What accountants need to know about artificial intelligence and machine learning: a review and call for future research Open Access

Artificial intelligence (AI) and machine learning (ML) are becoming core infrastructures of contemporary society (Kühl et al., 2022), used by individuals to search for information, consume news, interact with institutions and make economic decisions. Governments now deploy algorithmic systems extensively across the public sector, including in welfare administration (Zajko, 2023), predictive policing (Ferguson, 2017) and immigration control (Beduschi, 2021). Firms embed predictive models into marketing, logistics and pricing, and platforms rely on AI to curate content and match buyers with sellers (Bucher, 2018). Across these settings, the use of AI and ML has significant implications for efficiency, fairness and trust and raises questions about who, or what, is a credible source of knowledge in domains traditionally governed by expert judgement.

Public debates about AI therefore centre as much on epistemic authority, accountability and governance as on accuracy or automation: when algorithmic systems outperform human experts, whose judgement should prevail? How are errors justified, and where does responsibility reside? Accounting is an interesting context in which to examine these issues, given its core activities – measurement, recognition, analysis, assurance and compliance – are tightly coupled to evidentiary standards, regulatory oversight and professional accountability. AI and ML are increasingly embedded in accountants' work (Kommunuri, 2022). Audit firms deploy ML for engagement risk screening and anomaly detection, while corporations integrate predictive analytics into budgeting, forecasting and cost management (Kokina et al., 2025). Regulators and credit rating agencies likewise rely on algorithmic risk scores to identify high-risk filers. At the process level, tasks such as invoice coding, contract review and tax compliance are now routinely supported by AI-enabled automation (Pavlovic et al., 2024). More recently, generative AI has become a prominent tool in corporate communication, though its rapid adoption has not been without controversy.

In parallel, accounting research has experienced a methodological shift, with research across auditing, financial reporting, management accounting, taxation, financial distress, financial analysis, earnings forecasting, fraud detection, accounting misstatements and other areas adopting ML techniques to revisit long-standing empirical problems (Jones et al., 2023). Many studies document that ML models can uncover richer predictive structure, handle nonlinear interactions, exploit textual and unstructured data, and deliver substantially higher out-of-sample accuracy than traditional linear models. These methodological developments open new research questions and yield empirical findings with significant implications for professional judgement and authority.

At the centre of this paper is a simple but consequential proposition: AI and ML do not merely automate accounting tasks; they alter the epistemic foundations on which accounting judgements rest. By generating probabilistic assessments, uncovering patterns that elude human cognition, and, in some contexts, surpassing human performance in predicting corporate failure or detecting fraud (Rajaratnum et al., 2025), ML systems introduce a competing source of expertise into domains traditionally governed by professional judgement. Abbott's (1988) theory of the system of professions helps clarify why this shift is significant. Professional authority is sustained not by exclusive control over tasks but by control over legitimate chains of diagnosis and inference. Viewed through this lens, AI and ML function as exogenous epistemic technologies that reconfigure how evidence is produced, interpreted and validated, thereby challenging the profession's jurisdictional claims.

Against this backdrop, we make three contributions. First, we offer a conceptual exposition of AI and ML that links core model architectures to the epistemic foundations of accounting inquiry, showing how algorithms generate and structure evidence as emerging forms of knowledge production. Second, we present an integrated synthesis of research focused on the application of AI and ML across auditing, financial reporting, taxation, ESG measurement and public-sector analytics, highlighting where AI and ML outperform traditional statistical approaches and how methodological choices shape empirical findings. This cross-domain perspective reveals broader patterns in how ML influences risk assessment, misstatement prediction and evidence formation. Third, consistent with a human intelligence augmentation perspective that frames AI as complementary to human judgement (United Nations Development Programme, 2025), we introduce hybrid intelligence accounting to describe an emergent paradigm in which accounting judgement, evidence construction and decision processes are jointly produced by human expertise and ML systems.

The remainder of this paper is structured as follows. Section 2 outlines the foundational concepts of AI and ML, clarifying their epistemic logic and implications for accounting judgement and evidence. Section 3 synthesises the rapidly expanding literature on AI and ML across auditing, financial reporting, management accounting, taxation, ESG measurement, public-sector analytics and related domains, highlighting common methodological patterns and empirical insights. Section 4 examines how these developments reshape the accounting profession's jurisdiction, identity and claims to expertise, drawing on theories of professional authority and boundary work. Section 5 develops a theoretically grounded agenda for future research in an AI-enabled accounting environment. The final section reflects on the broader implications of AI and ML for the future of accounting expertise, professional judgement and institutional legitimacy.

Understanding how AI and ML reshape accounting judgement requires clarity about how accounting knowledge and evidence are formed in the first place. In traditional accounting analysis, judgement is grounded in models that reflect prior theory, professional standards and human assumptions about which relationships matter and how they should be specified. Evidence is generated by applying these predefined structures to data and interpreting the resulting outputs. ML represents a fundamental departure from this logic. Rather than imposing relationships in advance, ML systems infer patterns directly from data, learning which variables matter, how they interact and how strongly they predict outcomes. The credibility of these inferences is assessed not by their conformity with theory or interpretability alone, but by their ability to perform accurately when applied to new, unseen data.

This shift, from theory-driven specification to data-driven inference, matters for accounting expertise because it changes who, or what, is seen as a reliable source of judgement. When ML systems outperform human experts or traditional models in domains such as predicting fraud, misstatements or corporate failure, they challenge long-standing assumptions about how accounting knowledge should be produced, justified and authorised.

ML assumes that economic and organisational behaviour generates levels of complexity that simple models or unaided human judgement struggle to capture. Accounting data embody this complexity, combining features such as financial ratios, governance characteristics, journal-entry patterns, transaction flows and narrative disclosures that can interact in nonlinear ways. ML accommodates such complexity through iterative optimisation, producing predictive mappings shaped by the data rather than by accountants' ex ante judgements about which relationships should matter. In settings such as fraud detection or accounting misstatement prediction, these models often uncover combinations of signals that human experts or traditional methods overlook.

