From Enterprise Institutional Documents to Editable Process Models: A Hybrid LLM-Rule-Engine Architecture for Enterprise Governance

Enterprise institutional documents encapsulate substantial process‑oriented management knowledge. Transforming them into structured editable process models constitutes a foundational task for business process management (BPM) and governance‑risk‑compliance (GRC) digitalization. However, manual process modeling is time-consuming, inconsistent across practitioners, and difficult to maintain, while existing automated approaches—whether rule-based methods grounded in traditional natural language processing (NLP) or recent large language model (LLM) generation methods—cannot simultaneously satisfy the dual requirements of semantic understanding accuracy and output format conformance, and typically lack deep integration with enterprise approval systems. Following the design‑science‑research (DSR) paradigm, this paper proposes, implements, and deploys a hybrid LLM and rule‑engine architecture (LRHA) for automatically generating organization‑compliant editable process models from enterprise institutional documents. The architecture decomposes the transformation task into six stages—text parsing, element extraction, logic modeling, rule matching, JSON generation, and draft output—with the LLM handling the first three stages that rely on semantic understanding and the rule engine handling the latter three stages that rely on deterministic constraints, thereby separating semantic interpretation from format enforcement. For Chinese enterprise institutional texts, we derive a set of “textual cue to process element” mapping rules and define a JointJS-oriented process model representation specification together with ten executable quality validation rules. Distinct from generic text-to-BPMN methods, LRHA provides native support for enterprise governance capabilities at the architectural level, including institutional‑text ambiguity resolution, compliance constraint validation, asynchronous integration with existing approval OA systems, version traceability, and audit trail preservation. The system has been deployed in production within the risk control information system of a large energy enterprise group, achieving end-to-end automated integration with the group’s OA system through asynchronous task interfaces (task submission, status query, multi-file download, and proactive callback), underpinned by HMAC signature verification and idempotency mechanisms. Case evaluation shows that for complex processes with over 40 nodes and edges, manual modeling work is reduced from several days to approximately 20 minutes, requiring only minor manual fine‑tuning. Large‑scale quantitative benchmark evaluation remains as future work. This paper makes four contributions: (1) it proposes and validates the division of labour and collaboration mechanism between LLMs and rule engines in automated process model generation; (2) it constructs a mapping rule system and executable quality validation specification tailored to Chinese enterprise institutional texts; (3) it natively implements enterprise‑governance‑oriented architectural capabilities for institutional‑text ambiguity resolution, compliance constraint validation, OA‑system integration, version traceability, and audit‑trail preservation; and (4) it reports the complete deployment pathway and engineering experience of the hybrid architecture within an enterprise information system, offering a replicable reference paradigm for similar organizations.