AI Engineering Economics in a Minute
Cost Per Accepted Change
Attribute agent spend to changes that actually merge—so abandoned loops do not look like productivity.
In one sentence
Cost per accepted change divides attributable agent (and optional review) cost by changes that meet your acceptance definition—usually merged, reviewed work.
Why it matters
Agents can burn tokens on dead ends. Measuring only session spend without acceptance overstates productivity and understates waste.
How it works
- Define acceptance (merged PR, deployed commit, passing gated checks).
- Attribute token and tool cost to candidate changes; include failed attempts if you want fully loaded cost.
- Divide fully loaded cost by count of accepted changes in the period.
- Compare cohorts (model, repo, task type) only with the same definitions.
Example
Week spend $400 on agents. 8 PRs merged that you attribute to agent-assisted work. Fully loaded cost per accepted change = $50—before human review time. If 20 abandoned sessions caused half the spend, say so in the method.
What this proves
This metric proves a ratio of attributed cost to accepted-change count under your definitions.
What this does not prove
It does not prove those changes were the right product bets, low-defect, or cheaper than alternatives without a comparison group.
Last reviewed 2026-09-06. Title for citation: AI Engineering Economics in a Minute: Cost Per Accepted Change.