Artificial intelligence is increasingly used to support spoken language practice. However, different AI-assisted support systems do not support learners’ confidence equally. This explanatory sequential mixed-methods study examines associations among the perceived usefulness of AI feedback, AI-mediated argumentative scaffolding, self-reported cognitive engagement, and speaking self-efficacy in the context of spontaneous argumentation. The quantitative phase involved 130 university EFL learners who had used an AI-assisted argumentation coaching application for at least eight weeks. Regression analysis showed that perceived usefulness of AI feedback was positively associated with so-called speaking self-efficacy, that is, the confidence to perform spoken tasks successfully (β = .410, p < .001), as was AI-mediated argumentative scaffolding (β = .248, p = .002). In contrast, self-reported cognitive engagement was not significantly associated with speaking self-efficacy after the other predictors were taken into account (β = −.028, p = .725). The qualitative phase involved eight purposively selected participants representing variation in gender, year of study, IELTS target, and experiences relevant to the quantitative patterns. Thematic analysis of semi-structured interviews identified three explanatory themes: structural scaffolding, psychological safety, and cognitive limits. Participants frequently described organizational guidance and a nonjudgmental practice environment as useful, while some also reported repetition, fatigue, and dependence during extended AI use. Integrating the two phases suggests that perceived usefulness, scaffolding, and cognitive engagement should not be treated as interchangeable dimensions of AI-assisted speaking practice. The study proposes a provisional cognitive-threshold hypothesis for future testing, but does not claim a causal or nonlinear effect. The findings support strategically designed AI-assisted speaking practice that maintains opportunities for independent reasoning and human interaction.