AI moves into live bank processes

Banks and lenders are starting to put artificial intelligence into live credit decisions, moving it from test projects into underwriting and loan-approval workflows. (gao.gov) That shift is showing up in vendor and industry data. A November 2025 retail-lending survey of 106 United States banks, credit unions and consumer finance companies found 83% planned to raise generative artificial intelligence lending budgets in 2026, and 67% said their strategies were already complete or would be implemented by 2026. (businesswire.com) In practice, banks are using these systems to sort applications, pull data, score risk and route files to a human reviewer or an automated decision engine. FICO says its platform combines data, machine learning and human expertise into real-time decisions at scale, including human-in-the-loop controls. (fico.com, fico.com) Credit underwriting is the process banks use to decide who gets a loan and on what terms. Replacing manual review with software can cut turnaround times from days to minutes, which is why banks are pushing automation deeper into origination and servicing. (bai.org, fico.com) The push is also tied to cost pressure. PwC said in October 2025 that front-to-back artificial intelligence adoption in banking could improve a bank’s efficiency ratio by as much as 15 percentage points, a metric banks use to compare expenses with revenue. (pwc.com) Banks are not getting a free pass from regulators as these tools move into production. A May 2025 Government Accountability Office report said federal regulators have issued guidance on artificial intelligence in lending, conducted artificial-intelligence-focused examinations, and told auditors they may refine rules and oversight as risks evolve. (gao.gov) One pressure point is explainability: if a model denies credit, the lender still has to give the applicant specific reasons. The Consumer Financial Protection Bureau said in 2023 that lenders using artificial intelligence or other complex models cannot rely on vague checklist notices that fail to reflect the actual reason for denial. (consumerfinance.gov, consumerfinance.gov) Another is model governance, the internal process for testing and controlling automated systems before and after launch. Federal Reserve guidance that banks still use for model risk management calls for robust development, validation, implementation and oversight by senior management and boards. (federalreserve.gov, federalreserve.gov) Researchers and policy groups have argued that better models can widen access to credit, especially for borrowers who are hard to score with traditional files, but only if lenders can measure fairness and explain outcomes. FinRegLab said its work on machine-learning underwriting focused on model risk management, consumer disclosures and fair-lending compliance. (finreglab.org, finreglab.org) So the story in banking is no longer whether artificial intelligence can help with credit work. It is whether banks can run these systems at production scale, keep humans in the loop where needed, and still meet the disclosure, validation and fairness standards that apply to every loan decision. (gao.gov, consumerfinance.gov, federalreserve.gov)