In a virtual interconnected digital space, the metaverse encompasses various virtual environments where people can interact, including engaging in business activities. Machine learning (ML) is a strategic technology that enables digital transformation to the metaverse, and it is becoming a more prevalent driver of business performance and reporting on performance. However, ML has limitations, and using the technology in business processes, such as accounting, poses a technology governance failure risk. To address this risk, decision makers and those tasked to govern these technologies must understand where the technology fits into the business process and consider its limitations to enable a governed transition to the metaverse. Using selected accounting processes, this study aims to describe the limitations that ML techniques pose to ensure the quality of financial information.
A grounded theory literature review method, consisting of five iterative stages, was used to identify the accounting tasks that ML could perform in the respective accounting processes, describe the ML techniques that could be applied to each accounting task and identify the limitations associated with the individual techniques.
This study finds that limitations such as data availability and training time may impact the quality of the financial information and that ML techniques and their limitations must be clearly understood when developing and implementing technology governance measures.
The study contributes to the growing literature on enterprise information and technology management and governance. In this study, the authors integrated current ML knowledge into an accounting context. As accounting is a pervasive aspect of business, the insights from this study will benefit decision makers and those tasked to govern these technologies to understand how some processes are more likely to be affected by certain limitations and how this may impact the accounting objectives. It will also benefit those users hoping to exploit the advantages of ML in their accounting processes while understanding the specific technology limitations on an accounting task level.
Introduction
The recent COVID-19 pandemic led many businesses to undergo accelerated digital transformation (Deloitte, 2020; Lee et al., 2021; World Economic Forum, 2021), resulting in significantly faster implementation of long-term strategic technology plans than intended. This, in turn, may have increased the risk of a technology governance failure, one of the critical business risks identified by the Global Risk Report (World Economic Forum, 2021).
One of the aims of digital transformation is to benefit the way an organisation operates through, amongst other things, digital technologies. One example is enabling automation using artificial intelligence (AI) (Vial, 2019). According to Mardini and Alkurdi (2021), AI will automate most financial accounting tasks in the future. Machine learning (ML), a subset of AI, saw a massive increase in the development of tools and applications since 2019, with one of the top use cases being process automation (Algorithmia, 2020), such as capturing documents, classifying transactions, account reconciliations and preparation of financial reports.
The maturation of ML and AI is also one of the technologies enabling digital transformation to the metaverse (Lee et al., 2021; Blau et al., 2022), a virtual world where people can interact in various activities, including engaging in business activities that require accounting (Pandey and Gilmour, 2023). Commentators argue that the metaverse will be the culmination of various technologies to create an immersive platform which extends the way business is currently performed from the physical world to the digital world (Lee et al., 2021; Blau et al., 2022). Blau et al. (2022) identify four key factors that will impact the metaverse’s potential, one of which is governance.
The risks of technology governance failure, the impact of governance on digital transformation to a metaverse and the uncertainty surrounding how this transformation will achieve desired outcomes emphasise the need for adequate technology governance to ensure that digital transformation does, in fact, advance and achieve the objectives of the businesses, reducing the risk of technology governance failure. The risk of technology governance failure surrounding implementing new technologies, such as ML, prompted the King IV Report to require a business to govern technology to achieve its objectives (Institute of Directors of Southern Africa [IODSA]), 2016). Alreemy et al. (2016) concur in their description of the aim of technology governance, namely, to ensure compatibility between the goals of the business and a satisfactory level of risk with the use of emerging technologies.
There are two aspects to technology governance: governing business objectives using technology and the actual governance of technology to ensure the technology achieves its objectives (Wilkin and Chenhall, 2020). With a focus on the latter, obtaining buy-in from stakeholders and alignment within the organisation between various stakeholders were some of the challenges found when implementing ML tools (Algorithmia, 2020). Technology governance can be applied at a strategic and operational level (Goosen and Rudman, 2013). Our research focuses on strategic technology governance at a business process level, assisting accounting users to ensure business information technology (IT) alignment, as Goosen and Rudman (2013) recommended. The accounting process, one of the business processes most prevalent in organisations, can benefit from ML. ML enables the automation of many tasks in the accounting process, reducing the risk of human error and making these tasks more efficient (Fallatah, 2021). Tasks may include processing source documents and analysing business transactions (Cho et al., 2020).
