The ROI of AI coding tools survives or dies on rework

The debate about AI coding tools is usually conducted in the wrong unit. “We generate 40% more code” is not a business result. It is an input, and possibly a liability. The unit that matters is value delivered per unit of fully-loaded cost, and for AI-assisted work, that number is dominated by what happens after the first draft.

The rework tax

A line of generated code that is reverted next week cost more than it saved: it consumed a prompt, a review, a merge, a bug, and a rollback. A change that ships but doubles the review queue imposes a coordination cost on everyone behind it. So the honest ROI calculation is not “time to first draft” but roughly:

value of accepted, in-production output − (generation cost + review load + rework + downstream defects)

Most published productivity numbers stop at the first term of the numerator and ignore the entire denominator. So a tool can look like a 40% gain and be margin-neutral.

What to measure

An organisation that wants a defensible answer instruments a few things it usually does not:

• acceptance and revert rates by origin, human vs AI-assisted, tracked to production;
• review load, time-in-review and reviewer count per change, before and after adoption;
• rework, churn on recently-merged code, and the half-life of AI-assisted changes;
• defect attribution, incidents traced to recently-shipped work, by origin.

None of these require surveilling individuals. They are properties of the change stream. Together they turn “does AI coding pay?” from a slogan into a measurable, per-team economic question.

Why this is an idea and not yet a result

Because the answer is specific to your delivery, and we will not publish a universal figure we cannot defend. What can be said is where the number hides, in the denominator, and how to instrument it. That instrumentation is part of Gitrevio’s engineering-intelligence work, and a common first finding in an audit is that the “productivity” already claimed was never net of rework.