Large Language Models, Financial Reporting, and Sarbanes–Oxley Compliance

Abstract

The rapid adoption of Large Language Models (LLMs) across enterprise environments has generated substantial interest in applying generative artificial intelligence to accounting, financial analysis, regulatory reporting, audit support, and investor communications (p. 1). Yet the probabilistic architecture underlying modern LLM systems introduces serious governance and compliance risks when such systems are used directly for financial calculations, financial statement preparation, or regulated disclosure activities (p. 1). This paper examines the structural incompatibility between monolithic LLM architectures and the internal control requirements imposed by the Sarbanes–Oxley Act of 2002 (SOX), particularly Sections 302 and 404 governing executive certification and internal controls over financial reporting (ICFR) (p. 1). The paper argues that conventional end-to-end LLM deployment models create material auditability, explainability, reproducibility, and fiduciary liability concerns because they operate as probabilistic black-box systems lacking deterministic computational traceability (p. 1). In contrast, a hybrid architecture that separates deterministic financial computation from language-generation functionality offers a more compliant pathway for enterprise deployment (p. 1). Under such an approach, financial calculations are executed through auditable rule-based systems while language models are restricted to narrative synthesis and contextual explanation (p. 1). The paper further examines emerging regulatory guidance from the SEC, PCAOB, NIST, COSO, and global AI governance frameworks, and provides illustrative examples of AI-related reporting failures and governance breakdowns (p. 1). The analysis concludes that governance architecture—not merely model accuracy—will determine whether generative AI systems can be safely integrated into regulated financial reporting ecosystems (p. 1).

1. Introduction

Generative artificial intelligence has moved rapidly from experimental deployment into mainstream enterprise adoption (p. 1). Financial institutions, public corporations, accounting firms, and consulting organizations are increasingly evaluating Large Language Models (LLMs) for applications involving earnings analysis, management reporting, forecasting, disclosure drafting, compliance documentation, and financial planning support (p. 1). The attraction is understandable: modern LLMs can summarize complex documents, generate narrative explanations, automate repetitive writing tasks, and synthesize large volumes of information at unprecedented speed (p. 1).

However, the operational logic of LLMs differs fundamentally from traditional enterprise financial systems (p. 1). Conventional accounting systems are deterministic (p. 1). Given the same inputs, they are expected to produce identical outputs repeatedly and consistently (p. 2). Financial reporting infrastructure therefore depends on traceability, reproducibility, verification, auditability, and explicit computational logic (p. 2). By contrast, transformer-based LLMs are fundamentally probabilistic systems trained on statistical relationships within massive corpora of text (p. 2). Their outputs are generated through high-dimensional prediction mechanisms rather than explicit rule-based reasoning (p. 2).

This distinction becomes critically important within regulated corporate reporting environments governed by the Sarbanes–Oxley Act of 2002 (SOX) (p. 2). Following the accounting scandals involving Enron, WorldCom, Tyco, Adelphia, and Arthur Andersen, SOX established strict executive accountability and internal control requirements for public companies in the United States (p. 2). Sections 302 and 404, in particular, require that corporate executives certify the reliability of financial disclosures and maintain effective internal controls over financial reporting (ICFR) (p. 2).

The central argument of this paper is that directly embedding monolithic LLMs into financial calculation or reporting pipelines creates structural conflicts with the auditability and control requirements underlying SOX and COSO governance frameworks (p. 2). While LLMs may provide substantial value in drafting assistance, summarization, or decision support, their use as primary computational engines for regulated financial calculations introduces severe control vulnerabilities (p. 2). The paper further argues that a hybrid architecture—combining deterministic computational engines with constrained language-generation systems—provides a more viable governance framework for enterprise deployment (p. 2). In such architectures, numerical calculations remain under deterministic control while language models operate only within bounded narrative-generation functions (p. 2).

