Proposal for a Hybrid Audit Expert System
Susan Clemmons, Kenneth Henry
Decision Sysetms and Information Systems Department
Florida International University, Miami, Florida
Keywords: Expert System; Fuzzy Logic; Audit;
Abstract: This paper theorizes the use of a hybrid expert system to support
a complete audit of financial statements for
an enterprise. The expert system proposed would support the audit process by using two types of artificial
intelligence technologies: case-based reasoning and fuzzy logic technologies. The case base and automated
reasoning recommendations would give the auditing firm another insight on the audit. Unlike previous audit
expert systems, this system is intended to focus broadly on an enterprise’s entire financial statement audit
process; it combines a case based knowledge representation with fuzzy logic processing. The attempt at
capturing a wide domain is necessary to support organizational decision-making. Focusing on narrow
decision points within an audit process limits the users and usefulness of the system.
Accounting firms and researchers have devoted
significant effort to the use of decision support
applications to assist in audit work. Many of these
systems have been developed combining the
knowledge and rules of the practice in the form of
knowledge based expert systems. The use of expert
systems as decision support tools in narrow auditing
specialty areas is well documented (O'Leary, 1993).
However, a conceptual gap remains when the
outcome of several audit decisions are combined for
an overall opinion or outcome decision. It has been
noted that business expert systems lack a strategic
focus for organizational decision-making support
(Wong and Monaco 1995).
This paper theorizes the use of a hybrid
pert system to support a complete audit of
financial statements for an organization. The
financial statement audit consists of more than one
hundred action steps, and reviews four specific
financial documents. It is conducted through a
process that is triggered by many different decisions.
Some of the decisions can be supported through tax
codes and audit rules of generally accepted auditing
standards (GAAS). Other decisions rely on the
auditing firm’s knowledge of the client company,
the industry practices, and the firm’s prior
experiences, to assist in determining the proper
course of action. The audit of financial statements is
a complex series of judgments that leads to an
opinion about the financial health of organizations.
The audit is a critical financial information
validation process for all organizations.
As the use of expert systems started to
row, Bailey et al (Bailey, Hackenbrack et al, 1987)
outlined an expert system research agenda in the
accounting and audit area. They stress that the
contribution of academia is not one of software
development and creation, but one of thought
leadership about the decision making process used in
auditing. They advocate this direction as the
fundamental research objective of the academic
community. In this paper, we try to follow this
agenda and offer a method suited for the audit
environment to support the decision making process
of the auditing firm.
The expert system proposed would support
e audit process by using two types of artificial
intelligence technologies: case-based reasoning and
fuzzy logic technologies. The case base and
automated reasoning recommendations would give
the auditing firm another insight on the audit.
Typically, the domain of expert system applications
is very narrow. This allows for complete
exploration and rule development to support the area
of expertise (Giarratano and Riley, 1998). The
domain suggested for this hybrid system is much
Clemmons S. and Henry K. (2005).
REDUCING RISK IN THE ENTERPRISE - Proposal for a Hybrid Audit Expert System.
In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 260-266
DOI: 10.5220/0002540702600266
more general, and broader in scope, as it would
encompass the entire financial statement audit
process. The use of fuzzy logic or reasoning would
allow the representation of uncertainty or a degree of
certainty in the case data. Some of the decisions
used in the audit process are not simply “either or”
situations but use some gradation of the answer.
The use of fuzzy variables would assist in the
representation of this type of data (Fetter and
McMillan, 1987). This approach focuses on
organizational decision-making with the
combination of technologies, the wide decision
support domain, and makes this design unique.
The body of this paper is organized into
three sections. The first section gives a brief
explanation of expert systems use in business and
auditing. It highlights important elements that must
be considered when building an auditing expert
system. The second section describes the details of
the planning audit step as a case for processing in an
expert system. It uses pseudo code examples and
reasoning logic to describe the case structure and
definition. The last section gives a brief discussion
of the validation options for this approach.
