So much about artificial intelligence over the last two years has focused on GenAI that more “classic” uses of AI in credit analysis and credit management have been neglected. New research from Cornerstone Advisors indicates that the need for increasing efficiency and the potential for reducing losses will drive more community banks and credit unions to adopt AI lending tools in the year ahead. But regulatory headwinds must also be accounted for.
This rising interest in automating credit decision making and loan management — in a part of the industry that once took pride in every credit decision being a handcrafted matter — reflects the changing economics of the field. It also reflects the evolution of competition in both business and consumer lending, where fintechs engaged borrowers with automated credit decisions rendered quickly 24/7.
“It’s amazing — and not in a positive way — that financial institutions use of technology for lending is lacking,” writes Ron Shevlin, Cornerstone’s chief research officer, in the firm’s report. Cornerstone benchmark data indicates that institutions using AI tools can more than triple the credit analysis handled per underwriting full-time-equivalent employee.
However banks and credit unions can anticipate more attention from Washington in the wake of fact-finding by the Treasury Department over the summer. The department received voluminous input from more than 100 sources ranging trade associations, innovation-focused venture capital firms like Andreessen Horowitz, tech vendors like Fiserv, consultancies in regtech and more.
The deadline for input — some responses ran dozens of single-spaced pages — was August 12 and now Treasury analysts will dig into extensive commentary. The Biden administration has been working on AI issues on multiple fronts since the president issued an executive order handing out responsibilities on the matter in October 2023. In its 2023 annual report the interagency Financial Stability Oversight Council cited use of AI in financial services as a vulnerability.
Regulatory Concerns Have Held Back Adoption of Credit AI
Barring major change in Washington, more rules and regulation will likely come as artificial intelligence in all forms becomes a widely adopted tool in the industry. Potentially, lenders will gain greater clarity from the process — though perhaps not greater freedom — as bankers as a rule like definite bright lines governing their activities.
A key element in existing regulation and oversight is the expectation that lending institutions using AI tools will be able to explain how they work. This is part of regulators’ broader view on analytical models in general: Somebody in the bank is supposed to understand what happens inside the black box.
“Understanding how and why a model arrives at a particular decision or prediction can be difficult. … Financial service providers cited explainability as the top barrier to adopting machine learning,” writes Shevlin.
In fact, in its study, “Achieving High-Performance Lending: The Impact of AI on Lending Efficiency,” Cornerstone found that regulatory worries dominated the thinking on AI among the senior official responsible for credit surveyed in community financial institutions.
The top concern about using AI for credit modeling was regulatory scrutiny, cited by 70% of respondents. This was followed distantly by the following:
Cost to acquire: 45%
Transparency: 42%
Lack of knowledge: 42%
Bias: 42%
Fairness: 38%
Integration: 36%
Ongoing support/resources: 36%
Maintenance: 32%
In its comment letter to Treasury, the American Bankers Association pointed out one of the weaknesses seen in “explainability”: vendor noncooperation.
“Vendors rarely assist with bank efforts to validate the third-party product and may share necessary details only begrudgingly notwithstanding the increasing complexity of models,” ABA said. “Banks attempt to include contractual language requiring third parties to be transparent, but vendors often decline to do so in an effort to protect their proprietary information.”
Much confusion exists on what type of AI documentation examiners expect, ABA wrote. In its comments, Fiserv suggested adoption of a uniform set of definitions for various AI models and systems used by financial institutions to facilitate clearer discussion. This could be orchestrated through the federal National Institute of Standards and Technology, the vendor suggested.
The result of the regulatory concern has been that only 13% of the respondents in the Cornerstone study say that their institutions have deployed AI as part of their credit and lending activities. Another 17% are planning to invest in AI tools or use it to improve their credit processes in 2024.
Read more: How Banks Can Get Past the AI Hype and Deliver Real Results
Necessity is the Mother of AI Lending Action
Of those that do plan to invest in AI lending models, Cornerstone found that 54% plan to spend at least $100,000 on the technology over the next three years. And 12% plan to invest more than $500,000 on the AI tools.
Notably, 18% say that they will make significant investments in AI or, alternatively, enter strategic partnerships with providers. A much greater share of the sample anticipates making smaller investments in the tech or entering into partnerships for short-term experiments or trial runs.
A wrinkle here is the impact of money not spent. That is, expectations of cost can hold institutions bank. In its comments to Treasury, the Independent Community Bankers of America said that “responding to examiner scrutiny and showing compliance with third-party risk management guidance can be prohibitive. Simply said, it is costly for community banks to ensure and demonstrate compliance with relevant regulatory requirements when selecting and monitoring third-party relationships.”
In addition to those institutions moving ahead, 18% of the responding institutions say their boards have discussed potential adoption of AI lending technology. But 24% of institutions say the technology is only on their long-term road maps — and 29%, nearly a third of the sample, say they aren’t considering the tools at all.
Read more:
- Why the AI Revolution Is Being Led from Below
- Digital is Draining Banks’ Emotional Connections with Customers. GenAI May Make Things Worse
- An AI System Built for Everything is an AI Built for Nothing
Hurdles Keep Institutions from Realizing Potential AI Lending Benefits
Inadequate staffing and weak leverage with vendors are among the barriers that smaller institutions face when trying to move forward in this area.
A community banker told Cornerstone of these challenges: “We always feel we’re way behind the curve when it comes to emerging technologies like machine learning. Our executive team is small, and we watch what’s going on in the industry, but we feel like we don’t have the resources to really be state of the art.”
Continued the banker: “That’s why we rely so heavily on third-party vendors. If these vendors aren’t coming to us with these solutions, then there’s no real avenue for us to do it ourselves.”
In his report Shevlin underscores the potential gains that community financial institutions miss when they don’t utilize such technology.
One is the sheer volume of data AI can crunch — including both what banks already analyze and new, additional types making their way into credit analysis. It’s not just a matter of speed and labor saving, but quality of analysis.
Shevlin argues that machine learning can “identify patterns and correlations that may not be apparent through traditional methods.”
More nuanced pricing can also be tailored to borrowers when AI gets applied, he suggests. “By more accurately differentiating between low-risk and high-risk borrowers, lenders can optimize their loan offerings, potentially attracting quality borrowers while mitigating the risks associated with higher-risk applicants,” according to Shevlin.
Benefits don’t just appear on the front end of the lending process, he points out. Better credit monitoring through AI techniques can improve detection of problems and flag the best remediation steps. Shevlin, citing McKinsey data, points out that incorporating machine learning in collections can reduce net charge-offs by as much as 5%.