India's digital payments revolution - anchored by the Unified Payments Interface (UPI), the Account Aggregator (AA) consent framework, and the Jan Dhan Aadhaar Mobile (JAM) trinity - has created the world's most dynamic real-time credit ecosystem. Buy Now Pay Later (BNPL) transaction volumes surpassed USD 45 billion in FY2023–24, growing at over 30% CAGR since 2021, with more than 180 million active users. Roughly 60% of new digital borrowers in urban Tier-2 and Tier-3 markets are either absent from credit bureaus or carry scores below 650. The machine learning pipelines that underwrite these borrowers operate as autonomous, multi-step decision agents - querying AA APIs, processing UPI histories, running ensemble classifiers, routing decisions, and logging compliance records without human intervention at transaction scale. This is not AI in lending; it is agentic AI in lending, and its governance demands more than post-hoc explanation. This paper presents the first empirical evaluation of explainability frameworks applied to agentic AI credit scoring systems in Indian embedded finance. Three XAI methods - SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and a tabular adaptation of Gradient-weighted Class Activation Mapping (Grad-CAM) - are assessed across XGBoost and Random Forest classifiers on three contextualised benchmark datasets: the German Credit dataset (proxy for bureau-light consumer credit), the HMEQ dataset (representing high-missingness alternative data environments), and a LendingClub sample (capturing India's gig economy income-volatility profiles). Assessment follows a three-dimensional framework of fidelity, stability, and comprehensibility, complemented by structured fairness auditing and regulatory alignment mapping across RBI Digital Lending Guidelines (2022), the Digital Personal Data Protection Act 2023, and EU AI Act Article 13. SHAP achieves fidelity of 0.91 and stability rank-correlation of 0.97, outperforming LIME (0.76; 0.63) and Grad-CAM (0.68; 0.71) across all 18 experimental conditions. XGBoost outperforms Random Forest on AUC-ROC in every dataset, with the widest margin on the 43-feature LendingClub sample (0.821 vs. 0.798). Statistically significant demographic disparate impact is detected in two of three datasets; SHAP uniquely enables the decomposition of proxy discrimination pathways into regulatory-grade attribution evidence. A novel positive-gradient asymmetry is identified in the tabular Grad-CAM adaptation, with direct implications for protected attribute surfacing. The paper proposes the Autonomy–Accountability–Auditability (AAA) governance framework as an organising architecture for trustworthy agentic AI in BNPL credit decisioning, demonstrating alignment with India's evolving regulatory stack and EU AI Act transparency requirements.