The pattern under most production personalization infrastructure: call the model or scoring service, and if it doesn't return within the latency budget, serve a deterministic fallback (rule-based, default segment, or static experience). The fallback isn't a failure case, it's the design that makes the recipe robust enough to deploy in production at all.
*Retired at the recipe-50 lock audit (run at recipe 65). The pattern was held as provisional through 65 recipes without a single tagged use; the experimentation work the cookbook actually surfaced collapsed into threshold-based-routing and behavioural-segmentation. The definition is preserved below in case a future recipe needs it back under observation.*
The pattern where human-authored business logic and machine-learned outputs combine to produce decisions. Most real-world implementations live here rather than in pure-ML or pure-rules architectures. The orchestration question is which layer wins when they conflict, and how rule overrides feed back into model training.
The pattern for AI-orchestrated workflows where the system takes action autonomously within defined bounds, escalates ambiguous decisions to humans, and learns from the human resolution. Defines what the system can do alone, what it must ask permission for, and what it must report after the fact. The emerging architectural pattern for agentic martech, currently underdocumented and over-promised across the vendor landscape.
The pattern that groups visitors into editorial cohorts at request time based on observed behavior rather than declared preferences. Cohorts are defined upstream by the team (intent categories, lifecycle stages, persona-like buckets) and a runtime classifier assigns each visitor to one or more. Distinct from inferred attribute generation because the cohorts are deliberately authored as activation units rather than emerging from a model's clustering.
The pattern that defines the deliberate experience for visitors the personalization layer cannot place: no profile, low-confidence segment assignment, signal capture in progress, consent state restricting personalization, or model failure. Distinct from real-time decisioning fallback because the question is not which signal to use, it is what to show when no signal is usable. The default experience deserves the same editorial care as the personalized one.
The pattern that translates a continuous score into a discrete action by routing on threshold crossings: score crosses the handoff line and lead becomes SQL, propensity crosses the intervention line and customer enters the win-back journey, fatigue score crosses the suppression line and channel goes dark. Where rules-plus-model hybrid decisioning asks what to do, threshold-based routing asks when the underlying score warrants an action at all.
The pattern that varies the surface a visitor sees based on their current state: anonymous versus known, first-time versus returning, recent intent signal versus dormant, consented versus restricted. The activation-layer surface for the cohort assignment that behavioral segmentation produces. Centrepiece of any on-site personalization recipe and the surface that most often visibly breaks when the underlying signals are stale or unavailable.