Borrowing Reimagined: Impact of AI-Based Underwriting on the Future of Lending

Author: Bradley Tian, Graphics: Caroline Yee

The BRB Bottomline

As lenders begin to embrace artificial intelligence, new borrowing opportunities – as well as threats – emerge within the lending market. This is an exploration of the mechanism of AI-based underwriting, as well as its future ahead.


With the US housing debt balance exceeding $11 trillion and aggregate personal debt balance (student loan, auto loan, credit card, etc.) exceeding $4 trillion by Q4 2021, domestic banks and private lenders are turning to artificial intelligence (AI) to better navigate the massive market by harnessing the vast amount of borrower data available. Fueled by the development of machine learning algorithms and access to individual profiles, the implementation of AI programs in lending can unlock new opportunities both in optimizing internal efficiency and expanding services to unprecedented segments, but it also poses novel regulatory challenges in maintaining fair lending practices. 

Roles and Benefits of AI in Revolutionizing Lending

The concept of “artificial intelligence” in the context of lending in this article primarily refers to the implementation of machine learning (ML) algorithms for analyzing borrower data at scale. Currently, there are two main ML approaches in the industry: supervised and unsupervised learning. In the case of supervised learning, data are profiled and sorted based on pre-set classification protocols determined by human developers. For example, lenders may specify evaluation criteria, such as collateral value and criminal history, to instruct ML programs to sort out qualifying applications from large applicant pools. Thanks to its capability to filter large amounts of borrower profiles in a compact timeframe, this methodology is particularly useful in streamlining underwriting and evaluation processes, especially for larger lenders. 

Unsupervised learning, on the other hand, examines large quantities of raw, unlabeled borrower data to develop new patterns for assessing creditworthiness and other borrower metrics. This approach does not rely on existing criteria from the lender and is capable of identifying non-standard data signals that suggest underwriting risk. For example, ML analysis may reveal a correlation between borrowers’ behavior on social media platforms and their default risk. Unsupervised learning is especially important in the field of applicant screening, as it can provide underwriters with much more detailed borrower insights than traditional metrics, such as income reports and FICO scores. 

Combinations of these two methodologies lead to practical benefits for both lenders and borrowers. For lenders, ML-powered screening programs can significantly expedite underwriting and reduce latencies in the loan application process, thereby reducing overheads and saving operational costs for lending firms. Furthermore, non-standard methods of applicant screening can improve creditworthiness assessments for borrowers lacking FICO scores or credit history, potentially revealing a new customer segment. 

For borrowers, AI-based advisory services are emerging to help individual optimize their debt repayment schedules by offering recommendations on budget-saving, interest rate calculation, and long-term financial planning. This also helps more borrowers to become reliable customers for lending institutions.

Upstart: Using Automation to Democratize Loans Access

A pioneer in AI-based lending, Upstart, a rising technology firm based in California, has attracted quite some publicity in recent months with a revenue increase of 38% YoY and a rise in share price of over 103% YoY as of February 2022. In the past 15 months, Upstart has lent over $135 million to borrowers with negligible credit ratings – most of them being college students and new graduates without mortgages & significant collateral items. Upstart’s expansion into the young-generation segment is largely due to its novel, AI-enabled loan approval process: instead of assessing new graduates based on their FICO scores and other forms of traditional ratings, Upstart evaluate their “personality” and sense of responsibility through non-standard data signals, such as their SAT scores, colleges, majors, and GPA. According to Paul Gu, Upstart’s co-founder and head of product, “It’s not whether you can pay. It’s a question of how important you see your obligation.” 

This novel, character-based screening methodology has certainly helped democratize access to credit compared to traditional models. As shown by its latest quarterly lending report, the tested model lead to “near-prime” borrowers with FICO scores ranging from 620-660 being approved twice as frequently, with average APRs decreasing by 17%. In addition, applicants under the age of 25 are 32% more likely to be approved. Fair lending testing has shown no disparities between the new borrowers approved under the AI-based testing model and those approved by traditional models, thus proving the accuracy and practicality of Upstart’s non-standard underwriting algorithm. 

Biases, Risks, and Regulations

Orienting underwriting processes around alternative data also comes with inherent risks and imperfections. Due to the sheer novelty of the automated lending field and the high degree of experimentality of many new underwriting methods, policymakers have struggled to monitor AI-based lenders with proper regulatory oversight. While it is difficult to apply static policies to the use of alternative data, regulators have contrived a new system of supervision that grants a degree of flexibility to the lenders. Back in 2017, the Consumer Financial Protection Bureau (CFPB) issued its first-ever “No-Action Letter” to Upstart, enacting a new policy that exempted the fintech lender from regulatory fallouts incited by the Equality Credit Opportunity Act in exchange for Upstart’s quarterly reports on its handling of fair-lending practices. This policy illustrated CFPB’s acknowledgment of AI-based lending as innovative progress, but it also outlined the regulators’ consumer-first attitude in limiting experimental practices. 

Another aspect of risks brought about by pattern-recognizing systems is the potential development of a new set of discrimination and bias. Although lenders are aware of avoiding bias factors, such as race, gender, and credit scores, association-based profiling can nevertheless lead to the creation of “proxy discrimination.” As its name suggests, proxy discrimination occurs when certain aspects are explicitly avoided in profiling, but the next most-common proxy of these prohibited factors would be inherently used for biased classification. While one of the main benefits of AI-based underwriting is its ability to deploy objective decisions, proxy discrimination, if left uncontrolled, may result in the formation of a new set of facially-neutral prejudice against targeted consumer segments. 

Into the Future of AI-Based Lending

AI-based lending systems have the potential to provide unprecedented credit access to borrowers with limited credit backgrounds, such as college students and new graduates. In consideration of both supervised and unsupervised learning approaches, non-standard underwriting methods utilizing machine learning can provide immense benefit to both lenders, such as improving the efficiency of application processes, and to consumers, such as increasing credit opportunities and optimizing repayment schedules. As regulatory parties, such as the CFPB, demonstrate acceptance of this emerging technology, the prospect of AI-based lending’s eventual integration into the industry mainstream remains strong. However, to maintain its promise in bringing more democratized credit access to the public, automated lenders must remain proactive in identifying unintentional distortions – such as proxy discrimination – and constantly update their classification algorithm to optimize the trade-off between profiling accuracy and inclusivity. 

Take-Home Points

  • College students, as a primary beneficiary of AI-based lending, can receive immense credit-building benefits 
  • As underwriting processes pivot to become more character-oriented, it is important to start to cultivate a strong sense of responsibility
  • It is also vital to demonstrate that through disciplined behaviors that properly represent oneself on digital grounds

1 Comment

  1. Never heard of this idea before which is awesome. Love the unique topic and the clear description on why it matters.

Leave a Reply to Derrick Cancel reply

Your email address will not be published.