FACULTY WORKING
PAPER NO. 982 ocTHB
DEC. 5
3>S
Analysis of Audit Judgment Through
an Expert System
Christopher Dungan
John S. Chandler
College of Commerce and Business Administration
Bureau of Economic and Business Research
University of Illinois, Urbana-Charnpaign
BEBR
FACULTY WORKING PAPER NO. 982
College of Commerce and Business Administration
University of Illinois at Urbana-Champaign
November 1983
Analysis of Audit Judgment Through an Expert System
Christopher Dungan
University of South Florida (Sarasota)
John S. Chandler, Assistant Professor
Department of Accountancy
This research was supported by a grant from the Peat,
Marwick, Mitchell & Co., Research Opportunities in Auditing
Program.
This is a working paper and should not be quoted or reproduced
in whole or in part without the written consent of the authors
Comments and suggestions should be forwarded to the authors.
Digitized by the Internet Archive
in 2011 with funding from
University of Illinois Urbana-Champaign
http://www.archive.org/details/analysisofauditj982dung
Abstract
The goal of the exploratory research described in this paper was
to create an interactive computer program which would function as a
dynamic, operating model of auditors' judgment and provide a tool for
investigating the domain of audit judgment. The system would be operated
in a noisy, natural setting to make credible audit decisions. The
auditor's evaluation of the adequacy of the Allowance for Bad Debts con-
stituted the functional area of interest in this study. The result of
this work was the creation of AUDITOR — a model of an audit judgment in
the form of an interactive computer program called an expert system.
This paper describes how AUDITOR was designed and constructed. The
results of two successful validation exercises are presented. A discus-
sion of the insights into audit judgment provided by AUDITOR concludes
this paper.
INTRODUCTION
Interest in the methods by which accountants make professional
judgments has stimulated considerable research in recent years. Com-
prehensive reviews of this work may be found in Ashton (1982) and Libby
(1981). Whether effected via regression modeling or focused on the
probabilistic aspects of judgment, it is generally agreed that some of
the research to date incorporates in its design, features which may
operate as limitations on the external validity of the work. That is,
typically, a relatively passive subject has been requested to perform
a highly-structured and well-defined task via a simplified response
scale in a laboratory setting while utilizing a limited set of data
with which he has been provided.
Several approaches have been used to mitigate these limitations,
for example, by introducing less-restricted search behavior. Biggs and
Mock (1980) avoided the introduction of any restriction on subjects'
cue choices by studying their spontaneous verbal protocols as they per-
formed in an experimental setting. Abdel-Khalik and El-Sheshai (1980)
allowed subjects to expand the number of factors which might be examined
by choosing their own cues from among those provided in a shopping list.
Shields (1980) allowed subjects to make choices of cues from information
boards in a study of the effect of information load upon information
search patterns.
Like the work on verbal protocols, this present study utilizes
self-reports from the subjects themselves instead of inferring a deci-
sion model from the relationships between outputs and inputs. Although
the validity of self-reports of mental processes is the subject of on-
going debate, c.f. Ericsson and Simon (1980), Nisbett and Wilson (1977),
Payne, Braunstein, and Carroll (1978), Einhorn, Kleinmuntz, and Kleinmuntz
(1979), certain aspects of the present study distinguish it from previous
work.
The goal of this present, exploratory work was to create an inter-
active computer program which would function as a dynamic, operating
model of auditor's judgment. The system would be operated in a noisy,
natural setting to make creditable audit decisions. While operating
within the confines of the available programming technology, the expert
auditors who participated in building the system, themselves, freely
determined the cues, their weights, and the form and size of the system.
The result of this work was the creation of AUDITOR — a model of an audit
judgment in the guise of an interactive computer program of a type pro-
perly identified as an expert system.
Auditors' evaluation of the adequacy of the Allowance for Bad
Debts (ABD) constituted the functional area of interest in this study.
This particular audit judgment was chosen for study and modeling for
several reasons. It was one of the judgments most frequently named by
the experts themselves when asked, "What critical audit judgment
requires the attention of an experienced professional and is never
entrusted to 'green' auditors?"
-2-
Also, when auditors make judgments of the adequacy of the ABD they
are presented at a subsequent visit to the client's office with an oppor-
tunity for a referent outcome — the "outcome feedback" which is sometimes
considered to be a necessary condition to sharpen and calibrate the
expert's skills (Harrell, 1977; Ashton, 1981). That is, at a later
date auditors can observe from the client's records which accounts
have indeed proved to be uncollectible. Such a judgment seemed to the
researchers more likely to provoke the consensus which Einhorn (1974)
felt to be necessary requisite for expertise. Thus, in the absence of
any other procedure for assuring that the decision process chosen for
study would be one which would clearly be recognized as demonstrating
expert judgment, the researchers believed expert judgment in auditing
would most likely be fostered where the judgments were made under con-
ditions which provided the judges with the possibility of objective
outcome feedback. Finally, this area of -judgment gave promise of trac-
tability in initial interviews with the experts, in that they seemed
able to recite the cues to which they attend when making the judgment,
and the cues seemed similar among several experts.
EXPERT SYSTEMS
Expert systems are computer programs which offer consultative
advice in a bounded knowledge domain on a level of competence often
rivaling that of a human practitioner who is recognized as an expert in
his field. Such systems are constructed by a system-builder utilizing
the active collaboration of one or more of the experts themselves, who
are challenged to justify the manner in which they make their iudgments
These systems are now providing valuable consultative advice to clients
in real-world settings (Michie, 1980).
Expert systems trace their lineage to attempts in the 1950s to
create computerized problem-solving routines of broad generality, for
example the Logic Theorist program of Newell, Simon, and Shaw (1958).
This program and later generalized problem-solvers such as ARGUS
(Reitman, 1965) and GPS (Ernst and Newell, 1969) made contributions
to subsequent work but did not themselves display great breadth of
achievement. More recent work proceeds from a belief that the high-
level skill demonstrated by a human expert derives from his accumulated
experience of a concentrated nature which enables him to perform opti-
mally in situations to which that experience is pertinent but is of
little help in genuinely novel situations (Simon, 1978).
Although frequently applied to programs of medical diagnosis or
treatment, expert systems' technology appears to be appropriate wher-
ever evidence of less than certain reliability must be evaluated by a
skilled practitioner according to experientially based rules in order
to make a judgment. Thus, the field of auditing bears similarities
to domains already approached via the technology of expert systems.
Comments of workers in the field of "knowledge engineering," as it is
sometimes called, support this analogy.
-3-
The domain is one in which diverse factors must be
identified and synthesized to form judgments,
evaluate alternatives, and make decisions. Years
of experience are brought to the problem at hand;
experience and subjective judgment play a major
role. The domain is not easily amenable to precise
scientific formulation. (Duda et al . , 1979)
The domain lacks a strong mathematical structure, is
incorrigibly non-numerical, and is too complex for
adequate analytical specification. (Michaelson, 1982)
The knowledge which the expert brings to the task is
largely heuristic knowledge, experimental, uncertain —
mostly good guesses and good practice in lieu of
facts and rigor — much of this private to the expert.
How else explain internships of guild-like apprentice-
ship to a presumed master of the craft? What the
master really knows is not written in the textbooks.
(Feigenbaum, 1979).
Examples of successful, currently operating expert systems are
frequently reported in the popular press as instances of "artificial
intelligence." For example, MYCIN infers disease identity in blood or
meningitis infections and suggests antibiotic treatment (Shortliffe and
Buchanan, 1975). PROSPECTOR evaluates core samples to infer the pre-
sence of significant mineral deposits (Duda et al. 1979). AL/X diagnoses
causes of shutdowns occurring under the control of automatic safety
devices on oil production rigs (Reiter, 1980).
AL/X (as did MYCIN) included in its design features the goal of
separation of domain-specific knowledge from the computer control pro-
gram which utilizes the knowledge. Thus, while AL/X was originally
expert in only one realm of application its inference structure provides
a foundation upon which can be built expert systems in other domains.
AUDITOR utilizes the inference structure and control capabilities of
AL/X, described in more detail in Appendix [A] . For a fuller treatment
of its capabilities see Dungan (1983)."
AUDITOR FROM A USER'S PERSPECTIVE
The AUDITOR model was constructed in three distinct phases: initial
modeling, refinement, and validation, all of which will be described in
the course of this paper. However, certain aspects of this process of
system-building can best be understood after the system has been de-
scribed in use.
When a user operates the system interactively at a computer terminal
for purposes of consulting AUDITOR for its advice concerning the large,
delinquent receivable (which is the object of this inquiry), the
-4-
system asks him a series of questions about the extent of his personal
knowledge of the presence or absence of the evidence cues required in
the rules. AUDITOR begins each of its queries with the phrase, "How
certain are you..."? which is prefaced to each of the rules in the
rule base. For example, one query will be worded, "How certain are
you that recent collections toward the delinquent portions of this
account are proceeding satisfactorily?" The user's response provides
a means for the system to acquire data about the underlying cue,
suggested by the experts, which in this case concerns the debtor's
recent payment performance.
