Banks expanding their use of artificial intelligence agents face a cost calculation shaped by human supervision as well as computing, even as customers increasingly use AI for financial tasks.
For a banking customer-service task, McKinsey estimated that human oversight frequently makes up 70% to 75% of an agent’s variable running costs. Tokens account for 20% to 25% in the same example. Those proportions do not measure spending across a bank.
McKinsey’s 2026 survey covered about 46,000 retail banking customers in 13 countries. Nearly a quarter used generative AI for financial tasks, and 74% of those using AI for banking turned to ChatGPT, Claude or Gemini. Among surveyed AI users, 62% wanted AI-based financial help from their main bank.
Some banks have put teams they call AI FinOps in charge of controlling AI spending, Reuters reported, citing McKinsey senior partner Vik Sohoni. Sohoni told Reuters that methods for directing employees’ questions to different AI models become less useful when agents communicate with one another and consume more tokens.
A separate McKinsey analysis of product development estimated that tokens, computing, infrastructure and AI operations together could cost as much as 20% of existing labor spending as agents take on more work. That projection covers several expenses across organizations; it is not a measured bank token bill.
Royal Bank of Canada said May 28 that its large language model token use had risen more than 500% since fiscal 2025. The bank gave no cost figure for the increase.
The potential effect extends to bank revenue. McKinsey said agents could redirect low-cost deposits, reducing banks’ deposit profits. Banks may recover some of that value by offering customers advice and other services, the consultancy said.
