01
Prediction is only one layer
A propensity score does not create a customer strategy. Production systems also need identity rules, eligibility policies, channel coordination, experimentation, and a feedback loop that distinguishes correlation from incremental impact.
02
Build a consent-aware decision path
Every feature and action should trace to an approved purpose. Sensitive attributes, proxy risk, retention, and suppression logic belong in the architecture.
- Purpose-bound feature registry
- Time-aware training and leakage controls
- Calibrated scores with segment diagnostics
- Next-best-action policy with frequency limits
- Holdouts and causal measurement
03
Operationalize the learning loop
Log why an action was eligible, what the system expected, what was delivered, and what happened next. This creates evidence for monitoring, learning, privacy review, and commercial decisions.