ML treats prediction as empirical discovery rather than modelling with predetermined functional forms. Supervised learning, the predominant approach in accounting research, trains on labelled outcomes – such as fraud, bankruptcy, restatements or going-concern classifications – to identify distinguishing patterns. Techniques such as gradient boosting, random forests, AdaBoost and neural networks (NNs) differ in representation but share the premise that predictive structure is inferred from data (Hastie et al., 2009). Table 1 provides a comparison of alternative machine-learning techniques and highlights how they differ in methodology, predictive performance, interpretability and suitability for various data environments and accounting problems.

In contrast to supervised learning, unsupervised learning operates without predefined outcomes or labels, allowing models to explore the structure of data itself. In accounting settings, unsupervised learning is commonly applied to explore data for unexpected patterns, such as anomalies, clusters and irregular journal-entry behaviour that merit further professional investigation.

Recent advances in generative AI and large language models (LLMs) extend this approach to narrative data by converting unstructured text into numerical representations that can be used for prediction or classification, thereby expanding the evidentiary base of accounting. More broadly, generative AI – a specialised subset of ML – builds on these representations to autonomously produce content that reflects the statistical and semantic structures embedded in its training data.

Systems such as ChatGPT and Claude use deep learning architectures and advanced natural language processing (NLP) to generate coherent text, synthesise information and sustain interactive dialogue. Unlike conventional ML models that focus on classification or pattern detection, generative AI produces narrative explanations and context-sensitive outputs that approximate human reasoning (see Table 1). These capabilities are already reshaping practice within professional service firms, where they are used to draft reports, summarise technical guidance, support advisory work and augment analytical tasks. Their rapid adoption, however, is tempered by concerns about reliability, transparency and implications for professional judgement.

These developments point to a broader shift in how accounting judgements are formed and justified. This shift is also visible in accounting research, where the epistemic orientation of ML differs from that of the statistical models traditionally employed. Traditional econometrics used in accounting research impose theory-driven structures to ensure interpretability and causal mapping, but these assumptions often fail in irregular or strategically evolving reporting environments. ML suspends such constraints, allowing relationships to emerge from data. Model quality is assessed primarily through out-of-sample predictive performance and interpretive diagnostics such as variable-importance scores or partial dependence plots (Jones, 2017). Empirical research shows that ML models often outperform traditional approaches because they better accommodate nonlinearities, interactions and distributional shifts characterising accounting and auditing contexts (Bertomeu, 2020).

Because these models rely on flexibility rather than predefined structure, they create new demands for how evidence is evaluated in accounting and auditing. One consequence of this flexibility is heightened sensitivity to data quality. As shown in Table 1, label noise, missing values, economically irrelevant correlations or mechanical artefacts can undermine model generalisability. This sensitivity is amplified in dynamic reporting environments, where changes in accounting standards, reporting practices or macroeconomic conditions can further erode model performance over time. Evaluation challenges are compounded by the opacity of many modern ML models, as post hoc interpretability tools do not always yield stable or economically coherent explanations (Hastie et al., 2009). For professionals accustomed to linking analytical results directly to financial theory, these features necessitate a shift in evaluative practice – from asking why a coefficient is significant to assessing model stability, the defensibility of underlying assumptions and the conditions under which model outputs constitute reliable accounting evidence.

Recent methodological advances, including interpretable ML (Molnar, 2022), drift monitoring, enhanced documentation and improved data governance, partially address these concerns but highlight a broader point. ML constitutes a distinct mode of producing and validating knowledge within accounting, extending the evidentiary base and generating probabilistic assessments rather than deterministic classifications. It shifts evaluative criteria from structural interpretation towards robustness, stability and governance of AI systems. Understanding these conceptual features is essential for interpreting the literature review in Section 3.

To develop a comprehensive and conceptually grounded review, we focus on high-quality peer-reviewed research published in ABDC A and A* journals. However, because AI and ML evolve rapidly across computer science, information systems, economics and professional practice, we also incorporated selected influential studies outside the ABDC list, including SSRN working papers, professional reports and reputable non-ABDC journals. These sources were included only when they offered new conceptual, technical or empirical insights. This balanced strategy ensures the review represents leading accounting scholarship and cutting-edge interdisciplinary developments.

The review shows that ML does more than deliver incremental performance gains over classical statistical models. Instead, it alters the evidentiary foundations of accounting and auditing by enabling the systematic use of large-scale numerical, textual and behavioural data that were previously inaccessible or difficult to analyse. As a result, ML reshapes what is treated as relevant accounting evidence, how disparate sources of evidence are combined and whose judgement (human or algorithmic) carries authority in accounting decision-making. This integrated review therefore provides the empirical foundation for analysing how ML reshapes the informational basis of expert work and challenges established forms of epistemic authority in the accounting profession.

Auditing and big data. Auditing is one of the earliest adopters of AI applications. Early work on expert systems and rule-based algorithms (Baldwin et al., 2006) and on continuous auditing (Alles et al., 2008) anticipated the shift towards automated evidence gathering and exception-based procedures. Continuous auditing established the technological and conceptual groundwork for automated, near real-time assurance; while ML extends this foundation by enabling predictive risk assessment, adaptive anomaly detection and continuous analytical review. As these tools mature, auditors' work shifts from executing tests to governing, validating and interpreting intelligent systems. As client data environments grow in scale and complexity, traditional analytical procedures struggle to extract meaning from unstructured and high dimensional data.

Various conceptual studies posit that big data analytics can eventually alter the evidentiary structure of auditing by enabling (as examples) population level testing, granular anomaly detection and real-time risk assessment (Cao et al., 2015; Warren et al., 2015; Appelbaum et al., 2017). In practice, big data almost inevitably involves ML, as only ML-based methods can learn the complex patterns embedded in massive, fast-moving datasets. Despite this, Gepp et al. (2018) show that auditing lags other fields such as financial distress prediction, fraud detection, and quantitative finance in adopting these techniques, and call for stronger research–practice alignment and future work on real-time analytics and collaborative platforms.