2. Sarbanes–Oxley, COSO, and Internal Control Requirements

2.1 SOX Sections 302 and 404

The Sarbanes–Oxley Act of 2002 fundamentally transformed corporate governance and financial accountability in the United States (p. 2). Section 302 requires chief executive officers and chief financial officers to personally certify the accuracy and completeness of financial reports filed with the Securities and Exchange Commission (SEC) (p. 2). Executives must attest that the reports fairly present the financial condition and operational results of the organization (p. 2). Section 404 requires management and external auditors to evaluate and report on the effectiveness of internal controls over financial reporting (ICFR) (p. 2). Organizations must demonstrate that financial information is generated through systems capable of ensuring accuracy, completeness, consistency, authorization, segregation of duties, and audit traceability (p. 2). The Public Company Accounting Oversight Board (PCAOB) further reinforced these requirements through auditing standards emphasizing evidence, control testing, process documentation, and reproducibility (PCAOB AS 2201) (p. 3).

2.2 COSO Internal Control Framework

The Committee of Sponsoring Organizations of the Treadway Commission (COSO) provides the dominant enterprise framework for internal control design and governance (p. 3). COSO emphasizes five core dimensions, which include the control environment, risk assessment, control activities, information and communication, and monitoring activities (p. 3). To ensure compliance within a financial reporting system, the environment must deliver transparent calculation logic, reliable and repeatable outputs, stringent change management controls, strict access controls along with a proper segregation of duties, comprehensive audit trails with robust evidence retention, and proactive monitoring paired with explicit exception handling (p. 3). These core structural principles align naturally with traditional enterprise resource planning (ERP) systems, accounting ledgers, and deterministic financial software, but they align far less naturally with generative AI architectures (p. 3).

3. Structural Characteristics of Large Language Models

Modern LLMs are based primarily on transformer architectures introduced by Vaswani et al. (2017) (p. 3). These systems operate by predicting token sequences using learned statistical relationships derived from large-scale training data (p. 3). Several structural characteristics create profound governance challenges within financial reporting environments (p. 3).

3.1 Probabilistic Output Generation

LLMs do not calculate outputs using deterministic accounting logic (p. 3). Instead, they generate outputs probabilistically based on learned token distributions (p. 3). Even when provided with identical prompts, system outputs may vary unpredictably based on sampling temperature settings, random seed variations, and context-window differences (pp. 3-4). Furthermore, fluctuations are introduced through retrieval augmentation differences, baseline model version changes, fine-tuning updates, and hidden infrastructure modifications (p. 4). This inherent variability directly conflicts with financial reporting requirements demanding absolute reproducibility and computational consistency (p. 4).

3.2 Black-Box Decision Structures

Transformer models contain billions of parameters interacting through complex neural weight matrices (p. 4). While model interpretability research continues to advance, it remains extremely difficult to fully reconstruct why a specific output was generated (p. 4). In regulated financial reporting, auditors typically require the ability to trace data origin, calculation methodologies, transformation logic, authorization chains, reconciliation procedures, and exception handling (p. 4). Traditional accounting systems provide this necessary traceability through structured logs, static formulas, transaction histories, and documented business rules (p. 4). Monolithic LLM systems generally cannot provide an equivalent level of computational transparency (p. 4).

3.3 Hallucination and Fabrication Risk

One of the most widely documented weaknesses of generative AI systems is hallucination, which is the confident generation of inaccurate or fabricated information (p. 4). Numerous public incidents illustrate the severe risks of this problem (p. 4). For instance, in 2023, attorneys using ChatGPT submitted fictitious legal citations in Mata v. Avianca, leading to direct court sanctions (p. 4). That same year, Google’s Bard AI demonstration produced an incorrect factual answer regarding the James Webb Space Telescope, contributing to a substantial market-value decline in Alphabet shares immediately following the public broadcast (p. 4). Furthermore, multiple financial institutions conducting internal tests have reported that general-purpose LLMs occasionally fabricate accounting references, ratios, assumptions, or source attributions when prompts are ambiguous (p. 4). In financial reporting contexts, such hallucinated numbers or unsupported assumptions could easily constitute material misstatements (p. 4).

