Banks warned token bills could erode GenAI savings

Financial firms are being warned that rising consumption of generative AI tokens could weaken the cost case for wider deployment, even as per-token prices fall. Steve Morgan, banking industry market lead at Pegasystems, said in a Finextra opinion post that agentic AI can create unpredictable operating expense when banks use it for tasks that do not need higher-cost reasoning models.

Morgan said the first phase of GenAI adoption was shaped by subsidised pricing, large usage allowances and a push to encourage experimentation across industries. That has shifted, he argued, toward measuring and celebrating use itself, including what he called “tokenmaxxing,” where users or systems consume more tokens than may be needed for the business outcome.

In large language model systems, tokens are the units of text and reasoning that providers use to measure consumption and charge customers. Morgan said an AI agent may interpret a request, select tools, pull information, assess results, retry failed steps, validate an answer and then draft a response. Each step can add tokens, with more complex tasks requiring more computation and higher bills.

Rising bills despite cheaper units

According to Morgan, cheaper token pricing has not prevented enterprise AI bills from rising because companies are assigning more complex work to GenAI tools. He said the cost issue becomes sharper when organisations use agentic AI for processes where a narrower technology, lighter AI model or human review would be more appropriate.

Morgan cited reports involving JPMorgan, Uber, Walmart and Royal Bank of Canada as examples of token use exceeding budgets or forecasts. He referred to a Semafor report that JPMorgan employees were being encouraged to integrate AI into daily work and that some employees were spending more on tokens than their salaries. He also cited an American Banker report and said RBC’s token usage rose 500% from 2025 to 2026.

Those examples were used to support Morgan’s broader argument that banks can misread high AI usage as evidence of advantage while costs accumulate in less visible ways. He said repeated decisioning by AI agents, where a process does not require it, can add expense without improving the result.

Where GenAI fits in banking

Morgan said GenAI is well suited to summarisation and support tasks, particularly where it helps a professional review large volumes of material. In a legal context, he argued, an AI agent could summarise briefing documents for counsel to check, rather than requiring the lawyer to read every document from the start.

He also cautioned that regulated banking processes need clear controls, sign-offs and escalation paths. Card lending, in his example, is already highly automated, with 90% to 95% of the process requiring no human touch. Adding agentic AI to the remaining 5% to 10% may produce limited efficiency gains, while areas with lower automation, such as mortgages or commercial lending, may offer more useful applications.

Morgan distinguished between using reasoning AI at design time and at runtime. He said reasoning models can help design workflows, solve problems and improve processes, while runtime execution should emphasise reliability, governance and cost control. Using lighter AI to run workflows designed with more advanced models could reduce token spending and improve consistency, he argued.

The next phase of adoption in financial services, Morgan said, will depend less on technical capability alone and more on economic discipline, predictable outcomes and governance. His conclusion was that banks should match the type of AI to the use case rather than applying agentic systems across routine operations by default.

This story draws on original reporting from Finextra Research.