The user responds to the system's query by typing a numeral called
a certainty value (CV). CVs represent subjective certainty (or uncer-
tainty) on a scale from -5 to 5. The response, "5," is given by the
user of the system when he is certain that the matter referred to in
the query is true. When he is certain the matter is false his response
is "-5." In effect, the CV scale from .1 to 5.0 in absolute terms
(unsigned) communicates increasing amounts of certainty, to which a
positive sign is attached for truth and a minus sign for falsity. When
the user has no information about the matter, he is unable to judge its
truth or falsity, or he feels the question is irrelevant, he responds
with "0." The researchers did not investigate individual differences
in the use of this scale, although the process appears somewhat akin to
the extraction of a utility curve, c.f. Newton (1977).
AUDITOR utilizes the data in the user's response in two ways.
First, it updates the current strength of belief of the hypothesis,
which in AUDITOR is stated in degrees of belief (DB), (DB = 10*logl0 (Odds))
The nature of the updating depends upon the nature of the links between
the rules and the hypothesis: Bayesian (for IF: THEN links), Logical
(for AND, OR, NOT), or contextual (adaptable for unique linkages).
Second, the system inspects its rule base to determine which one of
the remaining questions to ask next. It does this by following criteria
built into the AL/X control (unless they have been modified by contex-
tual links). The criteria are simply stated: Ask that question next
which could possibly impact greatest upon the DB of the hypothesis , con-
sidering the inventory of questions, their diagnostic links, the current
DB of the hypothesis, and the CV responses which might be given by the
user. Thus, AUDITOR'S inquiry proceeds in economical fashion down that
line of questioning which could have the maximum impact upon the hypo-
thesis .
This process of query, response, and update continues until (1) all
questions in the system have been exhausted, (2) the user terminates
the session, or (3) the updating process has caused the DBs of the hypo-
thesis to reach a threshhold level beyond which further questioning
could cause no significant improvement, according to criteria adjustable
within the system. In any of these cases, the system can then report
its conclusion, which constitutes its "expert judgment" in the form of
a statement of the current degrees of belief to which the hypothesis has
been updated. In a typical consultation with AUDITOR this report might
read:
-5-
The delinquent portion of this account should speci-
fically be reserved-for in the allowance for bad
debts to a substantial extent: RESERVE. Prior degree
was 0.0, current degree is -19.5. At this point this
goal is probability .05 or less.
This report informs the user of the system that the data provided
by him has changed the degree of belief of the hypothesis from its ini-
tial state, 0.0 (.5 probability), to -19.5 (approximately .013 proba-
bility). At this point the user must decide for himself whether or not
-19.5 DB (.013 probability) dictates to him a course of action, since
there is insufficient experience with AUDITOR to provide guidance to
relate these DB and probabilities definitively to decision points in an
expert's judgment process.
CONSTRUCTION OF THE SYSTEM
A model of expert judgment, constructed in the form of an expert
system, is assembled from components consisting of:
One or more hypotheses stating the judgment which is
the end product of the expert's reasoning process.
Rules which express the relationship between the
evidence (cues) and the hypothesis.
Parameters which express (1) the diagnostic value
or impact which the expert believes are implied by
each cue, and (2) the strength of the expert's
beliefs prior to examination of any evidence, that
is, the initial state of the system.
These components were assembled and tested during three stages:
initial modeling, refinement, and validation. Initial modeling encom-
passed development of the hypothesis which is the goal of the system,
interviews with the auditing experts to elicit cues, conversion of
these cues into rules, and polling of the experts to elicit parameters
adequate to construct a preliminary, yet operating, system. In the
refinement stage, experts operated the model interactively on a com-
puter terminal and presented their suggestions for improvements to be
integrated into the system. These improvements consisted primarily of
changes in parameters, additions of interactions between the rules, and
a few instances of rewording of the cues. Finally, the completed sys-
tem was exposed to validation by testing it to learn the extent to which
a different group of experts would similarly acknowledge it to be expert
in its performance. (Further references within this paper to "experts"
will mean the several auditors who participated in the several stages
of building, refining, and validating the AUDITOR system.)
-6-
INITIAL MODELING: HYPOTHESIS
Expert systems in other knowledge domains are constructed under a
belief that an expert's information search and processing are focused
upon and organized around hypotheses, one or more of which ultimately
will express the judgment which he will render upon the data. There
is speculation that the predecisional behavior of accountants and audi-
tors is similarly organized around hypotheses (Libby, 1981). After
selection of the judgment area for modeling the next step in the con-
struction of AUDITOR consisted of formulating an hypotheses to reflect
the expert's judgment. In an expert system this judgment is referred
to as the goal hypothesis.
AUDITOR'S single, goal hypothesis is expressed, "The delinquent
oortion of this account should specifically be reserved for in the
allowance for bad debts to a substantial extent." (For ease of refer-
ence and recognition, the name of this hypothesis, RESERVE, as well
as the name of each rule in the system will be written in capital
letters.) The phrase, "this account" refers to the one individually
large account — or in some cases a single invoice — which is under scru-
tiny by the auditor. "To a substantial extent" was agreed among the
participants of this project to mean all cases in which they judged
recovery is likely to be no more than an amount considered insignifi-
cant in relation to the delinquent balance.
RESERVE is intended to reflect the approach which the experts
testified they visualized when scrutinizing the individually large,
delinquent accounts of a commercial, audit client. It can be labelled
a "worst-yet-possible-case" approach. This approach envisions that
each, individually large account (or invoice) be judged collectible or
uncollectible. The magnitude of the client's allowance for bad debts
(ABD) is then considered adequate or inadequate in comparison to the
aggregate of, the large, delinquent accounts which have been judged un-
collectible.
DEVELOPMENT OF THE RULE BASE
The rule base for AUDITOR was developed from the cues provided by
four practicing auditors chosen from among the eight who were consulted
initially when choosing an area for modeling. Those actively partici-
pating in the bulk of the project were three managers and one senior
on the audit staff of a CPA firm of international scope. (By compari-
son, for the development of expert systems in other fields the whole-
hearted cooperation of one practitioner who is clearly expert is often
considered adequate.)
In individual interviews these auditing experts recalled and named
the cues to which they said they attend when evaluating the collecti-
bility of a client's delinquent receivables. For example, one cue
-7-
relates to whether the debtor continues to be an active customer. Other
cues relate to the opinions of the client's credit manager as to collec-
tibility of the delinquent account, response to confirmation requests,
and contents of the client's credit file on the delinquent customer.
A total of twenty-five such cues were obtained, differing widely among
themselves in the diagnostic impact which they have upon the auditor's
conclusion. The complete list may be examined in Dungan (1983).
After eliminating from the list obvious redundancies and a few
errors in transcribing, the researchers then transformed each of the
cues into a simple rule in IF:THEN form. For example, a cue named by
one of the experts was: "The customer's stated intent regarding pay-
ment." This became the rule called NOTPAY: "The customer has stated
his intent to pay little or nothing of the delinquent account," which
is processed by AUDITOR utilizing a Bayesian revision as IF, NOTPAY: THEN
(to the extent determined by parameters), RESERVE. The connector AND
was later added (during the refinement stage of the work) in order to
provide linkages between rules said by the experts to be interrelated.
The connector NOT was used occasionally to achieve a more natural word-
ing of a rule. (Although available for use through AL/X, the connector
OR was not used. )
PARAMETERS
Next, parameters were developed for the rules — the diagnostic
weights which express what the experts believe to be the evidential
importance of each of the cues. This step began with a polling of the
four auditing experts. Each of their rankings of "strong," "moderate,"
"weak," and "no effect" for each of the rules were translated into a
quantitative scale on an expedient and preliminary basis. For example,
all four experts rated the impact of NOTPAY as "strong." These pre-
liminary values were later changed to the values contained in the final
version of AUDITOR when the experts themselves operated the system and
offered criticisms aiding in its refinement.
AL/X expresses the diagnostic impact of the evidence in the form of
Positive Weights (PW) and Negative Weights (NW). As the AUDITOR system
is operated interactively, the degree of belief attached to the goal
hypothesis (RESERVE) is incremented by a value which is a function of
the PW_ and the Certainty Value (CV) of the user's response if the user
has keyed-in a positive CV in answer to the system's query, and upon the
NW and the CV when the user's response is a CV in the negative range.
The role of PW and NW are explained more fully in Appendix [A] .
