The Submittals Agent

A Hybrid Workflow for Automating Submittal Extraction from Construction Specifications

Abstract

Construction specification documents encode contractual obligations and submittal requirements across documents often exceeding 1,000 pages. Manually extracting and organizing these requirements is labor intensive and error prone. We present the Submittals Agent, a two agent hybrid system that combines a conversational front end (Microsoft Copilot Studio) with a deterministic orchestration backend (Power Automate + FastAPI). Specifications are parsed via PyMuPDF and rule based Construction Specifications Institute (CSI) Master-Format segmentation; an LLM is invoked only for bounded metadata extraction, preserving human oversight over all contractual interpretation. Evaluation on 20 real world specifications demonstrates 94.3% F1-score, 94% time reduction, and 93% cost reduction versus manual baselines. The system has been deployed for six months with a construction contractor. Implementation details, a worked parsing example, and open source code are provided.