AI in action · Tools
Banks' AI Agents Could Cost 20% of Their Labour Bill
Token spending on AI agents could reach 20% of banks' total labour costs, McKinsey estimates, and some lenders have set up AI FinOps teams to control it.
The pitch for AI in banking has mostly been about savings. McKinsey's latest Global Banking Annual Review, reported by Reuters on November 10, 2026, looks at the other side of the ledger: the cost of running AI agents. It estimates that spending on tokens, the units AI providers charge for, could reach 20% of banks' total labour costs.
Agents are the reason the bill grows so fast. A chatbot answers one question; an agent breaks a task into steps, calls tools, checks its own work and tries again, and each step consumes tokens. Early cost controls sent simple employee queries to cheaper models and hard ones to expensive models, but McKinsey senior partner Vik Sohoni told Reuters that approach breaks down with agents because "tokens quickly add up." Some banks have handed the problem to new "AI FinOps" teams, borrowing the name from the groups that police cloud computing bills.
The response, according to the report, is selectivity. Banks are narrowing agent deployments to the areas with the strongest expected returns, and they are starting to judge AI projects on revenue growth as well as on productivity savings, which are easier to measure but may not justify the spending alone.
Customers are moving at the same time. In a McKinsey survey of 46,000 retail banking customers worldwide, most already use AI to compare banking services, look for higher-yielding savings accounts or refinance debt. Among AI users, 72% said they would let a banking agent complete a financial task for them, but would cap its autonomous approvals at $500.
That points to a risk for bank profits that goes beyond cost. Agents acting for customers could move money out of non-interest-bearing accounts into higher-yielding ones, squeezing a cheap source of funding. For now the effect is limited because most banks block outside AI agents from authorising transactions in their systems; McKinsey expects agents to be widely used within three years, much faster than the five to ten years mobile internet took.
Sources: Reuters via Investing.com — Banks seek to rein in soaring AI agent costs, McKinsey says (Nov 10, 2026)