All reports and guides
White paper No. 1 · October 2026

Cost per Accepted Outcome: the unit economics of AI workforces

Token prices are collapsing, yet agent bills are rising. Why, and how to manage an AI workforce by cost per accepted outcome instead.

About this white paper

Token prices are collapsing, yet agent bills and project cancellations are rising. This paper shows why: agentic work multiplies the tokens each attempt uses, reliability lags capability so attempts fail and retry, and human review is often the largest variable cost. In one banking workflow it was 70 to 75 percent of the run cost, against 20 to 25 percent for tokens.

It proposes cost per accepted outcome as the unit for managing an AI workforce, a matrix for deciding where people should approve agent work, a worked example and seven recommendations you can start tracking this quarter.

What it covers

  • The right unit of account. Why to manage agents by the cost of finished, approved work, not by the price of a token.
  • Where the tokens go. Why agentic tasks use about a thousand times the tokens of chat, and why list price misled in 32 percent of model comparisons.
  • Reliability over capability. Why most production agents run ten steps or fewer before a person steps in.
  • The biggest lever. How redesigning approval moved an illustrative cost four times more than halving token prices.
  • Memory, skills and routines. When they pay off, and how to count their tokens against a plain agent.
  • A readiness checklist. Ten questions to answer before an agent workflow goes live.
Pages
22
Reading time
About 25 minutes
References
30
Published
October 2026