3.4 Model Drift and Version Instability

Enterprise financial systems require absolute stability across reporting periods (p. 5). However, commercial LLM systems may evolve continuously due to vendor-side model updates, reinforcement learning adjustments, fine-tuning modifications, safety alignment changes, and background retrieval pipeline alterations (p. 5). Research has shown measurable performance shifts across successive model versions (p. 5). This creates severe governance concerns because the exact same prompt may produce entirely different outputs across quarters or audit cycles, undermining process consistency (p. 5).

4. The SOX Compliance Problem

4.1 Auditability and Internal Control Failure

Under SOX Section 404, enterprises must maintain effective internal controls capable of external audit verification (p. 5). A monolithic LLM-based financial reporting process introduces several foundational auditability failures (p. 5).

4.1.1 Lack of Explicit Computational Logic

Auditors generally expect the ability to inspect formulas, reconciliations, and rule structures underlying financial outputs (p. 5). In a neural transformer system, outputs emerge from distributed parameter interactions rather than explicit accounting formulas (p. 5). Consequently, calculation pathways cannot be fully reconstructed, output reasoning cannot be precisely validated, individual numerical decisions lack deterministic explanations, and independent replication becomes exceptionally difficult (p. 5). This baseline opacity fundamentally weakens the evidentiary basis necessary for ICFR certification (p. 5).

4.1.2 Inadequate Evidence Trails

Financial controls require durable audit trails documenting who initiated transactions, what calculations occurred, which rules were applied, when modifications were made, and which specific approvals were granted (pp. 5-6). Traditional ERP systems maintain these records naturally (p. 6). Conversely, LLMs often generate outputs dynamically without equivalent granular evidentiary logging, leaving a substantial gap in control validation (p. 6).

4.1.3 Control Environment Weaknesses

A core COSO principle requires management to maintain effective control environments (p. 6). Introducing a probabilistic black-box model into regulated financial calculations may create what auditors could classify as significant deficiencies, material weaknesses, an inadequate segregation of duties, or uncontrolled model risk (p. 6). Such adverse findings could materially affect external audit opinions and compromise the organization's regulatory standing (p. 6).

4.2 Executive Liability Under SOX Section 302

SOX Section 302 imposes personal certification responsibility on CEOs and CFOs, requiring executives to certify that disclosures are materially accurate (p. 6). The governance challenge is not merely whether an LLM is "usually accurate," but whether management can reasonably certify deterministic reliability, control completeness, process transparency, reproducibility, and the absolute absence of hidden manipulation (p. 6). If a generative AI system produces a materially inaccurate number—even on an infrequent basis—management could face severe SEC enforcement actions, shareholder litigation, restatement obligations, corporate reputational damage, and potential civil or criminal liability (p. 6). This risk exposure creates profound fiduciary concerns for corporate leadership (p. 6).

5. Emerging Regulatory and Governance Concerns

5.1 SEC Focus on AI-Related Disclosure Risk

The SEC has increasingly emphasized AI-related governance and disclosure integrity (p. 7). SEC Chair Gary Gensler has repeatedly warned that predictive analytics and AI systems can create conflicts of interest, systemic opacity, and broader market risks (p. 7). Reflecting this stance, the SEC has actively pursued enforcement actions involving misleading AI-related claims or inadequate controls surrounding automated systems (p. 7). Consequently, public companies utilizing artificial intelligence within financial reporting pipelines may face intense scrutiny concerning disclosure controls, model governance, data lineage, validation procedures, cybersecurity risks, and third-party vendor dependencies (p. 7).

5.2 PCAOB and Audit Implications

The PCAOB has heavily emphasized the importance of audit evidence reliability and management review controls (p. 7). If auditors cannot independently validate the underlying logic of AI-generated outputs, they may be legally unable to rely on those systems as part of substantive audit procedures (p. 7). This structural barrier could significantly increase overall audit costs, expand manual sampling requirements, demand extensive reconciliation efforts, heighten control testing intensity, and alter risk-adjusted materiality thresholds (p. 7).

