Banks Learn the Art of ‘Tokenomics’ to Make More Efficient Use of AI Tools

By Steve Cocheo, Senior Executive Editor at The Financial Brand

Published on November 5th, 2026 in Banking Technology

Simple Subscribe

Subscribe Now!

When companies first began using today’s frontier AI tools, flat-rate pricing was the usual arrangement. Banks were easing into the leading edge technology with the advantage of ordering off an all-you-can-eat buffet.

What changed: Then AI pricing approaches shifted. Major AI labs began charging by volume of usage of their most advanced models, measured in units called “tokens.”

The corporate appetite for AI continued to grow, and at first the consumption of tokens became a bragging point.

In a July analysis, Moody’s Ratings noted that AI providers were providing processing measured in quadrillions of tokens per month, with growth expected to keep climbing into 2027 at least.

But there was a catch.

“The corporate fad of ‘tokenmaxxing’ — maximizing token usage on the assumption that more usage signals more productivity — has backfired,” Moody’s observed.

The report also noted that as AI agents take over more tasks, usage will run even higher. Every time an agent reaches out for a tool, tries a task multiple times to refine results, works with other agents, and more, the token meter is running, just as it does when humans tap AI.

Need to Know:

“Tokenomics” is the discipline of measuring and controlling the cost of enterprise AI use now that major players charge by volume.
This pricing shift is putting more heat on management to show return for the money spent on AI.
An Accenture global survey of CIOs found that companies expect token consumption to grow by 78% over the next two years, even though they expect that token prices will fall by 19% over the same period.
One out of three respondents said they had exhausted their annual token budget before the year was over.
Accenture also found that agentic workflows account for only 11% of token usage currently, but is expected to ramp up significantly.

Why this matters: The issue is return on investment. Institutions have had to adapt to token-based pricing, balancing tasks against the level of sophistication of the model selected, in order to control costs and to avoid blowing through annual AI budgets in just a few months.

“As consumption scales, token economics is becoming a material risk and opportunity, drawing increasing attention across the C-suite and the board,” according to an Accenture report, “The CIO’s Guide to AI Tokenomics.” The economics of AI — benefits gleaned from its use versus the cost of obtaining those benefits — is also being examined by bank stock analysts.

Even as token prices fall over time, the cost of spending on tokens is expected to continue rising simply because more and more work will be turned over to AI.

“Tokens have always been part of any AI calculation,” says Ravi Khokhar, global head of cloud for financial services at Capgemini, in an interview. “What has made tokens a lot more a topic of discussion is that they are now being used as a measuring scale for commercializing usage.”

Key insight: “In the last 18 months the conversation has shifted from how much are we spending to what are we getting back?” says Khokhar.

What is a Token and What Does It Cost?

Before getting too far, let’s examine the terminology. Tokens are units of data processed by AI models. Tokens are not a standard unit like an inch, a pound or a gram, but can vary in size from one AI model and provider to another. The basic pricing of tokens can be found on the sites of AI providers.

The cost varies by whether the data is input that’s needed for the model to chew on (“input tokens”) versus data that provides the result the user seeks (“output tokens”). The latter cost more because they deliver results, a value, versus ingesting information to process.

“Input tokens represent most of the usage,” says Capgemini’s Khokhar. “Output tokens are usually the least .” He explains that input tokens account for all the raw material fed into the AI, and output tokens account for the desired content. As a rule of thumb, it’s a 95%/5% split. The broader the prompt given to a model, the more output tokens will be charged, he says, while narrower prompts will generally cost less.

Token pricing gets very specific, according to the sophistication of the model the user wants to access. A printout of the enterprise AI rate card for OpenAI’s ChatGPT runs for ten solid pages. Prices quoted are for million-token batches — it takes a huge number of tokens to get things done.

One of the ways that banks can control costs is deciding when a lower-level — cheaper — model will suffice, versus always using the state-of-the-art model.

