Key takeaways
- AI agent token spending could reach 20% of payroll
- Human oversight adds substantial costs to banking automation
- Autonomous financial assistants threaten traditional bank profit margins
Banks worldwide are facing an unexpected financial challenge as artificial intelligence agents become more expensive to operate, potentially consuming the savings institutions expect from automating customer service, compliance and administrative work.
McKinsey warns that spending on AI tokens, the units used to process information through language models, could reach 20% of banks' total labor costs. The warning comes as global banking profits reached a record $1.3 trillion in 2025, according to the firm's newly released Global Banking Annual Review.
AI Agent Costs Could Reach 20% of Bank Labor Spending
Banks initially adopted AI assistants to accelerate routine work, including answering customer questions, reviewing documents and assisting compliance teams.
However, autonomous agents create a different cost structure because they can execute multistep workflows, repeatedly query language models and communicate with other agents.
McKinsey senior partner Vik Sohoni explained that conventional cost controls become less effective when agents interact continuously, allowing token consumption to accumulate rapidly.
Some institutions have established AI FinOps teams to monitor computing expenses and direct workloads toward more economical models.
Yet token spending represents only part of the problem. McKinsey's earlier research found that human oversight accounts for approximately 70%–75% of variable costs in certain banking customer-service workflows, compared with 20%–25% for tokens.
That helps explain why banking AI regulation increasingly emphasizes human accountability and transaction-level controls.
Banks Face Automation Savings That May Never Reach Profits
The economics become more complicated when employee productivity improves without reducing actual operating expenses.
McKinsey's research found that AI agents reduced task completion times by approximately 60% across seven production systems at financial and professional services firms.
However, shorter tasks do not automatically translate into smaller teams or lower payroll expenses.
This distinction has become increasingly important as banks pursue workforce restructuring alongside technology investments.
A bank might save two hours per employee through automation while continuing to pay the same salaries, plus additional AI subscription, infrastructure and supervision costs.
AI Agents Could Also Threaten Banking Revenue
The financial risk extends beyond operating costs.
McKinsey surveyed approximately 46,000 retail banking customers and found that 23% already use generative AI for financial tasks.
Among customers using AI, 72% expressed willingness to delegate at least one financial task to an agent, although authorization limits remain a significant concern.
This shift could accelerate autonomous financial transactions, allowing customers to compare savings rates and move deposits toward higher-paying institutions.
For banks, that threatens a traditional source of revenue: inexpensive customer deposits used to fund interest-earning loans and investments.
The development also introduces new payment liability risks, particularly when software executes transactions without individual human approval.
Meanwhile, increasingly capable AI investment agents could expand automated financial decision-making beyond ordinary banking.
