- Key insight: After the pricing structure for AI changed from subscriptions to tokens, using the technology became much more expensive. But experts say there are ways to cut down the costs without losing the benefits.
- Supporting data: Royal Bank of Canada says its use of tokens, the new currency for AI, jumped 500% from 2025 to 2026.
- Expert quote: "Any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens." — PNC Financial Services Group CEO Bill Demchak
For years, AI has been pitched as a tool for cutting costs. Now, as the technology suddenly grows far more expensive, many banks are scrambling to cut the costs of using AI.
Over the past few months, the way tech companies charge businesses for artificial intelligence products has fundamentally changed. Prominent AI providers including Anthropic, OpenAI and Microsoft have switched their pricing models from subscriptions, which offered open-ended use over certain time
frames, to charging by "tokens," the units of data on which AI runs.
The change, for banks and other businesses, is akin to entering a restaurant labeled "All You Can Eat" and then being told to pay a la carte. As the checks begin to arrive, many businesses are suffering sticker shock.
"Sometimes people are surprised at how much money they're burning," Rob May, CEO of Neurometric, a company that consults companies on their AI use, told American Banker.
Amid the soaring costs, some banks have
If the expenses are not reined in, some are warning, the use of AI at banks could end up defeating its purpose.
"Any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens," Demchak
How can banks get their AI costs under control? One way is to deliberately step back from the cutting edge. Part of the problem has been that as AI models have reached the "agentic" stage — meaning they can now make decisions and complete tasks almost entirely by themselves — the rate at which they gobble up tokens has skyrocketed. But for many of the jobs banks use AI to do, this level of sophistication isn't necessary.
"If you're going to do simple tasks, those models are overkill," May said. "Why spend all the money on a giant model that can do that, but can also write you a great lasagna recipe and give you workout instructions? … You need a model that knows just enough to get done the task that you need to do."
The solution: Use less advanced models. In some cases, these are available as the older, supposedly obsolete products left behind by frontier AI companies. In other cases, banks are building their own "small language models" to handle such tasks. (For example, Bank of America's AI-powered virtual assistant,
This kind of careful selection of AI tools is already catching on at PNC.
"What we have learned is, you need to make sure that you've got the right tool for the right problem," Ned Carroll, PNC's head of data and automation, recently told American Banker. "I don't need a model to answer advanced calculus when I want to understand a policy or procedure around a check return."
Other token-saving tricks include tapping into open-source models, which are free to use, and relying on a simple but powerful habit: Saving the answers to old AI queries.
"If you're asking a lot of the same things of the models all the time, and the answers are coming back really similar, cache that," May said. "Hit the database first, see if it's there, and then you don't have to ask the model, and you don't pay any token costs."
Another strategy is for banks to build more of their AI infrastructure in-house. A growing number of banks own their own data centers, and PNC has publicly vowed to reduce its dependence on third-party tech vendors.
"We will have our own GPU compute," Demchak declared at the Morgan Stanley conference, referring to the graphics processing units that power modern AI. "We will not be as reliant on burning external tokens [as] what we will do internally for our own large language models."
Then there's another, more quaint solution: If using AI for a certain task is hugely expensive, it may actually be cheaper to use good old-fashioned human labor.
"There are still use cases for that," May said. "If you have a use case where the cost of failure is really expensive … you might just leave it as humans."
