[2Yin-A] Poster 3

[2Yin-A-20] Maintainable Workflow Automation via Contextual Knowledge-Augmented SOPs for LLM Agents

〇Hiromi Kawatsu¹, Mitsuhiro Nakayama¹, Toshiaki Sota², Takehiko Yamaguchi¹, Michiaki Tatsubori¹ (1. IBM Japan, 2. IBM Japan Systems Engineering)

Keywords:

Standard Operating Procedures,Contextual Knowledge (Tacit Knowledge),Large Language Model (LLM) Agents

Standard Operating Procedures (SOPs) are widely used to design reproducible business workflows, and recently have been leveraged by large language model (LLM)–based agents to generate executable skills for automation. However, to maintain readability and ease of maintenance, SOPs are typically kept concise and thus omit domain-specific common sense and tacit knowledge required for execution. As a result, such contextual knowledge is often embedded in ad-hoc skill files, leading to repeated human intervention as SOPs evolve. This paper proposes an approach that separately manages SOPs and Contextual Knowledge (CK), where CK captures tacit knowledge, implementation details, and environment-dependent information. This separation keeps SOPs concise while enabling flexible adaptation to changes through a dedicated knowledge layer. We evaluate the proposed method on three business workflows: expense approval, hiring, and copy editing. Experimental results show that the separated CK captures implementation knowledge without redundancy, and that SOP descriptions are reduced by 75–84% in file size while preserving functional equivalence, demonstrating improved maintainability of LLM-based workflow automation.