Generating Workflow DAGs from Natural Language with Non-Reasoning LLMs

2026-08-31

AI Summary

"This study addresses the challenge of transforming natural language routing rules, authored by business administrators, into executable workflow diagrams for enterprise contact centers. The project employs a neural-symbolic decomposition approach, utilizing a compact intermediate representation and a deterministic compiler to minimize the need for direct graph construction by large language models. A key innovation is the integration of a learned registry selection front-end, which enhances the generation of relevant vocabulary, significantly improving the quality and efficiency of Directed Acyclic Graph (DAG) workflow generation. Implemented across four distinct models, the system achieved an effectiveness score of approximately 89%, with condition accuracy around 90% and JSON format correctness rates of 99%-100%. Additionally, the token count per rule was reduced by half compared to monolithic prompting methods."

Abstract

This paper addresses the problem of translating natural-language routing rules written by business administrators into executable workflow graphs for enterprise contact centers. Each target is a directed acyclic graph (DAG) of conditional actions with parallel branches, hit-first fallback chains, and per-branch Boolean predicates, encoded in the JSON dialect of a commercial routing platform. We show that neuro-symbolic decomposition enables lower-cost, non-reasoning large language models to generate complex workflow DAGs at production-relevant quality without expensive extended-reasoning models. Our central diagnostic is an emission-density bottleneck: on a 635-rule benchmark of manufactured synthetic data, models select the correct graph nodes with high accuracy but increasingly misconfigure attributes and Boolean grouping as the number of interdependent nodes emitted in one pass grows. We therefore move combinatorial graph construction from the model into a deterministic compiler driven by a compact intermediate representation, with a learned registry-selection front end that focuses generation on relevant vocabulary. Across four models, the full system reaches approximately 89% LLM-judge validity, approximately 90% exact-match condition accuracy, and 99-100% valid JSON while using roughly half the per-rule prompt tokens of a monolithic prompt. On GPT-5.3-chat, the method improves judge validity by 24 percentage points and achieves statistical equivalence to a reasoning model's out-of-the-box quality, although an approximately 8-point frontier gap remains. We also present a deployment path and transferable lessons for structured-generation applications.

Problem

• natural-language
• workflow
• DAG
• enterprise contact centers
• routing rules

Innovation

• non-reasoning LLMs
• neuro-symbolic decomposition
• workflow DAGs
• deterministic compiler
• intermediate representation