What the Evidence Finds: More Reach, Better Screening

For U.S. bank small-business lending, the central empirical question is not whether AI sounds capable of underwriting. It is whether banks that use more AI actually lend farther, whether that change is more than a bank-size correlation, and whether the distant loans perform well afterward. The Census working paper answers those questions in stages. Its findings support a meaningful expansion of reach and a screening interpretation, but the strength of that conclusion depends on the data, the identification strategy, and the limits of the loan samples.

The order matters. Adoption tells us that the technology exists. A distance interaction tells us that lending geography changes. An identification strategy tells us how much of that change might be attributable to AI. Loan performance and pricing tell us whether the additional reach looks like improved selection or simply greater appetite for risk. No single result can carry all of that meaning.

More lending at a distance, not more lending everywhere

The study begins with the Census Bureau’s Annual Business Survey technology module, linked to banks’ financial data. Across roughly 1,500 bank-year observations from about 1,100 unique banks over the 2017–2019 period, AI use averaged 22%. The reported rate increased from 14% in 2017 to 17% in 2018 and 43% in 2019. Moderate or high use rose from 8% to 22% over the same span. Larger banks were more likely to use AI, while the other measured bank characteristics were not significantly associated with adoption.

Those numbers establish a period of rapid adoption in the study’s sample, not a current industry-wide rate. They also reveal why a simple comparison between AI and non-AI banks could mislead. Bank size might affect technology budgets, staffing, geographic coverage, borrower mix, and lending strategy at the same time. If AI users lend farther, the difference could reflect scale rather than the information-processing capability of AI itself.

To examine lending geography, the researchers merge the adoption data with Community Reinvestment Act records. The CRA data report the volume of small-business loans originated by a qualifying bank in each borrower county and year. The study measures distance from the borrower county’s geographic centroid to the bank’s closest branch and uses roughly 172,000 bank-county-year observations. It separates loans below $100,000 from loans between $100,000 and $1 million.

The basic pattern is familiar from relationship lending: loan growth generally declines as the borrower gets farther from the bank. The important result is that the decline is less steep for banks with greater AI use. AI is not significantly associated with more overall small-business credit when distance is not considered. Once AI is interacted with distance, however, the interaction is positive and significant for both small and larger loan categories. A one-standard-deviation increase in AI use attenuates the negative distance relationship by about 46% for loans below $100,000 and 64% for loans from $100,000 to $1 million.

“Attenuates the distance relationship” is more precise than “increases lending by 46%.” The estimate describes how much the usual relationship between distance and loan growth becomes less negative under a specified change in AI use. It is not a growth rate for an AI bank, a percentage increase in every borrower’s credit, or a guarantee that a bank’s total loan book expands.

The distance bins make the geography more concrete. The study compares borrowers under 50 miles away with those 50–250 miles, 250–1,000 miles, and more than 1,000 miles away. The AI interaction is positive in each group and rises with distance, but only the interaction for borrowers more than 1,000 miles away is statistically significant.