5.3 NIST AI Risk Management Framework

The National Institute of Standards and Technology (NIST) AI Risk Management Framework (2023) highlights several core dimensions relevant to financial reporting systems (p. 7). These key dimensions encompass validity and reliability, safety and resilience, explainability and interpretability, accountability and transparency, and privacy and security (p. 7). These principles strongly reinforce the notion that AI governance requires systematic architectural controls rather than an arbitrary reliance on model confidence or statistical accuracy alone (p. 7).

5.4 International Regulatory Trends

Globally, regulators are moving toward stronger, binding AI governance frameworks (p. 8). Prominent examples include the European Union AI Act, the OECD AI Principles, the Bank of England and FCA AI governance consultations, and the Basel Committee guidance on model risk management (p. 8). These frameworks consistently emphasize the necessity of human oversight, operational transparency, explainability, explicit risk classification, and clear lines of governance accountability (p. 8). Because of their systemic impact, financial reporting applications are highly likely to fall into high-risk regulatory categories that demand the strictest possible governance controls (p. 8).

6. Illustrative Examples of AI-Related Financial and Governance Failures

6.1 Zillow’s Automated Forecasting Failure

Although not an LLM case specifically, Zillow’s algorithmic home-buying program (Zillow Offers) clearly demonstrated the dangers of overreliance on automated predictive systems (p. 8). The company incurred substantial financial losses and was forced to close the business unit after its forecasting models failed to accurately predict housing-market dynamics at scale (p. 8). The example illustrates a broader governance lesson: sophisticated predictive systems can fail catastrophically when operating outside strictly validated control boundaries (p. 8).

6.2 Air Canada Chatbot Liability Case

In 2024, Air Canada was held legally responsible by a tribunal for inaccurate refund information generated by its customer-service chatbot (p. 8). The company attempted to argue that the chatbot was a separate legal entity endowed with its own liability, a defense flatly rejected by the tribunal (p. 8). This case is legally important because it reinforces a core governance principle: organizations remain fully responsible for AI-generated outputs deployed within enterprise operations (p. 8).

6.3 Financial Calculation Hallucination Risks

Internal enterprise testing across major accounting and consulting firms has demonstrated that general-purpose LLMs exhibit critical functional failures (pp. 8-9). Specifically, they frequently misapply complex accounting standards, invent entirely unsupported financial assumptions, produce mathematically inconsistent internal calculations, confuse GAAP and IFRS treatments, and misstate basic units, percentages, or scaling factors (p. 9). Even relatively low hallucination frequencies are entirely unacceptable within regulated reporting environments where precision requirements are absolute and unforgiving (p. 9).

6.4 Samsung Data Leakage Incident

In 2023, Samsung restricted employee use of generative AI systems after sensitive proprietary source code and internal information were inadvertently uploaded into external AI tools (p. 9). This prominent example highlights additional governance concerns involving confidentiality, data leakage, vendor exposure, intellectual property protection, and cross-border data processing (p. 9). These underlying privacy issues intersect directly with the strict mandates governing regulated financial data handling (p. 9).

7. Hybrid Governance Architecture: Deterministic Computation with Constrained Language Generation

7.1 Conceptual Framework

A more compliant enterprise design separates deterministic computation from language generation entirely (p. 9). Under this proposed hybrid architecture, financial calculations are executed exclusively by deterministic engines, structured databases remain the unalterable system of record, business rules are explicitly codified, and calculation logic remains fully auditable (p. 9). Concurrently, AI systems operate only as bounded narrative interfaces (p. 9). The language model therefore becomes a controlled presentation layer rather than an authoritative computational authority (p. 9).

7.2 Deterministic Calculation Layer

The deterministic calculation layer may include SQL-based calculation engines, ERP accounting systems, rule-based financial logic services, auditable APIs, reconciliation engines, core ledger systems, and controlled reporting registries (p. 10). By confining calculations to this layer, every single financial output can be precisely reproduced, systematically logged, automatically reconciled, independently tested, and seamlessly reviewed by external auditors (p. 10).