Expert systems are classified different ways. A
system may be classified by the way a problem is
addressed, the problem or expert domain, or by the
intended user group of the system. Generic problem
classifications can be overlapping due to many
expert systems containing more than one method of
problem resolution. The audit domain expert system
proposed in this paper combines problem
interpretation and prediction.
There are three commonly cited types of
expert systems: rule-based, model-based, and case
based. A fourth category, hybrid systems, has been
used recently that describes systems that overlap
these definitions. Hybrid systems combine different
types of technology, and may have several
knowledge presentation and inference modes. Most
of the business applications developed during the
1980s and 1990s were rule-based expert systems.
These systems represent knowledge in the form of a
rule stated by an IF-THEN format. When a certain
event occurs or relationship is present, then the
certain outcome is likely; and this outcome points to
appropriate actions. The application of sequential
rules leading to the most likely conclusion is
referred to as chaining (Turban, 1992). The chain is
formed as the action or result of a rule is linked to
the condition of the next rule or relationship.
Inference engines with a rule-based design have two
ways to proceed to an outcome. Backward-chaining
inference engines start with a goal and the search for
rules that will establish the facts to support the
conclusion. Forward-chaining inference engines are
data driven searches that arrive at a conclusion based
on the data presented. Typically, the auditing expert
system will fall into this category.
There are several advantages to using rule
representation to recommend a course of action.
Rules are easy to understand. They are a natural
form of knowledge and can be found in everyday
life. Inferences and explanations are also easily
derived from rules. Rules make modifications and
maintenance of the system simple. However, rules
also have limitations when used to represent
knowledge (Hayes-Roth, Waterman et al. 1983).
Complex knowledge may require thousands of rules,
as is the case with audit experience and judgment;
this could create a problem in the system
maintenance issue. Due to the common use of rules,
builders of expert systems tend to rely on this
knowledge representation, when other methods may
be more appropriate (Turban, 1988).
Model-based expert systems are based on
knowledge that represents the structure and behavior
of devices that the system is designed to understand.
They are useful in diagnosing equipment problems.
The system draws conclusions directly from
knowledge of the structure and behavior. One
feature of model-based expert systems is their
"transportability". A rule-based expert system may
be of no value for repairing a different type of
device that does not match its rules. A model-based
system could be used to diagnose or repair the
problem of any type of device based on the model
(Turban, 1992).
Case based expert systems use case based
reasoning to adapt solutions that were used to solve
old problems and use them for the basis of a new
solution for a new problem. Case based reasoning is
a problem solving approach based on the retrieval
and adoption of episodes with descriptions of
problems and their associated solutions. One
advantage of using case based reasoning is that the
existing data and knowledge is leveraged and can be
included in the database. The knowledge does not
have to be translated and coded into rules. This
makes knowledge acquisition a much faster process
(Kesh, 1995), critical for the audit domain expert
system. The system learns from both successes and
failures of cases. The more interaction or learning
that takes place, the richer the case database. Case
base reasoning systems provide information for
questions. Explanations then can be provided by
REDUCING RISK IN THE ENTERPRISE: Proposal for a Hybrid Audit Expert System
rule-based systems. A rule-based system provides
the explanation by the rules used to create the
solution. In a case based system, actual cases that
come close to matching the input case are used to
describe the solutions. Case based reasoning mimics
the human cognitive process for problem solving
better than other types of expert systems. Recall
usually takes the form of remembering the entire
case or episode rather than a set of rules. In this
way, case based expert systems are seen as more
flexible and friendly to system users (Kesh, 1995).
Hybrid systems are systems that use a
combination of knowledge base and reasoning
engines to derive a solution. They use the strengths
of each of the solutions to produce a result superior
to those of just a single method. Soft computing
techniques are being applied where uncertainty and
learning a part of the systems requirement. Soft
computing refers to techniques such as fuzzy logic,
neural networks, and genetic algorithms. Examples
of hybrid systems are expert systems utilizing
production rules in the knowledge base and fuzzy
logic as part of the inference engine. Nolan (Nolan,
1998) found that fuzzy technology enables the
improvement of approximate reasoning by three
different methods: (1) through efficient numerical
representation of vague terms, (2) through increased
range of operations in ill-defined environments, and
(3) by decreasing sensitivity to noisy data.
