Cost per outcome is the unit cost finance can set against a benefit. The numerator is everything it takes to run one AI use case for a month: model tokens, retrieval and compute, monitoring, evaluation runs, the staff who review and maintain it, and a share of the shared platform. The denominator is the result the use case exists to produce, measured in the same unit as its benefit.
The common mistakes sit at both ends. Counting only the provider invoice leaves out lines that are often larger than the tokens, and dividing by calls or drafts flatters the figure, because a draft nobody sends or a failed call still costs money. Failed calls, retries and evaluation runs stay in the cost.
Set beside the benefit per outcome, it gives a unit margin per use case. A use case whose cost per outcome rises above its benefit per outcome loses money on every transaction, however large the total benefit looks.