REFINEMENT
Armed with a basic set of rules and parameters the researchers
assembled a working model of AUDITOR with the aid of the AL/X software
package. For purposes of refinement, this working model was returned
to the experts to secure their suggestions for improvements. As a
-8-
result of this refinement stage, major changes were made in the PW and
NW, numerous interactions were added, and a few changes in the composi-
tion of the rule base were found to be desirable. An example of these
changes, the rule called WORKOUT was originally developed from a cue
which expressed the auditor's interest in a delinquent debtor's perfor-
mance toward fulfilling a "workout alreement" or similar negotiated
understanding between the client and the debtor. Before refinement the
rule was stated, "Payments are being received currently under a workout
agreement." As a result of reactions received during refinement of the
system, this rule was restated to read, "Recent collections toward the
delinquent portions of this account are proceeding satisfactorily," in
order to eliminate the restrictive nature of the reference to "workout
agreements . "
Also at this stage of the work interactions in the form of AND
statements were added to the rule base upon the suggestions of the
experts. For example, the individual effects of the rules called LEGAL
and NOTPAY are enhanced when both of these cues are found present at
the same time. That is, when a delinquent debtor has stated his intent
not to pay (NOTPAY) and the same debtor also presents a counterclaim
which would appear to make legal action fruitless (LEGAL), the AUDITOR
system operates by implementing each of these rules individually and
also implements a third rule, LEGAL&NPAY, which carries its own impact
(in the form of PW and NW) upon the degree of belief in the goal hypo-
thesis .
After interactions had been inserted, to accomplish further refine-
ment the experts operated the system in "full trace" mode (so called).
Although this mode was slow in operation, even tedious, it allowed the
experts to observe the effect upon the hypothesis which might be caused
by alternative possible responses of a user, thus stimulating them to
suggest changes in the values of the parameters, PW and NW. The goal
sought by these changes was to calibrate AUDITOR to the extent that a
probability of at least 90%, (DB 9.5) should be reported by AUDITOR at
the time of the expert's own report of a subjective feeling of "satis-
faction" in his judgment, despite the difficulties with this proce-
dure which might be predicted by a review of recent research, c.f.
Lichtenstein, Fischoff, and Phillips (1977), Crosby (1981), and others,
summarized both in Libby (1981) and in Ashton (1982). Although consensus
was found among the experts in the direction of the changes which they
suggested, the model proved somewhat insensitive to the exact values to
which che changes were made, a result reminiscent of the work of Dawes
(1975) and of Einhorn, Kleinmuntz, and Kleinmuntz (1979), a matter which
will be explored in a subsequent paper. The complete set of rules and
parameters utilized in the system may be examined in Appendix [C].
VALIDATION
The results of AUDITOR'S operations were compared against the
judgment of practicing auditors utilizing client audit data contained
in work papers. Two validation procedures were employed: "Open-book"
and "Blind."
-9-
OPEN BOOK: The validation procedure referred to as "Open-book" was
performed by an audit manager in a different office of the same CPA firm
which participated in the building of the system. This auditor, serving
as validator, selected work papers covering completed audits of two com-
mercial clients with which he personally was not involved. From the work
papers of each of the two clients the validator chose five individually
large, delinquent accounts. The criteria which he used for his selec-
tion was not communicated to the researchers. Each set of work papers
to which he referred contained the decisions made by the auditors during
the audit about their assessment of the collectibility of these five
delinquent accounts (and others). Since the validator was unfamiliar
with the two clients which he had selected, the work papers presumably
also contained all the information from which he determined his responses
to AUDITOR'S questions. Since to the validator it was obvious which
set of judgments appeared in the work papers and which "judgments" were
produced by AUDITOR, this procedure is referred to as an open-book
validation.
After a demonstration of operation of the system by the researchers,
the validator invoked the AUDITOR program and responded to its ques-
tioning by entering information which he gathered from the work papers.
That is, he responded to its inquiries by entering a Certainty Value on
a scale from -5 to 5 in answer to each question presented to him follow-
ing the prefix, "How certain are you...?"
In the case of each delinquent account, the validator himself
decided when to exit from the system, thus concluding one session with
AUDITOR. Presumably, that occurred either when he believed that he had
provided to AUDITOR enough data by which he personally could have made
a decision concerning collectibility of each account, or when all data
available in the work papers had been entered into the system.
At the conclusion of each session the validator relayed to the re-
searcher the judgment of the CPAs who had performed the audit. He also
examined AUDITOR'S report of degrees of belief and probabilities which
had been produced as a consequence of the interactive question and answer
session. On the basis of his own criteria he labeled each result a "hit"
(appropriate result) or a "miss." (In every case he commented that his
own judgment regarding the write-off of the delinquent account agreed
with that reported in the work papers.) The results are summarized in
Table 1.
RESULTS: In each case but one, when the validator chose to ter-
minate questioning and call for the report by the system, AUDITOR
reported a degree of belief which was the equivalent of a probability
of at least .865 in favor of the same decision as that called for by
the original audit team — that is, either in favor of, or against, the
need for an allowance for the delinquent account under scrutiny. The
validator and the researchers considered these results to be successful.
BLIND VALIDATION: A second validation procedure was carried out to
conform to the suggestions made by the mathematician A. M. Turing (1950).
a
•-4
CC
re
03
—
c
L
E
c
iH
0)
L
c
<2_
•H
•rJ
73
«
cu
co
4)
g
•1-1
4->
M
<
o
Li
u
o_
0)
o
c
u
c
c
0)
2
OS
c
H
1) C
<
z
c
I— I
E-
<
<
>
C/3
Cu
C/2
—
a
co
Ed
os
si
<
co
3 < 3
L
>> 03
<4-i J2 4J
O -H
73
• cu 73
o .* 3
z o re
>
c
>,
oi
-C
CO
73
r-"»
o;
E
2
u
0)
i-i
—
jj
c
CO
03
D«
>.
c
01
CO
•r^
:*:
CC
a:
be
GO
^-v
<
01
H
jj o u
•H 3 >
--i < a!
03
u
a.
0)
c
>
c
u
o
Ol
CU
4)
«
>
>
03
m
o>
u
L
OJ
i— )
>
0)
0)
od
<o-
Li
03
03
0)
0>
a
o
li
03
PA
os
c
>
0)
o
L
o3
c
c
•>
0)
z
z
en
03
0)
a!
sC
CJ>
C7*
I
a
X
O
c
c
c
CM
O
■H
73
-
•« 0!
•H
4-1 4-1
c o
L
««"«»
0> 0!
CJ
— •
3 <— I
03
—1
a- i-i
01
•H
c o
e
s:
•i- o
CO
0) >,
T3 >H
4J
c c
O 0)
f^ CU
o
o
o
<r
<o-
u_
C
OJ
cr
c
•H
a> c
73 >H
3
C C
o
o ^
• CCJ
X J-i
r^ o
CO- 4-1
m
03
0
o
bC
o
c -a
C 0)
• tc
in re
■H E
<j> re
j-> L
c o
3 n-i
o
CJ E
CJ -H
re re
l-H CJ
re
0)
C 3
O CT
C c
in cy
co- -c
0)
>
u
o
03
01
06
CN
in
oo
i->.
av
vD
CTn
CTN
&
o
CT>
01
J3
CJ
Li
re
01
CO
cu
J-I
•a
cu
73
L
o
u
0>
u
0)
73
co
CC
t-i
0)
CJ
u
re
cu
03
01
-
-c
Ui
c
o
cu
L
o>
T3
0
Z
c
re
c
CJ
E
C
•H
3
cr
re
0)
01
t:
0)
bC
Ui
re
c
01
CJ
X
0)
Ol
4-i bO
3 Li
o> re
U £.
L CJ
CN
* <
4-1
c
CU
0)
>
Li
01
03
Ol
OS
o
z
Oi
L
01
03
01
OS
CU
>
u
01
03
0)
OS
o
z
0)
>
Li
01
03
01
CN
in
o>
in
cc
en
cr
en
cu
CJ
c
re
re
J3
C
01
3
cr
01
CO
X3
3
03
XI
CU
u
cj oi
0> -u
^^ re
c
CJ J-l
0)
i— I cu
03
03
0)
c
•H
03
3
J3
cu
< 03 O
en
cr
c
01
c
c
c
<o-
c
3
L
C
re
03
Ol
Q
m
u X
C I •
0> Li U *C
03 re —• cu
Oi C 3 £
Li 03 CJ
o. co cu re
CU -H L CU
L x. U
re; -h jj
3
O
3
C
c
os re
o
H 3
Q -H
33 -a
< -H
>
c o
■H L
a
.c c
ju CU
E
u- be
O TD
3
Li '—
O
> u
re j3
C 03
CU
03
3 w-<
cu n
i—i CU
re cu
o c
03
cu
_ _
CU
i-i 03
i— I CU
0) 03
>
0 re
03 E
cu cu
CU 4-1
U 03
&C >^
CU 03
TJ
: as
O
OJ E-i
— —
33
E <
O
U 01
U-l £
4-1
T3
CU 03
4_l C
1 re
cu cu
> E
c
C 4-1
CJ 03
c
03 t-
o) re
._i tc
4-1 <
-J C»
■H C3>
X •
re
£} •>
O 03
L 3
a j=
E-
cu
s_
re •
03 4J
i-i c
re i-i
E re
L
&-S o
cr o.
o^ cu
L
a
3 :
C
03 C
^ -H
CJ CO
re 1-4
4-1 O
03 CU
CJ
B 03
0) -
4-1 E
03 U
>, -H
CO Cu
0> <
J3 CU
4-1 C_3
C T3
01
L 73
oi re
CO cu
3 J3
01 3
J3 E
4-J 3
i— (
>> c
x a
73 01
Ol j3
X E-
03
•H
3 •
L >>
3 !-i
u-i re
CO
CO CO
re cu
CJ
» cu
CU 3
CJ
X
■• c
4J 3
c
3 4-1 •
o re co
CJ J3 L
CJ 4-i 0
re 4J
•> i-i
L 03 73
re tj 3
—i co
CJ
cj u cj re oi
o a vi ^ i
~ CJ 4-1 4-1 4J
-11-
This procedure is considered to be a more stringent test of the validity
of an expert system and has become somewhat of a standard in the field
(c.f. Yu, 1979). Turing, when confronted with the issue of whether com-
puters and their programs could be said to think, proposed instead that
the testable question is whether an observer who is himself ignorant —
that is, blind — as to the identity of the source could distinguish
between the output of the machine and that of a human.
