Model
Token cost is not the unit that matters. What matters is what one decision costs you today, what it costs once agents carry part of it, and how long the change takes to pay for itself. Adjust the inputs to your operation.
Claims, submissions, alerts, files — one unit of the work you are automating.
Minutes of human time per decision today, including assembly and lookup.
Salary plus benefits, management, tooling, and facilities — not base wage.
Share of decisions that come back — errors, NIGO, reopened files.
Additional human time consumed each time a decision has to be redone.
Decisions an agent can close end to end within delegation boundaries.
On the remainder, how much human time the agent removes by pre-assembling the file.
Inference, platform, monitoring, and the operations owner's time.
Integration, evaluation set construction, and go-live.
Cost per decision today is average handle time plus the expected cost of rework, priced at your fully loaded rate. After deployment, fully automated decisions retain a residual 8% of that time for review and exceptions; the remainder keeps a human in the loop with the assembly work removed. Run cost is added back before savings are calculated — a model that ignores it is a sales deck, not a business case.