The idea of letting artificial intelligence models make lending decisions was once considered risky and prone to error, discrimination, bias, "black box" style obfuscation and lack of explainability. But now it is becoming an important tool for lenders competing with credit card issuers and fintechs that have been using this technology for years.
The latest financial institution to embrace AI-based lending is $9.8 billion-asset Teachers Federal Credit Union in Hauppauge, New York, which is letting AI approve credit cards, auto loans and personal loans.
Teachers began working with Corridor Platforms' Riskdecisioning.AI software a little over a year ago; it launched the platform in the fourth quarter. A primary benefit to the newly automated lending process is speed, the bank's leaders say.
"For our members and customers in the banking sector, the expectations continue to shift and accelerate, and they demand fast, seamless, personalized experience," Teachers CEO Brad Calhoun told American Banker. "There's an increase in the level of expectation around real-time answer, real-time response. It requires this type of investment in AI-powered lending for that speed, to be able to turn around the right decisions with the right risk management."
The need to compete with banks and fintechs that already have this technology was another driver of this project.
"Teachers as a credit union wanted to make sure they harness the power of all of their data and all of the experience with their members to start using the most advanced analytics to give the right price and the right product or even cross-sell," Corridor Platforms CEO Manish Gupta told American Banker. "National banks have spent a lot of resources on data analytics, modelers and platforms that do this, whereas credit unions have traditionally relied on members' loyalty."
When people apply for credit cards, personal loans and auto loans, the industry norm and the customer expectation is that they will get an answer and a price instantly, Gupta said. "Anytime it is manual, then it cannot be instantaneous," he said.
If competitors can give immediate answers "while you are saying, 'Let me get back to you,' then what essentially happens is all the good customers go to the instantaneous answers," Gupta said. Credit unions that can't keep up end up with customers who don't qualify for an instant online loan, he said.
"There's no choice," Gupta said. "Every bank needs to move to instantaneous decisioning, using all of the power of analytics and AI. Teachers has now got to that, and they are able to compete nationally with any of the top banks."
Automated loan decision platforms can be rules-based or AI-based.
"AI is not necessary to make an instant decision," Craig Focardi, principal banking analyst at Celent, now part of GlobalData, told American Banker. "Real-time instant decisioning is a technology issue, not an AI versus other analytics issue. Many lenders continue to make real-time decisions today with FICO scores or VantageScore in a real-time environment and do very well."
In his view, upgrading to the latest credit scores, like FICO 10 T or VantageScore 4.0, could give banks the same improved predictability of default of an AI-based lending model. Both FICO 10 T and VantageScore consider trended data, which reflects changes in credit behaviors over time, where older versions of credit scores rely on credit history records that represent one point in time.
However, for many banks, upgrading to a new credit score across all their lending systems is a challenge.
"It's expensive to recalibrate those models and to change their policies and their systems across loan origination, servicing and portfolio management," Focardi said.
Instead of upgrading to the latest credit score, banks typically keep their old scores and bolt on an AI model to the lending process to incorporate consumers' recent behavior into their decision, Focardi said.
Leading lenders use both modern versions of standard bureau scores and AI models, Focardi said. "They may also use internal custom scores developed internally or with a third-party vendor. Capital One Bank is one example."
How Teachers approaches AI-based lending
After Calhoun said he wanted to invest in AI-based lending, Suresh Renganathan, chief technology officer at Teachers, started to look for a platform that was purpose-built for financial institutions with the ability to offer loans nationally and settled on Corridor's software.
Large banks are ramping up AI investment at the same time they are reducing their workforces, though no one seems ready to publicly draw a connection between those two actions.
For the past year, the credit union has been shifting from manual to automated underwriting, with low-, medium- and high-risk tiers. It's now using automated underwriting for 80% to 90% of loans. Customers fill out standard loan and credit card applications on the Teachers website, Corridor's machine-learning models make the loan decisions and human credit experts validate them.
Calhoun hopes the software will enable staff to have more time to do things like cross-sell and build customer relationships. He also hopes using AI will allow Teachers to scale up its lending business without increasing headcount.
Another role for credit specialists at Teachers, Renganathan said, is to regularly recalibrate the decisioning models.
"We are focusing on post-decision audits," he said. "We are also looking into fair lending analysis in a regular way and stress testing our models under different economic conditions. We would not have been able to do that if it was still manual underwriting happening."
How the AI model works
Any external credit score, such as a FICO score, uses standard credit bureau data to try to predict a borrower's likelihood of repayment. It doesn't take into account the internal data a company like Teachers has about its customers, such as their responsible (or irresponsible) account management and their track record of paying bills and making rent payments on time.
Teachers' model uses FICO scores and credit bureau data from TransUnion to gain a standardized risk view across borrowers. It also creates a proprietary score for each customer by analyzing credit bureau data alongside customer account data. This custom score lets it take into account member-specific insights like how long a member has been a customer, payment behaviors, whether the member has a job and the size of the member's paycheck.
"The models have been trained not only using internal loan performance data, but also enriched with trusted bureau data from TransUnion in our case, so every decision is fully explainable, auditable, and aligned with fair lending standards," Renganathan said.
The net predictive power of the internal score is better than a standard credit score, according to Gupta. "That's why we would always encourage larger banks or credit unions to start incorporating their own data, because that is what you owe to your members, because they have developed a relationship with you, and FICO is always backward-looking and does not have the latest data."
"When we originally looked at this and did all the testing, it was very valuable for me to see that people that we otherwise, through a normal manual process, would have otherwise declined, would now be recommended for approval through the automated engine," Calhoun said.
To ensure the integrity and fairness of the model, Teachers engaged an external auditing company to validate it. "They rigorously assessed our model for bias, explainability, stability and regulatory compliance, including adherence to fair lending standards," Renganathan said.
Working through reservations
For years, banks worried that using artificial intelligence in their credit decisions would trigger scrutiny from regulators, who have warned in the past that credit decisions can't be made in a black box, they have to be transparent and explainable.
"Regulatory concerns never go away," Focardi said. "The main issues for regulators are credit risk, fairness, disclosures to consumers and the explainability of the loan decision, whether it's an approval or a denial."
For this reason, banks are not going to rip out and replace standard credit bureau scoring en masse in favor of AI/ML only modeling, Focardi said. "It's not going to happen, certainly not overnight. It's going to take maybe many years to evolve further in that direction as the regulatory framework becomes more accepting of AI modeling, and we're never going to get there in mortgage, because we have Fannie Mae and Freddie Mac automated underwriting systems."
But today, the biggest hurdle is resource constraints, Focardi said. Small financial institutions "typically lack the resources to do the model development," he said.
He sees banks using AI models for credit card and personal loan originations, "because those are the lower-loan-balance, easier-to-underwrite types of loan products that don't require collateral," he said.
Auto lending, home equity lending and mortgage lending have higher hurdles for credit policy, regulatory requirements and explainability to internal bank policymakers and regulators. Adoption of AI-based underwriting in these areas is smaller, but growing, Focardi said.
