Commercial banks that use artificial intelligence more intensively show higher returns on assets but also lend less to small and medium-sized businesses, according to a new study from the Federal Reserve Bank of San Francisco published in September 2026. The research found that banks with above-average AI adoption had returns on assets roughly 0.38 percentage points higher than banks with below-average AI use, while their share of loans to small businesses was nearly half that of low-AI banks—12% versus 21%. The findings suggest AI is reshaping which borrowers get credit and on what terms, with potentially significant consequences for small business financing.
The Fed's analysis tracked 1,006 banks representing over 87% of total banking system assets, using online job postings data as a proxy for AI adoption intensity. By the end of 2025, the share of AI-related job postings in the banking sector had climbed to 6.80%, up from less than 0.94% in 2015. That rate exceeded both the broader finance, insurance, and real estate sector at 3.20% and the economy-wide average of 2.69%. Large banks with assets exceeding $100 billion posted AI-related jobs at an 8.86% rate, compared with 4.48% for medium-sized banks and just 1.15% for small banks with assets below $10 billion. In the sample, 84% of small banks never posted any AI-related jobs during the study period. High-AI banks also carried a slightly higher average share of problem loans—those designated as substandard or with doubtful repayment—though the report notes the disparity was small. Among large banks specifically, the return on assets gap between high- and low-AI institutions reached 0.61 percentage points.
The difference in small business lending was most pronounced among small banks, where high-AI institutions devoted 13.3% of their portfolios to small and medium enterprise loans compared with 21.4% for low-AI banks. For medium-sized banks, the gap was 3.8% versus 7.8%, and for large banks it was 2.2% versus 3.3%. The authors write that "increased AI usage by banks has been associated with decreases in the shares of loans offered to small and medium enterprises," suggesting "more difficulties for SME financing going forward as AI adoption continues to grow." The study emphasizes these are correlations rather than causal claims, since banks with more complex loan portfolios would also have stronger incentives to invest in AI to manage riskier lending.
The report explains that AI and similar information technologies excel at processing hard data like credit scores, financial statements, and formal credit histories, but struggle with soft information such as personal relationships and knowledge of local business conditions. Small business lending has historically relied more heavily on that soft information, which gives smaller community banks a comparative advantage. As AI adoption tilts bank operations toward hard-data analysis, the research suggests, lenders may naturally shift away from the relationship-based loans that small firms depend on. Large banks have more resources to cover upfront costs like computing infrastructure upgrades, worker retraining, and hiring AI-skilled staff, which explains the adoption gap by bank size. Small banks also face financing constraints, challenges attracting AI-capable workers, limited access to necessary data, and reliance on older computing systems ill-suited for AI applications.
The authors caution it's too early to assess AI's full implications for economic activity and credit access, noting that generative AI might eventually help banks process soft information and bridge the gap for small business lending. Still, the early evidence points to measurable shifts in who gets loans and what kinds of risks banks are willing to take. Because commercial banks are industry leaders in AI adoption, their experiences may preview broader economic impacts as the technology spreads to other sectors. The bottom line: AI is making banks more profitable and more willing to take on complex loans, but it's also steering credit away from the small businesses that have traditionally been hardest to evaluate with data alone.