Firms that disclose greater use of generative AI face higher audit bills: a large panel of Chinese listed companies shows GenAI-related disclosures predict statistically significant increases in audit fees, particularly for firms with high institutional ownership, low analyst coverage, or non‑Big Four auditors.

Main Finding

The paper finds that firm-level adoption (or strategic emphasis) of generative AI (GenAI) significantly increases external audit fees for Chinese A‑share listed companies (2015–2024). This result is robust to multiple specifications. The authors attribute the higher audit cost to increased audit risk and greater audit effort driven by: weakened internal controls, organizational flattening (greater centralization risk), and valuation difficulty from a shift toward high-dimensional digital/intangible assets.

Key Points

• Effect size: baseline panel estimates (firm and year fixed effects) show a positive and statistically significant coefficient on the GenAI text-based indicator (e.g., ~0.034, p<0.01; ~0.026 after controls in robustness checks), implying a measurable upward effect on log audit fees.
• Robustness: results hold after (a) replacing dependent/independent measures (TF‑IDF), (b) excluding IT sector firms, and (c) excluding firms in technology-active cities.
• Three proposed mediation channels:
• Internal control deterioration — GenAI’s opaque processing reduces traceability and weakens control effectiveness, forcing auditors into more substantive testing.
• Organizational flattening / centralization — GenAI reduces middle-management/administrative layers (lower management expense ratio), concentrating decision power and increasing risk of management override.
• Asset‑structure & valuation complexity — GenAI shifts investment toward computing, data, algorithms and other digital capital whose fair-value measurement is subjective, increasing valuation difficulty for auditors.
• Mechanism tests: GenAI is associated with (i) lower internal control index scores, (ii) lower management expense ratio, and (iii) changes in inefficient investment (authors interpret the latter as evidence of investment reallocation toward digital assets and away from traditional inefficient projects).
• Heterogeneity: the audit‑fee increase is larger in firms with high institutional ownership, low analyst attention, and when audits are done by non‑Big Four firms (interpreted as lower technical capacity to handle complex algorithmic assets).

Data & Methods

• Sample: 36,533 firm‑year observations from 5,032 A‑share listed firms (Shanghai & Shenzhen), 2015–2024; financial firms, ST firms, and incomplete observations excluded. Continuous vars winsorized at 1%/99%.
• Dependent variable: log(audit fees + 1).
• Main independent variable (GenAI): firm‑level frequency of GenAI‑related keywords in annual reports using a custom dictionary; transformed as log(freq + 1). Alternative measure: TF‑IDF.
• Mechanism/proxy variables:
  • Internal control: DIB Internal Control Index.
  • Organizational flattening proxy: management expense ratio (administrative expenses / operating income).
  • Asset/valuation channel: inefficient investment measured as absolute residuals from an investment model (Richardson 2006); interpreted alongside qualitative arguments about asset composition shifting to digital/intangible capital.
• Controls: firm size (log assets), Tobin’s Q, largest shareholder share (Top1), leverage, SOE indicator, plus standard firm and year fixed effects.
• Estimation: panel fixed‑effects regressions; multiple robustness checks and subgroup (heterogeneity) analyses.

Implications for AI Economics

• Adoption costs and governance externalities: GenAI adoption can raise firms’ external governance costs (audit fees) even when internal operational efficiency improves. Cost–benefit assessments of AI adoption should incorporate these externalized governance and verification costs.
• Audit market and skill demand: demand for IT‑audit and algorithmic valuation expertise will increase, potentially advantaging large audit firms with in-house technical capacity and raising barriers for smaller auditors; this may affect audit market structure and pricing.
• Corporate governance models: organizational flattening and greater centralization risk pose trade‑offs—productivity gains from automation vs. higher monitoring and assurance costs. Models of organizational design and AI adoption should incorporate these governance externalities.
• Regulatory policy and disclosure: regulators may need to refine disclosure requirements for algorithmic assets, model risk, and algorithmic provenance to reduce audit uncertainty and lower pricing frictions.
• Research agenda: researchers should incorporate audit‑cost/assurance frictions into theoretical and empirical models of firm AI adoption and diffusion; quantifying net welfare (productivity gains vs. higher assurance costs) and exploring general equilibrium effects (e.g., labor for auditors, audit firm specialization) are important next steps.

Caveats / limitations to note - Measurement: GenAI indicator is text‑based (keyword frequency) and may capture strategic discussion rather than actual operational deployment. - Endogeneity: potential reverse causality or omitted variables (e.g., unobserved managerial quality or strategic shifts) could bias estimates; the paper relies on robustness checks but does not report IV or quasi‑experimental identification. - Mechanism proxies are imperfect: the interpretation that reduced inefficient investment implies asset‑structure shift toward intangibles requires further direct measures (e.g., changes in intangible asset shares, capitalized software, data/algorithm valuations).

Suggested follow‑ups for researchers and policymakers - Use instrumental variables, event studies (e.g., sudden GenAI investments or disclosures), or firm-level rollout data to strengthen causal claims. - Directly measure intangible/digital asset shares and audit procedures deployed (IT specialists, hours) to link GenAI adoption more tightly to audit resource allocation. - Assess cross‑country patterns and the role of audit regulation and audit‑market concentration in mediating the observed effects.