Definition
Proportion of AI-recommended corrective actions where the target deviation type did not recur in the same area within 12 months. Tests whether an action that worked at other sites actually fits local context.
The Measurement Gap
The AI quality management system recommends a corrective action for a recurring contamination deviation: install HEPA filter monitoring with continuous particle count trending. The recommendation draws from 12,000 historical CAPA records across the global network. The same action resolved similar deviations at three other sites. The team implements it. Six months later, the same contamination recurs: the source was a raw material supplier change, downstream of the air classification boundary.
The AI's recommendation was historically effective but contextually wrong. Legacy CAPA metrics measure process, not outcome. Closure rate tracks speed. On-time completion measures schedule adherence. The effectiveness check at 90 days asks "has it recurred?", a single check with a high false-negative rate for intermittent deviations.
Decision enabled:
When CCER drops below 75%, the organization revises how AI CAPA recommendations are evaluated. Requiring investigators to verify contextual match before implementation. When AI CCER falls below human CCER, the AI's pattern-matching advantage is being offset by contextual blindness.
Formula
CCER = CAPAs_without_recurrence / CAPAs_with_AI_recommended_action
Where CAPAs_without_recurrence counts AI-recommended corrective actions where the target deviation type did not recur in the same area within 12 months. CAPAs where the process changed during the monitoring window are excluded as confounders.
Worked Example:
67 AI-recommended CAPAs with completed 12-month windows:
48 without recurrence of the target deviation type (effective).
13 with recurrence (ineffective: corrective action did not prevent the problem).
6 excluded (process changed during window, confounding the assessment).
CCER = 48 / 61 = 78.7%
Human-recommended comparison from the same period: 33 / 40 = 82.5%.
Both above 75%, but the AI is not outperforming humans despite access to a larger historical database. The AI's pattern-matching advantage is offset by its inability to assess whether the current context matches the historical one.
Sign Convention:
Higher is better. Above 75% means mostly effective. Below 50% means more failures than successes. The AI-versus-human comparison is the most informative output.
Proposed formula, pending validation pilots.