7.3 Restricted Language Generation Layer

The language-generation component is architected to operate only after the underlying numerical values have been fully finalized by the deterministic layer (p. 10). For example, the system might output standard text templates such as: "The company reported quarterly revenue of and operating income of ." (p. 10) Because these variables are populated exclusively through deterministic systems, this architecture dramatically reduces the risk that the language model could invent financial values, alter calculations, misapply formulas, or introduce computational inconsistencies (p. 10).

7.4 Advantages of the Hybrid Model

The hybrid architecture offers several core governance advantages over monolithic models (p. 10). First, it ensures auditability, as all financial calculations remain completely traceable and reviewable back to source systems (p. 10). Second, it guarantees absolute repeatability, meaning identical data inputs will invariably produce identical financial outputs (p. 11). Third, it establishes a formal segregation of duties by logically isolating raw computational processing from subsequent narrative generation (p. 11). Fourth, it vastly reduces model risk, given that the AI system is structurally blocked from independently altering financial calculations (p. 11). Finally, it improves executive certifiability, empowering management to more credibly and safely certify the overall integrity of public financial reports (p. 11).

8. Broader Implications for Enterprise Governance

The ongoing corporate debate surrounding AI in finance is often framed primarily around statistical accuracy rates (p. 11). However, regulatory governance depends on more than simple statistical accuracy (p. 11). A system may appear highly accurate during testing while still failing baseline governance requirements because it lacks explainability, determinism, traceability, accountability, and reproducible control evidence (p. 11). This distinction is critical (p. 11). Financial regulation historically evolved around deterministic accounting infrastructures, yet generative AI introduces probabilistic reasoning systems into environments where legal accountability depends entirely on reproducibility and evidentiary traceability (p. 11). As a result, the central governance question facing organizations is architectural rather than merely algorithmic (p. 11).

The future of successful AI deployment in regulated finance will depend heavily on the institutionalization of human-in-the-loop oversight, layered governance controls, deterministic calculation segregation, formal model validation programs, independent auditability, dedicated enterprise AI governance committees, and continuous monitoring paired with real-time drift detection (pp. 11-12). Organizations that fail to establish these comprehensive architectural controls expose themselves to substantial operational, regulatory, legal, and reputational risks (p. 12).

9. Conclusion

Large Language Models offer significant opportunities for enterprise productivity, financial analysis support, and narrative automation (p. 12). However, their probabilistic architecture creates substantial conflicts with the deterministic governance requirements embedded within SOX, COSO, SEC oversight, and PCAOB auditing frameworks (p. 12). The primary problem is structural rather than incidental (p. 12). Monolithic LLM systems are not naturally aligned with the evidentiary, reproducibility, and auditability standards required for regulated financial reporting (p. 12). Their black-box characteristics, hallucination risks, model drift, and non-deterministic outputs create material internal control vulnerabilities (p. 12).

Accordingly, enterprises should strictly avoid allowing unconstrained LLM systems to function as authoritative computational engines within regulated financial workflows (p. 12). A more viable governance model separates deterministic financial computation from constrained AI-driven language generation (p. 12). Under such hybrid architectures, calculations remain fully auditable and reproducible while language models provide bounded narrative assistance (p. 12). Ultimately, enterprise adoption of generative AI in financial reporting will not depend solely on model sophistication (p. 12). It will depend on whether organizations can construct governance architectures capable of satisfying fiduciary accountability, regulatory transparency, and audit integrity requirements (p. 12).