Considering the link between achieving business objectives, technology governance and the digital technologies used, such as AI and ML, and in the context of accounting tasks, it would follow that it is necessary to understand what the ML technology should achieve and then to understand its limitations, that is, what would prevent the technology from reaching the stated objectives. Bavaresco et al. (2023) and Kommunuri (2022) identify the importance of understanding the technology’s limitations. In this way, the technology can be governed to reduce the risk of stated objectives not being achieved, as King IV requires. In considering the general limitations of ML that impact the accounting process, existing research focuses on the general limitations of using ML technology, such as the lack of interpretability of algorithms and algorithmic bias (Cho et al., 2020; Fallatah, 2021; PWC, 2019). However, to better understand the limitations of specific ML techniques in the accounting context, accounting decision makers need to know where ML fits into each accounting process and the specific limitations that may consequently arise. The problem is that the limitations of those ML techniques that can perform particular accounting tasks within the accounting process are not identified and, therefore, cannot be adequately considered and governed. An accounting task can be defined as any action to record an economic event, adjustment or modification according to accounting principles (Petkov, 2020).
This research aims to identify the limitations of some ML techniques that can perform specific accounting tasks within the accounting process. To achieve the research aim and answer the research problem, the authors formulated the following research questions to guide the research process:
Which tasks in the accounting processes can be assisted or performed by ML techniques?
What are the limitations associated with the identified ML techniques, and do these link to the accounting objectives?
This study finds that limitations such as data availability and training time may impact the quality of the financial information and that ML techniques and their limitations must be clearly understood when developing and implementing technology governance measures. This paper contributes to the field of technology governance research, specifically considering the limitations of using ML in accounting against the qualitative objectives of useful accounting information. Moreover, the study will benefit those accounting users hoping to exploit the advantages of ML in their accounting processes while understanding the specific technology limitations on an accounting task level. It will assist especially those accounting decision makers wanting to know how some processes are more likely to be affected by certain limitations and how this may impact the achievement of business objectives, specifically the qualitative accounting objectives.
The remainder of the article is laid out as follows: The next section presents a brief literature review to explain the aspects impacting the research problem, followed by the research design. Next, the analysis and findings are presented, followed by conclusions based on our research findings. The article closes with the limitations and suggestions for future research.
Literature review
Accounting objectives and quality of financial information
Gillion (2017) states that in all businesses, accounting processes aim to produce high-quality accounting information for decision-making and, therefore, high-quality financial reports. The Conceptual Framework for Financial Reporting describes and categorises the qualitative characteristics of useful financial information (International Accounting Standards Board, 2022) into fundamental and enhancing characteristics. For this study, the fundamental and enhancing qualitative characteristics are designated as the qualitative accounting objectives. Later, we will link any applicable limitations of a particular ML technique used to produce useful financial information to these qualitative accounting objectives. We briefly define and summarise the accounting objectives in Table 1.
This section has listed the qualitative accounting objectives of useful financial information. The tasks in these accounting processes will be set out next.
Accounting process and tasks to produce financial information
In the traditional accounting process, source documents received and generated are captured in an accounting record system, reconciled, and finally, the financial information produced is presented in financial reports. In this study, we used Deming’s (2024) description of the traditional record-to-report process to identify the three main accounting processes. Each main process is then broken up into tasks to enable us to identify ML techniques that could perform those task(s). Figure 1 illustrates Deming’s traditional record-to-report process and our summation into the three main accounting processes (processes 1 to 3).
Figure 1 illustrates how the record-to-report process commences with the external information sources (Process 1), followed by account reconciliations (Process 2) and then journal entries, month-end closure, analysis and reporting (Process 3). Our study does not address the performance of compliance and control procedures.