Auditor interaction with intelligent systems. A second wave of research examines how auditors interact with intelligent systems. For instance, Brown-Liburd et al. (2015) argue that big data can strain auditors' cognitive capacities, creating information overload and ambiguity that affect judgement. They also highlight practical challenges in integrating data analytics into audit processes. Similarly, Earley (2015) observes that while data analytics promises gains in testing coverage, fraud detection and audit quality; audit firms face substantial implementation obstacles including training gaps, data quality issues and evolving regulatory expectations.

Field evidence from Kokina et al. (2025) shows that large audit firms have widely adopted basic AI tools, such as optical character recognition and data extraction, but make more limited use of advanced ML applications. Their interview evidence points to persistent concerns about explainability, bias, privacy and overreliance, indicating a need for clearer safeguards and governance mechanisms in the use of AI. Zhang et al. (2022) also identify explainability as a central barrier to AI utilisation and show that post hoc interpretability techniques such as LIME and SHAP, which attribute model outputs to individual input features, enhance the transparency of AI-driven risk assessments. Behavioural dynamics further complicate adoption. Commerford et al. (2022) show that auditors exhibit algorithm aversion and are less willing to rely on AI-generated evidence than on human expert judgements. Together, these behavioural tendencies intersect with regulatory and standard-setting challenges, and an emerging stream of research therefore examines the implications of machine learning for audit regulation and professional standards.

Hope et al. (2025) find that convergence with International Standards on Auditing improves audit quality, with ML text analysis revealing which regulatory areas drive this improvement. Jiang (2024) shows that ML predicts cybersecurity breaches more accurately than logistic regression, and Küster et al. (2025) demonstrate that the semantic content of goodwill-related key audit matters (KAMs) predicts future impairments. Hunt et al. (2021) use ML to identify latent auditor switching risk, with higher predicted risk associated with more misstatements and abnormal accruals. Rajaratnum et al. (2025) show that gradient boosting and random forests can replicate auditor going-concern judgements with high precision, often outperforming auditors in predicting bankruptcies. These findings indicate that ML encroaches on judgement domains considered central to auditing expertise.

The Impact of AI on labour. The impact of AI on labour has generated substantial interest, with mixed and sometimes conflicting assessments. For instance, Frey and Osborne (2017) argue that auditors and accountants face a high risk of automation from AI, a view echoed in media commentary on Big Four audit firms' large-scale AI initiatives (Kapoor, 2020). Providing systematic evidence, Fedyk et al. (2022) analyse more than 310,000 résumés across 36 audit firms and find that AI-related roles are disproportionately staffed by younger, technically trained and centrally organised employees, who are more likely to be male. They further conclude that AI investment tends to improve audit quality and reduce fees but may displace human auditors over longer time horizons.

In contrast, Law and Shen (2024) show that offices hiring AI-specialised employees increase auditor employment, particularly at junior and mid-levels, and that AI adoption is associated with more accurate going-concern and internal control opinions. Partner interviews suggest that AI reshapes skills and workflows rather than replacing auditors. Generative AI exemplifies this reconfiguration, automating routine analytical tasks while increasing the importance of higher-order skills such as judgement, interpretation and professional scepticism. For example, large language models such as ChatGPT can extract and summarise risk factors from large datasets, generating preliminary risk outlines for audit planning (Vasarhelyi et al., 2023). Audit partners report that AI adoption frees staff from lower-level tasks, allowing greater focus on judgement-intensive analyses (Fedyk et al., 2022).

Empirical evidence nevertheless indicates that these changes are accompanied by workforce adjustments. Firms investing more heavily in AI experience reductions in auditor headcount, with a one-standard-deviation increase in AI staffing associated with a 3.6% decline in auditors after three years and a 7.1% decline after four, concentrated primarily in junior roles (Fedyk et al., 2022). Despite these displacement effects, employees often perceive AI as complementary, enhancing and supporting rather than replacing human labour (Wang and Lu, 2025). Overall, the literature suggests that generative AI reshapes audit tasks rather than eliminating audit jobs, automating data-intensive work while shifting auditors towards higher-value judgement and oversight roles.

Taken together, the auditing and assurance literature demonstrates how ML alters the nature of audit evidence, the cognitive demands placed on auditors, and the division of labour between human judgement and machine-based inference. These changes bear directly on evolving patterns of professional expertise and authority in an AI-enabled audit environment.

Parallel work on financial reporting examines how AI and ML transform the reporting and quality of accounting information. For instance, Anantharaman et al. (2023) show that firms adopting AI produce higher quality financial reports, with lower discretionary accruals, a tighter link between accruals and cash flows and more accurate estimates. They also document spillovers to audit quality, providing causal evidence that ML can enhance both reporting and assurance outcomes.

Accounting estimates. These are a natural target for ML, given their centrality and susceptibility to human bias. Ding et al. (2020) show that managerial estimates in financial reporting are often noisy or biased, and that ML models generate more accurate loss reserve estimates than managers across most insurance lines. Becker and Schölzel (2025) show that ML models predict warranty provisions more accurately than human experts, particularly in relation to overstatements when estimates are aggregated. Interviews and interpretability analysis trace human errors to aggregation bias, anchoring to past costs and organisational frictions such as learning barriers, auditor preferences and strategic motivations. Chen et al. (2022) develop an ML and text mining framework that interprets the economic substance of transactions and automates classification of preferred stock as financial liabilities or equity, illustrating how AI can support important recognition decisions.

Fraud detection. Fraud detection sits at the intersection of reporting quality, auditing and enforcement. Ramzan and Lokanan (2025) review 88 fraud detection studies, documenting a clear shift from traditional statistical techniques to ML, driven by fraud complexity and the limitations of standard audit procedures. Early work such as Kirkos et al. (2007) shows that decision trees, neural networks (NNs) and Bayesian belief networks can detect fraudulent statements using financial ratios. Perols (2011) finds that logistic regression and support vector machines (SVMs) perform competitively across fraud ratios and cost settings, and Perols et al. (2017) show that preprocessing methods addressing fraud's rarity and high dimensionality improve accuracy by about 10%. Bao et al. (2020) develop an ensemble ML model that integrates accounting theory with raw accounting numbers and outperforms benchmark approaches including Dechow et al. (2011) and Cecchini et al. (2010). Xu et al. (2023) apply ML to corporate fraud in China using GONE-based variables and find that random forests perform best, with exposure-related factors the strongest predictors. Further, Zhu et al. (2025) report that textual risk disclosures in annual reports significantly improve the early detection of accounting fraud, outperforming traditional MD&A analysis and their effectiveness is strengthened by SEC-guided disclosure standards.

