Definition
Degree of alignment between model-recommended capital allocation and independently assessed risk distribution across asset classes, measuring whether AI capital planning puts money where the risk actually lives.
The Measurement Gap
A New England water and wastewater utility manages $3.2 billion in infrastructure assets: 1,400 miles of water mains, 900 miles of sewer lines, 12 pump stations, 3 treatment plants, and 42,000 service connections. The utility invested $2.1 million in a predictive asset management platform that models failure probability across the portfolio and generates annual capital allocation recommendations for the 10-year CIP.
The model recommends concentrating 68% of pipe replacement capital on cast iron and ductile iron mains: asset classes with extensive break history, well-documented failure modes, and abundant training data. The remaining 32% is distributed across PVC, concrete cylinder, asbestos-cement, and prestressed concrete pipe.
An independent engineering assessment tells a different story. When risk is calculated independently (combining failure probability from condition assessments with consequence scores from criticality analysis) the actual risk distribution puts 41% in cast iron/ductile iron and 34% in asbestos-cement and prestressed concrete. The model is over-allocating to asset classes it knows well (abundant training data) and under-allocating to classes it knows poorly (limited failure records because these materials fail differently: silent cracking rather than dramatic breaks).
The result: $8.4 million per year in capital flowing to the wrong asset classes. Not wasted (cast iron replacement is valid work) but misallocated relative to where risk actually lives. The asbestos-cement mains that the model under-prioritizes are deteriorating without intervention, accumulating deferred risk that will eventually surface as emergency failures.
No existing water utility metric captures this capital allocation distortion. Utilities track total CIP spending, spending by asset class, and spending versus budget, but none compare what the model recommends to what an independent risk assessment says the allocation should be. Without CAFI, a utility cannot detect whether the predictive model is sending capital to where risk lives or to where the model has the most data.
Formula
Layer 1: Plain Language
Group assets into classes (by material, age cohort, diameter, or type). For each class, calculate what percentage of the model's recommended capital goes to that class and what percentage of the independently assessed actual risk lives in that class. Measure how closely those two distributions match.
Layer 2: Mathematical Notation
CAFI = 1 − ½ Σ |MA_c − RA_c|
Where:
- MA_c = model-recommended capital allocation fraction for asset class c = (model-recommended capital for class c) / (total model-recommended capital)
- RA_c = actual risk fraction for asset class c = (risk score for class c) / (total risk score across all classes)
- Risk_c = Σ (P_failure_j × C_consequence_j) for all assets j in class c
- c = asset class (pipe material × age cohort, or equipment type × criticality tier)
The formula is 1 minus half the total variation distance between the two distributions. Ranges from 0 (complete misalignment) to 1 (capital perfectly mirrors risk).
Confidence Grading (required alongside every CAFI value):
| Grade | Basis | Interpretation |
|---|---|---|
| A: Empirical | 5+ observed events in measurement window provide direct tail observation | Measurement grounded in observed outcomes; highest confidence |
| B: Hybrid | 1-4 observed events plus Monte Carlo simulation with Bayesian updating | Blends observation and simulation; moderate confidence |
| C: Simulated | Zero observed events; derived entirely from Monte Carlo simulation | Modeled, not observed; treat as indicative, not validated |
A CAFI of any value at Grade C is a simulation that agrees with itself. It tells you the model is internally consistent but not that it matches reality. Grade C scores should trigger proactive stress-testing against historical near-miss events and cross-asset analogues rather than acceptance as calibrated.
Proposed formula, pending validation pilots.