Blind validation of AUDITOR utilized two human auditors, one who
served as user of the system, the other as validator. Both auditors
were managers with an international CPA firm not previously involved in
the AUDITOR project. The auditor who performed as user selected an
audit client with which he was familiar. His choice was a NYSE listed
company which manufactures and markets through various distribution
channels consumer products such as cigarettes and candies. His audit
team, in their visit to the client's offices at an interim date, had
identified by computerized selection eleven invoices so large and
delinquent as to justify the auditors' individualized attention. The
criteria for this selection was not communicated to the researchers.
The user was familiarized with the AUDITOR system and then in-
structed to respond to its questions using the data contained in the
work papers. Since he himself had been a member of the audit team it
is not possible to determine if his answers were based solely upon the
contents of the work papers or perhaps also relied upon his memory for
information not in the papers. He answered the system's questions,
responding on the CV scale from -5 to 5. As in the first validation,
the user himself decided when to terminate each session. Concurrently,
he reported relevant data from the work papers to the researcher who
recorded it, such as age, dollar balance, confirmation results, nature
of customer's business, etc. This data — recorded on Fact Sheets — became
part of the raw data presented to the validator from which he made his
judgment. The user also reported the audit teams' judgment, just as it
was recorded in the work papers. This information was recorded by the
researchers on a Comparison Worksheet, on which was also entered the
"judgment" of the AUDITOR system.
Subsequently, the audit manager serving as validator was presented
with the Comparison Worksheets which contained the judgments from the
two sources and the Fact Sheets containing the data from which they
made their judgments. (Of course, the sources of each judgment were
disguised when presented to the validator; tbat is, he was "blind" as
to the source — auditor or AUDITOR — of each judgment.) The task which
he undertook was to study the same evidence on the Fact Sheets which
had been available to the Fact Sheets which had been available to the
two experts, auditor and AUDITOR, to form his own independent and pre-
sumably expert conclusion, and then to accept or reject each judgment
from each source on the basis of his opinion of the expertise which each
had demonstrated. That is, he accepted or rejected each of twenty-two
recommendations which were disguised as to source.
-12-
RESULTS: The validator accepted all of the audit team's judgments
and all but one of AUDITOR'S. These results are summarized in Tables
2 and 3. In sum, over the two validation procedures, open-book and
blind, the AUDITOR system scored nineteen hits out of twenty-one cases.
ANALYSIS
TAXONOMY
In AUDITOR a high score on one cue can offset a low score on a
different cue. Thus, AUDITOR can be called a compensatory, rather than
non-compensatory model. For example, credit to the delinquent customer
may not have been stopped by the client (CREDITSTOP) , and the customer
may have demonstrated a good record in the past of paying his account
(GOODRECORD) . However, both of these cues, favorable to an expectation
of collectibility of the account, may be offset — compensated for — by
evidence that the customer is in bankruptcy proceedings (BANKRUPTCY)
and the client's lawyer advises that recovery is unlikely (LAWYER).
AUDITOR incorporates probabilistic aspects of decision making, as
do other successful expert systems, recognizing that no evidence is
perfectly diagnostic of the condition which the expert must evaluate
and that real-world judgments are made on the basis of information
which must be probabilistically evaluated. The query format, "How cer-
tain are you...?" reminds the system's user to perform a subjective
assessment of the reliability of the evidence.
In contrast with Lens Model studies, and indeed with most studies
of auditors' judgments except those utilizing Verbal Protocol Analysis,
AUDITOR processes the independent variables with which it operates, in
the form of user's CV applied to the PW or NW of a cue, in a sequential
fashion. That is, AUDITOR continuously revises and expands the system
as its queries are successively put to and answered by the user. In
sum, AUDITOR incorporates subjective assessment of the evidence through
what is basically a Bayesian revision process operating on cues which
can compensate for each other. Therefore, the AUDITOR system can be
called a subjective, Bayesian, compensatory expert system, which func-
tions as a sequential model of auditors' judgment.
COMPARISON WITH OTHER STUDIES
Two previous studies have investigated auditors' perceptions of
what was called source reliability (Joyce and Biddle, 1980) or source
credibility (Bamber, 1980, quoted in Libby, 1981). Joyce and Biddle
explored whether auditors weighted differently the reliability of infor-
mation depending upon its origin with the client's credit manager or
with an outside credit reporting agency. Joyce and Biddle concluded
that unless both sources were called to the auditors' attention (as was
done in a within-subjects experiment) they did not differently weight
the sources. In the AUDITOR project no rule incorporating outside credit
reports was suggested by the experts, and apparently none was referenced
-13-
TABLE 2
CONVERSION OF DEGREE AND PROBABILITY INTO VERBAL JUDGMENTS
AUDITOR'S Reported
AUDITOR was Degree of Belief AUDITOR'S Result
Concealed as and Equivalent Was Reported to
Case # Expert No» Probability Validator as
"OK, no res. nor adj."
"No reserve nor adj .
needed"
"Yes , reserve or adj .
req 'd"
"No"
"No"
"Yes"
DB
%
3-1
II
-15.5
3.2
3-2
I
-19.5
1.3
3-3
II
15
97.0
3-4
I
-30
.01
3-5
II
-17
2.0
3-6
II
26
99.0
3-7
II
2
61.0*
3-8
I
-30
.01
3-9
II
-30
.01
3-10
I
-30
.01
3-11
I
33
99.9
'At most a partial res
or adj . req 'd"
"No"
"No"
"No"
"Yes"
*There were no clear guides for interpretation of the implications of
AUDITOR'S result at this level of likelihood (61% probability of need
for an allowance). Since the first validation procedure could be
interpreted as suggesting that partial allowances by auditors might
be associated with the AUDITOR'S results at this level of probability,
the researcher chose to state the expert's opinion as "At most a par-
tial res. (reserve) or adj. (adjustment) (is) req'd (required)".
-14-
TABLE 3
SUMMARY RESULTS BLIND VALIDATION
Case
#3- Brief Description
1 $19,292 over 90 da.
active gov't account, not
confirmed.
2 $17,553 over 90 da.
active gov't account, not
confirmed .
3 $16,346, 30-60 da., 100%
credit memo not posted. Not
confirmed.
4 $35,909, 30-60 da., not
confirmed. Workout agreement
in file.
5 $59,570, 30-60 da., not
confirmed. Bankrupt but letter
of credit in file.
6 $13,563, 30-60 da., not
confirmed. Customer request
to return goods, CM. not
posted.
$11,628, over 90 da.,
confirmed newer items no excep-
tion. Creditmanager calls
excellent account, but dispute
over price.
8 $12,498, 30-60 da., not
confirmed, later collected.
9 $14,655, over 90 da. not
confirmed, later collected.
10 $22,024, over 90 da. not
confirmed, later collected.
11 $36,763 delinq. and total.
Letter in file: "Financial
difficulty, no pay next six
months ."
Judgments
AUDITORS Auditor
Validator
Called
No reserve
Accept
No res.
A
No res.
A
No res.
A
Re s . or
adjust
Reserve
A
A
No res.
"
A
No res .
A
No re s .
A
No res.
A
Res . or
adjust
Res . or
adjust
A
A
At most,
partial
Reject
reserve
or adj .
No res.
A
No res .
No res.
A
A
No res.
No re s .
A
A
No res.
No res.
A
A
Yes, res.
:es, res
-15-
in the work papers — which covered large, commercial clients — consulted
by the auditors during validation of the system. Therefore, no direct
comparisons with Joyce and Biddle are possible.
However, several other outside data sources were included in rules
suggested by the experts involved in building AUDITOR. Among these
are PROBLEMS, relating to the results from confirmation requests,
LAWYER, concerning the opinion of outside legal counsel, NOTPAY, which
reports the debtor's stated intent not to pay (when that is known),
and NONCONTACT, which is a cue activated by the client's and the audi-
tors' inability to contact the debtor. These cues carry four of the
five heaviest, basic weightings (PW or NW) in AUDITOR. It seems
reasonable to believe that these relative diagnostic weightings carry
within them a factor which is dependent upon the experts' belief in
the reliability of the source of the information. Thus, these heavier
relative weightings evidently reflect the experts' belief in the
higher reliability of these outside data sources.
