As a crucial driving force for high-quality economic development, artificial intelligence (AI) attracts widespread investment. Based on panel data of Chinese listed companies from 2010 to 2023, this paper adopts two-way fixed effect, threshold and mediating effect models for empirical analysis. The results show that AI investment has a U-shaped relationship with enterprises’ total factor productivity (TFP) accompanied by a single threshold. Restrained by high adaptation costs and insufficient data accumulation in the initial investment stage, AI impedes TFP growth, while its effect turns positive after crossing the threshold. Capital utilization efficiency serves as a U-shaped mediating channel between the two variables. Heterogeneity analysis indicates that large-scale, highly digitalized and manufacturing enterprises reach the U-shaped inflection point earlier and gain more prominent productivity improvement, whereas SMEs, non-manufacturing and low-digital enterprises face higher transformation thresholds with differentiated mediating effects. This research provides empirical references for enterprises to optimize AI allocation and improve efficiency via capital management, as well as for governments to formulate targeted policies.