The Ninety Percent Trap: Empirical Documentation on Why Generative AI Requires Deep Domain Expertise

Abstract

This study examines how the dominant industry narrative presents generative artificial intelligence (GenAI) as a trigger for total software automation and positive exponential outcomes. Empirical evidence from codebases and recent software-engineering studies, however, suggests a paradox: rapid generation can increase review overhead, architectural decay, and critical edge-case failures. By analysing the ArtCraft case study alongside quantitative reports from GitClear, DORA, and the empirical software-engineering literature, this paper formulates the Ninety Percent Trap: GenAI can accelerate initial scaffolding from zero to one while shifting the burden of production dependability to human experts. The paper also presents the True Cost of AI Quality (TCAQ) framework, which evaluates GenAI productivity through generation time, expert review, edge-case resolution, and technical-debt cost rather than output volume alone. An illustrative calculation shows that a two-hour generation phase can become a 112-hour end-to-end engineering effort after review, edge-case resolution, and debt mitigation. The findings indicate that domain expertise remains essential for architectural reasoning, security verification, and dependable delivery.