Additionally, the low diagnostic weight accorded to the rule
CREDITMGR, which reports the opinion of the client's credit manager,
seems to indicate that such a source, although frequently resorted to,
carries low reliability. Bamber's 1980 study attempted to determine
if audit managers differentially weighted the results of work done by
audit seniors who were described in the experiment as being of dif-
fering reliability. No comparison with the AUDITOR project is
possible, since, as might be expected, no cues suggested by the
experts focused on the competence of the audit staff.
RELATIVE IMPORTANCE OF THE RULES
Cues frequently used which also have heavy diagnostic weight are,
by implication, the major diagnostic tools of the experts. Table 4
lists the cues in decreasing order of their importance in use as diag-
nostic indicators during the validation trials. This table has been
prepared, first, by totaling for each cue the largest of its Positive
Weight or Negative Weight together with the weights of the interactions
involving that cue (all in unsigned terms). Then, this combined weight
has been multiplied by the number of uses occurring during the valida-
tion trials. (Cues with zero usage during validation — implying that
no evidence concerning the cue was present in the working papers — are
excluded from this table.) The result provides a measure of the extent
to which the system utilizes the different cues.
Some comments are appropriate concerning the frequency of usage.
The figure for number of uses is a count of the number of times during
validation in which a non-zero Certainty Value constituted the response
by the system's user. Since twenty-one delinquent accounts were exa-
mined during the validation runs, that number is the maximum which could
appear as number of uses. The cue called COLLECTED serves less as a
diagnostic tool for the experts than as a means within the AUDITOR
system to identify which accounts continue to be doubtful, following
-16-
TABLE 4
CUES, THEIR WEIGHTS AND NUMBER OF USES, AS DIAGNOSTIC INDICATORS
Combined
Largest Weight No. of Combined
Unsigned W/Inter- Non-Zero Weight X
Cue PW or NW Actions Uses Uses
COLLECTED (Account is no longer
delinquent by audit comple-
tion date) 30
ECONOMIC (Economic factors hamper
ability to pay) 3
BANKRUPT (Debtor has entered bank-
ruptcy proceedings) 3
CORRESPOND (Recent data in credit
file supports collectibility) 4
ACTIVE (Debtor continues as active
customer) 3
WORKOUT (Recent collections are
proceeding satisfactorily) 3
NOTPAY (Debtor expresses his intent
not to pay) 5
PROBLEMS (Confirmations reveals
serious problems) 8
CREDITMGR (Client representative
expresses strong belief in
collectibility) 2
LAWYER (Lawyer expects poor prospects
of recovery) 5
LEGAL (Debtor has strong counterclaim) 3
WRITEOFF (Writeoff of this account
represents a material adjustment) 2
COLAGENCY (Collection has been
assigned to agency or lawyer) 2
G00DREC0RD (Debtor has good record
of paying in past) 2
NOPAYEVER (Debtor has made no pay-
ments on any invoice) 2
NONCONTACT (Confirmation request
was returned undeliverable) 4.5
30
19
19
11
15
20
14
11
21
9
11
10
9.5
13
13
600
171
152
143
105
78
70
55
45
42
36
22
21
20
20
9.5
-17-
their initial identification and listing as delinquent, perhaps during
an early stage of the audit.
Rules with the highest usage reflect aspects of the auditors'
apparent decision processes and usage of evidence. WORKOUT reflects
the significance to the auditor of his knowledge of recent cash collec-
tions from the delinquent debtor. The phrase, "proceeding satisfac-
torily," in effect asks the auditor to evaluate the recency, regularity,
and adequacy of the collections in comparison to what he judges is some
acceptable standard. Because of the crucial nature of the sales and
collection cycle, the inspection and testing of subsidiary accounts
receivable ledgers and cash receipts records normally would be part of
every audit. Thus, the information required in WORKOUT ordinarily would
be available in every audit.
The frequent usage of CORRESPOND similarly reflects typical audit
practice — that of examining whatever correspondence is available con-
cerning a delinquent debtor. To some extent, also, this cue consti-
tutes a catch-all. The great variety of different data which might be
included in correspondence files by different clients and for different
debtors discouraged the researchers from attempting to incorporate in
AUDITOR'S rule base each separate cue which might be discovered. For
example, one cue might indicate that a delinquent debtor has placed
with the client a bank's letter of credit, which effectively removed
any doubt about ultimate recovery of the account. A similar but dif-
ferent cue to be found in another debtor's files might refer to a
parent company's guarantee of indebtedness. To avoid a proliferation
of such rules in this first attempt at building an expert system of
auditor's judgments, CORRESPOND was formulated to allow the user to sum
up in one response the data which he gathered from the client's corre-
spondence files. Such a response constitutes in effect a "mini-judgment"
by the user, which more experience with the system may suggest should
be divided into several rules. ECONOMIC is also a rule of rather fre-
quent usage, perhaps as a result of concern about the economic condi-
tions which prevailed when this work was done in (1981 and 1982).
Usage of CREDITMGR reflects the common audit practice of discussing
delinquent accounts with a representative of the client.
On the other hand, several rules carry a relatively heavy impact
but evidently are infrequently applied. Auditors report they seldom
communicate with an attorney regarding a delinquent account, but would
give strong weight to a pessimistic prediction which he might make
(LAWYER). The creditor's stated intent not to pay could represent an
important cue but is rarely available (NOTPAY). The only instances
reported to the researchers of the presence of this cue occurred in the
case of debtors who were disputing the propriety of the charges in their
accounts, claiming errors in pricing. Similarly, if neither the audi-
tor nor his client were able to communicate with the debtor that would
represent an important cue (NONCONTACT) . No such situation arose between
the large commercial and government entities involved serving as test
cases in the AUDITOR studv.
-18-
FURTHER ANALYSIS
Three other aspects of the system and its usage are worthy of com-
ment. Michie (1980) and others have suggested that an experienced prac-
titioner's expertise consists of his ability to recognize clusterings
of cues — that is, patterns — which he has encountered previously and
found to be particularly diagnostic. Clusterings of cues may function
as mini-hypotheses, allowing the expert to focus his evidence-gathering
in an economical fashion. To the researchers, the most obvious of the
patterns apparently utilized by auditors in the valuation of accounts
receivable seems to be organized around a distinction between "slow-pay"
versus "no-pay" as diagnostic categories for each delinquent account.
Thus, a slow-pay customer, properly diagnosed, does not require a pro-
vision for loss. However, this categorization may hide several patterns
of less global impact, which may in reality be the organizers for a
practitioner's judgment, such as "the economic-problem" pattern or "the
legal-problem" pattern. This matter will be explored in a subsequent
paper .
The second matter concerns strength of belief and the meaning of
AUDITOR'S report. AUDITOR reports Degrees of Belief. This constitutes
its expert judgment. The user of the system must decide for himself
the significance of the report, for example whether 10 Degrees (.91
probability) mandates that an allowance be provided. The researchers
attempted to calibrate the system during its refinement stage so that
AUDITOR would produce a report of about 10 Degrees (positive or nega-
tive), coincident with the time a human expert auditor reported that he
had become "satisfied" with the data and had made his decision. How-
ever, many studies reflect unfavorably on humankind's ability to pro-
perly deal with data presented in a probabilistic format, c.f. Joyce
and Biddle (1981a). Thus, AUDITOR'S report of Degree of Belief can
best be viewed as an indication of relative strength of belief. No
claim is intended either that AUDITOR'S report reflects objective prob-
abilities nor that the researchers believe human experts process prob-
abilities similarly in the construction of their judgments. However,
while keeping this disclaimer well in mind, the researchers believe
that by the time the users terminated their validation sessions AUDITOR
can fairly be said to have reached an unambiguous result, at least in
terms of the refinement criteria, in the great majority of cases. For
example, for the third client, which comprised the blind validation
trial, in only one instance of eleven delinquent debtors was the session
terminated when AUDITOR was reporting less than four percentage points
from a value which would represent certainty (0% or 100%) Table 2. In
the ten open-book trials, eight reported within +/- seven points,
results which are within the criteria (Table 1). Thus, the researchers
believe this first complete version of AUDITOR can be said to diagnose
the collectibility of delinquent accounts in an unambiguous fashion,
which bodes well for the development of other expert systems for use
in audit situations.
The final matter for discussion concerns the sequence of inquiry
followed by the system. AUDITOR asks next that particular question
-19-
which potentially has the greatest impact upon the probability of the
hypothesis. Human auditors do not organize their own evidence-gathering
in such an efficient fashion, nor do they plan the audit in order to
investigate every source of highly reliable information. For example,
it is not a standard audit procedure to request confirmations from every
large delinquent debtor, even though responses might reveal evidence of
high diagnosticity through application of such rules as PROBLEMS,
NORESPONSE, NONCONTACT, LEGAL, and NOTPAY. Neither is it common prac-
tice uniformly to inquire of the client's attorney or collection agency
concerning a delinquent debtor, although evidence of high diagnosticity
might result (LAWYER, BANKRUPT, LEGAL, and various interactions).
Explanation for this apparent gaffe is easily found. Auditors are cost
conscious, AUDITOR is not. This expert system is designed to con-
centrate on the diagnosticity of the evidence while ignoring its cost.