To identify one or more ML techniques to perform the accounting process, the accounting processes need to be broken down into tasks within that broad process. According to Amani and Fadlalla (2017), there is a paucity of published research on applications that use ML techniques. While the reason for this is unclear, they speculate that it may be due to a lack of reporting on such applications because of the unwillingness to reveal details of these applications for competitive reasons (Amani and Fadlalla, 2017). Based on available research, we have identified the following tasks within the respective three accounting processes and summarised these in Figure 2.
Figure 2 describes the various accounting tasks in the accounting processes. In the next section, we explore ML and the techniques that could perform some accounting tasks.
Machine learning technology to automate tasks
As a technology, ML is a subset of AI, in which data patterns are learned and applied in a changing environment. The technology does not require all possible situations to be known during development (Ayodele, 2010a; Sainani, 2014) and is one of the technologies that can be used in the accounting process to assist in the automation of tasks (Everest Group, 2014).
Despite the uncertainty, ML technology can detect patterns or predict solutions (Valavanis et al., 1994). To detect patterns or predict solutions, ML uses algorithms. This feature enables the technology to automate routine tasks, making it so valuable for automating routine accounting tasks (Sapp, 2017).
Based on the objective of the algorithm, how the algorithm learns, as well as the structure and volume of the data used for learning, ML algorithms can be distinguished as supervised, unsupervised and semi-supervised learning (Ayodele, 2010b; Castle, 2018). For the reader’s benefit, the three categories are briefly described, as all three categories may have useful ML algorithms to automate accounting tasks.
Firstly, supervised learning algorithms require training. The algorithm is trained using a labelled data set consisting of examples of input data and labels indicating predicted targets or output data. Labels assist the algorithm in determining which features are essential. The algorithm then generalises the training set by mapping the inputs to the correct responses, which enables it to produce output for new inputs (Marsland, 2009; Ayodele, 2010b; Larsson and Segerås, 2016; Castle, 2018). The algorithm’s training on previous data sets, for example, the classification of accounting invoices and the respective financial fields, makes this type of ML particularly useful for accounting task automation, mainly due to the vast amount of data available and the rules-based nature of accounting.
Secondly, unsupervised learning algorithms do not require training. The input data are unlabelled, meaning the predicted values are not provided, which may be because they are unknown. The algorithm needs to determine the links between the inputs provided to identify patterns or commonalities that can be used to categorise new data or solve problems (Marsland, 2009; Ayodele, 2010b; Larsson and Segerås, 2016). An unsupervised algorithm is suitable where the action required, or outcome is uncertain for the task being performed or where unknown data needs to be grouped. This type of algorithm can be useful for automating accounting tasks such as error detection.
Thirdly, semi-supervised learning algorithms are trained using a combination of labelled and unlabelled data to generate an appropriate function. The labelled portion indicates patterns which may exist, while the unlabelled data, usually the larger portion of the data, are used to establish perceived or unknown patterns for the data (Ayodele, 2010b; Castle, 2018). It follows that where automation requires trained and untrained algorithms, such as allocating transactions to respective accounts, most transactions can be allocated using the prior knowledge from the training data. In contrast, a few transactions may not be known from the prior data, so a semi-supervised learning algorithm is preferable.
Machine learning techniques to solve problems
According to Someren and Urbancic (2006), the process of matching an identified problem to the technique to solve it is difficult. Firstly, the task must be understood. Secondly, a problem that ML can solve in that task must be identified and described. Thirdly, identifying the learning problem enables a developer to identify a solution, namely, the information and algorithm required to address the problem (Saitta and Neri, 1998; Someren and Urbancic, 2006).
Understanding the task and defining the learning problem is crucial, as many ML solutions are often available for addressing a problem (Someren and Urbancic, 2006). The types identified and described by Amani and Fadlalla (2017) and Larsson and Segerås (2016) are summarised in Table 2.
Table 2 shows the various ML solutions available to address an identified problem. One of the ML solutions is Classification (refer to Table 2). Kotsiantis (2007) explains that classification algorithms are supervised learners, and therefore, their development consists of a two-step training and testing process. As indicated earlier, supervised learning algorithms are particularly useful for accounting task automation.
