Abstract: Despite $30–40 billion in enterprise GenAI investment, 95% of organizations achieve zero measurable return, trapped on the wrong side of what we term the "GenAI Divide." This review synthesizes findings from MIT's Project NANDA research examining 300+ AI implementations and interviews with 52 organizations to identify why pilots stall and how exceptional performers succeed. The divide stems not from model quality or regulation, but from a fundamental learning gap: most enterprise AI systems lack memory, contextual adaptation, and continuous improvement capabilities. While consumer tools like ChatGPT achieve 80% exploration rates, custom enterprise solutions suffer 95% pilot-to-production failure rates. Organizations crossing the divide share three patterns: they partner rather than build (achieving 2x higher success rates), empower distributed adoption over centralized control, and demand learning-capable systems that integrate deeply into workflows. Back-office automation delivers superior ROI compared to heavily-funded sales functions, though measurement challenges persist. The emerging agentic web—enabled by protocols supporting persistent memory and autonomous coordination—represents the infrastructure required to bridge this divide at scale.