Separating Evidence from Action in AI-Assisted Procurement

This working paper examines how AI-assisted procurement systems can preserve useful evidence without allowing advisory signals to silently grant permission to act. It proposes separating evidence validity, policy eligibility and scoped authorization, drawing on dated QuickAIBuy engineering records and an illustrative reproducible Boolean model.The paper includes a conceptual architecture, a logical counterexample, a synthetic procurement-cost comparison and proposed implementation tests. It distinguishes reported historical observations from design proposals and illustrative calculations. It does not establish production-system safety, supplier reliability or improved commercial outcomes.Publication status: Author-approved working paper / technical design case study, version 1.0, dated 28 October 2026. Not peer reviewed.This Figshare record links to the manuscript hosted on Zenodo. Original DOI: https://doi.org/10.5281/zenodo.23715578AI assistance: ChatGPT assisted with source selection, literature lookup, drafting, diagrams and the illustrative executable model. Mohammad Maani reviewed and approved the manuscript. AI assistance and study limitations are disclosed in the paper.Competing interests: Mohammad Maani is the founder of QuickAIBuy, the project discussed in this case study.Project context: https://www.quickaibuy.com