A series for the agentic banking era
This is the second article in a seven-part series on tokenomics for banks. The next article explores outcome density, the metric that makes unit economics worth measuring at all.
Unit economics is the foundation of every AI investment decision
- As stated in the first article of the ‘Tokenomics’ series, Tokenomics: The new economic discipline for banking AI, tokenomics rests on five economic primitives.
- Unit economics is where the discipline begins because the input side is the foundation that every other primitive sits on.
- Outcome density tells the bank what the output is worth. Workload classification tells the bank how much governance to apply. The platform cost curve tells the bank how to bend the unit cost down over time. The agent-as-cost-center model gives the accountability structure that holds the rest together.
- None of those primitives function without unit economics underneath. A bank that cannot say what a workload costs cannot say what it returns, what governance is proportionate, or whether the platform investment is bending the curve.
- Banking’s asymmetric downside, its established model risk regimes, and its regulatory-grade data assets, all described in the first article of this series, only sharpen why unit economics matters more here than in any other sector.
What unit economics needs: Four questions to ask
Every meaningful AI workload in a bank reduces, at the level of consumption, to tokens. There are input tokens, output tokens, tokens consumed by foundation models, tokens consumed by smaller specialist models, tokens consumed by embedding and retrieval systems, and tokens consumed by orchestration layers. The discipline of unit economics requires the bank to answer four questions for any workload, at any time.
- How many tokens does the workload consume per unit of business work? Per customer interaction, per loan application processed, per complaint resolved, per code commit, per credit memo drafted.
- At what blended cost per thousand tokens? Across the mix of models, providers, and hosting modes the workload uses.
- At what expected utilization? Including idle time, retry overhead, evaluation overhead, and the consumption of safety and observability tooling.
- Attributable to which business unit, product, and outcome owner? With a chain of accountability that is auditable.
Most banks today cannot answer these four questions for most of their AI workloads. The State of FinOps 2026 report identified granular monitoring of AI spend, including tokens, LLM requests, and GPU utilization, as the single most-requested tooling capability in the entire survey, with 98 percent of FinOps teams now responsible for managing AI spend.
