Bankers Turn to AI Tokenomics to Control Rising Technology Costs

Banks are moving beyond artificial intelligence experiments and deploying AI across more business operations. As usage expands, financial institutions are paying closer attention to a critical question: what value are they receiving for every token consumed?

Early enterprise AI agreements often relied on flat-rate pricing, giving companies broad access to powerful tools for a predictable fee. That model has increasingly given way to usage-based pricing, with leading AI providers charging according to the volume of data processed by their models.

That data is measured in “tokens,” creating a new discipline known as AI tokenomics. The term describes the process of monitoring, managing and optimizing the cost of enterprise AI usage.

Why Token Usage Has Become a Banking Issue

AI providers are processing enormous volumes of tokens each month, and analysts expect consumption to continue increasing as businesses assign more work to AI systems and autonomous agents.

However, higher usage does not automatically translate into higher productivity. Moody’s Ratings has warned that the corporate practice of maximizing token consumption simply to demonstrate AI activity can increase expenses without producing comparable business benefits.

AI agents may drive costs even higher. Each time an agent calls a software tool, repeats a task, exchanges information with another agent or expands its reasoning process, additional tokens may be consumed.

• Accenture expects corporate token consumption to increase by 78% over the next two years.
• Token prices are expected to decline by approximately 19% during the same period.
• One-third of surveyed companies said they had used their annual token budget before the year ended.
• Agent-based workflows currently account for about 11% of token usage, but that share is expected to rise.

The result is growing pressure on bank executives and boards to connect AI spending with measurable returns. Even if the price of individual tokens declines, total spending can continue to rise as more processes are handed over to AI.

What Are AI Tokens?

Tokens are units of data that an AI model reads, processes or generates. They are not standardized measurements like pounds or inches, and their size can vary by provider, model and language.

AI providers generally separate token charges into two categories:

• Input tokens: The information, instructions, documents and context supplied to the model.
• Output tokens: The text, analysis or other result generated by the model.

Input tokens typically represent the majority of usage, while output tokens often carry a higher price because they produce the requested result. A broad prompt containing extensive background information may therefore cost more than a focused request.

Pricing also depends on the sophistication of the model. Frontier models deliver advanced capabilities but are usually more expensive than smaller or less powerful alternatives. For that reason, banks can reduce expenses by matching each task with a model capable of completing it without unnecessary processing power.

Costs can be difficult to predict, particularly when AI agents are involved. A short instruction may trigger a long chain of automated activities and consume millions of tokens. Without monitoring and controls, users may not know whether a request will cost a few dollars or hundreds of dollars.

Banks Are Asking What AI Spending Delivers

Concerns about AI costs are increasingly appearing in bank earnings discussions and industry conferences. Bank leaders have acknowledged that artificial intelligence can create meaningful productivity gains, but they also recognize that deployment, governance and ongoing usage carry significant expenses.

Bank of America Chairman and CEO Brian Moynihan has referred to the growing importance of tokenomics, emphasizing that AI spending cannot grow without limits and must be tied to business value.

Executives at First Horizon have similarly noted that early expectations that AI would immediately remove costs from the system did not account for the expense of operating AI programs and establishing appropriate governance.

U.S. Bank leadership has also highlighted the need to determine whether token expenses are outpacing returns. The institution has invested in capabilities that allow it to switch from costly models to more economical options when appropriate.

These comments reflect a broader change in the banking industry. The discussion is shifting from how much institutions spend on AI to what they receive in return for that investment.

Managing AI Costs in Production Environments

As banks move AI projects from pilot programs into full production, controlling token usage is becoming a core management responsibility. One challenge is changing employee behavior after years of using software under flat-rate or subscription pricing.

Users may be accustomed to entering lengthy prompts, attaching large files and supplying extensive context. Each additional word, document or repeated instruction can increase token consumption.

Techniques such as caching can help reduce unnecessary processing. By retaining relevant information from previous work, an AI system may avoid repeating the same analysis for a similar request. Banks must also manage the size of a model’s context window, balancing the quality of the response against the cost of processing more data.

Capgemini’s Ravi Khokhar compares the process to choosing the right vehicle for a task. A powerful, expensive model may be useful for complex work, but it is unnecessary for a simple request that a smaller model or software application can handle.

Training employees to write focused prompts and giving them clear usage guidelines can reduce token consumption. Banks that have adopted optimization measures may lower their token volume by 10% to 30%, with further savings possible through stricter controls and better prompt design.

Four Ways Banks Can Improve AI Tokenomics

1. Set Individual or Departmental Token Budgets

Banks should avoid providing unrestricted access to expensive AI models. Allocating token budgets to employees, teams or projects creates accountability and establishes a clear process for requesting additional capacity.

2. Use AI Model Routers

Model-routing technology directs each request to the most appropriate and cost-effective system. A simple task may be handled by a smaller model, while a complex analysis can be sent to a more capable platform.

For example, a request to retrieve the temperature in a specific city may be handled through an application programming interface rather than a large language model. This approach prevents expensive models from being used for routine tasks.

3. Introduce Showbacks and Chargebacks

Showbacks make users aware of the cost of their AI activity. The information can be compared with productivity measures, service levels and other key performance indicators.

Chargebacks go a step further by assigning AI expenses to a specific department, project or business line. This cloud computing practice encourages teams to consider cost before an AI application is expanded across the organization.

4. Consider Open-Weight and Open-Source Models

Not every banking task requires a frontier AI model. Accenture estimates that only about 10% of AI workloads need the most advanced systems.

Open-weight and open-source models may provide a more affordable option for routine activities, depending on a bank’s security, compliance and performance requirements. Selecting these alternatives where appropriate can reduce reliance on high-cost models.

The Future of AI Spending in Banking

Tokenomics is likely to remain a major management concern as financial institutions expand their use of generative AI and autonomous agents. Banks will need to track not only how much AI they use, but also which models are being used, which departments are responsible for the spending and whether the resulting benefits justify the expense.

The most successful institutions will treat AI usage as an operating discipline rather than an unlimited technology resource. Through budgets, model routing, employee education and detailed performance measurement, banks can pursue AI-driven efficiency while keeping costs under control.

Source: thefinancialbrand.com