AI is reshaping external auditing—boosting fraud detection and analytics—but the evidence is fragmented and thin on causal links to improved audit quality, and governance, transparency and bias concerns remain.

Main Finding

A systematic literature review and bibliometric analysis of 130 peer‑reviewed articles (2005–2025) finds rapid growth in research on AI in external auditing since 2018 and identifies three dominant, but poorly integrated, research streams—(1) machine learning & fraud detection, (2) audit analytics & big data, and (3) AI adoption & governance. Common AI techniques (ML, neural nets, NLP, RPA, expert systems) are shown to improve detection, risk assessment, and some dimensions of audit quality, but important theoretical, methodological, and governance gaps remain (notably algorithmic opacity, bias, professional‑skepticism erosion, data quality, and fragmented literature).

Key Points

• Scope and sample: 130 English, peer‑reviewed journal articles retrieved from Scopus and Web of Science; timeframe 2005–2025.
• Growth pattern: steep publication increase after 2018; leading contributing countries are the United States, China, and the United Kingdom.
• Three bibliometric/thematic clusters:
  • Cluster 1 — Machine learning & fraud detection: predictive models and anomaly detection dominate; ML often outperforms traditional statistical approaches.
  • Cluster 2 — Audit analytics & big data: continuous auditing and full‑population analytics; challenges include data quality, integration, and operational governance.
  • Cluster 3 — AI adoption & governance: ethical, accountability, transparency, regulation, and professional skepticism concerns; governance research is disconnected from technical research.
• Frequently used AI methods: machine learning (incl. supervised models), neural networks (high predictive power, low interpretability), natural language processing (textual disclosures), robotic process automation (efficiency), expert systems.
• Theoretical framing: fragmented use of Technology Acceptance Model (TAM), agency theory, and resource‑based views; rarely integrated into cohesive causal frameworks.
• Main benefits reported: improved fraud detection rates, enhanced risk assessment, potential increases in audit efficiency and thoroughness (e.g., continuous/full‑population testing).
• Main concerns reported: algorithmic bias, lack of transparency/interpretability, erosion of professional skepticism, accountability and liability issues, data governance and quality limitations.
• Contribution claim: the paper proposes an integrated framework linking AI applications to audit quality and outlines an interdisciplinary research agenda.

Data & Methods

• Review protocol: PRISMA 2020–compliant systematic review with preregistered protocol (inclusion/exclusion rules, search strategy).
• Databases: Scopus and Web of Science; search query combined AI terms (Artificial Intelligence OR Machine Learning OR Deep Learning OR Expert Systems OR Natural Language Processing) AND audit terms (External Audit OR Audit Quality OR independent audit), applied to titles/abstracts/keywords.
• Inclusion criteria: English language, peer‑reviewed journal articles, empirical and theoretical studies that explicitly address AI in external auditing. Exclusions: conference papers, book chapters, internal auditing-only studies, non‑English.
• Screening: duplicate removal, title/abstract screening, full‑text eligibility; two independent reviewers with discrepancies resolved by discussion/third reviewer; inter‑rater reliability measured (Cohen’s kappa — reported as high).
• Bibliometric analysis: VOSviewer used on Scopus export; keyword co‑occurrence (min occurrence threshold = 5), co‑citation and co‑authorship mapping; association‑strength normalization and clustering to detect thematic groups.
• Output: descriptive statistics (annual distribution), keyword clusters, co‑citation clusters (revealing theoretical streams), and qualitative synthesis of empirical findings.
• Limitations noted by authors: language and publication‑type restrictions (English, peer‑reviewed journals only), possible exclusion of practitioner literature and conference advances; bibliometric thresholds may omit emerging but low‑frequency topics.

Implications for AI Economics

• Productivity and cost structure:
  • AI enables full‑population testing and continuous auditing, potentially reducing marginal audit costs and sampling‑related inefficiencies; firms investing in AI can achieve scale economies in audit production.
  • Expect shifts in labor demand: lower demand for routine sampling tasks, higher demand for data science, model‑validation, and governance expertise—implying re‑skilling needs and wage skill premia.
• Market structure and competition:
  • Large audit firms (Big Four) have stronger capacity to develop or acquire AI capabilities, which may increase market concentration and raise entry barriers for smaller firms unless platforms or shared services emerge.
  • AI could be a strategic asset (resource‑based view) that differentiates firms by service quality, potentially altering competition on price vs. quality.
• Audit quality, signaling, and agency costs:
  • Higher detection probabilities from AI reduce information asymmetries and agency costs between managers and stakeholders, potentially lowering cost of capital for audited firms.
  • However, opacity/interpretability problems create new signaling frictions: stakeholders and regulators may discount AI‑based assurances unless transparency and validation mechanisms are observable.
• Regulatory and governance implications:
  • Economic returns to AI adoption depend on regulatory acceptance: standards for model explainability, validation, and liability allocation will influence investment incentives and market adoption rates.
  • Regulators and standard‑setters face a tradeoff between encouraging innovation (productivity gains) and enforcing controls that prevent systemic trust erosion from algorithmic errors or biases.
• Externalities and systemic risk:
  • Widespread use of similar ML models across auditors could create correlated detection failures or model‑driven blind spots, generating systemic audit‑market externalities.
  • Conversely, broad AI adoption raising baseline audit quality could reduce fraud incidence economy‑wide, with positive externalities for capital markets.
• Empirical research agenda (economics‑focused recommendations):
  • Causal impact studies: estimate AI adoption effects on audit quality metrics (misstatement detection, restatements, audit opinions) using quasi‑experimental methods.
  • Cost‑benefit analyses: firm‑level and market‑level ROI of AI investments in auditing, accounting for compliance, training, and governance costs.
  • Labor economics: measure heterogeneous labor displacement vs. task complementarities; wage and employment dynamics for auditors and data scientists.
  • Market structure modeling: analyze how AI adoption affects competition, entry/exit, and concentration among audit firms.
  • Regulation and incentives: evaluate how different regulatory regimes (transparency/explainability mandates, liability rules) alter incentives for AI investment and disclosure.
  • Measurement work: develop economic measures for algorithmic transparency, audit‑AI reliability, and governance quality to feed into applied studies.
• Practical policy insight: to capture positive economic gains while containing risks, coordinated policies are needed—standards for model validation and disclosure, incentives for independent model audits, and workforce transition support.

Limitations of the reviewed literature (relevant to economists): sparse causal identification, weak integration of governance and technical effects, limited cross‑country comparative studies, and lack of market‑level outcome measures. These gaps present clear opportunities for rigorous AI economics research tied to auditing outcomes.