According to Kotsiantis (2007), the different supervised classification algorithms can further be separated into three types of techniques:
logic-based;
perceptron-based; and
statistical (Kotsiantis, 2007).
Using the three different types, Kotsiantis (2007) gives examples of the various ML algorithms for each type of supervised classification technique. It is these algorithms that can be used to solve a problem.
To expose the reader to the possible ML algorithms, we use the classification of ML techniques adapted from Kotsiantis (2007) to identify several algorithms. Figure 3 illustrates the type of ML techniques grouped according to the layout in Table 2.
It is outside the scope of this article to consider the detailed technical features of each of the algorithms in Figure 3. However, we point the reader to Appendix for a brief description of the general characteristics of the algorithms.
Given the increasing reliance on ML within financial systems, algorithmic bias, even small, could scale into substantial errors. For the benefit of the reader and practical relevance of the study we acknowledge these concerns and risks in the following section:
Machine learning – general algorithmic bias
The increasing integration of ML in areas such as health care, education, employment and law raises the concern that unintended algorithmic biases can lead to adverse consequences (Barocas and Selbst, 2016; Kleinberg et al., 2016; Mitchell et al., 2019). In accounting, the general limitation of algorithmic bias, as highlighted in the research by Cho et al. (2020), is an important area that merits further exploration, albeit beyond the scope of this article. However, it is important for accounting users to be aware of this general limitation due to the associated risks of errors.
One illustrative example is in predicting accounting fraud, where Suresh and Guttag (2019) have identified different types of bias that may be embedded in ML models, along with potential solutions. Notably, many of these solutions focus on addressing the quality of the data used, aligning with Gillion’s assertion in 2017 that a successful machine-learning model requires high-quality data (Gillion, 2017).
Nevertheless, for a comprehensive overview of the types of general algorithmic bias and potential solutions, readers are directed to the excellent articles by Caton and Haas (2023), Cho et al. (2020) and Mehrabi et al. (2021). In this context, diverse and representative data sets with supporting documents (such as data sheets), labelling data, implementing statistical significance tests to detect discrimination and a commitment to ethical AI practices remain critical mitigation strategies to enhance fairness and reliability in ML applications.
The following section will explain how we analysed the phenomena to build a theory on the limitations of ML techniques that can perform specific accounting tasks within the accounting process.
Research design
Earlier, we noted that our research aims to identify the limitations of using ML technology for specific accounting tasks performed within the accounting process. To achieve this aim, we formulated research questions to focus on gaining theoretical insights into the limitations of ML techniques in an accounting process context. Therefore, the research is exploratory in nature.
An exploratory research design allows for the development of a grounded picture of the phenomena and the development of tentative theories or hypotheses (USC Libraries, 2024). Therefore, grounded theory (GT) will be used to achieve the research aim. GT is an inductive methodology used to develop a theory (Sutton et al., 2011; Wolfswinkel et al., 2013), and in this case, a theory about emergent technologies where there is not, as yet, an established theory. The GT method is particularly suited to information systems technology research (Bryant, 2002; Fernandez and Lehmann, 2011) and highlights the mandate of research to develop both an understanding of discovered facts and adequate models for specified purposes.
This study uses prior literature to identify the limitations of using ML technology for specific accounting tasks. We follow Wolfswinkelet al.’s (2013) five-stage process when conducting a GT literature review. The five stages are as follows:
define the scope of the review;
search the literature within the scope;
select the sample of literature to be analysed;
analyse the literature; and
present the results (Wolfswinkel et al., 2013).
As this study aims to identify links between new variables, as is the purpose of a GT literature review method (Wolfswinkel et al., 2013), a systematic approach was followed for each of the required variables. These variables are as follows:
the accounting tasks in the selected accounting processes;
the ML techniques available to perform the accounting tasks;
the limitations of the ML techniques which were identified; and
the accounting objectives.
The different stages of the GT literature review method were executed as follows.
Stage 1: define
As discussed earlier, the field of research is relevant to accounting practice, including information systems technology, financial processes and automation. Therefore, we considered literature on ML within an accounting context. However, if there were applicable examples and findings from practice, these were included in the scope of the research.
