Rule maintenance is the central weakness of the pre-AI era: every new vendor, fee structure or transaction type needs someone to notice a pattern and write a rule. The pattern is visible in the resolution data long before anyone acts on it.
Every proposed rule, before it takes effect. A rule change alters what clears without human review, which makes it a control change requiring approval and dating.
Create or modify a rule silently, or widen a tolerance. The reference case is instructive: widening a card tolerance from 0.50 to 25.00 absorbed roughly 160,000 a month of processor overcharging while the queue stayed empty.
Manually resolved exceptions grouped by vendor, account and variance band; existing rule library; tolerance history
Proposed rule with the pattern evidence, expected match impact and a simulated effect on prior periods