AI promises real‑time controls for corporate reporting but also brings model risk; firms that pair AI with independent model governance, robust validation, and human oversight secure better financial reporting quality.

Main Finding

The paper reframes AI in corporate reporting as embedded governance infrastructure rather than merely a tool. AI creates two countervailing pathways affecting financial reporting quality (FRQ): (1) control enhancement—greater scope, consistency, timeliness, and reallocation of skilled judgment that can improve FRQ; and (2) model risk—opacity, training-data drift, automation bias, and diffused responsibility that can worsen FRQ. The net effect depends on four mediating governance variables (model validation, human-in-the-loop review, audit-trail transparency, and organizational AI governance maturity), which interact substitutionally and complementarily (formalized in propositions P1–P4).

Key Points

• Conceptual reframing: AI systems that flag transactions, produce estimates, or draft disclosures perform COSO-like control activities and information functions and therefore should be treated as part of the control architecture.
• Dual mechanisms:
  • Control enhancement pathway: full-population anomaly detection, consistent rule application, near-real-time monitoring, and freeing humans for higher-judgment tasks → positive for FRQ (P1).
  • Model risk pathway: black-box opacity, dependence on historical training data and drift, automation bias, and diffuse accountability → negative for FRQ (P2).
• Four mediating governance variables determine outcome:
• Model validation (initial and ongoing testing/documentation)
• Human-in-the-loop review (depth and substance of human oversight)
• Audit-trail transparency (reconstructability and evidence retention)
• AI governance maturity (policies, roles, committees, ERM integration)
• Interactions and configurational logic:
  • Strong governance across these four dimensions tends to realize AI’s benefits (P3).
  • The marginal value of any single governance element depends on the state of the others (P4).
• Distinct application categories with differing governance needs:
  • Transaction-level processing, estimation (e.g., expected credit loss models), and narrative generation (LLMs).
• Practical recommendations for stakeholders:
  • Preparers: explicitly include AI in internal control frameworks and create independent model-governance functions.
  • Auditors: develop capacity to evaluate model governance and human-in-the-loop processes.
  • Standard setters / regulators: require disclosures about material AI uses in reporting, adapt prudential model risk templates, and align with regulatory direction such as the EU AI Act.
• Contextual note: relevance to Central and Eastern Europe (CEE) where firms often adopt AI tools from larger markets.

Data & Methods

• Methodology: Structured literature synthesis following Snyder (2019).
• Search strategy: Systematic searches in Web of Science and Scopus for publications 2010–2025 using combinations of AI/machine learning/algorithm terms with financial reporting quality/internal control/audit quality/earnings quality, plus parallel searches for algorithmic governance topics.
• Scope: Focus on peer‑reviewed articles in accounting, finance, business ethics, information systems, and regulatory publications from standard-setters/supervisors. Snowballing used to capture foundational work predating 2010.
• Coverage: About 60 sources synthesized; ~35 cited as directly supporting the framework.
• Theoretical integration: Agency theory (incentives), COSO internal control framework (control architecture), and literature on algorithmic accountability (opacity, explainability, auditability).

Implications for AI Economics

• Information production and market effects:
  • AI can reduce information asymmetry and thus lower cost of capital if governance captures control benefits; conversely, poor governance can increase misreporting risk and raise information uncertainty and risk premia.
  • The balance affects market pricing, liquidity, and asset allocation—empirical variation across firms should reflect governance configurations, not just AI adoption.
• Incentives and investment in governance:
  • Firms face potential underinvestment externality: benefits of accurate reporting accrue to investors but costs of governance (model validation, audit trails, skilled reviewers) fall on firms. Policy or regulation may be needed to internalize these externalities.
  • Governance variables are complements/substitutes—optimal firm-level investment is a configurational decision; simple one-dimensional measures of AI adoption will misstate economic effects.
• Audit market and costs:
  • AI shifts auditor demand for skills (data science, model validation) and may alter audit pricing and scope. Auditors unable to evaluate model governance create market frictions and potential assurance gaps.
• Regulatory and disclosure policy:
  • Standardized disclosure of material AI uses and model risk (drawing on prudential model risk frameworks) can improve market discipline and comparability; the EU AI Act provides a regulatory direction that economists should model for compliance costs and effects on competition/entry.
• Measurement and empirical research agenda:
  • Empirical tests should move beyond binary AI-adoption indicators to measure governance maturity and the four mediating variables, and examine interactions (configurational analysis, e.g., fsQCA or interaction terms in regressions).
  • Natural experiments (regulatory changes, auditor interventions, high-profile model failures) and differences across application types (transaction-level vs. estimation vs. narrative) are promising identification strategies.
  • Outcomes to study: accrual-based and broader FRQ proxies (earnings persistence, predictability, timeliness), cost of capital, audit fees, misstatement incidence, and market reactions to AI-related disclosures.
• Systemic risk and externalities:
  • Widespread use of similar models and shared training data could produce correlated model failures across firms, raising systemic reporting risk—important for macroprudential consideration.
• Policy trade-offs:
  • Regulators must balance disclosure/validation requirements (improving trust and FRQ) against compliance costs and potential stifling of beneficial adoption—economists can quantify welfare trade-offs and optimal regulatory design.

Suggested short research questions for AI economics: - How does governance maturity moderate the effect of AI adoption on cost of capital? - Do firms that invest in the four governance dimensions realize persistent FRQ gains relative to AI adopters with weak governance? - What is the market pricing of disclosed material AI use and model-risk events?