projection · not a quote

Your AI spend, read through rates we have measured.

Type a monthly spend. The calculator applies rates measured on real agents in production, each dated below next to the record it comes from. It shows ranges, because one measured project is a small sample. It prints no saving until a change has held on 3 re-measures after it merged.

Every rate today comes from case study two: our founder’s own agents, not a customer’s. Your jobs will split differently.

Rates read 2026-11-01 from production

/ month

A projection from 1 measured project. Nothing you type leaves this page.

In your 5 to 10 costliest jobs
$16,000 to $19,000In the measured project, 63% to 77% of 30 days’ spend sat in its top 5 to top 10 jobs. The reading
In costly jobs with no eval
$9,100 to $11,000There, 7 to 8 of the 10 costliest jobs had no eval to test a change against. The reading
Saving from a change that held
Not yet
  • Pull request #7, cost cut 95% on 3,446 matched live calls: did not hold, so it carries no saving.
  • Pull request #12, cost cut 95% on 55 matched live calls: did not hold, so it carries no saving.
How a change counts as held

the readings · case study two

Three rates, each with the date it was read

RateReadingRead on
Where spend sitsThe top 5, 10 and 20 of 119 jobs with spend carried 63%, 77% and 91% of 30 days’ spend2026-11-01
Costly jobs with no eval8 of the 10 costliest jobs, 44% of spend, before Margin added evals; 7 of 10, 36%, after2026-10-31, 2026-11-01
Pull request #7Cost cut 95% on the same 3,446 live calls priced on the old model; its eval went from 86% to 80% on 32 cases. Did not hold: 0 of its last 3 re-measures cleared the bar.merged 2026-10-17
Pull request #12Cost cut 95% on the same 55 live calls priced on the old model; its eval went from 69% to 89% on 37 cases. Did not hold: 0 of its last 3 re-measures cleared the bar.merged 2026-10-31

the gate

A change counts as held after 3 re-measures in a row.

The cut is the change’s live calls since the switch, priced against what the same calls would have cost on the old model. It counts only cost.

Quality is the project’s own scheduled eval suite. Each scheduled run after the switch is one re-measure, and it holds when it grades at least 10 cases and scores at least 97% of what those same cases scored before the merge. Only after the last 3 re-measures all hold does the cut become a saving in the calculator, and then it shows as a range: from the 10th costliest job’s share of spend to the share of the job the change actually moved.

A change whose re-measures stop holding leaves the sum on the next reading.

how to read it

A direction from one measured project, not a quote.

These are shares of one personal agent set’s spend, applied to yours. They say where money tends to sit and how often costly work goes untested. They do not say what your jobs are, and they are not a promise of what Margin will save you.

Your own numbers come from your own calls. Connect a repository and the same readings run on your estate.