Banking AI learned to refuse improper requests 80% of the time after training.

Banks need chatbots that answer questions and help with accounts safely. Everyday artificial intelligence models fail at this because they make mistakes with real facts, digital tools, and sensitive bank rules. To fix this, researchers built FiMI Banking, a controlled test setting based on vetted Indian retail-banking records, tools, and synthetic customer profiles. They tested two ways to train models after initial learning. First, preference optimizationpreference optimizationA training method that adjusts AI behavior based on preferred responses. trains an AI on good and bad responses. This raised safe refusals of out-of-scope requests from 52% to 80%. Second, reinforcement learning rewards the AI when it completes multi-step tool actions correctly. This boosted edge-case scores from 0.509 to 0.718, raised order-sensitive task scores from 0.590 to 0.679, and used 29% fewer generated tokens.

Why it matters

Banks can deploy AI that safely handles customer interactions, reducing errors and ensuring regulatory compliance. This protects both the bank and its customers from mishandling of sensitive requests.

Questions answered