Definition
The ratio of quality-adjusted output value to AI compute cost, measured per task and aggregated by complexity tier, revealing where expensive model routing produces disproportionate cost without corresponding quality benefit.
The Measurement Gap
There is no metric that tells you whether expensive model routing produces proportionally better outcomes than cheaper routing would have, which means AI compute spend can grow without corresponding quality improvement.
Formula
Plain language: For each complexity tier, calculate the average AI compute cost per task and the average quality score per task. Divide quality by cost to get quality-per-dollar. Compare across tiers. If Tier 1 tasks (simple) are being routed to expensive models and producing the same quality as when routed to cheap models, that tier's RCER will be low: signaling over-routing waste.
Notation:
RCER_tier = (Q_avg_tier / C_avg_tier) × 100
Overall efficiency:
RCER_overall = Σ(RCER_tier × W_tier) where W_tier = T_tier / T_total
Variables:
- Q_avg_tier = average quality score for tasks in this complexity tier (dimensionless 0–1, DERIVED using MRAI's quality composite formula)
- C_avg_tier = average AI compute cost per task in this tier (USD, NATIVE from Bobalytics or equivalent cost tracking)
- W_tier = weight for this tier = proportion of total tasks in this tier (dimensionless, DERIVED)
- T_tier = task count in this tier (count, NATIVE)
- T_total = total tasks across all tiers (count, NATIVE)
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Are we spending AI compute budget efficiently, or are tasks being over-routed to expensive models where cheaper models would produce equivalent quality?