--- title: "McKinsey: agent token bills could reach 20% of bank labor costs; JPMorgan caps Claude Code at $2,000 a month" description: "McKinsey's Global Banking Annual Review warns that agent token bills could reach a fifth of banks' labor costs, and the caps are already policy at JPMorgan, Cisco and Rippling. The money moves from whoever sells tokens to whoever meters them, the week Anthropic set its IPO terms." publisher: "The Inference" section: "Economy" published: 2026-11-11T15:08:26.793Z modified: 2026-11-11T15:08:26.793Z canonical: https://theinference.org/article/mckinsey-agent-token-bills-could-reach-20-of-bank-labor-costs-jpmorgan-caps-claude-code-at-2-000-a-month language: en keywords: "AI Costs, Agents, Banks, Anthropic, CrowdStrike, Enterprise Software" --- # McKinsey: agent token bills could reach 20% of bank labor costs; JPMorgan caps Claude Code at $2,000 a month > McKinsey's Global Banking Annual Review warns that agent token bills could reach a fifth of banks' labor costs, and the caps are already policy at JPMorgan, Cisco and Rippling. The money moves from whoever sells tokens to whoever meters them, the week Anthropic set its IPO terms. [McKinsey](https://theinference.org/markets/companies/mckinsey) put a number on the bill this week: the tokens that run banks' AI agents could reach a fifth of total labor costs, [its Global Banking Annual Review says](https://www.channelnewsasia.com/business/banks-seek-rein-in-soaring-ai-agent-costs-mckinsey-says-6539721) [1]. Banks are treating that as a procurement problem, and the procurement is already staffed. Responsibility for controlling AI spending has shifted to new teams, which some banks call AI FinOps, whose job is to hold the agent bill down [1]. ## Why chat-era cost controls fail on agents The first round of controls routed simple employee queries to cheaper models. That stops working once agents run, Vik Sohoni, a McKinsey senior partner and co-author of the review, told Reuters, because agents talk to each other and the tokens compound [1]. The spread defeats forecasting: the cost of the same task can [differ by as much as 30 times depending on which agent runs it, McKinsey found](https://www.businessinsider.com/mckinsey-ai-agent-costs-variable-expensive-2026-9) [3]. So banks ration instead, deploying agents where returns are clearest. Wholesale credit monitoring is one of those places, with human credit officers keeping the final lending calls [1][8]. ## The caps are already policy JPMorgan [capped some engineers' use of Claude Code](https://www.businessinsider.com/cisco-goldman-jpmorgan-reconsidering-ai-access-work-2026-11), [Anthropic](https://theinference.org/markets/companies/anthropic)'s coding tool, at $2,000 a month [2]. [Cisco](https://theinference.org/markets/companies/cisco) is handing token budgets to managers across a 32,000-person organization [2]. Rippling, a workforce-management software startup, [added monthly caps and model routing](https://theinference.org/article/accenture-funds-29-of-claude-a-month-per-person-then-guides-margin-up-every-year-to-fiscal-2029) after projecting its uncapped token bill could reach 40% of its research and development headcount budget [2]. *Chart: **Rippling cut its projected token bill from 40% to between 10% and 15%** Bar chart of projected token spending as a share of people-cost budgets: Rippling's uncapped projection at 40% of its R&D headcount budget, banks' agent token spend per McKinsey at 20% of total labor costs, and Rippling's projection with caps and routing at between 10% and 15% of its R&D headcount budget.