What is cost per decision?

A decision here is a single unit of judgment, and it is usually small: scoring a lead, routing a ticket, approving a discount inside a defined threshold, deciding whether a record needs a human. Measuring it requires naming the category first, because a blended figure across unlike decisions describes nothing.

The reason it sits next to CAC rather than in a separate operations report is that agents absorb work that used to be part of acquisition cost. What was a rep's hour becomes an agent's decision. If nobody measures what that decision costs, acquisition cost looks lower than it is—the spend moved rather than disappeared.

Escalation is the line item that decides whether the number is honest. An agent that hands off to a human whenever it is not confident is not cheap; it is a routing layer with a human cost attached. Counting the escalations, and the review time they consume, is what separates a real cost per decision from a compute bill.

The metric only works on top of three other things. It needs a baseline—the current cost of making that decision the old way—or there is nothing to compare against. It needs "decision" defined the same way across teams, or two departments will report different numbers for the same work. And it needs edge cases logged, because an unrecorded bad decision is not counted anywhere and quietly flatters the average.

Used well, it answers a question CAC alone cannot: whether the automation is cheaper per unit of judgment than the process it replaced, and at what point escalation volume erases the saving.