We formalize Prescriptive Artificial Intelligence as a distinct paradigm for human-AI decision collaboration in high-stakes environments. Unlike predictive systems optimized for outcome accuracy, prescriptive systems are designed to recommend and audit human decisions under uncertainty, providing normative guidance while preserving human agency and accountability. We introduce four domain-independent axioms characterizing prescriptive systems and prove fundamental separation results. Central among these is the Imitation Incompleteness theorem, which establishes that supervised learning from historical decisions cannot correct systematic decision biases in the absence of external normative signals. Consequently, performance in decision imitation is bounded by a structural bias term epsilon_bias rather than the statistical learning rate O(1/sqrt(n)). This result formalizes the empirically observed accuracy ceiling in human decision imitation tasks and provides a principled criterion for when automation should be replaced by epistemic auditing. We demonstrate the computational realizability of the framework through an interpretable fuzzy inference system, applied as a stress test in elite soccer decision-making, where it reveals systematic decision latency and risk states obscured by outcome and status quo biases. The proposed framework establishes Prescriptive AI as a general, realizable class of decision-support systems applicable across safety-critical domains in which interpretability, contestability, and normative alignment are essential.

翻译：我们将规范性人工智能形式化为高风险环境中人机决策协作的一个独特范式。与针对结果准确性进行优化的预测系统不同，规范性系统旨在不确定条件下推荐和审计人类决策，在保留人类自主权和问责制的同时提供规范性指导。我们提出了刻画规范性系统的四个领域无关公理，并证明了基本的分离性结果。其中的核心是模仿不完全性定理，该定理确立了在缺乏外部规范性信号的情况下，基于历史决策的监督学习无法纠正系统性决策偏差。因此，决策模仿的性能受限于一个结构性偏差项 epsilon_bias，而非统计学习率 O(1/sqrt(n))。这一结果形式化了在人类决策模仿任务中经验观察到的准确性上限，并为何时应以认知性审计替代自动化提供了原则性判据。我们通过一个可解释的模糊推理系统证明了该框架的计算可实现性，并将其作为压力测试应用于精英足球决策分析，揭示了被结果偏差和现状偏差所掩盖的系统性决策延迟与风险状态。所提出的框架确立了规范性人工智能作为一类通用且可实现的决策支持系统，适用于可解释性、可争议性和规范性对齐至关重要的各类安全关键领域。