of auditing that focuses on the implementation of
machine learning identifies five issues: data quality
and visualization, model interpretability and
transparency, overfitting and generalization,
regulatory compliance, and ethics, the skills and
training gap among auditors. The section proposes the
measures to improve data governance, model
interpretability, overfitting control, compliance and
ethical concerns, and auditor's technical skills all of
which would contribute to the solution of these issues.
Such methods are likely to contribute to the
subsequent diligence of machine learning in auditing,
resulting eventually in the growth of the quality and
efficiency of audits. This article tries to deliver
valuable findings and to present some
recommendations of efficiently using ML for raising
the level of audits’ quality and providing compliance
with the laws of compliance.
2 CHALLENGES
2.1 Data Accuracy and Precision
Another major feature concerning the auditors is the
incomplete data that they get. The summary part of
the report may be distorted in most cases when the
transaction data, such as date, time, nature of
transaction, cost of goods sold, etc., are revealed
incompletely or inadequately. These areas inhibit
machine learning from discerning anomalies, fraud
and compliance violations; hence, the level of audit
clutter cannot be trusted and lacks reassurance (Office
of Inspector General, 2012). The utilization of
machine learning on unsuitable and incomplete data
sets will not only yield inaccurate insights, but also
miss out on knowing the profound risks and
anomalies the artificial intelligence should have been
dealing with.
Discrepant data is one of the major difficulties
confronting practitioners. Disparities in data entry
systems, human mistake, and variable specificity of
data sets could be the cause of drifted information
between data source systems. Such differences can
disrupt the engine of machine learning algorithms that
form the basis of this whole idea, thereby provoking
errant conclusions about risk assessment or wrong
identification of mischief during the audit (Stefanov
et al., 2012). For instance, labelling the same financial
transactions originated from one source into different
categories could mean that the machine will not be
able to notice the patterns of fraud or error.
Additionally, machine learning is further
complicated because of wrong data, like
discrepancies in the transaction amounts or
accounting transactions that are not in the right
account. Such errors will result in a situation where
the models are used to learn from the wrong
information, and their outputs may not be useful
inputs to even monetizing the said organization.
Maintaining high quality and integrity of data
becomes increasingly challenging but critical,
particularly with businesses generating huge volumes
of complex data.
In an environment where auditing becomes too
demanding for the teams due to these data quality
matters, the entire audit process may fail leading to
serious mistakes in audit conclusions, and, moreover,
it can result to failing to meet the compliance and
regulatory requirements. Hence, the pillar that goes
hand in hand with the capacity to utilize machine
learning so as to improve audit accuracy, efficiency,
and effectiveness is data quality and integrity
(Stefanov, 2014). If these problems are not solved,
many of the advantages of ML-driven practices will
remain unexploited, and the overall audit result
quality will be compromised.
2.2 ML Interpretation and Describe
An enormous difficulty to use machine learning in the
auditing field is the issue of machine learning models
transparency and interpretability. The advanced
machine learning models, especially deep learning
that are the most robust algorithms nowadays, tend to
be “black boxes” thus having no knowledge of the
way the decision-making is based on (Carvalho,
Pereira, & Cardoso, 2019). These models utilize data
processing as well as decision making systems that
enable the algorithms to successfully do that, but the
humans, including the auditors, cannot follow the
internal procedures used for arriving at the decisions
because such procedures are not clearly shown to
them.
In auditing, this kind of uncertainly may be
problematic since the auditors need to know the path
of how decisions have been taken to achieve every
conclusion. In finance, traditional auditing entails the
following and logically verifying procedures and
steps that are easy to understand. However, the
process of machine learning has been found to be
complex, and it is not easy to follow the road from
input data to output decision, especially when there
are deep neural networks with numerous layers in
between. This complexity impedes auditors from
interpreting the reasoning behind the model outputs
to validate and attest audit findings which is one of
the most important steps in the whole audit process.