A series for the agentic banking era
This is the third article in a seven-part series on tokenomics for banks. The next article explores the structural lever that makes outcome density scale, the platform cost curve and the agent as cost center.
Why cost per token is the wrong primary question
In our 2nd article of the ‘Tokenomics’ series, What bank CFOs must know about AI unit economics, we established that unit economics tells the bank what a workload costs. In this article, we pick up the harder half of the equation: what does a workload actually produce? The pivot matters because banks default to the wrong reflex. When finance sees an AI run rate, procurement is asked to negotiate it down, not asked what it produces. Once a bank accepts that cost per token is the wrong primary question, the real question surfaces immediately, and it needs its own metric.
The reflex is familiar. A generation ago, banks measured offshore captives by per-full-time equivalent (FTE) cost arbitrage and missed the strategic question of what those captives could become. In the early cloud era, banks demanded vendor invoices be cut by 20 percent and missed the question of what cloud-native architectures could unlock. The cost-obsession trap fails for the same reason every time. It asks the wrong primary question.
What does a token produce? The right economic question for banks
The right primary question is not “what does a token cost”. The right primary question is “what does a token produce”. Outcome density is defined as the business value produced per unit of token consumption, measured in the currency of the workload’s intended outcome. The currency is workload-specific. The same wealth advisor copilot we costed in the previous article produces value in one of two currencies: qualified lead conversion, or client meeting preparation time saved. Outcome density measures which. That measurement is what allows a bank to decide whether the workload is worth the tokens it consumes.
For a fraud triage agent, the currency is loss avoidance per thousand tokens. For a complaint resolution agent, it is regulatory escalation rate and customer satisfaction per thousand tokens. For a developer copilot, it is shipped engineering throughput per thousand tokens, net of defect rate. Outcome density is harder to measure than input cost. That is precisely why most banks default to measuring input cost only. The result is a distorted economic picture.
