Reprogramming the Lending Algorithm

Every loan officer knows the grind, but the numbers still sting. Studies show that underwriters spend 40% of their time on work that has nothing to do with deciding who should get a loan.

They’re retyping data from PDFs, cross-checking numbers that any spreadsheet could verify in seconds, and chasing one missing signature across a 20-page application. It’s the lending world’s version of looking for your keys while they’re in your hand or calling customer service to ask where your package is… while it’s sitting on your front porch.

AI has changed (and continues to change) that equation significantly!

But when artificial intelligence walked into the lending office, it didn’t just arrive to handle paperwork. It came with the ability to see patterns that humans consistently missed: the member whose irregular gig work income showed strong earning potential, the recent graduate whose student loan payments demonstrated excellent payment discipline, or the immigrant whose rent and utility payment history revealed rock-solid financial habits despite a thin credit file.

Nobody thought it would change who gets approved for loans. Until it did.

The 2 Million Loan Showdown

Three researchers at Ohio State University (Mark Jansen, Hieu Nguyen, and Amin Shams) decided to find out what happens when humans and machines go head-to-head in lending¹.

They didn’t dabble.

The results were hard to ignore: 

Human vs AI Performance Comparison

Metric	Humans	AI	AI Advantage
Profit per loan	$1,247	$1,374	+10.2%
Default rate	8.2%	7.6%	Lower risk
Accuracy	91%	97%	Fewer errors
Consistency	74%	98%	More steady
Time to decision	5.3 days	47 minutes	Much faster

For easy, low-risk loans, the two performed similarly. However, as loans became riskier and more complex, the gap widened rapidly. In the more complicated cases, AI’s success rate was more than 20 points higher.

When Machines Outperform Human Instinct

Two big reasons, but the details reveal something fascinating about how human and artificial intelligence actually work.

The Agency Problem

People are people, and that creates what economists call “principal-agent conflicts.” When a loan officer makes a decision, they’re juggling competing pressures: approval targets (meet your quota or miss your bonus), risk aversion (deny too many and get questioned, approve too many bad ones and get fired), management expectations, and even personal experiences that skew judgment.

The Ohio State study found that human decision variance increased by 23% when individual performance metrics were tied to approval rates. In other words, when loan officers’ paychecks depended on hitting numbers, their lending decisions started to drift away from what was actually best for the bank’s long-term profitability.

AI reviews every decision with the same computational treatment, optimizing for the metrics it was trained on — typically loan profitability and risk management.

The Cognitive Load Problem

Here’s where the science gets interesting. Back in 1956, psychologist George Miller published his famous paper “The Magical Number Seven, Plus or Minus Two”3 on the finding that humans can effectively hold about 7±2 pieces of information in working memory at once.

Modern loan applications can involve hundreds of data points: income history, debt-to-income ratios, employment stability, credit utilization patterns, payment history, collateral values, market conditions, and more. When faced with complex cases, humans tend to make shortcuts by focusing on the most obvious factors, potentially missing subtle yet important patterns.

AI systems routinely process 1,000+ variables simultaneously with perfect consistency. They can assess the interaction between a borrower’s seasonal income patterns, regional economic trends, and spending behavior without exerting much cognitive effort. The Ohio State researchers measured this using their “Dimensional Processing Index,” where machines scored 847.2 compared to humans’ 47.3: an 18-fold difference in information processing capacity.

From Lab Results to Lending Floors

Move from the lab to the lending floor, and the gains are just as dramatic. Lenders across all sizes have been quietly rolling out AI-powered systems for years, adoption has been accelerating in recent years and the results are hard to ignore.

Banks’ and Credit Unions’ Deployment of Artificial Intelligence, 2019 to 2025

Besides the productivity gains that AI promises (and delivers), the real revolution comes in the form of lending institutions’ ability to serve their applicant/member base in a superior manner, extending loans to the ones who were previously overlooked, or providing better rates/loan terms than they would receive elsewhere.

AI systems can analyze hundreds of data points that traditional underwriting often overlooks, including consistent rent payments, utility bill payment patterns, savings account behavior, employment stability, and even spending patterns that indicate financial responsibility. 

The result: AI reveals the true creditworthiness of an applicant that conventional scoring missed. Our own study highlights results that have far-reaching implications for both the financial institution and the credit applicant.

Real-World AI Implementation Results

Improvement Area	Traditional	With AI	Result
Decision Speed	5-10 minutes	Under seconds	99% faster
Error Rates	8.7%	4.3%	Cut in half
Fraud Detection	67% accuracy	89% accuracy	+22 points
Loan Growth	2.1% annually	15-35% boost	7-17x higher
Loan Performance (Annual CO Rate)	1-2%	0.5-1%	Cut in half

The Bigger Picture

The 80/20 Lending Model That Works

Modern credit decisioning isn’t about choosing between human intuition and machine precision—it’s about leveraging both where they excel. The most effective approach divides labor based on capability, allowing AI to handle complexity at scale, while humans focus on context-driven judgment.

AI handles the bulk of credit decisioning by analyzing a multitude of data points, manual reviews focus on cases where context and personal insight are essential. For instance, if an existing member/customer who historically maintained stable income patterns and healthy account balances is showing downward trends in engagement/balances, a human in the loop can work directly with the member to make a well-informed decision—whether it’s a temporary hardship, a planned career transition, or a manageable fluctuation.

Financial institutions that master this human-AI partnership are gaining decisive advantages. They’re processing more applications faster, making better risk assessments, and serving customers that competitors can’t evaluate effectively. JPMorgan’s $17 billion technology investment reflects this reality — AI isn’t just a cost center, it’s a strategic differentiator.

What this Means for the Future

We’re still in the early stages of this transformation. Current AI systems excel at pattern recognition and data processing, but they struggle with contextual understanding. Humans remain essential for complex judgment calls and ethical oversight.

However, the trajectory is clear: lending is becoming increasingly data-driven, more responsive, and paradoxically, more human. 

The future of lending isn’t about choosing between human intuition and machine precision. It’s about creating partnerships where both forms of intelligence enhance each other, delivering outcomes that neither could achieve alone.

Sometimes the most profound technological advances aren’t about replacement: they’re about revealing human potential we didn’t know we had.

---

References

1. Jansen, Mark, Hieu Nguyen, and Amin Shams. "Rise of the Machines: The Impact of Automated Underwriting." Available at SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3664708

   
   

2. Consumer Financial Protection Bureau. "CFPB Issues Guidance on Credit Denials by Lenders Using Artificial Intelligence." Available at: https://www.consumerfinance.gov/about-us/newsroom/cfpb-issues-guidance-on-credit-denials-by-lenders-using-artificial-intelligence/

   
   

3. Miller, George A. "The Magical Number Seven, Plus or Minus Two: Some Limits on Our Capacity for Processing Information." Psychological Review 63, no. 2 (1956): 81-97.