This study investigates the paradox of generative AI adoption in Business English writing instruction. While tools like ChatGPT and Doubao enhance efficiency, they may simultaneously inhibit the development of independent writing skills and critical judgment. Through a mixed-methods design, the study analyzed 107 usable text segments from open-ended survey responses of 80 Chinese undergraduates, capturing authentic descriptions of AI use cases and primary concerns. Findings reveal a consistent efficiency-critique gap. Students proficiently employ AI for rapid, surface-level text optimization using simple commands (e.g., “formalize this email”), yet over 60% express acute anxiety about dependency, skill atrophy, and loss of authentic voice. Their instrumental use aligns with the “identity–ideology–capital” framework, where AI becomes the dominant agent in meaning-making, displacing student agency. While students recognize risks of homogenized output and contextual misalignment, they lack structured practices for critically interrogating AI suggestions. The study concludes that current pedagogical models insufficiently bridge functional tool proficiency and critical digital literacy. The study argues for intentional instructional design that transforms AI from a crutch into a cognitive scaffold—compelling students to slow down the “optimize-submit” cycle, deconstruct algorithmic choices, and actively rewrite outputs for specific rhetorical contexts. By foregrounding students’ own fragmented yet candid voices, this research grounds the AI-in-education debate in empirical learner practices and points toward pedagogical frameworks that reconcile instrumental efficiency with value rationality.