Keep the core. Close the intelligence loop.
A small architecture change creates a large operating question: can the bank measure more evidence-backed decisions without moving the system of record?
Understand above the core. Record back into the core.
90 days to earn the right to scale
One state · one bank · one product category · one AI credit workforce. Measure a baseline first, run the same work with the workforce, and let the bank’s own committee judge quality.
Lock definitions, consent path, human authority, current TAT and fully-loaded cost.
Shadow then operate on the selected product. Retain every conflict, correction and quality score.
Blind-score CAMs, review economics and decide: stop, adjust, or scale.
Measure value without hiding the denominator
Days from complete evidence pack to sanction
Include officer, review, rework and tooling
Count only decisions that pass quality review
Bank credit committee scores without knowing the author
Caught before sanction, with retraction retained
Pathways identified and actually claimed
Signal to human intervention, not a default prediction
Surfaced per anchor sanction, with evidence path
Organic origination, not PSLC purchase