A mixed-methods evaluation of an AI-powered adaptive learning system for professional english training

Artificial intelligence is increasingly used to personalize language learning, but evidence from workplace-oriented programs remains limited, especially regarding the mechanisms through which adaptive systems influence learning and implementation. This study evaluated an integrated AI-powered adaptive learning framework for occupational English development in which learner modeling, analytics, content adaptation, autonomous pacing, and reinforcement-learning-informed sequencing operated as a coordinated intervention. A 16-week quasi-experimental mixed-methods design compared the adaptive platform with non-adaptive e-learning and instructor-led provision. Condition allocation was non-random and used stratified matching within partner-organization and cohort constraints; attrition was low and similar across conditions, and the intended weekly learning dose was standardized. Outcomes combined CEFR-aligned proficiency assessment, authentic communication rubrics, interaction logs, motivation and stakeholder measures, and qualitative evidence. The adaptive condition produced a statistically significant and large advantage in overall proficiency ( p < .001; η² = 0.19), with particularly strong improvements in domain-specific vocabulary and functional communication, as well as better retention. Documented routine instructor and administrative workload, assessment and scoring, content preparation and assignment, learner follow-up, progress monitoring, and reporting, was also lower than under instructor-led provision. These findings contribute system-level evidence that integrated adaptive delivery can connect fine-grained learner information with timely instructional adjustment in organizational language education. Because the intervention bundled multiple adaptive and implementation components, it does not identify reinforcement learning as the sole causal mechanism. Nevertheless, it provides a practical basis for designing human-supervised, privacy-conscious adaptive programs and for testing individual mechanisms through factorial or ablation studies.