From Fraud Signals to Audit Response: A Human–AI Decision Governance Framework for Explainable Financial Statement Fraud Risk Assessment

This article has been Reviewed by the following groups

Read the full article

Abstract

Financial statement fraud research has progressed from rule-based and statistical screening toward machine learning and explainable artificial intelligence (XAI). Recent finance and auditing literature has also begun to frame AI as part of human–AI decision and governance systems rather than as a stand-alone predictive tool. Yet a narrower audit-specific problem remains unresolved: how a model-generated fraud-risk signal should be evaluated, challenged, and translated into a proportionate audit response when explanations may be unstable, error consequences are asymmetric, and audit resources are constrained. This conceptual article develops a Human–AI Decision Governance Framework for financial statement fraud risk assessment. Using theory synthesis and conceptual model development, the framework integrates six layers: fraud-risk signaling, explanation assurance, professional judgment, decision utility, audit response, and governance with feedback. It treats AI outputs as decision inputs rather than fraud conclusions and distinguishes model performance from explanation reliability, professional validity, and decision utility. Six propositions specify testable relationships among predictive performance, explanation quality, professional skepticism, resource-sensitive thresholds, human override, and lifecycle governance. The contribution is deliberately domain-specific: it connects fraud analytics and XAI to the auditor’s assessment and response process, including the risks of material misstatement due to fraud under ISA 240 (Revised), rather than claiming a generic theory of human–AI governance. The framework provides a structured research agenda for behavioral, archival, simulation, and field validation.

Article activity feed

  1. This Zenodo record is a permanently preserved version of a PREreview. You can view the complete PREreview at https://prereview.org/reviews/24269867.

    1. Does the introduction explain the objective of the research presented in the preprint?

    • Selection: Yes

    • Comments:

      Reviewer: Julian Rodriguez, Jr., FRSA, MRes, M.ISRM | ORCID: 0009-0007-9332-0140

      The introduction sets up the problem clearly and logically. It walks through how fraud detection evolved from basic financial ratios (Beneish, Dechow) to machine learning and explainable AI (XAI), and nails the core practical gap: getting a fraud score and a local explanation still does not tell an auditor how much to trust that explanation, how to document an override, or how much extra testing is justified. The three research questions and four contributions match that objective cleanly.

    2. Are the methods well-suited for this research?

    • Selection: Neither appropriate nor inappropriate

    • Comments:

      For a conceptual framework paper, pulling together audit standards, XAI reliability research, and human-factors literature makes sense, and keeping the scope focused specifically on financial statement fraud rather than generic AI governance is a smart choice.

      Three brief refinements would strengthen the design:

      • Justify the six-layer structure: Briefly explain why these specific six layers—and this sequence—were chosen over alternative groupings.

      • Align the Layer 4 equation with earlier layers: The expected-utility formula relies directly on the raw model probability (pip_i) without formally incorporating explanation assurance (Layer 2), epistemic risk (Rodriguez, 2026a), or professional judgment (Layer 3).

      • Add one worked example: Include a short, end-to-end case walkthrough (e.g., a revenue-recognition flag where SHAP and LIME conflict) to show how an engagement team moves through all six layers in practice.

    3. Are the conclusions supported by the data?

    • Selection: Somewhat supported

    • Comments:

      Because there is no empirical dataset here, the conclusions are conceptual arguments rather than tested results, and the author is upfront about that. The central takeaway—that XAI outputs have to be checked for stability and filtered through professional skepticism before changing audit procedures—is well backed by the cited literature on class-imbalance distortion and automation bias.

      Two claims need tighter grounding:

      1. Tie directly to specific ISA 240 / audit documentation rules: Instead of referencing ISA 240 (Revised) broadly, link Layers 2, 3, and 6 directly to concrete audit documentation and quality-control rules (such as how an engagement team documents contradictory evidence or escalates to specialists).

      2. Account for hidden false negatives in Layer 6: The feedback loop assumes audit outcomes can recalibrate the model over time. In real audits, you immediately see false positives when extra testing turns up nothing, but missed frauds (false negatives) often stay hidden for years until a restatement or enforcement action. The paper should note how governance avoids building false confidence from missing labels.

    4. Are the data presentations, including visualizations, well-suited to represent the data?

    • Selection: Somewhat appropriate and clear

    • Comments:

      The summary tables—especially the breakdown of failure modes, empirical measures, and follow-on research designs—are practical and easy to scan.

      A few quick cleanups would polish the presentation:

      • Introduce Figure 1 in the text before it appears, and make sure the placement of "human override" matches between the diagram (Layer 3) and the body text (Layer 6).

      • Fix duplicate variable symbols in the equations (for instance, using the same letter for both cost and context).

      • Clarify in the risk-response table that a "low" model fraud score never allows an auditor to skip mandatory baseline procedures required by auditing standards (like journal-entry testing).