Human expert auditors, at least those who work for profit-oriented
firms, are expected to be time and cost conscious.
Additionally, of course, an eclectism in auditors' pursuit of evi-
dence is motivated by factors such as the unpredictable availability of
clients' records and employees, conflicting schedules necessitated by a
desire to service several clients, and a need to accomplish several audit
objectives in a limited time. Also, since both client and auditor seek
more from the audit than a satisfactory valuation of the allowance for
bad debts, other tasks may intervene.
In only one case during validation did sequencing of inquiry cause
difficulty. In that case, #1-2, the user serving as validator ter-
minated the validation session immediately prior to a question which
had great bearing on that particular debtor and which would have put
AUDITOR more in agreement with the judgment of the human auditor, that
is, at a higher Degree and probability level. In all other instances,
the sequence of greatest-impact-first, with termination controlled by
the user, appeared to be a satisfactory method of handling the evidence
contained in working papers without provoking a complaint by the user
of the system. Thus, the sequencing in AUDITOR seems to be at least
satisfactory to its users in the majority of cases. Whether it is rea-
sonable to expect auditors to follow a more efficient search pattern is
beyond the scope of this paper.
SUMMARY
Expert systems such as AUDITOR can constitute a model of auditors'
judgment, answering the researchable and interesting question said to
be at the heart of all investigation of diagnostic judgment in the face
of uncertainty: what cues do the judges use, what are their weights,
and how are they assembled into a judgment model. Thus, subject to the
influence of the researchers' intervention, AUDITOR demonstrates the
feasibility of a new approach to the descriptive study of auditors'
judgments .
-20-
LIMITATIONS
All of the auditors who participated in the initial system-building
were members of one office of one large public accounting firm. To a
great extent all had been exposed to similar professional training.
Additionally, they perform under policies which might perhaps be unique
to one firm. However, auditors from other firms were involved in the
refinement and validation stages of the project — which lends confidence
in the universality of the system as a model of auditors' judgment and
reduces the chances of parochialism.
Further refinement of the system may produce better results. Rules
might be refined to reduce them more nearly to the elemental cues,
reducing the need for the mini-judgments presently contained in such
rules as CORRESPOND and LEGAL. Such a revised rule base might be more
reliably applied by an unexperienced auditor — one who had not developed
competence in the mini-judgments which a few of the rules call for. On
the other hand, such changes increase the size of the rule base and
lengthen the list of questions to which a user is exposed — factors which
tend to decrease the convenience of working with the system. However,
work papers used by the auditors who were involved in various stages of
this work evidently contained very little relevant data beyond that
called for in the rules of the system, with the exception of the age of
the delinquent balance which was under scrutiny. Some indication during
validation was gained that the absolute age of the account may have to
some auditors a significance not reflected in the system. Therefore,
a rule relating to age of the delinquent balance might be useful in the
system. However, for Che large, successful, client companies whose
audit work papers were referred to in various stages of this work it
seemed likely that no delinquent account was allowed to gain more than
a few months' age before some finality was forced in its disposition.
(These can be presumed to be clients with relatively good internal con-
trols over sales and collections.) However, there is the possibility
that alternative systems of rules and weights may produce equally suc-
cessful systems which demonstrate audit judgment.
The influence and beliefs of the observer — the researchers — cannot
be ruled out in project of this kind. Also, the subjects' desires to
please and accommodate the researchers may work to the detriment of
effort to create and particularly to validate an effective expert system.
Criticism can be directed against the system for its use of Bayes '
revision. It is commonly believed that auditors are no more Bayesian
in their processing of evidence than are other humans (Libby, 1981).
However, the point is worth emphasizing that the use of Bayesian revi-
sion in computerized expert systems operating in fields other than
auditing produce results similar enough to those produced by human
experts as to satisfy these experts themselves.
Further refinement of the system may be called for, particularly,
regarding instances in which practicing auditors conclude that partial
rather than 100% allowances are proper to provide against a delinquent
-^1 -
account. In the Open-Book Validation procedure, auditors called for a
partial allowance against an account for which AUDITOR'S processing
produced (upon replay by the researcher) a result of approximately 70%.
Further experience with the system may serve to clarify the set of cir-
cumstances under which human experts call for partial allowance and the
probabilities which AUDITOR might be expected to report in the same
circumstances.
On the other hand, the judgments of expert auditors may be suf-
ficiently idiosyncratic and inconsistent that attempts to refine the
system any further will be fruitless. Changes to the system will be
carried out only after an analysis of the objective to be sought. For
example, if enhanced convenience of the system for use by practicing
auditors were to be an objective of modification, a desirable change
might be to engineer a revision in the order in which certain questions
are asked. Certain evidence, while of lower diagnostic impact, seems
always to be available, and might be asked about early in a consultation
to allow the user promptly to contribute his knowledge early in ques-
tioning. On the other hand, already existing capabilities within AL/X
permit the user to volunteer evidence prior to questioning. For example,
if no request for confirmation was sent and no information was received
from any outside source, the system could be modified to accept this
data even before questioning began. These capabilities have not been
exercised yet in AUDITOR.
-22-
FOOTNOTES
Named second in frequency was the judgment process leading to the
budget of audit time, the study of which might have allowed useful com-
parisons with previous research (c.f. Joyce, 1976). However, the widely
differing descriptions of this process which were given by the experts
discouraged the researchers from attempting to extract their expertise.
The judgment process involved in determining the threshhold for "report-
ing materiality" was similarly considered then rejected for study because
the aspects of negotiation evidently resorted to between client and
auditor have not been amenable to expert system technology.
2
AL/X (Advice Language / X) was developed for Intelligent Terminals
Ltd. by John Reiter, Steve Barth, and Andy Paterson in association with
the University of Edinburgh and was supported by BP Petroleum Develop-
ment Ltd. It is a Pascal system based upon the Prospector consultant
system for mineral exploration developed at SRI International by Richard
Duda, Peter Hart, and others, see Duda et al. (1979).
3
The capabilities of AL/X allow formation of queries in a form
calling for a "Yes" or "No" answer. This capability was installed after
the work on AUDITOR had begun. Because of the apparent ease with which
users of the system had already adapted their responses to the CV scale,
this added capability was not used in AUDITOR.
RESERVE does not make provision for situations in which the audi-
tor concludes the amount to be provided for potential loss of the account
should be more than zero but less than 100% of the delinquent balance.
One such case was encountered among the twenty-one delinquent accounts
reviewed during validation of the system, and the validator ruled that
AUDITOR did not perform satisfactorily.
Four separate sets of cues were accumulated from these experts.
The nature of differences between these cue sets will be explored in a
later paper. The final AUDITOR system was an aggregation of the cues
provided by the four experts.
Rules related by AND linkages are processed according to the "fuzzy
logic" rules of L. Zadeh (1979): P(A1 AND A2 AND... AND An) = minimum
[P(A1), P(A2) , . . . ,P(An) ] , where P is probability and the An are eviden-
tial cues.
It was also necessary to "prime" the system by providing the
Bayesian portions of it with initial (prior) degrees of belief. In the
absence of any definitive guidance in the literature or from the experts
all such initial values were set to reflect 0.0 degrees of belief, which
is a probability of 0.5. At the beginning of any session with AUDITOR
these values can be changed if desired.
-23-
g
In both open-book and blind validation procedures the assumption
was made that the work: papers contained both the auditors' judgments and
the support therefore, in conformity with Generally Accepted Auditing
Standards, AU 338.
9
Since AUDITOR reports its results in degrees of belief and in prob-
abilities, which would be a novel way for a subjective audit judgment
to be recorded in work papers, it was necessary for the researchers to
interpret and in effect to translate AUDITOR'S report before its presen-
tation to the validator in order to maintain anonymity of the source of
each judgment appearing on the Comparison Worksheet. This interpretation
was easily made in all but one case. In every case but one, AUDITOR'S
report produced a DB of at least 15, positive or negative, that is,
probability of at least .97 either in favor of or against the hypothesis
which called for an allowance. These results appeared to the researchers
to be unequivocal and were reported on the Comparison Worksheet as that
of an expert auditor who called for "No reserve nor adjustment needed,"
(in the cases which reported negative DBs) or "Yes, reserve or adjust-
ment needed" (in cases which reported positive DBs). (In keeping with
common practice the term "reserve" was freely used as a synonym for
allowance for bad debts among the researchers and the practitioners.)
However, one case, that of number 7, produced a DB of 2 (probability
.61) — an equivocal result having neither precedent nor clear interpre-
tation within the researchers' experience with the AUDITOR system. The
researchers reported this on the Comparison Worksheet as, "At most a
partial reserve or adjustment is required." On the same case, the audit
team had found no need for an allowance; the validator scored this judg-
ment at a hit. The validator initially scored AUDITOR'S result as a
miss and that is the way it was recorded by the researchers. Later,
however, when reviewing his work, he referred to AUDITOR'S result on
number 7 as a "close call." The issue of partial allowance will be
studied further in later applications of AUDITOR.