Accounting misstatements. A related stream of research examines the prediction and interpretation of accounting misstatements. Dechow et al. (2011) developed a structured model for predicting material misstatements, which subsequent studies extend using more flexible ML techniques such as boosting, random forests and neural networks. Building on this work, Bertomeu et al. (2021) apply ML to high-dimensional accounting, market, governance and audit data to detect and interpret patterns in material misstatements. They show that interactions between accounting variables and audit or market indicators are particularly informative, enabling the prediction of irregularities up to two years in advance.

Other studies focus on specific aspects of misstatement behaviour. Hayes and Boritz (2021) use ML and textual analysis to classify restatements by management intent based on announcement disclosures, achieving performance comparable to traditional methods without relying on predefined dictionaries. Hunt et al. (2022) show that ML-based estimates of misstatement risk closely align with auditors' own risk assessments and influence audit fees, auditor turnover and the behaviour of Big N auditors, even though overall audit quality does not differ markedly by auditor type. More recently, Pham et al. (2025) examine how big data analytics affect accounting manipulation, further illustrating the expanding role of ML in identifying and interpreting misstatement risk.

Recent research applies ML to sustainability reporting, ESG measurement and non-financial disclosure. Lim (2024) maps the growing intersection of ESG, AI and finance, showing that ML techniques increasingly underpin sustainability analysis. Kumar et al. (2025) provide a large-scale review of sustainable finance and identify major themes such as climate finance, impact investing and governance, highlighting the need for additional work on predictive analytics and data quality.

A smaller set of studies demonstrates how ML interrogates proprietary ESG metrics and disclosure quality. Del Vitto et al. (2023) reverse engineer Refinitiv ESG ratings, finding that both transparent and black-box models closely reproduce the scores, while also revealing substantial noise and inconsistent factor weightings. Frost et al. (2023) use gradient boosting and random forests to predict voluntary carbon reporting, showing that ML materially outperforms logistic regression, offering different interpretations of variable importance and exposing limitations in traditional methods.

Overall, the ESG literature shows that ML expands accounting's evidentiary reach into sustainability domains by extracting structure from complex, inconsistent and often opaque non-financial data. These tools reveal where disclosure quality varies, where proprietary ESG indicators may be unstable and how alternative data can refine sustainability measurement, while raising important questions about transparency, reliability and bias in ESG reporting.

Earnings forecasting and profitability prediction are amongst the most quantitatively intensive domains of accounting research and have seen significant ML adoption. Zarowin (2019) observes that modern ML methods hold considerable promise for earnings prediction, an observation increasingly borne out in recent literature. For instance, Anand et al. (2019) use random forests to forecast the sign of five profitability metrics, reporting success rates that outperform a random walk. Similarly, Chen et al. (2022) predict the sign of next period earnings changes using gradient boosting and random forests with detailed financial data. Binz et al. (2025) apply a deep learning model to Nissim and Penman's (2001) structural profitability framework to forecast return on capital employed, finding that ML outperforms linear models and random walk benchmarks (see also Easton et al., 2024; You and Cao, 2021).

Jones et al. (2023) show that ML models forecast next period profitability changes more accurately than regressions, as well as revealing nonlinearities and interactions among predictors. They show that ML methods confirm and extend Penman and Zhang's (2004) model, particularly the value of both DuPont components, while demonstrating that the original variables capture most explanatory power even in high-dimensional settings. Binz et al. (2025), Chen et al. (2022) and Easton et al. (2024) show that ML-based forecasts often outperform analysts and econometric models, although portfolio returns based on these signals are not consistently higher, perhaps suggesting that improved ML predictions are incorporated into prices and contribute to informational efficiency.

Management accounting research increasingly examines how ML supports internal decision-making, planning and performance management. Ranta et al. (2022) review ML's emerging role in management accounting and conclude that, while adoption remains in its formative stages, there are substantial opportunities to analyse textual data, convert unstructured information into new measures, improve predictions and deploy explainable AI in management accounting contexts. Complementing this perspective, Mahlendorf et al. (2023) map novel data sources, including text, crowdsourced ratings, video, geolocation and satellite imagery, to management accounting research questions, highlighting promising avenues for work on decision-making and control. Pavlovic et al. (2024) similarly argue that AI can enhance management accounting by improving predictive analytics, automating reporting and strengthening real-time decision support, while also emphasising challenges related to data quality and system integration. Zhang et al. (2026) further show that AI's growing sophistication is reshaping managerial accounting, creating both significant opportunities and new organisational challenges.

Empirical studies provide evidence of tangible performance gains from ML adoption. Kureljusic and Metz (2023), for example, evaluate ML algorithms for forecasting customer payment dates using more than one million invoices and find that neural networks deliver the most accurate predictions, enabling more proactive cash flow management. Ranta and Ylinen (2024) use social media data and ML to show that firms offering more generous family-related benefits exhibit higher employee satisfaction and stronger performance, with high-growth firms providing broader benefits and highly profitable firms offering more targeted packages. At the same time, Korhonen et al. (2021) highlight the limits of automation in management accounting. Studying an attempted automation project in manufacturing, they document that many tasks initially deemed programmable were better performed by humans, and that misclassifying complex, expertise-driven work as automatable can undermine automation initiatives. Taken together, these studies suggest that ML has clear value in prediction and optimisation, but that professional judgement remains essential in many management accounting activities.