Some research (Lenard, Alam et al. 2001)
suggests the use of fuzzy clustering applied to
qualitative questions asked during the audit can be
successfully used in a hybrid system. Their work
focused on combining fuzzy clustering and a proven
statistical model to support an auditor’s decision
about going concern. Their expert system hybrid
model provides statistical support and expert
knowledge for use in the audit opinion. The success
of their system with bankruptcy predictions indicates
using both quantitative and qualitative information
has the potential for better accuracy than each model
being used separately. Strategic expert systems is
still an under addressed topic in business (Wong and
Monaco, 1995). This type of expertise is difficult to
extract, and due to wide domain areas, the issues
may be very complex and interrelated. While
researchers have recognized the importance of these
systems, there is a void in the business literature
with regard to this topic.
The domain of auditing is defined as: "a systematic
process of objectively obtaining and evaluating
evidence regarding assertions about economic
actions and events to ascertain the degree of
correspondence between those assertions and
established criteria and communicating their results
to interested users" (Concepts, 1973). There are
three types of audits: (1) financial statements audit,
(2) compliance audits, and (3) operational audits.
The financial statements audit encompasses the
process of collection and evaluation of evidence
about an organization's financial statements. Its goal
is to express an opinion as to the statements’ fair
representation of the financial position, results of
operations, and cash flows of the organization; and
whether they are prepared in conformity with
Generally Accepted Accounting Principles (GAAP)
and other applicable criteria.
Briefly described, the financial statement
audit process consists of four phases: (1) Planning
and design of the audit approach, (2) performing
tests of controls (TOC) and substantive tests of
transactions (STOT) (3) Performing analytical
procedures and tests of detail balances (TDB),
(4) Completing audit fieldwork and issuing the audit
report. This formal structure lends itself easily to
the application of a case-based approach. Each set
of case data generated by the performance of the
annual audit for a given client is conveniently stored
in a matrix format, wherein a specific set of tasks
must be performed in a specific order. This is done
overall for audit planning purposes, and more
specifically for each “audit cycle” performed for the
financial statement line item classification.
Each audit performed by the audit firm will
generate data and expert system recommendations in
each of the four phases for a wide variety of
circumstances. The collection of facts, rules,
inferences, and conclusions will be represented by
one case in the expert system’s case database. Every
successful audit firm will normally perform multiple
audits during the course of a year, with each audit
generating a new case for the database.
Furthermore, as the years pass, additional cases are
generated for new conditions as a given audit
client’s financial statements undergo the annual
auditing process.
Within each cell of the defined case matrix,
a specific set of data (facts) must be gathered about
the planning for that phase or about the financial
statement line item for the other phases. Also for
each cell, a specific set of rules (production rules)
must be applied to the facts (asserted or bound). The
inference engine of the expert system must then
apply the rules to the facts gathered, typically using
fuzzy logic algorithms, and generate
recommendations for specific audit conclusions
derived, and procedures to be performed.
For illustration purposes, the next section of
the paper gives a detailed description of the facts to
be gathered (audit evidence) and the rules to be
applied (generally accepted auditing standards) in
phase I of the plan and design of an audit: making
the audit decision on acceptance/renewal of a client.
Selected sets of facts and rules are shown using
CLIPS-like pseudo-code to demonstrate how the
facts are asserted or bound, and how the rules are
applied to produce expert audit conclusions.
3.1 Accepting or Rejecting a Client
The expert system, being used as an audit firm’s
assistant, has four decisions to make at this early
stage of the audit. First, the audit firm must decide
to accept (or not) a new client or continue (or not)
serving an existing one. An individual auditor will
review and recommend the course of action to the
audit firm’s management. Typically, the
recommending auditor is experienced and in a
position to make important decisions. This decision
needs to be made early, before the audit firm incurs
any significant audit costs that cannot be recovered.
