Artificial Intelligence for Competitive Exam Preparation in India: A Learning Yield Framework for Personalised Practice, Independent Performance, and Cost-Effectiveness

Competitive-exam aspirants must repeatedly decide what to study, which errors to revisit, when to seek an explanation, and how to divide scarce time across a broad syllabus. Artificial intelligence (AI) can analyse practice histories, recommend a next task, generate feedback, adapt explanations, and support coaches at far lower marginal cost than one-to-one instruction. Those capabilities do not by themselves establish learning. The relevant test is whether AI improves what learners can later do without assistance, across the syllabus, and at a lower full cost per unit of retained learning. This paper develops an AI Learning Yield Framework for that test, using India’s competitive-exam ecosystem as the application context. We conducted a focused structured synthesis of AI tutoring, learning analytics, feedback, retrieval and spaced practice, technology-aided instruction in India, and direct exam-preparation evidence. No participant data were collected and no new pooled effect was calculated. The evidence is promising and sharply conditional. In a Turkish high-school field experiment involving nearly 1,000 students, unrestricted GPT-4 raised assisted practice grades by 48% but lowered subsequent unassisted exam grades by 17%; a teacher-designed GPT tutor raised assisted practice grades by 127% and avoided the decline, but did not significantly improve the exam outcome. In a 194-student Harvard physics crossover trial, a carefully designed AI tutor produced more than twice the median learning gain of an active-learning class; the estimated effect was 0.73 to 1.3 standard deviations, and median AI study time was 49 minutes versus an assumed 60 minutes in class. A human-AI system tested with 900 tutors and 1,800 underserved K-12 students increased topic mastery by four percentage points overall and by nine points among students of lower-rated tutors, at an author-estimated annual API cost of US$20 per tutor. These studies differ in population, treatment, comparator, and outcome timing. They show that instructional design and the removal of assistance at assessment matter as much as model capability. India-specific evidence establishes scale and data opportunity but does not yet establish an AI effect. UPSC2M describes 1,962,573 attempts by 46,235 aspirants, while its repository states that the full dataset is proprietary. Official NTA data record 1,539,848 unique JEE Main 2025 Paper 1 registrations. The proposed AI Learning Yield Loop combines syllabus coverage, learner modelling, an explainable next action, scaffolded help, delayed unaided retesting, and plan revision. A four-arm field design compares usual preparation, a descriptive dashboard, a rules-based loop, and the same loop with bounded generative AI. Primary outcomes are independently scored unseen items and delayed retention; secondary outcomes include coverage, time, dependence on assistance, full programme cost, implementation fidelity, and equity. Direct causal evidence in Indian competitive-exam preparation remains limited. Momentum is a prospective testbed, and the paper reports no Momentum outcomes.