    5. How clearly do the authors discuss, explain, and interpret their findings and potential next steps for the research?

    • Selection: Somewhat clearly

    • Comments:

      Breaking out separate takeaways for researchers, audit firms, regulators, and AI developers works well, and the proposed empirical tests give other researchers clear starting points.

      Three additions would make the discussion much sharper:

      1. Differentiate static financial-statement fraud from continuous transactional gating (KYC/KYT): Treating "financial fraud" solely through periodic financial-statement screening overlooks how global financial governance—including recent IMF and World Bank policy discussions shifting from static Know Your Customer (KYC) checks toward continuous Know Your Transaction (KYT) monitoring—relies on real-time human gating under high alert volume. Explicitly situating the framework's human-gating boundary within this broader shift toward continuous transaction-level scrutiny would strengthen its practical relevance.

      2. Reframe over reliance through the "Zero-Defect Paradox" rather than cognitive bias alone: The manuscript frames overreliance and under reliance primarily as individual cognitive heuristics (citing Parasuraman & Riley, 1997; Skitka et al., 2000; Dietvorst et al., 2015). In high-stakes financial compliance and auditing, however, the breakdown of human gating is driven by institutional incentives—specifically the Zero-Defect Paradox and Systemic Intent Shadows (Rodriguez, 2026b, 2026c). When regulators impose zero-tolerance penalties for a single missed fraud while rarely penalizing false positives, practitioners face severe compliance latency and administrative paralysis: rational auditors rubber-stamp noisy AI flags ("defensive over-auditing") rather than risk regulatory second-guessing on a human override. Incorporating institutional inspection liability into Layer 5 and Layer 6 would make the governance model far more realistic.

      3. Map the four types of validity to the six layers & add a 2026 comparison grid: Explicitly linking predictive, explanatory, professional, and decision validity across Layers 1–6—and adding a compact comparison table against the cited 2026 hybrid human–AI governance frameworks—would clarify the paper's distinct contribution at a glance.

      References

      • Rodriguez, J., Jr. (2026a). An Analytic Grammar for Diagnosing Epistemic Risk in Generative AI. Zenodo. https://doi.org/10.5281/zenodo.22780948

      • Rodriguez, J., Jr. (2026b). The Zero-Defect Paradox: Systemic Lags, Administrative Paralysis, and Compliance Latency in Automated Risk Governance. FAIEMA 2026 / Zenodo.

      • Rodriguez, J., Jr. (2026c). The Brights in the Fog: How AI-Driven Compliance Amplifies Systemic Intent Shadows (SIS). SSRN / ResearchGate.

    6. Is the preprint likely to advance academic knowledge?

    • Selection: Moderately likely

    • Comments:

      Yes—it tackles a real operational bottleneck just as firms are deploying XAI tools and preparing for ISA 240 (Revised). Treating "explanation assurance" as a mandatory checkpoint before an auditor relies on an XAI output is a genuinely useful idea. Cleaning up the Layer 4 utility equation, situating human gating within broader KYC/KYT and regulatory inspection dynamics, and adding a worked forensic walkthrough will make this a solid reference for both researchers and audit teams.

    7. Would it benefit from language editing?

    • Selection: No

    • Comments:

      The writing is clear, well-structured, and uses auditing and XAI terminology accurately throughout.

    8. Final Recommendation & Disclosures

    • Would you recommend this preprint to others? Yes, it's of high quality

    • Is it ready for attention from an editor, publisher or broader audience? Yes, after minor changes

    • Competing Interests: The author declares that they have no competing interests.

    Competing interests

    The author declares that they have no competing interests.

    Use of Artificial Intelligence (AI)

    The author declares that they used generative AI to come up with new ideas for their review.

  2. This Zenodo record is a permanently preserved version of a Structured PREreview. You can view the complete PREreview at https://prereview.org/reviews/24244231.

