The Commonplace
Home Papers Evidence Explore Syntheses Digests About 🎲 Workforce Futures
← Papers
Direction, evidence grade, and study type are AI-generated labels (gpt-5-mini), not human-verified. Syntheses are LLM-written. "Tensions" are machine-detected candidates, not confirmed contradictions. A research-acceleration tool, not peer review. How this is built →

AI investment has a U-shaped link to internal-control problems: modest AI spending appears to strengthen controls, but beyond a turning point higher AI investment is associated with more control deficiencies — a pattern strongest for software AI and softened by CIOs, IT-industry firms and greater audit scrutiny.

Too Much of a Good Thing? AI Investment and Internal Control Deficiency
Wentao Ma, Bingjie Deng, Xiaofan Chen, Wanyun Li · Fetched June 16, 2026 · Journal of the Association for Information Systems
openalex quasi_experimental medium evidence 7/10 relevance Source PDF
Using 41,725 Chinese firm-year observations, the paper finds a U-shaped relationship between AI investment and internal control deficiency risk: low-to-moderate AI investment reduces ICD risk while high AI investment increases it, particularly for software AI and absent strong governance (CIO, IT industry, audit attention).

This study examines whether artificial intelligence (AI) investment is associated with internal control deficiency (ICD) risk in a linear or regime-dependent way. Using 41,725 firm-year observations from Chinese A-share listed firms, we measure AI investment with capitalized AI-related assets identified from financial-statement footnotes. The results show a U-shaped association: AI investment is associated with lower ICD risk at lower levels of exposure but higher ICD risk beyond a turning point. Spline regressions, the Lind–Mehlum U-test, an instrumental-variable analysis using leave-one-out peer AI investment, and entropy balancing support the non-linear pattern. The pattern is concentrated in software-based AI applications rather than supporting hardware. Moderator tests further show that the increase in ICD risk at higher levels of AI investment is weaker among IT-industry firms, firms with CIO presence, and firms with above-normal external audit attention. The study highlights internal control reliability as an important condition for understanding AI-related organizational outcomes.

Summary

Main Finding

AI investment and internal control deficiency (ICD) risk exhibit a U-shaped relationship in Chinese A‑share firms: at low-to-moderate levels of AI investment ICD risk falls, but beyond a turning point higher AI exposure is associated with increased ICD risk. The non-linear pattern is robust to several econometric checks and is driven primarily by software-based AI applications rather than supporting hardware.

Key Points

  • Sample: 41,725 firm-year observations for Chinese A‑share listed firms.
  • AI measure: capitalized AI-related assets identified from financial‑statement footnotes.
  • Core result: a U-shaped association — protective effects of AI at lower exposure, harmful effects at higher exposure.
  • Robustness and identification: spline regressions, Lind–Mehlum U‑test for U‑shape, instrumental-variable (IV) analysis, and entropy balancing all support the pattern.
  • IV instrument: leave-one-out peer AI investment (to address endogeneity).
  • Heterogeneity: the increase in ICD risk at high AI investment is attenuated for:
    • IT-industry firms,
    • firms with a Chief Information Officer (CIO),
    • firms receiving above‑normal external audit attention.
  • Effect is concentrated in software-based AI applications, not in AI-supporting hardware.

Data & Methods

  • Data: hand-collected capitalized AI asset entries from footnotes; assembled into firm-year panel (41,725 observations).
  • Outcome: internal control deficiency (ICD) risk (incidence of internal control weaknesses/deficiencies).
  • Empirical strategy:
    • Flexible spline regressions to allow non-linearity.
    • Lind–Mehlum U‑test to formally test for a U-shaped relationship.
    • Instrumental-variable analysis using leave-one-out peer AI investment as an instrument for own AI investment to mitigate reverse causality/omitted variables.
    • Entropy balancing to achieve covariate balance between different AI-exposure groups.
    • Moderator (interaction) tests to probe institutional contingencies (industry, CIO presence, audit attention).
  • Robustness: multiple approaches converge on the non-linear finding; separate analyses for software vs hardware AI investments.

Implications for AI Economics

  • Non-linear returns/risks: AI adoption produces decreasing ICD risk at early stages but rising governance risk when exposure becomes large — implying marginal effects of AI investment are regime-dependent, not monotonic.
  • Complementarity with governance: strong internal governance (CIO role, industry expertise, high audit scrutiny) dampens the adverse governance consequences of high AI exposure — highlighting complementarities between AI and organizational institutions.
  • Investment strategy and risk management: firms should weigh not only productivity gains but also governance costs that escalate past an investment threshold; staged adoption and investment in governance/specialized human capital can shift the turning point.
  • Measurement and empirical practice: capitalized AI assets from footnotes can serve as a practical firm-level AI exposure measure for macro/firm-level analysis.
  • Policy and auditing: regulators and auditors should monitor firms with large AI footprints for internal control risks; audit attention and governance requirements could mitigate emergent ICD risk.
  • External validity & future research: results are based on Chinese A‑share firms — replication in other institutional contexts is needed. Future work should (a) quantify the turning point and its firm-level determinants, (b) unpack mechanisms (complexity, model risk, data/process integration), (c) analyze dynamic effects over adoption stages, and (d) study welfare and productivity trade-offs of high AI investment accounting for governance costs.