The initial search was broad and targeted to online databases. The search terms included keywords such as “machine learning,” “artificial intelligence,” “algorithm,” “accounting,” “financial,” “source document,” “invoice,” “reconciliation,” “reporting” and “automation” on the Scopus, EBSCOhost, IEEE and AAA digital library databases as recommended by Sutton et al. (2016). Following that, the search was expanded to websites and resources offered by accounting software providers (e.g. SAP and XERO), as we found that academic literature on accounting tasks was limited.
The final variable for inclusion was to identify the limitations of ML from existing literature. Here, the focus was first on the accounting literature and then expanded to information systems literature using the IEEE Xplore database. Search terms were based on the identified ML techniques combined with keywords in the search such as “risks,” “disadvantages” and “limitations.”
Stage 2: search
Having prepared the criteria and selected the appropriate sources and search terms, the searches were performed systematically. The researchers ensured that essential synonyms of search terms were included (Wolfswinkel et al., 2013) if identified.
Stage 3: select
The literature’s abstracts identified in the search were read to determine if they were relevant to the aim of the study. In certain instances, the search criteria needed to be refined to find the relevant literature (Wolfswinkel et al., 2013). Those papers that were then found to be in line with the aim of this research were downloaded for analysis.
Stage 4: analyse
The three selected accounting processes were investigated as a starting point to determine which tasks can be performed using ML techniques. To do this, each accounting process has been presented by separating it into its respective accounting tasks. The literature was then evaluated to identify which tasks offer learning problems: the task the ML should perform (Someren and Urbancic, 2006). Following that, the different types of ML techniques available to address each learning problem were identified and linked to the applicable accounting task.
Finally, the literature was reviewed to provide a list of the limitations of each ML technique. These limitations were considered in the context of the respective accounting tasks and linked to the applicable accounting process objectives. This link is made by considering the description of each qualitative accounting objective. This paper presents the Analysis (Stage 4) and Findings (Stage 5) together.
Stage 5: presentation
The last stage of the GT literature review method is to present the findings and insights gained in a structured manner. The following section presents the analysis (stage 4) and the findings (stage 5). The analysis and findings are linked and presented according to the particular research question.
Analysis and presentation of findings
Findings related to RQ1: which tasks in the accounting processes can be assisted or performed by machine learning techniques?
We discussed earlier that the process of matching each task to a technique involves, firstly, understanding the task, secondly, defining the learning problem and finally, identifying the information and algorithm required to address the problem (Saitta and Neri, 1998:137; Someren and Urbancic, 2006:366).
To match an identified learning problem to the technique to solve it, we explored each task in the three accounting processes to find whether a suitable ML technique can be applied to perform the identified task. We found and noted six learning problems in Process 1 (tasks 1.3, task 1.4 and task 1.6), one learning problem in Process 2 (task 2.3) and three learning problems in Process 3 (tasks 3.3, 3.7 and 3.8). We then consulted existing literature to find a suitable ML solution and type of ML technique(s) to address these learning problems. In Table 3, we summarise the learning problem identified per accounting process and task, and for each learning problem, the solution and the specific technique that achieves the solution.
The findings in Table 3 demonstrate that there may be more than one ML technique available to address a specific learning problem (Someren and Urbancic, 2006), and it is critical to understand the accounting process, the tasks within the process, the potential learning problems and what ML techniques can solve those problems. As there may be different limitations for each specific technique, it also follows that these need to be placed into the context of each accounting process, depending on the objectives.
Findings related to RQ2: what are the limitations associated with the identified machine learning techniques, and do these link to the accounting objectives?
To ensure adequate technology governance, as described by King IV (IODSA, 2016), the accountant would need to determine, for example, whether accuracy is more important than cost saving. Therefore, we considered the limitations of each ML technique that may impact the respective qualitative accounting objectives and are identified. We present our findings in Table 4.
Having considered the respective limitations of each ML technique we identified from the existing literature, Table 5 links these limitations to (an) applicable qualitative accounting objective(s). Moreover, we table the applicable tasks in the accounting process for each ML technique for the reader’s benefit.
