Skip to content

Whitepaper · Measurement model

AI ROI for engineering

A measurement model for the AI coding spend already on your books, and a reference architecture for tying every dollar of it to work that shipped.

  • One management unit: AI cost per accepted outcome
  • Attribution from the token to the pull request
  • 8 questions to ask about your own organization

28 pagesFrom the creators of Kyverno

Get the whitepaper

Free PDF · Instant download · 28 pages

Loading form…
Get the free whitepaper

October 2026 · Anthropic's S-1

Nine times the spend is a fact. Nine times the output is not.

Sources: Anthropic S-1; Menlo Ventures; Jellyfish

~900%
Net Revenue Retention reported in Anthropic's S-1, as of 30 September 2026
55%
Of departmental AI spend went to coding in 2025
35x
A 99th-percentile developer's monthly AI spend, against the median

The unit

The token is the billing unit, not the value unit

Finance manages seats and invoices. Platform teams manage tokens and rate limits. None of those is a unit of value. The paper proposes one management unit, read against guards for quality and review load.

  • Median PR size
  • Review time per PR
  • Change-failure rate
  • Test share of changes

AI cost per accepted outcome

attributed AI spend in a periodaccepted outcomes in the same period
An accepted outcome: a pull request that merged to a protected branch and was not reverted or more than half rewritten within 14 days.

What's inside

  • The bill changed shapeSection 2
  • What makes agent spend differentSection 3
  • Cost per accepted outcomeSection 5
  • Attribution: token to pull requestSection 6
  • Levers that move costSection 8
  • 8 questions for your organizationSection 9
Get the whitepaperAlready read it? Request an AIControls ROI assessment or see AIControls