Assessment

Paper Typequasi_experimental Evidence Strengthmedium — Large sample and multiple complementary methods (spline tests, IV, entropy balancing, moderator checks) provide consistent support for a U-shaped association, increasing confidence that the pattern is not a simple artifact; however, causal interpretation depends on the validity of the leave-one-out peer instrument and on accurate measurement of AI investment from footnotes, both of which leave room for residual endogeneity and measurement concerns. Methods Rigormedium — The study applies a suite of appropriate empirical techniques (nonlinear specification tests, IV, balancing, heterogeneity checks) and a very large firm-year panel, which indicates careful empirical work; nonetheless key limitations remain—potential violations of the IV exclusion restriction, possible omitted variables that vary with AI diffusion, and reliance on capitalized AI assets extracted from footnotes that may incompletely capture AI activity—so methods are rigorous but not definitive. Sample41,725 firm-year observations from Chinese A-share listed firms; AI investment proxied by capitalized AI-related assets identified in financial-statement footnotes; analyses include splits by AI type (software vs hardware), industry (IT vs non-IT), presence of a CIO, and level of external audit attention; time period not specified in the summary. Themesgovernance org_design IdentificationObservational firm-year analysis using capitalized AI-related assets from financial-statement footnotes as the treatment measure; nonlinearity tested with spline regressions and the Lind–Mehlum U-test; robustness via entropy balancing; instrumental-variable strategy using leave-one-out peer AI investment as an instrument for firm AI investment; moderator analyses (software vs hardware AI, CIO presence, IT industry, audit attention). GeneralizabilitySample limited to Chinese A-share listed firms — results may not generalize to private firms, SMEs, or non-Chinese institutional contexts., Measure of AI investment based on capitalized assets in footnotes may omit uncapitalized R&D or service contracting and may vary in disclosure quality across firms/countries., Instrument (leave-one-out peer AI investment) could reflect common regional/industry shocks, limiting causal generalization., Findings concentrated in software-based AI applications, so conclusions may not transfer to hardware- or infrastructure-focused AI investments., Unclear timeframe; effects might differ under later waves of AI (e.g., generative AI) or different regulatory environments.

Claims (11)

ClaimDirectionOutcomeConfidence & EvidenceDetails
The study uses 41,725 firm-year observations from Chinese A-share listed firms. Regulatory Compliance null_result internal control deficiency (ICD) risk
Reading fidelity high
Study strength high
n=41725
0.8
AI investment is measured using capitalized AI-related assets identified from financial-statement footnotes. Adoption Rate null_result AI investment (capitalized AI-related assets)
Reading fidelity high
Study strength high
n=41725
0.8
There is a U-shaped association between AI investment and internal control deficiency (ICD) risk. Regulatory Compliance mixed internal control deficiency (ICD) risk
Reading fidelity high
Study strength high
n=41725
0.8
At lower levels of AI exposure, AI investment is associated with lower ICD risk. Regulatory Compliance negative internal control deficiency (ICD) risk
Reading fidelity high
Study strength high
n=41725
0.8
Beyond a turning point, higher AI investment is associated with higher ICD risk. Regulatory Compliance positive internal control deficiency (ICD) risk
Reading fidelity high
Study strength high
n=41725
0.8
Spline regressions, the Lind–Mehlum U-test, an instrumental-variable analysis using leave-one-out peer AI investment, and entropy balancing all support the non-linear (U-shaped) pattern. Regulatory Compliance mixed internal control deficiency (ICD) risk
Reading fidelity high
Study strength high
n=41725
0.8
The U-shaped pattern is concentrated in software-based AI applications rather than supporting hardware. Regulatory Compliance mixed internal control deficiency (ICD) risk
Reading fidelity high
Study strength medium
n=41725
0.48
The increase in ICD risk at higher levels of AI investment is weaker among IT-industry firms. Regulatory Compliance negative internal control deficiency (ICD) risk
Reading fidelity high
Study strength medium
n=41725
0.48
The increase in ICD risk at higher levels of AI investment is weaker among firms with CIO presence. Regulatory Compliance negative internal control deficiency (ICD) risk
Reading fidelity high
Study strength medium
n=41725
0.48
The increase in ICD risk at higher levels of AI investment is weaker among firms with above-normal external audit attention. Regulatory Compliance negative internal control deficiency (ICD) risk
Reading fidelity high
Study strength medium
n=41725
0.48
Internal control reliability is an important condition for understanding AI-related organizational outcomes. Regulatory Compliance null_result internal control reliability as moderator of AI outcomes
Reading fidelity high
Study strength speculative
0.08

Notes