Therefore, the more information about this type of
decision in the case database, the better decision the
expert system can make. Second, the audit firm
identifies why the client needs or wants the audit.
This information is likely to affect the remaining
parts of the planning process, because it directly
affects the audit scope. Third, the audit firm obtains
an understanding with the client about the terms of
the engagement to avoid any misunderstandings.
Fourthly, the audit firm must select staff for the
engagement, including any required audit specialists.
Stated in terms of audit risks, an audit
expert is unlikely to accept a new client or continue
serving an existing client if acceptable audit risk
(AAR) is higher than the CPA firm’s risk threshold.
This suggests the first of the facts to be asserted and
rules to be defined by the audit expert system
(assert AAR-threshold)
(bind ?AAR-value
(defrule accept-or-reject1
(> ?AAR-value AAR-threshold)
(assert (reject-client)))
3.2 New Client Investigation
An audit firm should evaluate a prospective client’s
standing and reputation in the business community,
its financial stability, and the relations with its
previous CPA firm. For example, many audit firms
are very cautious in accepting new clients in newly
formed, rapidly growing businesses. From
experience, many of these businesses fail
financially, and expose the audit firm to significant
potential liability. For prospective clients previously
audited by another CPA firm, the new successor
audit firm is required by SAS 84 (AU §315)
communicate with the predecessor audit firm”, to
help the successor audit firm evaluate whether to
accept the engagement. Communications may
inform the successor audit firm that the client lacks
integrity, or that there have been disputes over
accounting principles, audit procedures, or fees.
Even when another CPA firm has audited a
prospective client, other investigations are often
made. Sources of information include local
attorneys, other CPA firms, banks, major suppliers
and customers, and other resources. In some cases,
the audit firm may hire a professional investigator to
obtain information about the reputation and
background of key members of management. A
more extensive investigation may be necessary when
there is no previous audit firm, when a predecessor
audit firm will not provide the requested
information, or when any problems arise from the
communication. In expert systems terminology:
(bind (?client-integrity “high”))
(bind (?disputes “none”))
(defrule accept-or-reject2
(client-integrity = “high”) AND
(disputes “none”)
(assert (accept-client)))
The series of audit judgments made is illustrated in
this section using pseudo-code. This is intended
merely to convey the facts to be considered and
rules to be applied, and is not intended to represent
portions of a syntactically correct CLIPS program.
AICPA Codification of Statements on Auditing
Standards AU §315.01 to §315.23
REDUCING RISK IN THE ENTERPRISE: Proposal for a Hybrid Audit Expert System
3.3 Continuing Clients
Many audit firms evaluate existing clients annually
to determine whether there are reasons for not
continuing to do the audit. Previous conflicts over
such things as the appropriate scope of the audit, the
type of opinion to use, or professional audit fees,
may cause the audit firm to discontinue association.
The audit firm may also determine if the client lacks
integrity and therefore should no longer be a client.
If the client files a lawsuit against an audit firm or
vice versa, the firm cannot do the audit. Similarly, if
there are unpaid fees for services performed more
than one year previously, the CPA firm cannot do
the audit. To do an audit in either of these
circumstances violates the AICPA’s Professional
Conduct Rules on independence.
Even if none of the previously discussed
conditions exists, the audit firm may decide not to
continue doing audits for a client because of
excessive risk. Just as for new clients, excessive risk
for a continuing client is when acceptable audit risk
(AAR) is above the audit firm’s threshold. For
example, a CPA firm might decide that the client’s
tax position vis-à-vis changing IRS regulations gives
rise to considerable risk of regulatory conflict
between the IRS and the client, which could result in
financial failure of the client, and ultimately lawsuits
against the CPA firm. Even for a profitable
engagement, the risk may exceed the short-term
benefits of doing the audit.
