AI Document Authoring: Structure for Reading, Prose for Writing

This research examines a multi-agent system designed for authoring formal documents, specifically tender responses, and uncovers a crucial "asymmetric structural conditioning" principle. The study found that while rendering documents as structural markup (like XML) significantly improves the system's ability to extract information during the reading phase, the opposite is true for writing. When instructing the AI to write, converting bid instructions from prose to nested XML actually degraded answer quality. The system, running an open-weights model under sovereignty constraints, performed comparably to human-written bids in blind comparisons, with most identified gaps stemming from a lack of source knowledge rather than writing quality. The key takeaway is that structure is beneficial for AI when it reads and processes information, but prose is more effective for conditioning its writing output, and explicitly naming forbidden constructions can paradoxically concentrate them.