Is AI the Cornerstone for Financial Inclusion?

35

By Dr. Josep Gisbert, Assistant Professor of Financial Economics, IE University

Can artificial intelligence (AI) be the disruptor that finally makes financial services more accessible? In this article, we will explore AI’s transformative potential for improving access to credit by drawing on insights from top journals and taking a historical perspective. We will also examine whether AI is likely to level the playing field in financial inclusion and discuss how historical technology disruptions have shaped access to credit. We will wrap up this article by exploring AI’s challenges and opportunities.

AI’s role in financial inclusion can be better understood by focusing on the key technological disruptions that have transformed loan assessments. If we focus on the primary methods of making loan decisions, three distinct historical periods become apparent.

The Soft Information Era

During the early 20th century and prior, loan decisions were predominantly manual, heavily relying on human judgments and qualitative assessments. Computers were not mainstream, and the available technology still needed to be more advanced. Therefore, lending decisions were influenced by personal relationships and the borrower’s character; local bankers often knew their clients personally, relying on their knowledge of their reputations and integrity. The borrower’s social standing and community ties were also crucial factors in assessing the likelihood of loan repayment. Extensive interviews were conducted to understand the borrower’s needs and repayment ability. Physical collateral inspections, such as property and inventory, were standard to assess values and conditions. References from respected community members provided additional assurances of the borrower’s reliability.

The Hard Information Era

In the mid-20th century, as businesses grew more complex, financial statements became significant in lending decisions. Bankers manually analyzed balance sheets, income statements and cash-flow statements. Fundamental financial ratios, such as debt-to-income (DTI) and loan-to-value (LTV), were calculated to evaluate repayment ability. The establishment of credit bureaus allowed the collection and sharing of credit information. Lenders accessed credit reports containing borrowers’ credit histories, including loans, repayments and defaults. These reports, however, were less comprehensive than modern credit scores, and the soft information collected through bankers’ interactions was still essential to facilitating credit.

The late 1950s and 1960s saw the development of statistical models for credit scoring, pioneered by companies such as Fair, Isaac and Company (now FICO). These models used historical data to predict default likelihood. Despite the introduction of these models, calculations were often manual or aided by simple mechanical calculators, laying the groundwork for today’s automated systems. However, during the 1980s, 1990s and early 2000s, the transition to a more quantitative and less qualitative credit-assessment trend continued. We call this era the Hard Information Era.

This was especially true in heavily standardized banking segments, such as retail banking and payment methods. Financial-services providers increasingly relied on what became known as “hard” (quantitative and verifiable) information rather than “soft” (qualitative and subjective) information. In some of these segments, nearly complete algorithm-based credit decision-making was implemented. This new focus on hard information crucially excluded groups that relied on exchanging soft information to obtain credit. Erik P. Gilje and coauthors, in a 2016 article in the Journal of Finance, and Hoai-Luu Q. Nguyen, in a 2019 paper in the American Economic Journal, documented that minorities, lower-income individuals and other traditionally vulnerable population groups were the most affected segments of the population.

The Financial Technology Era (Is fintech the solution for access to credit?)

During the Financial Technology Era, between the mid-2000s and the mid-2010s, the disruption of financial technology (fintech) lenders, characterized by their intensive use of technologies to provide financial services, was seen as the promise for a future society with a non-existent credit-access gap. This was especially the hope for traditionally financially excluded population segments historically dependent on relationship banking. Fintech’s promise of introducing the latest advances in information processing, cutting-edge machine-learning (ML) models, access to big data and innovative data-driven culture promised that the boundaries between hard and soft data might disappear. So, the question is: Will the advanced technologies, such as machine learning and big-data frameworks, introduced by fintech companies be sufficient to guarantee access to credit for all solvent borrowers while charging fair interest rates? Below, I will try to answer some critical questions in this regard.

Do fintech companies provide credit faster than traditional lenders?

Yes! In 2019, Andreas Fuster and coauthors documented in the Review of Financial Studies that fintech lenders process mortgage applications significantly faster and provide credit more elastically in response to exogenous mortgage-demand shocks.

Do machine-learning-based fintech credit-scoring models predict borrowers’ defaults better than traditional empirical models?

Yes! Amir E. Khandani and coauthors back in 2010 in the Journal of Banking and Finance documented (by combining customer transactions and credit-bureau data from January 2005 to April 2009) out-of-sample forecasts based on machine-learning techniques to construct nonlinear, nonparametric forecasting models of consumer credit risk that significantly improved the classification rates of credit-card-holder delinquencies and defaults.

