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

Get the white paper

Tell us who you are and the download opens straight away.

First name *

Last name *

Business email *

Job title *

Country *

Phone number *

I'd like a demo of the Office

We use these details to send you this paper and, if you ask, to arrange a demo. See our Privacy Policy.

Your copy is ready

Thank you. Download the PDF below, and keep it to share with your team.

Download the white paper (PDF, 1 MB)

More from Manelink

See all research

LandscapeWho builds frontier AI? A map of the labs and playersWho trains the most capable AI models, how close the race really is, where the real gaps lie in compute and money, and where Canada fits. ApplicationsAI in sales: what the evidence shows it can and can't doWhere AI measurably lifts sales, where it backfires, and what buyers now expect from the humans who are left in the loop.