Recent work also highlights the growing potential of generative AI for planning, forecasting and performance analysis, while underscoring the continuing need for human oversight. In a public-sector forecasting setting, Chung et al. (2025) show that combining ChatGPT outputs with expert judgement produces highly accurate revenue forecasts, whereas reliance on the LLM alone yields substantial errors. This evidence suggests that hybrid human–AI approaches can deliver viable, low-cost forecasting solutions for budget planning. Consistent with this view, Abbas (2025) finds that AI technologies, including LLMs, are automating many traditional budgeting and forecasting tasks, shifting the role of management accountants away from manual data processing towards interpretation and strategic analysis. In practice, generative models can support decision-making by providing scenario analyses, narrative explanations and variance commentaries; for example, ChatGPT can rapidly generate risk-responsive planning outlines that managers then refine (Vasarhelyi et al., 2023). Overall, the literature indicates that LLMs can enhance internal analytics by improving forecasting speed and insight, provided accountants apply professional judgement to validate and contextualise their outputs.

Financial distress and credit risk prediction. Financial distress prediction and credit risk modelling represent one of the most mature domains for ML in accounting. Early studies, such as Anandarajan and Anandarajan (1999), show that neural networks and support vector machines outperform logistic regression in bankruptcy classification. Reviewing 35 years of corporate failure research, Jones (2023) identifies two major developments: the adoption of increasingly sophisticated ML methods, including gradient boosting, random forests and deep learning; and the expansion of predictor sets beyond financial ratios to incorporate market signals, earnings management measures, governance attributes and auditor going-concern opinions (GCOs).

Building on these developments, Jones et al. (2015, 2017) provide systematic comparisons of advanced ML methods and conventional classifiers in the context of credit rating changes and corporate failure. Evaluating 16 statistical learning approaches, ranging from linear models such as logit, probit and LDA, through semiparametric methods including mixed models, multivariate adaptive regression splines (MARS), generalised additive models (GAM) and generalised lasso, to fully non-parametric techniques such as gradient boosting and random forests, they find that classifiers capable of modelling nonlinearity and heterogeneity, particularly tree-based ensembles, deliver superior predictive performance. Subsequent studies report further gains from advanced ensemble methods (Barboza et al., 2017; Jones, 2017; Jiang and Jones, 2018), with similar advantages documented in credit rating transition prediction (Jones et al., 2015).

A more recent innovation extends this line of research by incorporating textual and audit-related information into bankruptcy and credit risk prediction. For example, Muñoz-Izquierdo et al. (2022) examine whether KAMs disclosed in extended audit reports help predict credit ratings, finding that identifying KAM topics alone yields approximately 74% accuracy, rising to 84% when combined with accounting ratios. Both external risks, such as going-concern issues and internal factors, such as debt-related KAMs, contribute meaningfully to explaining credit ratings.

Going-concern opinions. GCOs provide a rare setting where expert auditor judgements and objective outcomes are both observable. A growing body of work shows that ML can replicate or exceed auditors' performance in predicting failure, sometimes years in advance (Anandarajan and Anandarajan, 1999; Barboza et al., 2017; Hsu and Lee, 2020; Rajaratnum et al., 2025). Martens et al. (2008) use SVMs and rule-based classifiers, including AntMiner+, to generate interpretable decision rules for auditor going-concern assessment. Stanisic et al. (2019) evaluate 12 models for predicting auditor opinions, showing that when prior audit history is unavailable, tree-based ML models, especially random forests, outperform traditional techniques. Hedback (2025) shows that logistic regression and extreme gradient boosting can accurately identify going-concern modified reports and detect audit report pages within SEC filings.

Other work directly questions the informativeness of GCOs themselves, situating them within a broader predictive ecosystem. For instance, Wang et al. (2025) use extensive data and several ML models to test whether GCOs improve bankruptcy prediction, finding that they add little incremental value, even across auditor types and cost settings (see Rajaratnum et al. (2025) for more detailed analysis). This casts doubt on the traditional assumption that the auditor's opinion provides a distinct, decision-useful signal about failure risk. Evidence from regulatory changes reinforces this concern by showing how institutional framing shapes audit judgements. Using ML and logistic regression, Sánchez Medina et al. (2019) find that auditors became more willing to issue going-concern warnings following a Spanish reform that reduced the perceived severity of such classifications.

A parallel stream of work explores whether alternative information sources offer richer or earlier signals of distress than auditors. Condie and Moon (2025) show that a deep learning measure of social media “bearishness” predicts failure beyond traditional market signals and is associated with more accurate GCOs, implying that public sentiment captures emerging risks that auditors only partially incorporate. Similarly, Mayew et al. (2015) find that management's stated going-concern assessments and the linguistic tone of MD&A narratives are useful for predicting whether a firm will cease as a going-concern, providing incremental information over financial ratios, market variables and the auditor's opinion. Together, these studies suggest that auditors' going-concern judgements represent only one, and often not the most informative, predictive signal in a landscape increasingly populated by alternative data and ML-based models.

ML has become increasingly important in taxation, regulation and public sector analytics, where anomaly detection and the analysis of large administrative datasets are central. In taxation, Guenther et al. (2023) show that an ML model predicts next-year effective tax rates more accurately than rates implied by analysts' earnings forecasts. Using Shapley value–based explanations, they identify which financial statement and footnote disclosures are most informative and where analysts systematically misweight information. Similarly, Battaglini et al. (2025) apply ML to Italian tax return data and show that replacing the least promising audits with ML-selected cases increases detected evasion by up to 39% and collected evasion by 29%, indicating substantial efficiency gains. Consistent with this evidence, an OECD (2020) report documents how tax authorities increasingly deploy analytics, text mining and anomaly detection to support compliance monitoring.

Related applications appear in public sector accounting and performance measurement. Duan et al. (2023) demonstrate how social media data from Twitter and Facebook can be integrated into government accounting systems to construct alternative performance measures, illustrating this approach through an ML-based assessment of street cleanliness in New York City. Their study shows how ML can enrich public sector performance measurement and accountability by incorporating non-traditional data sources.