Investigation of new clients and re-
evaluation of existing ones is an essential part of
deciding acceptable audit risk. Assume a potential
client in a reasonably risky industry, where
management has a reputation of integrity, but is also
known to take aggressive financial risks. If the CPA
firm decides that acceptable audit risk is extremely
high, it may choose not to accept the engagement. If
the CPA firm concludes that acceptable audit risk is
high but the client is still acceptable, that is likely to
affect the fee proposed to the client. Audits with a
high acceptable audit risk would normally result in
higher audit costs that will be reflected in higher
audit fees.
(assert (AAR-value “low” ))
(assert IRS-regulation)
(assert tax-position)
(bind ?industry “risky” )
(bind ?client-integrity “high”)
(bind ?management-aggressive
(defrule accept-or-reject3
(<> tax-position
(assert (AAR-value “very
(defrule accept-or-reject4
(= ?industry “risky” ) ( =
?client-integrity “high” )
(= ?management-aggressive
(assert (AAR-value “high”)))
(defrule accept-or-reject5
(= ?AAR-value “low”)
(assert (accept-client)))
(defrule accept-or-reject6
(= ?AAR-value “high”) )
(assert (accept-client))
(assert (increase-fee)))
(defrule accept-or-reject7
(= ?AAR-value “very-high”)
(assert (reject-client)))
One of the most important steps in expert system
development is the validation of the decision model
and domain boundaries. Typically, validation
entails comparing outcome measures between the
computer model and that of the experts used during
the knowledge acquisition process. If differences
are found, developers must fine-tune the model to
reflect accurate representation of the expert’s
knowledge. The next step in the process adds
additional expert opinion to address the same
questions and outcomes. Outcomes are compared
and if any significant variations are found between
the original expert opinion and the secondary
experts, information is sought to explain the
differences. The validation process occurs through
two methods. The first method is using a statistical
approach to analyze judgment outcomes. The
second approach is tracing the process for
understanding the sequence and relationships.
Process tracing is used to capture the outcome of a
judgment leading to the outcome used in the
problem domain. Each of these methods generally
explains significant levels of variation and yield
interesting results. However, the methods do not
provide satisfactory explanations for the observed
differences. Understanding the domain judgment
requires more than knowledge. To generate possible
hypotheses, to recognize potential relationships, and
to create strategies for using the knowledge, a
greater understanding of the domain is necessary.
Four integral steps are used in the proposed
Expert System Research Approach (ESRA):
knowledge acquisition, knowledge representation,
computational modeling, and theory validation. In
the first two stages, the researcher identifies and
organizes the many processes that constitute the
domain expertise. The modeling stage then takes
this expertise and translates it into a computer
compatible representational framework. Once
confidence is achieved in the computer model
adequately emulating the experts’ process, the
theory validation phase has begun. The computer
model is seen as a validation of the researchers
understanding of the emulated domain. Any
additional efforts to extend this degree of confidence
can only improve the representation of the experts’
judgment. This approach does not rely on a single
expert's opinion. It uses many sources of expert
domain knowledge to create the system. The breath
of the domain is critical to the depth of
understanding achievable in the system [2].
Typically, the broader the domain the more shallow
representation of understanding the judgment
process is achieved. However, if the domain has a
well-established process and high degree of
governance, the domain may be more accurately
represented. Having a well-defined domain
establishes much of the necessary knowledge for the
experts to use in their judgment making. The expert
system described in this paper uses the well
established auditing process and auditing
governance to represent the domain.
Instead of focusing on a single decision that
an expert's judgment is critical for successful
outcome, the auditing expert system focuses on the
entire domain and guides the user to a logical result
based on the past audit decisions. Using a case
based system the validation of the correct outcome is
found in prior decisions made with similar domain
parameters. If the theory validation process suggests
agreement between the computer case model and the
test cases, a researcher can be reasonably assured
that the domain and judgment tasks have been
represented correctly.
