Unit Economics: Cost per Successful Task

The cost-driver inventory identifies what to measure. It does not convert those measurements into a figure that a business stakeholder can use. That requires a unit economics model. Cost per token is an input to a business decision, not a business metric by itself. Instead, evaluate cost per successful task under a defined quality service-level objective (SLO). Include the costs of retries, fallbacks, policy-based refusals, and any human oversight required to execute write actions in compliance with policy.

Define the unit of work and what counts as success

The unit of work is a case-resolution recommendation. Producing it may require retrieval, one or more tool calls, and, for some write actions, human approval. Define binary success criteria for this unit of work. Every execution path that fails to meet these criteria increases the numerator through additional costs and reduces the denominator by producing fewer successful outcomes.

Combine quality and operational requirements into an explicit success gate. A successful task produces a grounded recommendation with traceable provenance, meets the defined latency target, and causes no policy violation. Specify whether a refusal or escalation counts as a successful, policy-compliant outcome or an unsuccessful task, based on the product requirements and risk posture.

The diagram shows where costs accumulate along each execution path and where the success gate classifies an outcome as either a success or an exception.

Build an expected value cost model over execution paths

Start from expected cost per attempted task: a weighted sum across mutually exclusive execution paths. For a path set $P$, the expected cost is

where every path cost $C(p)$ sums the variable costs on that path, plus any allocated fixed overhead the organization charges back to the workload.

Decompose every path cost into token, retrieval, tool, and other per-request variable costs, reusing the same decomposition across every path. A compact generic decomposition is

where other might include a content filtering call, a feature store lookup, or a per-request orchestration charge, if those ...