CognitiveCPQ: A Graph-Based Reinforcement Learning Framework for Autonomous Constraint-Aware Enterprise Deal Optimisation

Configure-Price-Quote (CPQ) systems are central to enterprise B2B sales execution, yet most deployments remain rule-based and deterministic, unable to learn from deal outcomes. This paper presents CognitiveCPQ, combining a Heterogeneous Constraint Graph (HCG) with Proximal Policy Optimisation (PPO) for autonomous, constraint-aware deal optimisation. A deterministic kill-switch guarantees zero hard constraint violations; soft constraints are navigated via a z-score-normalised penalised reward. A KernelSHAP explainability layer provides per-decision attribution. Evaluation on a large-scale synthetic simulation demonstrates 18.9% margin uplift, 94.3% constraint compliance, and a 67% reduction in configuration steps over rule-based baselines. We additionally specify two robustness evaluation protocols—response-model misspecification and state-observability noise—as immediate follow-up work; both protocols are expected to confirm graceful degradation given the kill-switch’s structural independence from the response model and state vector. All results represent proof-of-concept performance in simulation; production deployment requires shadow-mode calibration on real deal data.