References

• Air Canada v. Moffatt. 2024. BCCRT 149 (Civil Resolution Tribunal) (p. 12).
• Bommasani, R., D. A. Hudson, E. Adeli, et al. 2021. On the opportunities and risks of foundation models. Working paper, Stanford Center for Research on Foundation Models (p. 12).
• Chen, L., Y. Zou, S. Arora, et al. 2023. How is ChatGPT’s behavior changing over time? Working paper, arXiv preprint arXiv:2307.09009 (p. 12).
• Committee of Sponsoring Organizations of the Treadway Commission (COSO). 2013. Internal Control—Integrated Framework. New York, NY: COSO (p. 12).
• Lipton, Z. C. 2018. The mythos of model interpretability. Queue 16 (3): 31–57 (p. 12).
• National Institute of Standards and Technology (NIST). 2023. Artificial Intelligence Risk Management Framework (AI RMF 1.0). Washington, D.C.: U.S. Department of Commerce (p. 13).
• OpenAI. 2023. GPT-4 technical report. Working paper, arXiv preprint arXiv:2303.08774 (p. 13).
• Public Company Accounting Oversight Board (PCAOB). Auditing Standard AS 2201: An Audit of Internal Control Over Financial Reporting That Is Integrated with An Audit of Financial Statements. Washington, D.C.: PCAOB (p. 13).
• Sarbanes–Oxley Act of 2002. Pub. L. No. 107-204, 116 Stat. 745 (p. 13).
• U.S. Securities and Exchange Commission (SEC). Various public statements and enforcement guidance relating to artificial intelligence, disclosure controls, predictive analytics, and cybersecurity governance. Washington, D.C.: SEC (p. 13).
• Vaswani, A., N. Shazeer, N. Parmar, et al. 2017. Attention is all you need. Advances in Neural Information Processing Systems 30 (p. 13).
• Zillow Group, Inc. 2021. Shareholder letters and public disclosures regarding Zillow Offers operational shutdown. Seattle, WA: Zillow Group (p. 13).

---

(Your Name/Corresponding Author Name)
(Your Affiliation/University Department)
(Your Institution’s Address)
(Your Email Address)

May 27, 2026

Editorial Board
Accounting Horizons
American Accounting Association

RE: Manuscript Submission for Consideration in Accounting Horizons

Dear Editors,

I am pleased to submit our conceptual manuscript titled "Large Language Models, Financial Reporting, and Sarbanes–Oxley Compliance: Structural Risks, Regulatory Implications, and Governance Requirements" for consideration for publication as an original article in Accounting Horizons.

Consistent with the unique mission of Accounting Horizons to bridge academic research with professional accounting practice and regulatory policy, this paper addresses an immediate, top-of-mind vulnerability confronting the corporate financial ecosystem. The rapid enterprise adoption of Large Language Models (LLMs) for automation and narrative drafting has inadvertently outpaced corporate governance boundaries. This paper provides a rigorous analysis of the fundamental, structural mismatch between the probabilistic architecture of monolithic LLMs and the deterministic compliance mandates of the Sarbanes–Oxley Act of 2002 (SOX)—specifically Sections 302 and 404 governing executive certification and internal controls over financial reporting (ICFR) (p. 1). Rather than focusing solely on downstream algorithmic accuracy metrics, this paper argues that the central risk is architectural (p. 11). We establish the legal and financial compliance breakdowns that occur when black-box neural networks are introduced into audited calculation workflows (p. 1).

Crucially, to provide actionable value to practitioners, regulators, and auditors, the paper proposes a "Hybrid Governance Architecture" that cleanly segregates deterministic computation layers (such as standard ERP environments) from constrained, template-driven language generation layers (pp. 9, 12). This architecture preserves absolute auditability and computational reproducibility while still capturing the operational efficiency gains of generative AI (p. 12).

We believe this paper will be highly impactful to the readers of Accounting Horizons, including accounting researchers, corporate executives (CEOs/CFOs), auditors, and regulatory bodies like the SEC and PCAOB. It provides a timely framework to guide enterprise AI risk-management strategies under current regulatory scrutiny (p. 12).

This manuscript is original work, has not been published previously, and is not currently under consideration for publication elsewhere. All authors have reviewed the manuscript and approved its submission. Thank you for your time and consideration of our work.

Sincerely,

(Your Name)
(Your Title)
(Your Institution)