* *Projected token spending as a share of people-cost budgets; all three bars are projections, not reported results, and the two denominators differ. The Rippling bars are measured against its R&D headcount budget at near-peak token use; the banks bar is McKinsey's estimate of what agent token spend could reach, against total labor costs. Sources: Reuters via CNA; Business Insider. [1][2]* ## The cap is now a product The buyer capping tokens still buys software that polices agents. On November 3 the cybersecurity firm [CrowdStrike](https://theinference.org/markets/companies/crowdstrike) [agreed to buy Fiddler](https://cyberscoop.com/crowdstrike-acquires-fiddler-dreadnode-ai-security/), which records what agents do, and Dreadnode, which attacks them with more than 100 known methods before launch [4]. It also shipped an AI Kill Switch, a control that lets a human stop any agent; the Wall Street Journal put the two deals near $495 million [4][5]. "The more certain a company is that it can stop any agent, the more agents it can put to work," chief executive [George Kurtz](https://theinference.org/markets/people/george-kurtz) said [4]. The shares still fell almost 10% within two days, in a pullback across richly valued software [5]. Its December 1 report should be the first clean read on whether enterprises keep buying agent capability or only the oversight around it [5]. ## The shop that sold the case now supplies the ceiling McKinsey built the adoption case, calculating that agent deployments pay for themselves in about ten months [7]. Its review now supplies the figure a bank can reach for when a vendor asks for a bigger budget. That lands the same week Anthropic, the AI lab behind Claude, which primarily sells to other businesses, [set IPO terms at $70 to $78 a share](https://www.cnbc.com/2026/11/10/anthropic-sets-price-range-of-70-to-78-in-upcoming-ipo.html) [6]. At the top it would be worth up to about $1.4 trillion, on customers who are installing meters on its product as they buy it [6]. The bottleneck has moved from what agents can do to what a task costs. Pricing a task is now real work, done by AI FinOps teams inside banks and by vendors selling routing, caps and oversight, and it decides where agents run next. Whoever does that work well is turning a capability into something a bank can afford to run, and that is where the value now sits. ## Takeaway McKinsey puts agent token bills at up to 20% of bank labor costs, and JPMorgan, Cisco and Rippling already cap and meter them. The money now flows to whoever sells the controls, the same week Anthropic set IPO terms at up to $1.4 trillion. ## Sources 1. [Reuters via Channel News Asia, Banks seek to rein in soaring AI agent costs, McKinsey says, 10 November 2026](https://www.channelnewsasia.com/business/banks-seek-rein-in-soaring-ai-agent-costs-mckinsey-says-6539721) 2. [Business Insider, Bosses are rethinking whether you get access to certain AI tools, 11 November 2026](https://www.businessinsider.com/cisco-goldman-jpmorgan-reconsidering-ai-access-work-2026-11) 3. [Business Insider, McKinsey warns that AI may get a lot more expensive for companies, 22 September 2026](https://www.businessinsider.com/mckinsey-ai-agent-costs-variable-expensive-2026-9) 4. [CyberScoop, CrowdStrike moves to acquire Fiddler and Dreadnode, adding AI observability and testing to its security platform, 3 November 2026](https://cyberscoop.com/crowdstrike-acquires-fiddler-dreadnode-ai-security/) 5. [Yahoo Finance, CrowdStrike Sinks 10% Two Days After a $495 Million Bet on Policing AI Agents, 10 November 2026](https://finance.yahoo.com/technology/ai/articles/crowdstrike-sinks-10-two-days-135321360.html) 6. [CNBC, Anthropic sets IPO price range of $70 to $78, looks to start trading Nov. 19, sources say, 10 November 2026](https://www.cnbc.com/2026/11/10/anthropic-sets-price-range-of-70-to-78-in-upcoming-ipo.html) 7. [finews.com, Swiss banks especially profitable, but pressure is growing in AI, 11 November 2026](https://www.finews.com/news/english-news/74432-global-banking-sector-new-record-profits-swiss-institutions-study-mckinsey-artificial-intelligence-access-to-bank-customers-new-challenges) 8. [ET BFSI, Banks are using AI to identify corporate loan risks before defaults: McKinsey, November 2026](https://bfsi.economictimes.indiatimes.com/articles/banks-are-using-ai-to-identify-corporate-loan-risks-before-defaults-mckinsey/135719400) --- McKinsey: agent token bills could reach 20% of bank labor costs; JPMorgan caps Claude Code at $2,000 a month — The Inference. Canonical: https://theinference.org/article/mckinsey-agent-token-bills-could-reach-20-of-bank-labor-costs-jpmorgan-caps-claude-code-at-2-000-a-month