Cost per outcome is the metric that converts AI spending data from an accounting record into a business intelligence asset. Every other AI cost metric describes the expenditure. Cost per outcome describes the return. Without this metric, AI spending decisions are made on cost alone, which produces cost optimization at the expense of value optimization. With this metric, decisions can be made on efficiency, the relationship between what you spend and what you get, which is the rational basis for investment decisions.
What an Outcome Is
For the purpose of AI cost measurement, an outcome is a discrete unit of value-generating work that an AI workflow completes. For a customer support AI workflow, an outcome might be a ticket resolved without human escalation. For an AI-assisted code review workflow, an outcome might be a review completed that results in accepted changes. For an AI document classification workflow, an outcome might be a document correctly classified on the first attempt without human correction. The key characteristic of an outcome is that it represents value delivered, not just work performed.
How to Calculate It
Calculating cost per outcome requires connecting cost data to outcome data, which are typically stored in different systems. The connection is made through a shared identifier that appears in both systems. When a customer support ticket is processed, the ticket ID should appear both in the AI infrastructure logs and in the support system records. With this connection in place, computing cost per outcome is a straightforward aggregation across the joined datasets.
What Cost Per Outcome Reveals
Cost per outcome reveals things that aggregate cost metrics cannot. Two workflows with identical total monthly costs may have very different cost-per-outcome profiles: one workflow is completing outcomes efficiently, the other is consuming the same tokens with a much lower success rate. Cost per outcome also reveals the impact of optimization efforts more accurately than cost-level metrics. A prompt optimization that reduces token consumption by twenty percent while holding outcome rates constant has produced a twenty percent reduction in cost per outcome. A model routing change that reduces cost by thirty percent but also reduces outcome rate by fifteen percent has produced a smaller improvement. Evaluating optimization efforts against cost per outcome prevents the common mistake of optimizing for cheap AI rather than efficient AI.