This cue contributes to the researchers ' impression that the
expert auditors' basic diagnostic task consists of an attempt to clas-
sify each delinquent account as "slow-pay" or "no-pay," that is, these
categories constitute competing hypotheses for classification of each
delinquent account.
Technical data for this appendix also came from "AL/X USER MANUAL,"
Andy Paterson, Intelligent Terminals LTD. Oxford, England, 1981.
-24-
APPENDIX A: Description of AL/X and AUDITOR11
The structure of an expert system, in general, and AL/X, in parti-
cular, parallels the structure of human decision making. Data is
gathered, analyzed, evaluated with respect to some criteria and used,
and, if more data is required, what additional data to gather is deter-
mined. Figure A-l shows how the specific step in the operation of AL/X
map to the generic decision making phases.
The two key operational areas are revision and selection. Revision
refers to the process of updating the Degree of Belief associated with
the spaces, that is, the rules and hypotheses. Figure A- 2 depicts this
revision as a two stage process. At all times the hypothesis, which is
the goal of the system's inquiry process, has associated with it a
Degree of Belief (DB) or simply degree. Degrees of Belief measure
strength of belief on a scale having a range from -100 to +100 but,
AUDITOR only uses a range of -30 to +30. The initial ("prior") DB
value within AUDITOR of the hypothesis, RESERVE, is set to 0.0 but can
be easily set to any initial value. The DB scale is derived from prob-
abilities as:
Degree of Hypothesis(H) = 10 log10(Probability(H)/(l-Probability(H))
Similarly, every space carries a DB which reflects its prior strength
of belief. The user's response (CV) to each question initiates stage 1;
revision immediately of the DB for that space (unless the response is 0) .
The value of CV revises the degree in the following manner. A user's
response of +5 assigns to that space the maximum DB permitted by the
systems, i.e., +100. Similarly, an answer of -5 results in a degree of
-100. An answer of zero leaves the degree unchanged at its prior value.
An answer other than +5, -5 or 0 results in interpolations between these
three points. A linear interpolation is performed on probability values.
A positive CV causes interpolation between the prior probability and 1.
A negative CV value results in interpolation between zero and the prior
probability. The resulting probability value is then converted back to
a degree value. The effects of the user's answer are immediately pro-
pagated through the inference net to update the DB of all spaces for
which the selected question is evidence, including the hypotheses.
The second stage of the two-stage process of revision involves cal-
culations of the incremental weight which will be added to the present
degree of the hypothesis (and to the present degree of any other spaces
which, in a particular system, might be consequences of the immediate
question to which the user is responding). First, we will speak about
revisions involving spaces connected via IF:THEN inferences. A process
of interpolation is again used employing the PW or the NW of the space,
together with the degree calculated from the user's CV response. If
the evidence is true then a positive weight (PW) of one will increase
the degree of belief of the hypothesis by one. Similarly if the evidence
-25-
GATHER DATA
£
SYSTEM
ASKS
QUESTION
USER'S
RESPONSE
(CV)
ANALYZE
REVISION
UPDATE DI
"1
[REPORT TCl
USER
EVALUATE
DETERMINE NEXT QUESTION
MORE
UESTION
SYSTEM SELECTS
NEXT QUESTION
( STOPj
Figure A-l Overview of Expert Systems
-26-
USER'S
RESPONSE
ON
CV SCALE
r
TRANSFORMATION
OF CV TO
PROBABILITY
EQUIVALENT
CONVERSION
OF PROBABILITY
EQUIVALENT TO
DEGREE OF BELIEF
( DB)
|_R EV I SIO N_: STAGE J,
r
____j
PW OR NW
OF SPACE
V/HICH IS
CAUSING REVISION
1
CALCULATION OF
INCREMENTAL
WEIGHT ( IVV )
IW=f(DB,PW,NW)
HYPOTHESIS
NEW DB=
OLD DB + IW
j REVISION: S TAG E_2
j
Figure A-2 Overview of the Process of Revision
-27-
is false and the negative weight (NW) of one then the degree of the hypo-
thesis will decrease by one. The greater the positive weight then the
more strongly does the presence of the evidence imply that the hypothesis
is true. For negative weights, the corresponding implication is that
the absence of the evidence implies that the hypothesis is false. In
the situation where the presence of evidence implies that a hypothesis
is false rather than true it is perfectly legitimate to have a negative,
positive weight and a positive, negative weight (i.e., PW = -4 and NW =
+5).
If the degree of the space, as a result of the user's response is
greater than the prior degree of that space, then the incremental weight
to be added to the degree of the hypothesis is calculated as
Incremental weight = PW/(PW* min(PW, current - prior))
If the degree of the space as calculated is less than the prior degree
of that space, then the incremental weight is:
Incremental weight = NW/(NW* min(NW, prior - current)).
AND spaces are composed of the connection by AND of two or more IFrTHEN
spaces. Each AND space has its own degree of belief. Also, each AND
space has its own PW and NW. The weight which an AND space has on the
hypothesis is dependent on the degree of that AND space and its PW and
NW, just as if it were an IFrTHEN space. The degree of an AND space is
a function of the degree of the components and is calculated as
degree (Al AND A2 AND ... AND An) = min(DB(Al), DB(A2), ..., DB(An))
Selection refers to the manner in which AL/X chooses which, of many,
questions is appropriate to ask next. (See Figure A-3). The manner in
which this is done is important because human experts are characterized
by their ability to focus quickly on the most efficient line of ques-
tioning. AL/X chooses and investigates that particular hypothesis which
is most likely out of a group of perhaps several alternative hypotheses.
The choice criterion is: Select that hypothesis which currently has
the highest degree of belief. Once the hypothesis has been chosen,
questions are selected which will speed resolution of the diagnosis.
In AUDITOR, the process of selection consists of choosing which ques-
tion to ask of the user, i.e., which question has the highest potential
incremental weight impact on the hypothesis. The incremental weight of
each of the remaining questions is calculated prospectively by the
system. AL/X carries this out by scanning through all of the, as yet,
unanswered questions, calculating the incremental weights which could
-28-
PW AND NW
OF UNANSWERED
QUESTIONS
PW AND NW
OF AND
SPACES
SYSTEM ASKS
QUESTION
USER
RESPONDS
SYSTEM
CALCULATES
.IW*OF ALL
UNANSWERED
QUESTIONS
SYSTEM
IDENTIFIES
UNANSWERED
QUESTION WITH
HIGHEST IW
IW= incremental v/eight
Figure A-3 Overview of the Method for Selecting the Next Question
-29-
arise from each question, as a result of a CV response by the user of
either -5 or +5. (AND and NOT spaces are included in the calculation.)
That question is chosen next which could conceivably cause the greatest
impact on the strength of belief of the hypothesis. Since the poten-
tial impact of an AND space depends upon the user's answers to the com-
ponent spaces, the order of questioning may vary from one consultation
to the next depending upon the user's responses to the components of
the AND spaces. This knowledge is utilized by the system in the form
of rules often in the style I_F: Evidence , THEN (to a specified
extent ) :Hypothesis . The connectors AND, OR, and NOT may also be used.
The data are immediately incorporated into the system by the con-
trol and processed through the rule base. This action updates the rule
base to reflect the user's knowledge and revises the degree of belief
in the hypothesis which has been the focus of the expert's inquiry pro-
cess. It is the degree of belief in the hypothesis (and its transla-
tion into probabilities) which constitutes the goal of the system and
its expert judgment.
Goals with degree > 0.0 are:
The delinquent portion of this account should specifically be reserved
for in the allowance for bad debts to a substantial extent (RESERVE).
Prior degree was 0.0. Current degree is 13.8. At this point the goal
is certain: Probability .95 or greater.
The above is AUDITOR'S report that the user's answers to its ques-
tions have resulted in increasing the strength of belief of the truth
of the hypothesis called RESERVE from 0.0 to 13.8 on the degree (of
belief) scale. Prior to questioning of the user, RESERVE is presumed
as likely to be true as false, that is, to have prior degree (or belief)
of 0.0. A degree of belief (DB) or, simply, degree measures strength
of belief in the goal hypothesis. The possible range is from -100 to
100. Since AUDITOR'S report is based upon its Bayesian processing of
subjective probabilities, the researchers utilized the report of
degrees and probabilities merely as a guide to imputing a judgment from
AUDITOR, and initially operated under the assumption that a probability
of .90 (degree of 9.5) would indicate the need for an allowance.
A degree of zero indicates that the hypothesis is equally likely
and unlikely, i.e., probability of 0.5. The mid-range of the degree
scale from about -10 to 10 DB represents a large range of probability
(about .10 to .90). Thus, small increments in degree may represent
large changes in probability. This effect is considered desirable in
this expert system since small changes in degree are considered to be
-30-
relatively more diagnostic when the current truth or falsity of the
hypothesis lies in a doubtful or ambiguous area.
Only rarely, however, are the cue weights symmetrical. For example,
the rule stated as, "This debtor is in bankruptcy-type proceedings"
(BANKRUPT) carries PW = 3.0, NW = -1.0. These values indicate that to
the experts the debtor's bankruptcy is considerably more diagnostic —
reflected by the PW = 3.0 — of uncollectibility than the absence of bank-
ruptcy is predictive of collectibility.
