Next-Best-Action Engine

The engine must ingest a streaming event, arbitrate between candidate actions, and serve a decision within strict millisecond budgets to avoid disrupting the user experience.

• Target Latency: < 50ms at the 99th percentile (p99) for synchronous web interactions.
• Architecture: Relies on in-memory feature stores and pre-computed embeddings rather than heavy transactional database queries.
• Mechanism: Uses data locality by co-locating the model serving layer with the decisioning service to eliminate network hops.

The engine uses contextual bandits to dynamically balance the exploration of new offers with the exploitation of known high-performers, maximizing long-term cumulative reward.

• Exploration: Allocates a controlled percentage of traffic to random or uncertain actions to gather training data.
• Exploitation: Routes the majority of traffic to the action with the highest predicted reward given the current user context.
• Algorithms: Common implementations include Thompson Sampling, Upper Confidence Bound (UCB) , and LinUCB for linear payoff models.

Before any predictive scoring, the engine applies a deterministic filter to ensure actions comply with business rules, legal regulations, and contact policies.

• Global Suppression: Automatically excludes users on legal hold, GDPR deletion requests, or global opt-out lists.
• Fatigue Rules: Enforces frequency caps (e.g., max 3 marketing contacts per week) and channel-specific time windows.
• Inventory Eligibility: Removes offers for products that are out of stock or services unavailable in the user's geography.

The engine scores each eligible action by calculating a weighted utility score that maps a predicted outcome to a specific business objective, not just click probability.

• Multi-Objective Optimization: Combines disparate goals like Customer Lifetime Value (CLV) , short-term revenue, and retention risk into a single scalar score.
• Weighting: Business stakeholders configure the relative importance of each objective (e.g., 70% revenue, 30% retention).
• Formula: Utility = w1 * p(conversion) * avg_order_value + w2 * p(churn_reduction)

The engine must consume the downstream outcome of its own decisions to correct future predictions, creating a continuous feedback loop.

• Event Ingestion: Listens for conversion events, churn signals, and unsubscribes from the operational data stream.
• Delayed Reward Handling: Updates bandit model parameters with the actual reward, even if the conversion occurs days after the initial action.
• Counterfactual Evaluation: Uses techniques like Inverse Propensity Scoring (IPS) to estimate what would have happened if a different action had been chosen, enabling offline policy evaluation.

Every decision must be logged immutably with the full context, winning score, and runner-up actions to satisfy model risk management (MRM) and regulatory requirements.

• Decision Audit Log: Records the user context vector, all candidate action scores, the applied eligibility rules, and the final selected action.
• Feature Attribution: Uses SHAP (SHapley Additive exPlanations) values to explain which user features most influenced the final decision.
• Compliance: Provides the data lineage required for SR 11-7 compliance in financial services and consumer protection audits.