Banks Work to Control Rising AI Agent Costs as Token Spending Grows

Banks around the world are trying to control the fast-rising costs of artificial intelligence agents because spending on the tokens that power them could soon reach 20% of total labor costs, a new McKinsey report shows. To handle this spending, banks are setting up new groups known as AI FinOps. Early efforts to save money involved routing employee questions to different AI models based on how hard the task was. However, those methods do not work well when AI agents run tasks on their own because the agents talk to each other and token use adds up fast, according to Vik Sohoni, a senior partner at McKinsey and a report author. Because of these high costs, banks are now being more careful about where they use agentic AI and are focusing on areas that bring the best returns. While saving time is easy to measure, banks are also looking at how AI can help grow revenue to justify the higher spending. Customers are taking up AI much faster than past technologies like the mobile internet, which took five to ten years to become common, while AI agents are expected to see wide use in retail banking within three years. A McKinsey survey of 46,000 retail banking customers across the world showed that most users already use AI to compare services, find better savings accounts, and manage debt. Out of those AI users, 72% said they would let a banking agent do one financial task, but they would set a limit of $500 for automatic approvals. Meanwhile, Meta's Muse agent hit 5 million downloads in the United States during its first 22 days after launch.

Context

Financial institutions constantly look for ways to balance the high costs of new technology with the productivity and revenue gains it provides. As artificial intelligence tools become more common in everyday consumer banking, banks must balance customer demand with security rules and potential financial risks.