No one wanted to hand the technology team a blank check. Yet when companies first rushed to embrace AI, many chief financial officers were effectively asked to do exactly that. Now, after receiving several unexpected bills, keeping costs under control has become an urgent priority.
This issue is gradually building within the banking sector. Bank technology budgets were already squeezed by competing demands, from modernizing legacy IT systems to cybersecurity, all vying for resources. One technology that the industry holds high hopes for in boosting productivity is the AI agent, which can autonomously handle complex tasks, but such AI is inherently difficult to budget for, because how much computing power each task will consume is often unpredictable.
Bhavi Mehta, a partner at Bain & Co. and global head of advanced analytics for financial services, said he does not actually think anyone in organizations has a real handle on it; if you want to do XYZ, what exactly will the cost be. He said that as AI moves toward more complex reasoning and agentic workflows, the number of tokens needed to complete a specific task can vary widely.
As banks scale up AI adoption, mastering costs while measuring business value is becoming increasingly important. According to a latest report by AI evaluation firm Evident, among the 50 largest banks it tracks globally, only 12 have publicly explained the expected or realized returns from AI adoption. Yet nearly every chief executive is talking up AI's potential benefits.
Several US banks, from Citigroup to Goldman Sachs, discussed increased AI investment on their third-quarter earnings calls, but few executives could provide exact figures on AI spending or the financial benefits it brings. As the banking industry moves from the experimentation stage to actual deployment, Bloomberg interviewed more than a dozen engineers, corporate executives and consultants to learn about their experiences and coping strategies.
While the practice of tokenmaxxing, in which engineers use as much AI as possible, has become notorious, runaway token usage is not always caused by this. Christoph Rabenseifner, Deutsche Bank's chief strategy officer for technology, data and innovation, said the bank's monitoring system found that after individual employees consumed unusually large amounts of tokens, it turned out many of them were actually doing work with real benefits. As a result, the bank did not cut off their access, but guided them to use AI in a more efficient way.
Rabenseifner also said the bank once shifted a task to a smaller, specially trained model instead of using a large frontier model, cutting costs to a fraction of the original. Banco Santander SA only first uses large frontier models to test whether an application can meet quality standards, and once feasibility is confirmed, it moves the work to lighter, lower-cost models. Ignacio Bernal, the bank's chief AI officer, said this approach reduced the execution cost of each loan approval process from about 5 euros ($5.61) to 0.03 euros.
Bernal said that when a global bank moves AI from the experimental stage into formal operations, inference costs are no longer a negligible small figure, but become a real operating expense item. He noted that some processes run millions of times a year, so computing spending can increase substantially. Canadian Imperial Bank of Commerce (CIBC) found that the most powerful models, including Anthropic's Opus, can cost as much as C$3,000 ($2,130) per employee per month; the combined expense for thousands of employees can be quite startling.
To control costs, the bank limited access to several hundred employees, while its latest internal work platform still under testing, CAI 2.0, is capped at 250 users, each allowed to spend C$4 to C$250 per month. AI agents are becoming a major driver of rising costs. According to Evident, large banks announced 19 AI agent use cases in the first quarter of 2026, up from 8 in the previous quarter. The number halved in the second quarter, but this was not because banks' enthusiasm for AI agents cooled, but rather because they began asking them to take on more complex work.
For example, Wells Fargo wants AI agents to master the 1,100 steps in the mortgage process before taking on various workflows. Unlike traditional AI tools, agent systems can autonomously perform complex multi-step tasks, including analyzing financial statements, tracking borrowers, verifying documents and detecting money laundering activity. This kind of processing work requires substantial computing resources, meaning more token consumption and further driving up monthly costs. A senior data executive said that after some teams used up their allocated token quotas in the first week, the bank decided to centralize token quota management and build an AI agent to streamline prompts before they are sent to the model. The executive asked for anonymity when discussing internal matters.
Comments