The Three-Layer Architecture

Every decision engine has three layers. The context layer: everything known about the current state — account health, recent interactions, product usage, stakeholder engagement, competitive signals, support history. The context layer aggregates data from every available source into a unified decision input. The logic layer: the rules, models, or algorithms that transform context into recommendations. "If churn probability exceeds 60% and no QBR in ninety days, recommend scheduling a QBR." The logic can be rules (explicit, transparent), models (learned from data, more nuanced), or a hybrid. The action layer: the interface that presents the recommendation to the decision-maker and captures their response. The recommendation must be specific ("schedule a QBR with this account by Friday"), not vague ("engage this at-risk customer").

1. Context Layer Aggregate all relevant data into a unified entity profile. For an account-level decision engine: health score, usage metrics, engagement history, billing status, support tickets, stakeholder map. The context refreshes in near-real-time as new data arrives.
2. Logic Layer Start with rules — they are transparent, debuggable, and easy to modify. "If X and Y and not Z, recommend A." Graduate to ML models when you have enough outcome data to train one. Rules first, models later, hybrid eventually.
3. Action Layer Present recommendations in the workflow where the decision is made — the CRM, the inbox, the dashboard. Include the recommendation, the rationale, and a one-click action to execute. Every friction point between recommendation and action reduces adoption.