Accurate prediction of carbon monoxide (CO) concentration at the main exhaust outlet of the sintering process is important for combustion state assessment and CO emission control. Existing data-driven methods treat process variables as homogeneous inputs, neglecting their process affiliations and cross-system interactions. This study proposes a process-guided Kolmogorov–Arnold network (PG-KAN) for CO concentration soft measurement. According to process relationships, 230 process variables are organized into material, flue gas, and auxiliary systems, and independent radial basis function-based Kolmogorov–Arnold network (RBF-KAN) encoding branches extract system-specific nonlinear representations. Cross-system interaction modules capture interactions among these systems. A tail coverage-aware loss reduces large errors and underestimation bias in high-CO concentration intervals. Experiments on 18,749 industrial samples show that PG-KAN achieves a mean absolute error (MAE) of 107.700 mg/m³, a coefficient of determination (R²) of 0.9105, and a Pearson correlation coefficient of 0.9542, outperforming six baseline models. Ablation studies verify the effectiveness of process-guided variable organization, cross-system interactions, and the proposed loss. Edge-function and SHapley Additive exPlanations (SHAP) analyses provide insights into nonlinear mappings and feature contributions learned by the model. The proposed method provides an interpretable, process-guided framework for CO concentration soft measurement in industrial sintering processes.