Ensuring that enterprise operational activities align with officially registered business scopes is essential for compliance and risk management. However, evaluating business consistency remains challenging due to heterogeneous data sources and the semantic complexity of unstructured enterprise documents. To address this, we propose a Hybrid Large-Small Model Framework (HLSMF) that integrates a large language model (LLM) for prompt-enhanced semantic extraction with a small-model ensemble for multi-source data fusion and quantitative consistency scoring. The LLM captures deep semantic cues from supplier contracts, project reports, and financial records, while embedding encoders and a graph attention network consolidate structured and unstructured features into a unified representation for consistency evaluation. Experimental results show that HLSMF achieves an AUROC of 89.16%, a recall of 82.67%, and an F1-score of 78.35%, outperforming all baseline methods. The results demonstrate the framework’s effectiveness in detecting off-scope operations for compliance assessment in enterprise environments.