Measuring AI Agent Costs by Finished Job Rather Than Per Token

The article argues that teams track AI agent expenses at the token level but overlook the cost that most affects budgets: whether output is accepted. Per-token pricing can differ roughly seventy-fold across models, yet agent loops, retries, and tool-call round trips often make real bills far higher than unit-price estimates suggest. In one team's October accounting of 1,840 accepted deliverables, model fees made up 18% of total cost while human review and rework accounted for 41%. The author proposes a vendor-neutral schema that logs attempt cost, which must be metered at a single client wrapper with a mandatory job ID to avoid undercounting. It also recommends capturing review time passively, as a side effect of the reviewer's existing workflow, so that the full cost per finished job can be measured.