This paper examines when and how artificial intelligence (AI) systems create economic value in decision-making processes. Through a decision-theoretic model, we show that an AI system's accuracy alone is insufficient to determine its optimal usage. While conventional wisdom suggests AI is deployed once it reaches a threshold accuracy level (the ``95\% Rule"), we demonstrate that economic value depends crucially on verification costs, the stakes of incorrect decisions, and the availability of complementary information sources. We identify conditions under which high-accuracy AI may not be used at all, or may be used only with human verification, despite impressive benchmark performance. Our analysis explains observed patterns in AI adoption across sectors and provides guidance for AI evaluation that better aligns with economic value creation. In particular, we demonstrate that AI capabilities are most valuable when their prediction errors are detectable through cost-effective verification methods, especially in high-stakes domains where errors are costly. The economic perspective on AI reveals that ``95\% accuracy is not all you need."