Ben Danner, Senior Analyst, Credit and Commercial, looks at what changes lie in store for card issuers. “Generative AI is changing the way financial institutions analyze data and is streamlining customer service operations,” Danner said. “But it also comes with considerable risks.”
The Existing Use Cases
The most visible example of AI now is the chatbot, which we can expect to get more intelligent as AI capabilities expand. Instead of a basic chatbot that sends you through a link list or a hierarchical checkbox list, the improved bots will use natural language processing to have more intelligent and human-like responses. The enhanced intelligence comes with some challenges, presenting the prospect of an untamed chatbot going off the guardrails and saying all sorts of strange things to customers.
In the credit scoring and decisioning spaces, AI has been used for a while to work through unstructured data. Generative AI can output new modes of information based on what it’s been learning. But there are potential regulatory hurdles limiting how that data can be used for scoring and decisioning. Credit scoring is tightly regulated, with a variety of laws that have been on the books for years and haven’t caught up with some of the advances in AI tech.
Companies like FICO say they’re not using AI at all right now in their credit scoring. But other companies that provide data to FICO are leveraging AI technology. They are using it to analyze unstructured data, like social media, email, and even tax returns and rental agreements.
“A rental agreement or an invoice might come to you in a PDF, for example,” Danner said. “But if you need to provide that to your credit agency, a human would have to sit there and look through that document, find what you owed and if you paid it on time, and all that. AI can look at those unstructured invoices, aggregate all the data together, and build that profile for you.”
Unstructured data has a lot of promising uses for evaluating creditworthiness. But regulatory concerns have limited its use when it comes to actually constructing a credit score.
Problems to Be Solved
As a rising and rapidly changing technology, AI still has several kinks to be worked out. By now, everyone has become familiar with AI’s problems with hallucinations.
“I used ChatGPT this morning when I was trying to analyze a certain graph,” Danner said. “I asked it to spit me back three sentences on what it thought this graph was about, and it sent me back numbers that were incorrect. I think it interpreted an 8 for a 6 on one of the charts and sent back data that was completely wrong, but it defended it like it was correct. That’s been one problem that’s plaguing data.”
Another concern is the transparency of the model. AI tends to be a black box, which makes explaining how some of the algorithms arrived at their choices difficult. A credit regulator needs to know how the model comes up with its decisions.
“If you can’t explain the result to me, then we can’t use that,” Danner said. “That’s something all the AI companies are trying to figure out. That’s why there’s all this verticalization of AI and using their own data internally, so that they can fully explain their model. They’re not just going out and getting data from all over.”
Finally, there is algorithmic bias. Training an algorithm from data collected by humans will introduce biases, and those biases will be reflected in the outputs from the algorithm. A study from Lehigh University looked at racial disparities in large languages models and found these disparities persisting in mortgage underwriting.
“It’s perpetuating these social inequalities,” Danner said. “The banking industry’s been trying to correct those mistakes, especially in credit. Those are things that need to be solved for with these models before a wider application.”
Personalizing Loyalty and Rewards
Credit card companies have also begun incorporating AI into their rewards programs. Much of the data they’re using is derived from transactions. Every time a shopper swipes a credit card, the issuer is collecting that data, then using it to offer different merchant rewards.
For example, Chase has its Chase Offers platform built into its mobile app. Every swipe builds another piece of a huge transaction history. AI has the ability to take a large data set like that, with thousands and thousands of transactions, and personalize it to just one individual.
“Let’s say I know Ben likes to buy coffee in the mornings at 8 a.m.,” Danner said. “Should we present some type of offer to him at 7:45? If a human had to do that, you would have to hire a whole team of people to sit there and figure all that out. We can now have AI analyze all that transaction data. That’s an opportunity for card issuers that are historically sitting on millions of data points but don’t have a good way to analyze or leverage that information.”
The Next Steps
The new agentic AI shopping models will make the world even more complicated. We will soon have AI agents making payments on behalf of customers. Consumers will eventually figure out how to use that system to find the best deal for hotels, for example, but issuers will also use it to garner more usage from their cardholders.
“Visa gave us a little bit of a hint into their how their AI analytics is going to work,” Danner said. “They presented a picture of a cellphone with a person requesting a hotel, saying, ‘Could you find me the best hotel in the area?’ And it popped back and said, ‘Sure, would you like to add your card to this?’”
