Definition
The degree to which final performance review narratives retain the framing, language, and sentiment of AI-generated draft text rather than reflecting independent manager judgment.
The Measurement Gap
Performance management AI now drafts review narratives from data inputs: completion metrics, peer feedback snippets, goal progress indicators. Managers are presented with pre-written text and asked to review, edit, and finalize. The intention is efficiency. The risk is anchoring: managers accept the AI's framing of an employee's performance rather than forming their own narrative from direct observation and judgment.
This anchoring is invisible to every legacy performance management metric. Review completion rates measure whether reviews were submitted but not whether the manager actually authored them. Rating distribution analysis detects central tendency and leniency but cannot distinguish between a manager's own assessment and an AI-drafted assessment that the manager rubber-stamped. Calibration session metrics measure agreement across managers but not whether that agreement was driven by all managers accepting the same AI-drafted framing.
The consequence is a gradual hollowing of the performance review as a management tool. If every review reads like a standardized AI output with minor edits, the review ceases to contain the observational specificity and contextual judgment that makes it useful for development conversations, promotion decisions, and performance improvement plans. The organization retains the administrative burden of reviews while losing the management signal they were designed to produce. Employees receive feedback that sounds professional but may not reflect their manager's actual experience of their work.
Decision Supported:
Are managers writing substantive performance reviews, or are they accepting AI-generated drafts with minimal modification?
Formula
Plain language: Compare the final submitted review text against the AI-generated draft text across three dimensions: semantic similarity, topic overlap, and sentiment alignment. RNAI is a weighted composite of these three measures.
Notation:
RNAI = 0.4 × SemanticSim + 0.3 × TopicOverlap + 0.3 × SentimentAlign
Variables:
| Variable | Definition | Unit |
|---|---|---|
| SemanticSim | Cosine similarity between sentence embeddings of the AI draft and the final submitted text | Score (0–1) |
| TopicOverlap | Proportion of topic clusters present in the AI draft that are retained unchanged in the final text | Score (0–1) |
| SentimentAlign | Correlation between section-level sentiment scores in the AI draft and the final text | Score (0–1) |
Proposed formula, pending validation pilots.