Scholarlink

<p>Automated workflow generation is considered a&#xa0;promising approach to involve domain experts without deep programming skills in process automation. However, existing approaches fail due to a&#xa0;fundamental problem: they generate syntactically correct processes that abort at runtime because required master data is missing or hallucinated. The authors develop a&#xa0;multi-agent system that resolves this dependency dilemma through intelligent division of labor. Specialized agents generate workflow structures, while a&#xa0;Master Data Agent checks which master data exists in the target application and identifies missing records. The evaluation using real-world administrative workflows shows that the system achieves complete syntactic correctness and high semantic quality for standard processes. However, it reaches its limits when dealing with complex dependencies. The pragmatic solution is a&#xa0;hybrid approach: the system handles technical complexity, while the human validates functional plausibility and creates master data manually. This division of labor saves, using the more efficient Instruct-Model, one-third of the creation time and transforms domain experts from creators into reviewers. The paper provides recommendations for action and demonstrates when the use of agentic systems for workflow generation is worthwhile.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Agentic Workflow Generation: Mit Agentic AI von der funktionalen Beschreibung zur ausführbaren Prozesslogik

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月29日

查看全文
  • Sebastian Schuppe,
  • Björn-Lennart Eger,
  • Barbara Dinter

摘要

Automated workflow generation is considered a promising approach to involve domain experts without deep programming skills in process automation. However, existing approaches fail due to a fundamental problem: they generate syntactically correct processes that abort at runtime because required master data is missing or hallucinated. The authors develop a multi-agent system that resolves this dependency dilemma through intelligent division of labor. Specialized agents generate workflow structures, while a Master Data Agent checks which master data exists in the target application and identifies missing records. The evaluation using real-world administrative workflows shows that the system achieves complete syntactic correctness and high semantic quality for standard processes. However, it reaches its limits when dealing with complex dependencies. The pragmatic solution is a hybrid approach: the system handles technical complexity, while the human validates functional plausibility and creates master data manually. This division of labor saves, using the more efficient Instruct-Model, one-third of the creation time and transforms domain experts from creators into reviewers. The paper provides recommendations for action and demonstrates when the use of agentic systems for workflow generation is worthwhile.