Rework rate
Rework rate in AI-assisted work is the share of completed units that must be reopened, repeated, or materially repaired because the original work did not remain acceptable. It captures downstream quality cost that an initial completion metric can miss.
An AI-enabled workflow may finish work faster on the first pass while creating more reopened cases, revised analyses, defect fixes, or repeated reviews. Rework rate tests whether the apparent gain survives after completion.
How is AI rework rate calculated?
A basic formula is:
Rework rate = reworked completed units / total completed units x 100
Define the observation window and rework event. A support case may reopen within an agreed period. A code change may require a corrective change after merge. A reviewed document may return because a material error escaped the original process.
Use a stable unit of work and acceptance rule. If the denominator includes drafts or partial work in one period but only approved work in another, the comparison will mislead.
Record rework severity and effort where possible. A brief clarification and a complete redo should not carry the same cost without explanation.
Rework rate versus correction rate
Correction rate in AI-assisted work usually measures changes made during review before work reaches its accepted completion state. Rework rate measures work that fails or reopens after that point.
Both belong in the quality view. If teams measure only corrections, they may miss errors that escape review. If they measure only rework, they may miss a large manual burden that reviewers absorb before completion.
NIST's AI Risk Management Framework calls for deployed systems to be monitored, for errors and emerging risks to be tracked, and for measurable performance improvements or declines to be documented (NIST AI RMF Core). Rework records provide one operational signal of that decline.
How does rework affect AI productivity?
Rework consumes labor, model usage, review capacity, and elapsed time. Count it in the complete workflow rather than treating the initial completion as final.
A useful companion measure is:
Net accepted output = completed units - units invalidated by rework
The exact formula should fit the workflow. Some reworked units remain valid after repair, so teams may instead track rework hours, cost, severity, and delay beside the rate.
Include the labor and system cost in the fully loaded cost of AI. Otherwise, a workflow can appear cheaper as it shifts effort from production into later repair.
How should rework be attributed to AI?
Do not label every reopened unit as AI-caused. Record whether AI influenced the original work, what failed, and which other factors contributed. Compare eligible AI-assisted work with a baseline or suitable comparison.
Segment by model, workflow version, case complexity, reviewer, and failure type when the sample supports it. Preserve the share of work with unknown cause instead of forcing every case into a precise category.
Rework can reveal AI value leakage when an earlier speed or capacity gain disappears through downstream repair.
How can teams reduce rework?
Use failure patterns to change workflow eligibility, prompts, retrieval, validation, escalation, or review design. Some work should require stronger approval, while other work may not justify AI involvement.
Review the complete outcome after each change. A stricter review may reduce downstream rework while increasing pre-completion correction and cycle time. The investment decision should consider the combined cost and quality result.
Common rework-rate mistakes
Common failures include using an observation window that is too short, excluding abandoned work, confusing planned revisions with failures, counting repeated work more than once, and comparing teams with different case complexity.
Another mistake is optimizing the first-pass completion rate without following the work downstream. Rework rate exists to test whether completed work stays complete.