Projects can breed surprises. “One sentence of instruction can trigger a chain of agent actions that burns millions of tokens, and there is no preview,” says the Accenture report. “You have no gut sense of whether a task will cost one dollar or a hundred.”

-- Article continued below --

‘What Are You Getting for the Tokens You Spend?’

During bank earnings briefings in 2026, the rising cost of AI versus results began coming up.

• In June, during a Morgan Stanley conference, Hope Dmuchowski, senior EVP and CFO at First Horizon, said that, “Early on, there was a lot of expectation that AI was going to take all these costs out of the system, and it was just going to fall to the bottom line. But there is a cost to running AI. We’re having to set up AI governance.”

• In May, at a Bernstein conference, Bank of America Chairman and CEO Brian Moynihan spoke of the increasing need to manage AI expense. “You’re going to start to have the thing they call ‘tokenomics’,” said Moynihan. “The amount of money we spend on this can’t be infinite because it’s got to have value.”

• In October, during third-quarter earnings, responding to questions from Wells Fargo analyst Mike Mayo, Moynihan spoke of 130 projects implemented. None were yet huge contributors to results, he said, but the benefits seen so far were two to three-fold the costs, over time.

• In another third-quarter briefing, U.S. Bank’s president, CEO and chairman, Gunjan Kedia, acknowledged that, “whether the token costs are overtaking returns is a relevant discussion. We have invested in developing the capabilities to swap out expensive models for cheaper models, and we have some confidence over time that we’ll continue to get positive returns from our AI program on productivity.”

Tokenomics Will Keep Management Eyes on Costs

Such considerations will intensify as more banks move from proof of concept projects with AI towards actual production with AI tools, according to Capgemini’s Khokhar. The industry’s pioneers are focusing on managing the costs in production scenarios now, he says.

One of the challenges of controlling costs is changing users’ mindset, according to McKinsey‘s Kaavini Takkar. The days of flat-rate pricing let users grow “used to not prompting efficiently,” said Takkar, associate partner.

“With every extra word I’m typing, every extra file I’m putting in, I’m forgetting how to cache and forgetting how to reduce my context window,” she said.

“Caching” refers to such practices as retaining past work by a model so that a new, similar project doesn’t have to repeat that part of the work again, which repetition adds to costs. “Context window” refers to the amount of data a model can hold and work with at a given time. Users have to balance completeness, which can increase quality of results, versus length, which increases costs in tokens.

“It’s the difference between driving a Cadillac to go buy milk from the grocery store versus driving an EV,” says Khokhar. “Do you really need a Cadillac to do that?”

Khokhar says optimizing AI token spending depends on educating users in how to use the technology economically and imposing tools and rules that automatically apply discipline to usage. He says that banks that have started to optimize can save between 10%-30% of their token volume. He adds that banks that make a real religion of controlled usage, especially stressing more care in setting the context for prompts and better focusing prompts, can expect to save even more.

How to Get Better at Tokenomics

Some measures recommended by Khokhar and other experts:

1. Don’t give users a blank check for AI, instead allocating each a token budget. The budget represents a hard stop and users have to ask for more tokens before they can exceed that.

2. Adopt model router technology. Simply put, this tool shunts each piece of an AI request to the most economical option for the task.

“The decision is taken away from developers and users around which model to use, with a central intelligent engine deciding the right way to serve a request,” says Khokhar. For example, if one piece of a task requires finding out the temperature in a given city, the router will access an application programming interface, rather than incurring the expense of sending that task to a large language model.

3. Use “showbacks” and “chargebacks.” Showbacks inform users what their queries are costing the bank and can be correlated with measures like key performance indicators, according to Accenture. Chargebacks are where those costs are allocated to specific projects or departments. This is an example of a practice from cloud computing that can be adopted. Accenture recommends that the costs of a task be clear before it scales to production.

4. Explore open-weight and open-source models. Accenture estimates that only about 10% of AI tasks really require the frontier models. Some tasks can be handled by models other than these token-guzzlers. Open-weight models and open-source models are similar in concept, but differ in how they can be used and what comes with the model.