The use of AI-powered chatbots has recently been extensively examined for second language (L2) learning. While their positive effects have been widely reported regarding L2 English learning, studies involving less commonly taught languages (LCTLs) are scant. The current study incorporated an AI chatbot called Iruda in L2 Korean teaching, to improve Korean lexico-grammar, while examining moderation of proficiency levels and task types. Sixty-six students engaged with AI chatbot-incorporated tasks with half of the grammar points of the course curriculum throughout the semester. During their final exam, students' grammar was assessed through selected-response (SR) and constructed-response (CR) tasks (sentence completion and sentence composition). Using a paired-sample t-test, we compared item mean differences between grammatical features practiced with chatbot-incorporated and conventional methods. Items related to chatbot practice showed a significantly higher mean (t(65) = 4.29, p <0.0001). Follow-up regression and repeated-measures ANOVAs revealed greater changes by chatbot practice among students with lower pre-existing Korean proficiency and significant moderation of task types with higher SR item means. Students' perceptions were also positive. The study demonstrated the effectiveness of chatbot-enhanced instruction for LCTLs, while suggesting nuanced application for learners of varied proficiency levels and different task types.