    Does the introduction explain the objective of the research presented in the preprint? Yes The introduction explains the objective clearly and in a logical order. 1. It first describes the background and the fact that fraud detection has moved from statistical models to machine learning, explainable AI (XAI), and decision-oriented evaluation (pp. 1–2). 2. It then names the problem. A fraud score with an explanation does not tell the auditor what to do next. The paper calls this the "prediction–decision–governance gap" (p. 2). 3. It states the aim directly, which is to develop a Human–AI Decision Governance Framework for fraud risk assessment. 4. It lists three research questions and four contributions.
    Are the methods well-suited for this research? Neither appropriate nor inappropriate This is a conceptual paper, so the methods section draws from existing literature and no data, sample, or statistical analysis is involved. What works well: 1. A conceptual design fits the aim of the paper, which is to connect several separate research areas and to propose ideas for later testing. 2. The paper follows a recognized approach for conceptual articles as recommended by Jaakkola (2020). 3. The author states openly that the literature was selected for a specific purpose and that the paper is not a systematic review. This honesty is good practice. 4. The author explicitly states boundary conditions, so readers know where the framework applies. What to Improve: 1. The steps from the literature to the six layers are described only briefly. The paper does not explain why there are six layers, why they are in this order, or what other designs were considered. 2. The equations do not match the framework. The decision formula uses only the model probability and leaves out explanation assurance and professional judgment: The expected-utility formula in Layer 4 uses only the model probability pᵢ. Explanation assurance (EAᵢ, Layer 2) and professional judgment (Jᵢ, Layer 3) do not appear in it. The formula therefore goes directly from probability to action. This is exactly the shortcut that the paper argues against, and it does not agree with the paper's own statement, which asserts that "pᵢ ≠ audit action" (p. 9 of the draft). 3. The literature review for RQ1 is short and selective. It gives much weight to two studies from one country and does not cover text-based fraud detection. 4. There is no worked example and no expert feedback to show that the framework can be applied. The overall scholarly contribution of this conceptual paper would be even more significant if the revision documents how the framework was derived, aligns the equations with the text, and adds a worked example.
    Are the conclusions supported by the data? Somewhat supported The paper has no empirical data, so the question is whether the conclusions follow from the literature and the argument. What is well supported: 1. The main conclusion is that an AI prediction is an input to professional judgment and not a replacement for it. This follows from the cited research on automation misuse, algorithm aversion, and professional skepticism. 2. The conclusion that an explanation is not automatically reliable is supported by the cited systematic review of fraud-XAI studies. 3. The paper is careful in its wording. It states that the propositions are not tested and that the framework still requires empirical validation. It does not report effects that were not observed. Where the conclusions go further than the paper shows: 1. The paper states that it is built on ISA 240 (Revised). However, it discusses the standard only in general terms. No specific requirement is cited or connected to a layer. 2. The conclusion that feedback from audit outcomes can improve the system assumes that fraud outcomes can be observed. In practice, only detected fraud becomes known. Missed fraud usually stays unknown.
    Are the data presentations, including visualizations, well-suited to represent the data? Somewhat appropriate and clear The paper contains no data. Its presentations are one figure and five tables that summarize ideas. What works well: 1. Table 1 gives a clear overview of the six stages of fraud assessment research. 2. Table 3 is especially useful. For each concept it gives a definition, a typical failure and a possible measure for future studies. 3. Tables 4 and 5 link each proposition to a suitable research design, which will help other researchers. 4. Figure 1 is simple and easy to read. The feedback arrow shows the lifecycle idea well. 5. Table 2 is correctly labelled as an illustration and not as a fixed rule. What should be improved: 1. Figure 1 is not mentioned anywhere in the main text, and it appears before the section that explains it. 2. The title of Figure 1 differs slightly from the framework name used in the text. The figure also places "human override" in Layer 3, while the text discusses it in Layer 6. 3. Table 1 has no citations for its stages. 4. Table 2 suggests that low risk needs only routine procedures. This should be checked against procedures that auditing standards require in every audit. 5. Some symbols are used twice. For example, C means both cost and context.
    How clearly do the authors discuss, explain, and interpret their findings and potential next steps for the research? Somewhat clearly What is clear: 1. The implications are organized for four groups of readers, including researchers, auditors and audit firms, regulators and standard setters, and AI developers. Each group receives specific points. 2. The limitations are stated honestly. The paper says the framework is not validated and that no universal threshold is proposed. 3. The next steps are concrete. Table 5 gives a research question, a design, a comparison and an outcome for each proposition. What is less clear: 1. The four types of validity are presented as a main contribution, but they are not connected to the six layers. 2. The next steps do not address how fraud outcomes will be observed, although the proposed field studies depend on this.
    Is the preprint likely to advance academic knowledge? Moderately likely Reasons for a positive view: 1. The research question is important and timely, because audit firms are adopting AI tools and a revised fraud standard is coming into effect. 2. The idea of checking the reliability of an explanation before an auditor relies on it is useful and can be studied in experiments. 3. The link between decision thresholds and graded audit responses connects two research areas that are usually separate. 4. The research agenda gives other researchers clear starting points. Reasons for the rating: 1. The author claims that several 2026 frameworks have already narrowed the gap. Most of the six layers have close relatives in that work, and the paper does not yet show the difference in a direct comparison.A comparison table would address this concern. The earlier frameworks could be placed in the rows and the framework elements in the columns, with marks showing what each one covers. 2. The two most original ideas are stated but not developed in detail. These are: (a) explanation assurance as a check before reliance, and (b) the link from decision utility to graded audit responses. These two ideas are the strongest material in the paper. 3. No empirical evidence or worked example is provided, so the practical value of the framework is not yet shown.
    Would it benefit from language editing? No The English is clear, fluent and consistent. Sentences are well formed and technical terms are used correctly. Professional language editing is not needed.
    Would you recommend this preprint to others? Yes, it's of high quality
    Is it ready for attention from an editor, publisher or broader audience? Yes, after minor changes

    Competing interests

    The author declares that they have no competing interests.

    Use of Artificial Intelligence (AI)

    The author declares that they did not use generative AI to come up with new ideas for their review.