This paper has set forth a design for an expert
system in the audit domain. Unlike previous audit
expert systems, this system is intended to focus
broadly on the entire financial statement audit
process and combines a case based knowledge
representation with fuzzy logic processing. The
attempt at capturing a wide domain is necessary to
support organizational decision-making. Focusing
on narrow decision points within an audit process
limits the users and usefulness of the system.
Narrow domain systems typically support only
individual decision-making. By widening the
domain, the judgment process is also widened. This
holistic approach to organizational decision-making
strives to support the audit process and the audit
organization. In addition, the case based model
allows the system to store results of multiple experts
with the firm so that as the case base grows and
knowledge increases, the quality of the decisions
made by the system will improve. This heuristic
component of the proposed expert system is yet
another significant improvement over previous audit
expert system designs.
A number of factors remain unaddressed in
this approach. The complexity and size of the expert
audit system may make it too difficult and
cumbersome to process outcomes in a useful timely
manner. Development of the actual system would
be necessary to understand the limitations of
computing power and processing for this wide of a
domain. After development, empirical testing will
need to be employed to validate the approach and
attempt to duplicate the domain. The audit domain
needs to confirm the usefulness of this type of
system. Reality of the interaction complexity of the
process steps needs to be accounted for in the
design. Most expert system use a hierarchical
decision making process. Being organizational
driven, the decision making approach needs to be
more matrix in nature.
The value of using this audit system will be
represented in many ways. The first way is the
support and assists individual auditors gain from
using the system. It may assist them in feeling more
confident about their decisions and create new
environments to learn the audit process within
without having to be in a real-world situation.
Another contribution may be in the reduction of
professional errors in judgment with regard to audit
conclusion. This is a well-documented phenomenon
and the negative results have resulted in resent
industry and government corrective action. The last
and foremost contribution of the use of this system
REDUCING RISK IN THE ENTERPRISE: Proposal for a Hybrid Audit Expert System
may be in organizations making better and more
effective decisions using the captured knowledge of
experts over time. Such a system that can support
organizational decision making by approximating
process and judgment of a human expert would be a
valuable contribution to real-world application areas.
Although this system will never replace the actual
audit firm, it may increase its ability to better service
clients and the industry, and improve on the
judgment capabilities in the audit process.
Bailey, A. D., K. Hackenbrack, et al. (1987). "Artificial
Intelligence, Cognitive Science and Computational
Modeling in Auditing Research: A Research
Approach." Journal of Information Systems (Spring):
Concepts, C. o. B. A. (1973). A Statement of Basic
Auditing Concepts. Sarasota, FL, American
Accounting Association.
Fetter, R. B. and C. McMillan (1987). Business Expert
Systems. Homewood, Illinois, Irwin.
Giarratano, J. and G. Riley (1998). Expert Systems
Principles and Programming. Boston, PWS
Publishing Company.
Hayes-Roth, F., D. A. Waterman, et al. (1983). Building
Expert Systems. Reading Mass, Addision-Wesley
Publishing Company.
Kesh, S. (1995). "Case Based Reasoning." Journal of
Systems Management: 14-19.
Lenard, M. J., P. Alam, et al. (2001). "Decision-Making
Capabilities of a Hybrid System Applied to the
Auditor's Going Concern Assessment." International
Journal of Intelligent Systems in Accounting,
Finance, and Management 10: 1-24.
Nolan, J. R. (1998). "An expert fuzzy classification
system for supporting the grading of student writing
samples." Expert Systems with Applications 15: 59-
Turban, E. (1988). Decision Support and Expert Systems.
New York, Macmillan Publishing Company.
Turban, E. (1992). Expert Systems and Applied Artificial
Intelligence. New York, New York, Macmillan
Publishing Company.
Wong, B. K. and J. A. Monaco (1995). "Expert System
Applications in business: A review and analysis of
the literature (1977-1993)." Information and
Management 29: 141-152.