The primary limitations identified from the research for the respective ML techniques suitable for use in the selected accounting processes have been summarised in Table 5 and linked to the appropriate accounting objectives.
As noted in Table 5, some ML techniques limit interpretability, as the knowledge that the ML technique uses or discovers to perform a particular task may not always be available to the user (The Royal Society, 2017). This would mean that the user does not know how data input A resulted in information output B. This impacts the accounting objectives of verifiability and understandability.
The above could be a problem when using, for example, an artificial neural network ML technique to perform matching during a reconciliation accounting task, especially if it is unclear why the application matched one transaction with another.
Another limitation of some ML techniques is that they are quite computing intensive. This aspect of ML and AI has been identified as one of the challenges for the metaverse, as it impacts access to users on mobile devices, for example (Lee et al., 2021). This limitation will need to be considered in terms of whether the costs of incorporating this technology in the accounting process may exceed the financial benefits to the business as these ML techniques require advanced data integration tools and infrastructure (Gillion, 2017; Sapp, 2017). Blau et al. (2022) advised users wanting to incorporate the metaverse into their strategy to consider such costs from a broader digital transformation perspective, not just the financial benefit of investing in ML technology.
The limitation of adequate data being required for the training of algorithms may not be applicable in the metaverse, as one of the characteristics of the metaverse would be the availability of immeasurable amounts of structured data (Lee et al., 2021).
Conclusion
This paper aimed to identify the limitations of those ML techniques that can perform specific accounting tasks within the accounting process. We achieved our aim by identifying the accounting tasks that ML could perform, which techniques would be able to perform which functions in the accounting tasks and then identifying the limitations associated with specific ML techniques. Finally, these limitations were linked to qualitative accounting objectives, which may be impacted.
We find that there were limitations to the ML techniques, which may impact the achievement of the qualitative accounting objectives when using ML in the accounting process. Some ML limitations identified in this research are barriers to entry into the metaverse, such as the extent of the computation abilities required to apply ML. In contrast, other limitations, such as adequate data, may be addressed by immersion in the metaverse.
Our research highlights the need for accounting users to understand where ML technology can be used in the accounting process and to be aware of the limitations of ML techniques that may impact the ability to achieve qualitative accounting objectives. As tasks are transformed from the physical to the digital world, we, as accounting users, can lead in enabling a governed transition to the metaverse.
Limitations and future research
The study does not intend to cover all the areas of accounting in which ML intervention is possible; it only addressed the three accounting processes: the translation of manual and electronic documents into accounting software information, the reconciliation of financial information and the preparation of management accounts. Areas for future research could consider the use of ML in accounting estimates, such as the estimate of expected credit losses of loans in the financial services industry.
Only those tasks for which a suitable ML technique can be found during this research were addressed. Furthermore, the study only considered ML techniques appropriate for addressing the identified accounting tasks in the accounting process and, therefore, does not intend to present an exhaustive list of ML techniques.
Our study does not explain how each ML technique functions. The limitations identified for the ML techniques are those unique to the technology and not those about the environment in which ML operates, such as a database or accounting software limitations or limitations about supporting technologies such as cloud platforms. Hence, these limitations are not addressed in this paper. Furthermore, considerable research has been performed in the areas of ML applied to auditing and the detection of fraud using such technologies. Therefore, these areas were not considered for this research.
Future research could look at how these limitations can be addressed to ensure that the technology is adequately governed. The question of whether standards (such as specifications regarding training data, ML model, performance and updating of models) need to be set to regulate the ML tools used by accountants (Cho et al., 2020) has also been asked, considering the limitations and risks faced when using ML technology and the need for technology governance. Considering the risks, opportunities for further research may also include a better understanding of the risks when implementing ML technology in an accounting system. Research into the benefits of ML can moderate future research into the limitations, as explained above.
Finally, in this study, we consider ML as one of the technologies enabling rapid digital transformation to the metaverse, explicitly focusing on the limitations of ML techniques in an accounting context. Future research could explore some of the challenges and opportunities that the rapid pace of digital transformation presents for technology governance in the accounting field.
