01
Data readiness
Repair history before training models
Mark outages, closures, stockouts, unusual campaigns and missing records so the model does not learn constrained demand as normal behavior.
ARKASH forecasting governance
Make every cash proposal explainable and reviewable.
AI forecasting should combine transaction history with calendar, salary, event, location and anomaly signals while preserving human approval for operational cash orders.
Acceptance checkpoint
Forecast accuracy is useful only when it improves a defined operating decision.
02
Objective
Balance availability, idle cash and movement
Define service and cost weights by location type instead of optimizing a single network-wide target.
03
Control
Capture overrides as learning evidence
Require a reason when planners change a proposal, then compare the override and final outcome to improve policy and model behavior.
Turn the page into a testable decision.
Capture the workload, constraint and required evidence, then review the fit with an operations specialist.
