Separating Evidence from Action in AI-Assisted Procurement

AI-assisted procurement combines supplier observations, product matching, cost estimates and human judgments. This working paper examines how evidence, advisory outputs, eligibility checks and permission to act can be separated. A retrospective documentary case from QuickAIBuy motivates a proposed control architecture and a reproducible Boolean counterexample. Exhaustive enumeration finds advisory-dependent permission changes in three of eight paired comparisons under the deliberately coupled rule and zero under the separated rule. These are properties of an illustrative model, not measured procurement outcomes. A synthetic supplier-quote example, implementation test plan and reusable buyer review card translate the analysis into practical checks. Scope: author-approved working paper, not peer reviewed. Historical internal records were not independently rerun; no customer outcomes or production safety are established. ChatGPT assisted with synthesis, drafting, diagrams and model construction, as disclosed in the manuscript. Mohammad Maani approved the content and public deposit on 28 October 2026. Affiliation and project context: https://www.quickaibuy.com Included materials: manuscript PDF, standalone Python reproducibility script and Boolean truth-table CSV.