Auditing AI in Finance: Risk, Fairness, Robustness and Compliance: A Systematic Review

This systematic review synthesizes the emergence and maturation of algorithmic auditing and assurance frameworks for AI-driven financial systems, integrating conceptual, technical, regulatory, and ethical literatures to produce a coherent approach for practice and governance. Using PRISMA and a structured search of major databases from January 2015 to October 2025, we screened 1,243 records and retained 40 peer-reviewed studies for detailed synthesis and appraisal. The review organizes findings across four core dimensions: risk assessment, fairness, robustness, and regulatory compliance. In the risk domain, the study contrasts traditional measures such as value at risk and expected shortfall with AI-aware formulations that incorporate model uncertainty, predictive miscalibration, distributional drift, and adversarial vulnerabilities. Hybrid approaches that combine statistical risk models are highlighted, with machine learning forecasts and recommended adversarial risk testing, comprehensive sensitivity analysis, and routine back testing to detect tail exposures and manipulation vectors. This study synthesizes operational metrics, including demographic parity and equalized odds, and highlights intersectional fairness to capture overlapping vulnerabilities that single-attribute measures may miss. The robustness analysis emphasizes adversarial training, systematic sensitivity testing, and hybrid scenario-based evaluations designed to probe both technical fragility and economic instability. For compliance and governance, established practices that embed explainability, comprehensive documentation, model cards, provenance tracking, and automated compliance checks into production pipelines are reviewed, with attention to alignment with Basel III, the GDPR, and other cross-jurisdictional regulations. Building on this synthesis, an integrated Assurance Index is proposed that aggregates normalized scores for risk, fairness, robustness, and compliance into a decision-ready composite with configurable weights that reflect institutional priorities. The framework pairs quantitative diagnostics with participatory auditing and stakeholder engagement to strengthen legitimacy and surface social concerns that technical metrics alone may overlook. The paper concludes with actionable research and policy recommendations, including benchmarking assurance practices, harmonizing metrics across jurisdictions, improving adversarial evaluation methods, incorporating industry experience into scholarly research, and supporting regulatory pilots to operationalize assurance mechanisms. Adoption of these measures can help ensure that AI-driven finance remains resilient, fair, transparent, and accountable. Future work should focus on open benchmarking datasets, collaborative audits, and practitioner-oriented toolkits for routine assurance deployment globally.