When the token bill balloons
At first, banks thought they had solved the cost problem: an automated routing system sent employees' simple questions to cheap AI models and reserved expensive models for complex ones. But once AI stopped merely answering and began acting on its own through agents, that arithmetic quickly broke down.
But those models are ineffective when AI agents are working, because they communicate with each other and tokens quickly add up.
Sohoni, a senior partner at McKinsey and one of the report's authors, explained to Reuters that each agent can call another, double-check the result and send yet another request, causing token consumption to multiply. Responsibility for controlling spending has therefore shifted to new teams that some banks call AI FinOps.
Selective bets
The direct consequence is that lenders are becoming choosier. Rather than spreading agentic AI across every department, they are concentrating on areas with the highest expected returns.

