Is the information content of non-traditional data sources useful to make better predictions?

Yes! Recent work by Leonardo Gambacorta in the Journal of Financial Stability in 2024 documented that non-traditional information sources, such as digital applications on mobile phones and e-commerce platform data, significantly enhanced credit models’ predictive power. Previous work by Julapa Jagtiani and Catharine Lemieux in 2019 in the Financial Management Journal had already established the tendency of technology-intensive lenders to use social-media activity; Tobias Berg and coauthors in the Review of Financial Studies noted users’ digital footprints, such as orders, transactions and customer reviews, which Jon Frost and coauthors documented in Economic Policy.

Do fintech lenders guarantee higher financial inclusion?

There is mixed evidence! Gregory Buchak and coauthors in the Journal of Financial Economics and Andreas Fuster and coauthors in 2019 in the Review of Financial Studies and 2022 in the Journal of Financedocumented that there is no evidence that fintech lenders target borrowers with low access to credit and even documented that the opposite is the case. Fintech lenders seemed to cream-skim the market, targeting the most profitable borrowers in a risk-adjusted return case. Still, other authors, such as Marco Di Maggio and Vincent Yao in the Review of Financial Studies in 2021, have shown that, at least initially, fintech lenders may increase the financial inclusion of higher-risk borrowers not typically served by incumbents. They also documented that fintech companies relied mainly on hard information to make credit decisions.

Perhaps the most surprising to date was Emilio Calvano’s and coauthors’ 2020 paper in the American Economic Review. To analyze the possible consequences of replacing humans with algorithms in pricing goods and services, they experimentally studied the behaviors of algorithms powered by artificial intelligence (Q-learning) in a workhorse oligopoly model of repeated price competition. They found that the algorithms consistently learned to charge supra-competitive prices without communicating with one another! The high prices were sustained by collusive strategies with a finite phase of punishment followed by a gradual return to cooperation. In other words, AI algorithms autonomously learned to collude to charge higher market prices.

The Generative AI Era

Advanced generative AI (GenAI) technologies can potentially be the ultimate tool that blurs the traditional boundaries between hard- and soft-information analyses. These technologies are embedded in new models for computer vision, used to analyze videos and pictures; natural language processing (NLP), used to analyze text and audio; large language models (LLMs), used to facilitate live interactions with customers; and big-data capabilities, used by most lenders to automate both soft- and hard-information discovery and analysis. While these new technologies and the unstructured data (e.g., text, video, audio, pictures) fed to them were previously used by a minority of lenders, they have been widely implemented across the lender spectrum in the new Generative AI Era. If we used the same methods to classify financial providers as either fintech or non-fintech today that were used in the most influential financial economics papers, all the leading lenders in an advanced economy, even the incumbent, mainstream banks, would be classified as lenders making intensive use of technology.

The historical perspective of the Generative AI Era still needs to be analyzed. The financial-economics community focuses on analyzing the influences of these novel technologies. However, we still need to wait for rigorous evidence to declare that GenAI is the cornerstone of financial inclusion.

Conclusion

The use of GenAI across the industry can transform access to finance. However, it is still being determined whether it will lead to better inclusivity and efficiency than previous technologies. What is clear is that, given our experience with previous disruptive technologies, we should not expect that these technologies are going to be the solution per se. They tend to be opaque, and the rational goal of profitability embedded in the models can lead to undesirable outcomes for financial inclusion.

In the years ahead, the advent of more advanced AI, such as cutting-edge AI agents, will be able to act directly in the world to achieve mid- and long-term goals, with little human intervention or specification of how to do so. AI agents can interact seamlessly with humans, process unstructured information quicker than humans and do so autonomously. The boundaries between the typical soft-information-based relationship lending based on the bank-branch-network model and the mostly hard-information-based model may disappear. For a careful review of the topic, I recommend reading the recently published BIS (Bank for International Settlements) working paper titled “Intelligent financial system: how AI is transforming finance” by Iñaki Aldasaro and coauthors.

In an optimistic scenario, AI has the potential to significantly enhance financial inclusion by improving access to credit, reducing transaction costs and providing personalized financial education. In a pessimistic scenario, with the birth of general intelligence AI, the shock in our society could completely transform financial intermediation as we know it. If the majority of human labor is replaced, our society will be fully transformed, and in turn, this will completely transform financial intermediation as we know it today. Still, addressing the challenges and risks associated with AI is essential to ensuring its benefits are equitably distributed. Continued research and careful implementation of AI in financial services and its supervision will be vital to achieving inclusive economic growth.