Banking and Artificial Intelligence: Near-Universal Use, Thin Evidence on Outcomes, and a Supervisory Agenda in the Making

Abstract

Purpose. Banks were among the first industries to use machine learning at scale, in credit scoring and fraud detection, and are now among the first to deploy generative AI. This article asks what is known about the results, what slows the next stage, and where banking may be by about 2035. Design and methods. A review combining the share of the banking research literature that involves AI (OpenAlex, 2010–2025), supervisors' surveys of AI use in banks (Bank of England and Financial Conduct Authority, 2024; European Central Bank, 2026), peer-reviewed studies of AI in credit markets, asset management and customer service, and international regulatory reports. Claims from the literature were checked against the cited sources, and the scenarios are judgements built on them. Findings. The share of banking works that mention AI rose from 2.1% in 2018 to 15.5% in 2025, 7.5 times as much, and language-model works from 15 to 652. Supervisory surveys show use that is close to universal: 75% of UK financial firms were using AI in 2024, with foundation models in 17% of use cases, and more than 85% of the large banks under European supervision used AI in some form in 2026. The evidence on outcomes is much thinner and mixed. Machine-learning credit models predict default better but, in US mortgages, Black and Hispanic borrowers were disproportionately less likely to gain; the staggered introduction of a generative assistant in customer support raised issues resolved per hour by 15% on average. The binding constraints named by firms are safety and robustness, skills, explainability and data protection, and supervisors point to third-party concentration and correlated behaviour. We expect generative AI to be routine in internal operations and customer service by 2028, credit decisions to be the main arena of regulatory friction as EU high-risk rules apply from December 2027, and the longer-term picture to depend on governance more than on models. Originality. The article puts the most-cited figures on AI use in banking next to the much smaller body of outcome evidence, distinguishes uses that are well evidenced from uses that are only widespread, and gives signposts for the supervisory and competitive questions that will decide the next stage.