Banks Set Up AI FinOps Teams as Agent Tokens Head Toward 20% of Labor Costs

Banks are starting to give AI-agent tokens their own team and their own budget line.

McKinsey's Global Banking Annual Review, reported by Reuters on November 10, projects that agent token spend could reach 20% of total bank labor costs. Vik Sohoni, a McKinsey senior partner and one of the review's authors, said routing work by query complexity breaks down once agents work together, because "they communicate with each other and tokens quickly add up." Reuters reports that responsibility for the spend has moved to new teams that some banks call AI FinOps.

A FinOps team gives agent tokens an owner on the buyer's side. For a usage-priced AI vendor, growth inside a bank likely depends on getting a spending limit raised as well as on usage growing.

JPMorgan capped Claude Code at $2,000 a month

JPMorgan has set a $2,000 monthly limit on Claude Code for some of its engineers, Tokenpost reported on October 22. The limit resets each month, and engineers can ask for more. On the October 13 earnings call, CFO Jeremy Barnum called the bank's earlier estimate of about $700 million a year in AI expense rough and stale, according to Tokenpost. Barnum said token costs could absorb part of the productivity gains.

The cap is a per-person ceiling, much like the $29 Claude allowance Accenture set. Anthropic sells the controls that make such ceilings easy to set. Its Claude Code cost documentation describes three kinds of spend control:

• workspace spend limits
• the seat allowance as the default ceiling
• spend limits at the organization, group and member level

The cap sits well above typical use

Anthropic puts the average enterprise cost at about $13 per developer per active day, or $150 to $250 a month, with 90% of users below $30 a day. JPMorgan's $2,000 is 8 to 13 times that monthly average. A limit that high, raised on request, likely catches only the heaviest users and leaves typical use alone.

Tokens may also be the smaller cost. McKinsey's earlier research, as reported by Coinpaper, found human oversight made up about 70% to 75% of variable cost in some banking customer-service workflows, against 20% to 25% for tokens.

Agent spend now passes a FinOps check before it grows

If AI FinOps teams set ceilings the way JPMorgan did, a vendor's expansion in an account could run through a request to raise a limit, approved by that team. If the ceilings stay far above typical use, usage keeps growing on its own and the request matters only for the heaviest engineers.

they communicate with each other and tokens quickly add up.

The number to watch is where banks set their AI FinOps ceilings. JPMorgan's $2,000 monthly limit sits 8 to 13 times Anthropic's $150 to $250 average.

By Business of GTM