Definition
Measures the degree to which adjusters' reserve estimates represent genuine independent assessment versus marginal adjustments anchored to the AI's initial reserve recommendation, distinguishing real human oversight from cosmetic adjustment.
The Measurement Gap
When AI sets an initial reserve estimate and a human adjuster reviews it, the review is supposed to provide independent professional judgment. But anchoring bias is one of the most robust findings in cognitive psychology: people presented with an initial number systematically adjust insufficiently from that anchor. Over time, adjusters learn that the AI is "usually close enough" and reduce their adjustments to a narrow band around the AI's recommendation, typically 5-10% in either direction. The reserve review becomes cosmetic. Legacy metrics cannot detect this erosion of independence. Reserve adequacy measures whether total reserves are sufficient at the aggregate level but cannot distinguish independently set reserves from anchored reserves. Reserve development (favorable/adverse) captures outcomes but only years later, after the damage is done. Adjuster productivity metrics track how many reserves an adjuster reviews per day but not whether the reviews involve genuine judgment. Override rate tracks how often the adjuster changes the AI's number but not whether the magnitude of change reflects independent assessment or anchoring behavior.
Decision Supported:
Is the human-in-the-loop reserve review actually exercising professional judgment, or have adjusters learned to nudge the AI's number by a small margin and call it a review, and should the reserve review process be restructured to restore independence?
Formula
Plain language: For all claims where the AI set an initial reserve and the adjuster subsequently adjusted it, measure the standard deviation of the percentage deviations. If reviews are independent, the distribution will be wide and vary by claim characteristics. If anchored, deviations will cluster in a narrow band around the AI's number.
Notation:
RAIS = sigma(dev%) / sigma_baseline(dev%)
Where: dev%_i = (Reserve_adjuster_i - Reserve_AI_i) / Reserve_AI_i x 100
Variables:
| Variable | Definition | Unit |
|---|---|---|
| Reserve_adjuster_i | Adjuster's final reserve for claim i | Dollars |
| Reserve_AI_i | AI-recommended initial reserve for claim i | Dollars |
| dev%_i | Percentage deviation of adjuster's reserve from AI recommendation for claim i | % |
| sigma(dev%) | Standard deviation of deviation percentages in the current period | % |
| sigma_baseline(dev%) | Standard deviation in the baseline period (first 6 months post-AI deployment, or pre-AI adjuster-to-adjuster variance) | % |
RAIS of 1.0 means current deviation spread equals baseline (independence maintained). Below 1.0 means deviation spread is narrowing (anchoring). A RAIS of 0.50 means adjusters are deviating only half as much as they did initially.
Proposed formula, pending validation pilots.