Sales compensation has long resisted systematic optimization despite its central role in driving organizational performance. Traditional approaches rooted in historical benchmarks and managerial intuition struggle with the mounting complexity of modern B2B sales environments. Machine learning now promises to revolutionize incentive design by processing vast datasets to identify patterns invisible to human analysts and generate recommendations that supposedly balance competing objectives. Yet amid the enthusiasm, a troubling question persists: does the technology actually deliver? This review critically examines what we know and more importantly, what we don't about AI-powered sales incentive systems. Drawing on empirical studies, theoretical frameworks, and implementation experiences across behavioral economics, organizational psychology, and computational intelligence, we find a substantial gap between predictive capability and prescriptive value. While algorithms can forecast performance with reasonable accuracy, evidence that AI-optimized compensation improves business outcomes remains surprisingly thin. More concerning, we identify serious risks around algorithmic bias, unintended behavioral consequences, and over-optimization that organizations have barely begun to address. The field stands at a critical juncture where sober assessment matters more than technological optimism.