The rapid integration of generative artificial intelligence into higher education necessitates rigorous pedagogical frameworks to foster self-regulated learning without inducing cognitive over-reliance. This quasi-experimental study investigates the impact of an adaptive AI-driven scaffolding tutor on student metacognitive awareness, conceptual problem-solving efficacy, and self-directed study behaviors in undergraduate STEM physics courses. Over a sixteen-week academic semester, 380 undergraduate students were assigned to either an experimental group utilizing Socratic-guided AI conversational scaffolding or a control group utilizing standard static LMS resources. Quantitative pre- and post-intervention assessments combined with learning analytics telemetry reveal that students in the AI-scaffolded environment demonstrated statistically significant gains in problem-solving performance (F = 18.42, p < 0.001) and higher-order metacognitive self-regulation scores. Qualitative focus group interviews indicated that proactive cognitive prompt intervention encouraged reflective error analysis rather than passive answer retrieval. The paper articulates practical instructional design guidelines for educators and academic policymakers aiming to deploy generative AI technologies responsibly to enhance deep learning outcomes.