Fixing AI's work is still work

AI drafts a change, the developer adjusts it, and the reviewer sends back comments. Another attempt means another round of checks. A comparison that stops at the first answer leaves all that work out.

I trace a specific task through to acceptance with the team. We look at who reviewed the code and what needed redoing. We record time waiting for review separately from time people spend working. A queue and an expensive review need different fixes.

We use changes from the repository and the comments left on them. We check what AI knew about the task, which constraints it received, and what the reviewer rejected. The missing piece might be system context. Or the team may disagree about what counts as finished.

We change one part of the workflow based on what we find. That could mean making requirements clearer before generation or automating a repeated check. An engineer still accepts the work. For consequential changes, we agree in advance what that engineer needs to inspect personally.

We compare similar tasks using an agreed acceptance rule. The comparison includes context preparation and time spent working with AI, followed by human review and corrections. We record initial setup separately so we can see how often the task would need to repeat to recover that effort.

And we keep the cases that got worse. One successful change isn't enough evidence to redesign the whole process. If the tasks differ too much, the comparison remains provisional.

The work leaves review rules for the chosen workflow, observations about time spent, and a decision on what to try next. If there's no saving, that's what the finding will say.