Chief financial officers at global banks have begun clamping down on the cost of adopting artificial intelligence, as bills running well above projections turn spending control into a pressing task.
Banks are using smaller AI models and Chinese models with cheaper token pricing to trim their technology budgets, Bloomberg reported on the 21st.
Banco Santander runs large frontier models only to check whether an application clears its quality threshold, shifting validated tasks to lighter and cheaper models. Ignacio Bernal, the bank's chief AI officer, said the approach cut the running cost of its loan approval process to about 0.03 euro per case from roughly 5 euros.
In a July report, J.P. Morgan Private Bank said the GLM model from Beijing-based Z.AI delivered performance comparable to the best model available in March while costing more than 80% less than the most expensive frontier model at the time. Large Singaporean banks are also testing multiple models. United Overseas Bank said its vendor-neutral AI framework is open to Chinese suppliers, while DBS Group said it avoids excessive reliance on any single source, whether American or Chinese.
Cost is the reason banks are willing to consider Chinese models. Canadian Imperial Bank of Commerce found that using Anthropic's top-performing model could run up to 3,000 Canadian dollars (about 2.83 million won) per employee a month. The bank cut the number of staff with access to the hundreds from several thousand. Its in-house workspace, CAI 2.0, which is still being tested, limits access to 250 employees and keeps spending within 4 to 250 Canadian dollars (about 3,800 to 240,000 won) per person a month.
Agentic systems autonomously carry out multistep tasks such as analyzing financial statements, tracking borrowers, verifying documents and detecting money laundering. They consume more computing resources and burn through more tokens, pushing up monthly costs. At one Asian bank, some teams used up their entire token allocation in the first week, after which the bank decided to pool tokens and build an agent that trims prompts before they are sent to a model, a data executive there said.
Banks Eye Chinese Models at 80% Discount as Regulators Push Back
Regulators are uneasy about the use of Chinese models. Canada's banking regulator recently instructed staff to draw up minimum standards, including on security, for how banks use frontier AI models. Michelle Bowman, vice chair for supervision at the U.S. Federal Reserve, warned banks to harden their systems against AI-enabled threats.
Beyond cutting prices, global banks' strategies for using AI remain thin. According to the latest report from AI benchmarking firm Evident, only 12 of the world's 50 largest banks have disclosed expected or realized returns from AI adoption. U.S. banks including Citigroup and Goldman Sachs cited expanded AI investment in their third-quarter earnings presentations, but few executives put numbers on spending or on the financial effect of that spending. Bobby Mehta, a partner at Bain & Co., said the more banks scale up AI, the more important it becomes to look at costs alongside it.
Even so, banks cannot afford to let go of AI, with some projecting that AI agents could be entrusted with bond trading. Michael Roberts, chief executive of HSBC Holdings' corporate and institutional banking division, said in an interview with Bloomberg TV that quantum computing will be able to examine highly complex market movements, and that as AI accelerates that process, agents will be seen using quantum computing to trade bonds.