Building on these developments, generative AI also shows growing potential in taxation, regulation and public sector analytics. Prior audit research highlights that AI substantially improves fraud and anomaly detection in large datasets, enabling more effective risk assessment (Fedyk et al., 2022). By analogy, tax authorities could use LLMs to flag suspicious filings or generate risk scores for audit selection, while generative models may also assist with policy design and budget analysis. For example, Chung et al. (2025) show that LLMs can produce accurate government revenue forecasts when used judiciously. Although peer-reviewed research on generative AI in tax and regulation remains nascent, existing evidence suggests that these tools can complement human expertise in monitoring compliance and analysing administrative data (Fedyk et al., 2022), pointing towards a future in which AI supports regulators' and public officials' data-driven decision-making.

Text analysis and NLP use computational methods to extract and interpret information from unstructured text such as annual reports, earnings calls, analyst reports, news, social media, audit workpapers and regulatory filings. These techniques support applications including sentiment and tone measurement, deception detection, disclosure classification and assessments of clarity, complexity and latent themes that numerical data cannot capture. Loughran and McDonald (2016) identify five main methodological categories in accounting research that use text mining and NLP: targeted phrase extraction, sentiment analysis, topic modelling, document similarity and readability analysis. Textual cues from MD&A narratives, conference calls, analyst reports and news improve forecasts of stock returns, volatility, earnings, distress, credit risk, bankruptcy and corporate innovation (e.g. Li, 2010; Tennyson et al., 1990; Nousiainen et al., 2024; Liu et al., 2024).

NLP is also widely used to detect problematic corporate behaviour such as fraud and deception by identifying linguistic markers, tone shifts or obfuscation (Larcker and Zakolyukina, 2012). These techniques further support measurement of disclosure quality, transparency and boilerplate language (Brown and Tucker, 2011; De Franco et al., 2015). In auditing, NLP methods assist misstatement-risk assessment and help identify inconsistencies or anomalies in client disclosures. Sustainability research similarly illustrates NLP's growing relevance. Maibaum et al. (2024) show that LLMs outperform traditional text-mining methods in extracting sustainability-related information. This work underscores NLP as a methodological frontier where ML increasingly complements, and sometimes rivals, traditional professional judgement.

Three cross-cutting themes emerge from this literature review. First, ML-based models often outperform traditional statistical learning models on prediction tasks ranging from fraud detection, accounting misstatement and restatement prediction, financial distress forecasting, earnings prediction, operational planning, tax compliance monitoring and sustainability reporting. These benefits reflect ML's ability to capture nonlinear interactions, high-dimensional structures and complex features that conventional models cannot always accommodate, even when based on well-specified, theory-driven inputs.

Second, ML expands evidentiary boundaries. It enables auditors, managers, analysts, regulators and public sector entities to exploit large-scale numerical, textual, behavioural and alternative data, extending the evidentiary base of accounting beyond aggregated financial indicators and small sample judgement cues. By identifying latent patterns, semantic signals and relational structures in unstructured or high-volume data, ML uncovers signals that human judgement either overlooks or processes inconsistently. Examples include KAM narratives, MD&A tone, social media sentiment, cyber risk disclosures, ESG ratings and administrative records.

Third, ML has the potential to substantially reshape professional jurisdictions. As ML systems increasingly deliver superior predictive performance in domains historically governed by expert judgement – such as accounting estimates or GCOs – accounting professionals shift from being primary producers of judgement to supervisors of machine-generated inference. Expertise becomes centred on model design, interpretation, validation, explainability and governance, rather than on unaided predictive judgement. This reconfiguration echoes Abbott's (1988) account of jurisdictional realignments, in which new knowledge systems and technologies challenge established forms of professional authority.

The conceptual, empirical and practice-based developments outlined above have potentially profound implications for the accounting profession's identity, jurisdiction and authority. Accounting is widely recognised as highly exposed to automation risk, particularly in routine and rules-based work. Professional and consulting firm reports consistently identify functions such as accounts payable and receivable, reconciliations, expense processing and monthly closing as among the most automatable tasks (CPA Australia, 2019; ICAEW, 2022; IMA, 2024). While such narratives often frame automation as a technical efficiency challenge, they also reflect a deeper risk: that large portions of accounting work may be removed from professional control altogether.

A more nuanced view recognises that AI and ML are likely to reshape accounting work rather than eliminate the profession outright; however, this outcome is contingent, not guaranteed. AI systems already dominate routine activities such as transaction coding, reconciliations, anomaly detection and basic forecasting, reducing demand for clerical and process-oriented labour as machine-readable information enables automation of baseline roles or outsourced functions. Without deliberate jurisdictional repositioning, this process risks hollowing out the profession, leaving accountants confined to residual or compliance-oriented roles. At the same time, AI creates new domains requiring professional oversight, and industry reports argue that automation can shift work towards analytics, scenario modelling, data governance and the interpretation of machine-generated insights (EY, 2025; ICAEW, 2022; Institute of Management Accountants (IMA), 2024). Whether this shift strengthens or weakens the profession depends on whether accountants successfully claim authority over these emerging domains.

The profession's jurisdiction is therefore being actively contested and reconfigured. Research on professions shows that when expertise is threatened, professionals engage in boundary work to defend and extend their jurisdiction; however, success in these efforts is neither automatic nor guaranteed. Accountants have historically defended their jurisdiction in response to technological and regulatory incursions, including the expansion of consulting services (Greenwood and Suddaby, 2006) and the protection of audit territory (Radcliffe et al., 2018). Recent studies document similar strategies in response to digitalisation, including expanding into business-partner roles, guarding core work, collaborating across functions and bridging new domains (Wanderley and Horton, 2024). These strategies reflect an awareness that failure to adapt risks jurisdictional erosion rather than mere task reallocation.

These dynamics align with Suddaby and Viale's (2011) argument that professions simultaneously draw on established expertise while pursuing jurisdictional extension. Crucially, however, this process is inherently unstable: where professions fail to align new technologies with legitimate claims to expertise, technological change can erode or even displace professional jurisdictions. AI therefore poses a genuine threat to accounting's traditional domains, intensifying boundary work as accountants seek to reassert claims over financial expertise while negotiating emerging work territories shaped by algorithmic systems (Faulconbridge et al., 2023).

