Correction rate
Correction rate in AI-assisted work is the share of AI-influenced outputs that require a person to fix, revise, or replace them before the work reaches its accepted completion state. It measures the review burden hidden between model output and completed business work.
A low generation time can look productive while correction work rises. Tracking correction rate keeps speed claims tied to the quality of accepted output.
How is AI correction rate calculated?
A basic formula is:
Correction rate = corrected AI-assisted units / reviewed AI-assisted units x 100
Define a unit before calculating the rate. It might be a drafted response, classified case, contract clause, code suggestion, or another output that receives a consistent review decision.
Also define what counts as a correction. A spelling change, factual repair, policy edit, complete rewrite, and rejected output may have very different operational costs. Record severity or correction type when those differences matter to the decision.
Use reviewed units in the denominator when not every output receives review. If review coverage varies across teams or periods, report it beside the rate so a lower review rate does not appear as better quality.
Correction rate versus error rate
Error rate measures outputs that fail a stated test or requirement. Correction rate measures outputs that people change before acceptance. The measures can overlap, but they answer different questions.
A person may edit a correct answer for tone or formatting. An error may also go undetected and receive no correction. Teams should therefore pair correction data with defect sampling, incidents, escalations, or other quality checks relevant to the workflow.
NIST's AI Risk Management Framework recommends documenting test sets, metrics, deployment conditions, errors, and performance changes over time (NIST AI RMF Core). Correction rate contributes field evidence from actual use but does not replace technical evaluation.
How does correction rate differ from rework rate?
Rework rate in AI-assisted work measures completed work that must be reopened or repeated because it failed after the initial acceptance point. Correction rate usually captures changes during review before completion.
The boundary should match the workflow. For a support case, edits before sending may count as corrections, while reopening the case after the customer replies may count as rework. Keeping both measures prevents effort from moving across the acceptance boundary unnoticed.
How does correction affect AI cost and productivity?
Correction consumes reviewer time and delays completion. Include that time in the fully loaded cost of AI and in any productivity calculation based on labor input.
Segment the rate by work type, complexity, team, model, and workflow version when those factors influence the result. A blended rate can improve simply because users route easier work to AI.
For cost analysis, a useful companion measure is:
Correction cost per accepted unit = total correction labor cost / accepted completed units
Use the organization's approved labor-cost method. Do not substitute a generic wage estimate for observed internal cost.
How can teams use correction rate?
Use correction patterns to decide where to improve instructions, retrieval, model choice, review rules, training, or workflow eligibility. The objective is not always the lowest correction rate. Some high-value work may justify extensive review because the consequence of an undetected error is high.
Track the rate before and after changes using stable definitions. Pair it with cycle time, accepted throughput, severity, and downstream AI value leakage so local quality improvements remain connected to the end-to-end outcome.
Common correction-rate mistakes
Common failures include counting edits without defining severity, measuring only accepted outputs, excluding abandoned drafts, changing review coverage, and comparing workflows with different complexity.
Another mistake is treating fewer human edits as proof of better quality. Reviewers may trust the system more, face higher workload, or skip checks. Correction rate needs review-coverage and independent quality context before it supports an investment decision.