Real-World Adaptation and Enhancement of Recurrent Marginal Structural Networks to Estimate Causal Impact of Sales and Account Management Teams

In industrial settings like Google Customer Solutions (GCS), estimating the incremental value of human and agentic-AI sales interventions is challenging because treatments are dynamic and sequential, and their assignment is influenced by complex time-varying confounders that also affect advertiser outcomes. To address these challenges, we introduce several methodological extensions to RMSN: attention-augmented sequence modeling with layer normalization, long-horizon weight stabilization, multi-action counterfactual prediction with scalable Shapley attribution, hyperparameter ensembling, and explicit handling of structurally untreated control accounts. The resulting framework supports scalable estimation and attribution of longitudinal treatment effects across heterogeneous sales programs. We demonstrate improved multi-step prediction performance over the original RMSN baseline on two public benchmark datasets.