AI rework is paid through payroll and never appears on the AI bill. A practical method for estimating it, pricing it and putting it next to token spend.
AI rework is any time a person spends because an AI answer was not good enough to use as it was. It includes noticing the problem, working out what went wrong, asking again, editing the output by hand, and redoing the work entirely when the AI could not get there. It also includes the quieter cost of reviewing everything carefully because the output cannot be trusted.
Rework is spread thinly across many people. A few minutes here, half an hour there. Nobody logs it, and it is paid from payroll rather than the AI budget. Yet on short tasks it is usually the largest cost, and one hour at $100 dwarfs the few cents a cheaper model saved.
Once rework has a dollar value, add it to token spend for each task. That total is the cost to finish. Compare it across models and teams, and the expensive habits become visible: the cheap model that needs four attempts, the task type nobody should be using AI for yet, the team that reviews everything twice.
CompletionPrism™ estimates rework from its Rework Task Taxonomy and the labor rates each company sets, then predicts it before every request so the model choice accounts for it. Every receipt shows the rework line next to the token line, labelled estimated, and the two-week baseline reports rework reduction alongside cost per completed task.