Importantly, the high predictive accuracy of ML models does not by itself confer institutional legitimacy or legal authority in professional domains such as auditing or going-concern assessment. Algorithms may replicate or even outperform human experts in predicting outcomes, yet technical performance alone does not grant the authority to issue sanctioned judgements. In regulated professions, legitimacy rests on socially sanctioned jurisdiction, accountability structures and legal mandates rather than predictive accuracy alone (Abbott, 1988; Freidson, 2001). However, these institutional protections are not immutable. If professional standards, education and governance fail to adapt, authority may migrate towards technology vendors, platform providers or regulators rather than remaining within the profession.

In auditing and assurance, authority is currently anchored in professional standards, statutory obligations and liability regimes, such that accountability attaches to licensed practitioners even when judgement is informed by algorithmic outputs. As Abbott (1988) emphasises, professional jurisdictions endure not through exclusive task performance but through control over legitimate chains of diagnosis and inference. A central mechanism through which this control is maintained is interpretive authority: the claim that accountants, rather than machines, are best positioned to interpret financial information. Despite AI's analytical sophistication, professionals frame AI as a generator of data and recommendations rather than a decision-maker, reinforcing interpretation, trust and ethical reasoning as core professional competencies (Issa et al., 2016).

Abbott (1988) and Freidson (2001) similarly identify the synthesis of ambiguous information as a defining professional skill. Radcliffe et al. (2025) illustrate this dynamic in tax advisory work, where practitioners invoke a “commitment to craftsmanship” to signal that expert judgement resists technological encroachment. Through such claims, accountants position their highest-value contributions as lying beyond the reach of algorithmic automation (Issa et al., 2016; Radcliffe et al., 2025). Yet these claims must be continuously reinforced through practice, standards and education if they are to remain credible.

At the same time, studies consistently find that AI generates new forms of work and services rather than simply displacing existing ones (Faulconbridge et al., 2023). Accountants increasingly shift towards advisory and analytical functions (Wanderley and Horton, 2024), with work organised through hybrid teams in which technologists conduct initial analyses and accountants exercise higher-level judgement. This reconfiguration reshapes governance structures, as technology vendors, platform providers and data owners gain influence over the design, training and deployment of AI systems used in accounting (Faulconbridge et al., 2023). Professional authority is therefore no longer contested solely within the profession but increasingly negotiated across organisational and technological boundaries.

These interactions between human expertise and machine capabilities can be clarified by mapping their respective contributions across phases of the accounting knowledge process. Table 2 summarises how machines and accountants contribute at each phase and how these shifting roles reshape professional jurisdiction.

Overall, AI's impact on the accounting profession is mediated through boundary work, interpretive authority and evolving role configurations. The profession is not inevitably displaced – but neither is its survival assured. Sustaining professional authority will require active jurisdictional maintenance, including redefining core competencies, embedding AI governance and interpretive responsibility within professional standards, and ensuring that education and certification align with hybrid human–machine expertise. As Faulconbridge et al. (2023) argue, AI is producing new forms of professional work; whether these remain under professional control is the central challenge facing the accounting profession.

Our review points to an urgent need to integrate AI and ML into accounting research not merely as novel analytical tools, but as epistemic technologies that reconfigure how accounting evidence and expertise is produced, how professional judgement is exercised and how assurance is legitimised. Beyond discrete technical adaptations, future research must interrogate how algorithmic systems unsettle foundational accounting constructs and reshape the profession's claims to epistemic authority. Consistent with Abbott's (1988) system-of-professions framework, these developments call for renewed attention to the chains of inference through which accounting knowledge is generated, validated and governed in an increasingly AI-mediated environment.

AI and ML are fundamentally expanding the evidentiary boundaries of accounting by enabling the use of large-scale numerical, textual, behavioural and alternative data in audit and reporting contexts. Advances in NLP allow systematic extraction of sentiment, semantic structure and forward-looking signals from disclosures (Bochkay et al., 2023), while ML models can process entire populations of transactions to detect misstatements and anomalies beyond the reach of traditional sampling-based procedures (Appelbaum et al., 2017). Recent studies show that continuously updating ML-based fraud and risk models outperform static benchmarks (Zhang and Zhou, 2026), reinforcing earlier arguments for continuous auditing and real-time assurance (Alles et al., 2008; Warren et al., 2015). These developments suggest a shift in audit practice from retrospective verification towards ongoing algorithmic inference, raising fundamental questions about what constitutes audit evidence in an algorithmic environment.

Future research should therefore move beyond integration questions and directly interrogate the epistemic status of algorithmically generated evidence. For instance, when audit conclusions rely on probabilistic model outputs rather than observable source documents, how should concepts such as evidence sufficiency, corroboration and reliability be redefined? How should materiality be reconceptualised when AI systems surface large numbers of statistically significant but individually immaterial anomalies? Building on prior work on audit risk and judgement, scholars can examine how traditional constructs travel (or fail to travel) into AI-enabled inference regimes. Research is also needed on how explainable AI techniques (e.g. SHAP, LIME) translate model outputs into auditable artefacts that satisfy regulatory expectations (Zhang et al., 2022). More broadly, this line of inquiry positions audit evidence as an evolving construct rather than a fixed technical category, with important implications for standard-setting and professional accountability.

A second research theme concerns how professional judgement and scepticism operate when accounting decisions are jointly produced by human expertise and algorithmic inference. Behavioural accounting research already demonstrates that auditors' reliance on decision aids depends on tool design, ownership and perceived accountability (Brown-Liburd et al., 2015; Commerford et al., 2022). Recent evidence shows that auditors who inherit analytics tools exhibit reduced scepticism and weaker follow-up on red flags, suggesting that AI may inadvertently erode professional vigilance if not carefully governed (Li et al., 2025). At the same time, other studies document algorithm aversion, where professionals discount superior model outputs in favour of human judgement, even when doing so reduces accuracy (Dietvorst et al., 2015; Commerford et al., 2022). Together, these findings indicate that hybrid intelligence environments generate new cognitive and organisational tensions around trust, responsibility and judgement.

