As AI is moving beyond the pilot stage, the question arises whether banks can turn growing AI consumption into results that are predictable, repeatable and defensible.
The numbers suggest most cannot yet. The 2026 State of Tokenomics study of 472 organisations found that three in four could not confidently prove AI business outcomes to the CFO.
Closer to home, 54% of Asia-Pacific respondents in an IDC study for Expereo reported cost overruns or missed ROI, 80% in Malaysia.
In a Cognizant survey of 250 FinOps leaders, 72% had exceeded their AI budgets in the past year.
The issue isn’t the price of a token. It’s the operating system around AI: who consumes it, for which outcome, under what policy, on which model, and with what evidence of quality and return.
AI spend enters a bank through many routes, from AI model APIs to embedded software features and self-hosted infrastructure. The model invoice is only part of the bill: vector databases, embeddings, orchestration, observability and data movement can account for 40% to 60% of an AI feature’s cost.
Cost per onboarding file processed, servicing request resolved or alert reviewed tells a bank far more than cost per million tokens. Failed and discarded runs must stay in the denominator.
Buyers already know this: 23% of State of Tokenomics respondents asked providers for more granular usage data. Only 4% asked for lower prices.
Once consumption is visible, it can be steered. Not every task needs a frontier model—and one can cost 50 to 100 times more per token than the smallest model from the same provider.
Source link







