Help-seeking is a central self-regulated learning (SRL) process; however, generic scaffolds often misalign with learners’ heterogeneous needs. This systematic review addresses the lack of consolidated evidence on how artificial intelligence (AI), including large language models (LLMs), customises help-seeking scaffolds, the effects thereof, and the methodological and ethical constraints. Guided by SRL theory, this review examined empirical studies on AI-driven customised help-seeking support, focusing on evidence availability, comparative impacts versus standard scaffolds, evaluation approaches, and reported gaps. Database searches (IEEE Xplore, Web of Science, ERIC, and ScienceDirect) conducted on 28 May 2025 yielded 3,576 records; blinded multi-reviewer screening in Rayyan following PRISMA procedures identified 31 eligible studies for narrative synthesis. Data extraction by three reviewers showed high raw agreement (87.3%) but low chance-corrected agreement (Cohen’s κ ≈ 0.05; Fleiss’ κ ≈ 0.17), resolved via structured adjudication and full consensus. The included interventions (2017–2025) ranged from rule-based intelligent tutoring and learning-analytics dashboards to LLM-based textbooks, writing feedback, laboratory guidance, and coding assistants, enabling conversational, intent-sensitive, on-demand scaffolding. Most studies reported improved academic performance and engagement, particularly for lower-performing or lower–prior knowledge learners, although null or mixed effects and inconsistent satisfaction outcomes were common. The recurrent limitations concerned sample scope, duration, measurement heterogeneity, usability, and risks related to hallucinations, bias, privacy, equity, and overreliance. The review concludes that although AI-customised help-seeking scaffolds are promising, they are context-dependent and delineate methodological and design priorities to advance SRL-aligned, responsible AI support for help-seeking.