Future research should therefore shift from asking whether professionals rely on AI to examining how accountability and scepticism are enacted when judgement is distributed across human–machine systems. What does professional scepticism mean when the object of scrutiny is a non-human agent whose reasoning is opaque, probabilistic and data-dependent? How is responsibility determined and allocated when an AI-informed judgement proves incorrect, particularly in high-stakes domains such as going-concern assessments or accounting estimates? (Ding et al., 2020; Rajaratnum et al., 2025). Drawing on theories of accountability, expertise, and professional identity (Abbott, 1988; Freidson, 2001), scholars can examine how accountants maintain interpretive authority by governing model selection, validation and use rather than by generating unaided judgements. This perspective also opens avenues for studying how accounting education, training and career socialisation evolve when core judgement skills increasingly involve interrogating and contextualising algorithmic outputs (Issa et al., 2016; Sutton et al., 2016).

A third theme concerns the implications of AI for assurance, professional jurisdiction and legitimacy. The increasing use of AI-enabled analytics challenges the traditional conception of reasonable assurance, which is grounded in sampling, professional judgement, and procedural compliance rather than continuous probabilistic inference. While AI systems promise more comprehensive risk detection and predictive accuracy, they also introduce new sources of uncertainty related to hallucination, model bias, data drift and opacity (Bertomeu, 2020; Molnar, 2022). This raises a critical question for assurance theory: can reasonable assurance be sustained or enhanced when audit conclusions depend on continuously learning systems whose internal logic may be difficult to document or explain to regulators and courts?

From a jurisdictional perspective, AI also redistributes epistemic power beyond the accounting profession to technology vendors, platform providers and data owners. As prior research shows, professions under technological threat engage in boundary work to preserve interpretive authority and redefine their expertise (Suddaby and Viale, 2011; Faulconbridge et al., 2023). In an AI-enabled accounting field, this boundary work increasingly involves asserting control over model governance, validation and ethical oversight rather than over routine analytical tasks. Future research should examine how assurance frameworks, auditing standards and regulatory regimes adapt to this redistribution of authority, including the emergence of algorithmic audits, model certification schemes and AI governance structures (Nwachukwu et al., 2025; Tian, 2025). Institutional and legitimacy theories offer a valuable lens for analysing how trust in financial reporting is maintained as algorithmic processes become embedded in accounting practice, and how the profession renegotiates its social licence in an era of hybrid human–machine judgement.

Table 3 synthesises these three thematic areas into a set of theory-driven research questions that foreground the disruption of core accounting constructs rather than incremental task change. Together, they outline a research agenda that addresses both the practical consequences of AI adoption and the deeper epistemic, institutional and jurisdictional transformations reshaping contemporary accounting.

AI and ML have become core epistemic technologies in accounting rather than mere efficiency tools (Kühl et al., 2022), directly challenging the profession's traditional authority over judgement and evidence. This transformation does not automatically render accountants obsolete, but it introduces a genuine risk of professional erosion if control over accounting knowledge production shifts away from the profession. As Autor et al. (2003) show, automation disproportionately affects routine work; however, when AI systems begin to encroach on non-routine cognitive tasks – such as estimation, risk assessment and narrative explanation – the boundary between assistive technology and substitute expertise becomes increasingly fragile.

Abbott's (1988) theory helps explain both the danger and the potential response. Professional jurisdictions endure not by monopolising tasks, but by controlling legitimate chains of reasoning. In an AI-rich environment, this control is no longer assured. If ML models are treated as authoritative decision-makers rather than evidentiary tools, epistemic authority may migrate to technology vendors, platform providers or regulators, leaving accountants responsible for outcomes without retaining interpretive control. Preserving jurisdiction therefore requires accountants to master and govern expanded chains of reasoning that incorporate ML systems, rather than deferring judgement to them. We use the term hybrid intelligence accounting to describe this contingent outcome: an epistemic regime in which human expertise and algorithmic inference jointly produce accounting judgements under professional oversight.

This outcome, however, is not inevitable. Professions that fail to integrate new epistemic technologies into their knowledge base risk marginalisation or displacement. While automation historically affects routine work most strongly, AI' s capacity to generate predictions, narratives and recommendations means that even traditionally “expert” domains are vulnerable. Absent deliberate intervention, accounting risks being reduced to a residual monitoring and compliance role, layered on top of opaque algorithmic systems designed and governed by non-professional actors. Noordegraaf's (2020) notion of connective professionalism highlights the challenge: expertise and technology must co-evolve, or professional authority dissipates.

Empirical studies show that firms respond through boundary work, emphasising ethical discernment, contextual interpretation and professional scepticism as AI automates analysis (Suddaby and Viale, 2011; Bucher et al., 2016). Many audit and advisory teams now operate through hybrid arrangements, where machines process large datasets and flag anomalies, while accountants apply judgement to interpret, validate and explain results (Faulconbridge et al., 2023). In practice, hybrid intelligence accounting manifests in cross-disciplinary teams in which technologists manage models and data infrastructures, while accountants retain responsibility for inference, accountability and client-facing explanations. These strategies aim to preserve jurisdiction, but their success depends on whether interpretive authority remains credible to regulators, courts and stakeholders.

The implications for governance and standard setting are therefore substantial. Accounting standards, auditing guidance (e.g. ISA 500; PCAOB AS 2501) and regulatory frameworks must evolve to address algorithmic evidence explicitly, including requirements for model documentation, validation, bias monitoring and explainability. Professional bodies increasingly recognise this need (ICAEW, 2022), yet failure to embed AI governance within professional standards risks ceding epistemic authority to opaque systems beyond professional control.

Ultimately, the profession faces a conditional future. Accountants who develop AI literacy, data governance expertise and the capacity to critically interrogate algorithmic outputs can retain jurisdiction over accounting judgement. Those who treat AI as a black box or cling to legacy practices risk becoming peripheral actors in a knowledge system dominated by machines and their designers (Abbott, 1988). AI thus forces a choice: either accounting evolves into a hybrid intelligence profession that governs human–machine judgement, or it risks erosion of its authority as epistemic control migrates elsewhere.