THE THESISThe best automation sends fewer, more relevant actions because it decides from customer state—not campaign calendars.
Model customer state
Define lifecycle stage, recent behavior, product ownership, service eligibility, value potential, risk and channel permission. A state model turns scattered events into a small set of commercially meaningful conditions.
Create the action library
For each state, list eligible actions: educate, remind, assist, cross-sell, recover, suppress or route to a human. Attach expected value, cost, urgency, fatigue and policy constraints.
Decision hierarchy
Apply consent and safety first, then service recovery and contractual obligations, then high-value commercial actions, then engagement. Resolve collisions centrally so customers do not receive competing messages from separate teams.
Learn by policy
Measure incremental response, margin, unsubscribe risk and downstream quality. Use holdouts at journey or policy level, not only subject-line tests, so optimization rewards long-term customer value.
Practical questions
Do we need machine learning?+
No. Start with explicit state rules and eligibility. Add models only when data volume and decision complexity justify them.