Cost analysis

The curve

Two lines over the same 17 weeks. The lower one is the recorded work at flat rates, $3,576 a week. The upper one is the modelled team, $32,910 a week. Both are straight because both are rates, not invoices. What separates them is the slope.

What the weeks held

A cost line at a flat rate says nothing about whether anything was built. These two do. Neither involves the model: both are counts of things that happened.

The two columns

Scope: Campaign Brain (cb*) repositories, 2025-12-01 to 2026-03-26.

The in-window commit count is not the one recorded on 2026-03-26. That version was generated before the window’s last day had ended: its commit source stopped at 2026-03-26 05:00 UTC, and it counted 5,315. The count taken on 2026-04-03, after the window closed, is 5,411. A count of the same days on 2026-10-02 gives 5,362, because one repository’s history was rewritten in between.

The modelled team

10 people at $1,806,000 a year fully loaded. Salary figures are market rate; the 1.4x loading covers benefits, tax and overhead. That is every seat at full time; the cost model counts one of them at half time.

The gap

95.3% below the modelled cost of the whole job: $60,800 recorded over the window against $1,282,500, the midpoint of a modelled $855,000 to $1,710,000 (9.5 people for 6 to 12 months at $15,000 per person-month). That assumes the work of the window was the whole job. When the window closed the modelled line stood at $559,469; against that, the gap is 89.1%. Either figure inherits the model’s assumptions entirely.

The pipeline also computed a time compression of “9 months to 34 days” and a velocity multiplier of 50.3x. Neither is a headline here. Both divide by 34 active days, the days with a recorded AI session, in a record that only begins 2026-02-09, while commits by all authors landed on 116 of the window’s 116 days. A ratio built on 34 overstates.

What the work covered

The model prices 8 job titles. This is the same ground scored from the other side: how much of each domain’s vocabulary turns up in project notes, plan files and the technologies named in session logs. It shows breadth of work touched, not depth of expertise.

What the research says

Published findings on AI-assisted development, for calibration against the single project above.

What this is evidence for

The numbers are here. The argument they belong to is on the shift: what changed about teams, cadence, coordination and context when the tooling changed, and what one case study can and cannot show about it.