The reader may notice two consequences of these relationships.
First, an apparently small portion of the available range on the DB
scale, say from about -10 to 10, encompasses the rather large proba-
bility range .09 to .91. Additionally, since the effect of an eviden-
tial cue is additive upon the hypothesis, the diagnostic or inferential
strength of any piece of evidence is at its highest when it impacts an
hypothesis of neutral degree and decreases as more evidence is accumu-
lated. This is believed to be a desirable feature in an expert system.
-31-
APPENDIX B:
ILLUSTRATION OF AUDITOR'S OPERATION
A hypothetical case situation is used to demonstrate the operation
of the system. First, the facts of the situation, as they might be in
the working papers, are described. Second, the consultative session
between AUDITOR and the expert is shown, including annotations to high-
light the features of AUDITOR. And, third, the results of the AUDITOR
session are given in terms of degrees of belief (DB) and probability,
which represent the revised likelihood of the hypothesis under scrutiny.
SITUATION
The delinquent amount owed by this customer, a regional, publicly
held hardware chain, is $82,000. This represents about two-thirds of
the balance of one invoice approximately 75 days overdue on which the
customer has recently paid $40,000. In total, the customer owes
$320,000, the rest of it classified as current (45 days and under).
The overdue amount, while large among delinquent accounts, is clearly
not material to the client's financial statements. They show accounts
receivable of $20 million and current assets of $100 million. The
accounts receivable subsidiary ledger shows the most recent three-year
history of the account, in which the customer has become overdue only
one other time. He ultimately paid the full amount owed after an
adjustment was made which represented about 2% of the invoice. Sales
this year are above those of a year ago, probably reflecting a trend
toward "do-it-yourself" repairs by hardware customers. Correspondence
and other files indicate a long-term, on-going business relationship
with no suggestion of any problems with the account.
The auditors did not attempt to confirm the account and there is no
information available from an attorney or any other source outside of
the client. The customer is not bankrupt. The client continues to
make credit sales to this customer, who has paid other invoices within
their due dates. According to the credit manager who has talked with
them, the customer states emphatically he will not pay the balance
which is overdue. The credit manager goes on to say, "Sometimes we
give them a small adjustment when they put up a fuss like this, but
they have been good customers for 15 years and they always pay." The
"fuss" referred to in this case evidently resulted from a late delivery
CONSULTATIVE SESSION
The user first signs onto the system. AUDITOR then begins ques-
tioning the user about the facts of the case. The order of the ques-
tions depends on the user's responses. Material in brackets represents
AUDITOR output, in parentheses the user's responses, and in asterisks
the researchers annotations. The result of this session is that the
-32-
systera determines that there is only .013 probability (-19.5 DB) that
an allowance need be provided for this account. Other, similar accounts
would go through a similar set of interactions.
■33-
APPENDIX C: NETWORK DESCRIPTION FILE
MODEL 3ADDE3T
VERSION A3DMA
SPACE RESERVE
TEXT DESCRIPTION
/* THE DELINQUENT PORTION OF THIS ACCOUNT SHOULD SPECIFICALLY 3E RESEEVED-FOR IN
THE ALLOWANCE FOR 3 AD DE3TS TO A SUBSTANTIAL DEGREE */
INFERENCE
PRIOR 0.0
SPACE OUTSTAND
TEXT DESCRIPTION
/* WHILE SMALL PAYMENTS ARE 3EING RECEIVED ON THIS ACCOUNT. THE OUTSTANDING
DELINQUENT BALANCE IS GROWING LARGER */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 1 NW 0 ) .
SPACE AVEAG2
TEXT DESCRIPTION
/* THE AVERAGE AGE OF THE UNCOLLECTED PORTIONS OF THIS ACCOUNT IS INCREASING */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 0.5 NW -0.5 )
SPACE ALL3UT0NE
TEXT DESCRIPTION
/* THIS CUSTOMER'S ACCOUNT 3ALANCS IS ALL CURRENT EXCEPT FCR ONE I^RGS AND
DELINQUENT OiJ-JRGZ */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE FW 0.5 NW 0 )
SPACE AL3T1&NOT?
TEXT DESCRIPTION
/* ALL3UT0NE AND NOTPAY */
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION AND ( ALL3UTCNE NOTPAY )
RULES CONSEQUENTS ( RESERVE PW 2.0 NW 0 )
SPACE CREDITS TOP
TEXT DESCRIPTION
/* CREDIT TO THIS CUSTOMER HAS 3EEN STOPPED BY THE CLIENT */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 1 NW -0 . 5 )
-34-
4. Kac! been assigned to a collection age,.cy or
C.-that this account ha^ DeB" °^ a
lawyer? ^3 (-5)
C .that this debtor has a ^V"^""^!* U>timately payi"9
^stantially all delinquent balances? 71 (5)
«.,- •,,««■ ever been received from this customer?
C.bhat no payments havs ever u
?3 (-5) '
*.~ «r. h^d stated his intent to pay little op
C.that the customer haS sw , C.that despite tha
nothing of the e hnquent ^ ance s .^ ^ ^ ^ny
presence of « deh ^ 3uditor's call to the system to
paid? ?1 (PG) * PG is Li either than an adequate amount of
present its conclusion. £ailable Qvidence. has been entered into
evidence* or all of the 3v<s*
the system.*
•«««««""**"""to dea^e cf belief > -100.0 are: .-..none at
Clnvest19ated goals with deg indicating that the goal
the moment. 3 *This should ^e r other words, thera
hypothesis has not been £hUX ™v£ ^ of the system prefers
are still questions "nasked^ but ^ ^^ y^ .fc £g
that they not be asked. Also, t, V uheneVar '.t is over
for the hypothesis whenever
t
set to report the degree
-100 DB-*
COther goals with degree > ~100 areS
u • „ ~r this account should specifically bs
The delinquent Pj; ^, a^Sfor bad debts to a substantial
reserved- for in tne aixuv
degree :RESERVe.
n a rurrent degree is -19.5. At this point this
Prior degree was 0 . 0 . cui / | probt3bility is only. 013
goal is P^^i1^-9^;' ,ccount will become uncollectible:
SP:;l'; user will Undoubtedly conclude no provision need b,
rr.ade for loss from this account.*
^jc******** ***************************************
««*«******-«**************
. fh.t despite the presence of this delinquent
CHow certamareyo hat £ esp P ^ # ^ ^^ ^
item, newer items have bj.«n * y P unanswered, and the user^
rp3pat5 the last question wnx.ul . .f •. • • ~ *
?! !u * ^^ r-^^e or Quit its questionn ing.*
tells the system to ceasF?
■35-
CThe current goal is whether cr not the ce] inquent portion of this
account should specifically be reserved-for in the allowance for
bad debts to a substantial degree. RESERVE!. *7he system is
announcing its goal hypothesis.*
CHow certain are you that your te"sts of cash receipts indicate
that the delinquent portions oP the account in question have been
collected substantially in their entirety subsequent to the date
of the balance sheet? ?3 (-5) *0n this line the system has given
a "prompt" in the form of a question mark ■ and the auditor has
replied with "-5" which indicates he is certain the account haa
not been substantially collected.*
CHow certain are you that serious problems with the delinquent
portions of this account were revealed through confirmation
causing you to believe they are invalid?
?3 (0) *No confirmation was requested* so the user responds with a
"O" which means* in this case* the question is irrelevant.*
C...that legal counsel gives poor prospects of any significant
recovery from this debtor? ?3 (0) #There is no information
available from a lawyer.*
C...that the confirmation request was returned by the postal
service as undel iverable and the client for several months has
been unable to communicate with the debtor? 71 (0) -*The auditor
responded to this question as irrelevant.*
•
C...that recent correspondence and other data in the customer's
credit file supports your belief in the collectibility of this
account? 71 <5) *The auditor believes the long-term business
relationship will be continued by the customer* or that for other
reasons the account will by collected.*
C...that recent collections toward the delinquent portions of thits
account are proceeding satisfactorily? 71 (5) *A substantial
amount has recently been received.*
C...that the credit manager expresses a strong belief in the
ultimate collection of substantially all this account? ?3 (5)
C...that although a portion of this customer's total balance is
still delinquent* he continues to be an active customer? 71 (5)
C...that this debtor is in bankruptcy-type proceedings? 71 (-5)
C...that economic factors are causing particularly bad effects on
this customer's 3bility to pay? 71 (-5)
C...that the .r.erits of this debtor's likely counterclaim agiansfc
your client's suit indicate that a legal action would be
fruitless? °3 (0) *The auditor has no information about a
counterclaim.*
*»
-36-
SPACS NRES&NTACT
TEXT DESCRIPTION
/* NORESPONSE AND NOT ACTIVE */
i .«- ERENCE
PRIOR 0.0
LOGICAL DEFINITION AND ( NORESPCNSE NOTACTIVE )
RULES CONSEQUENTS ( RESERVE PW 2.0 NW -1.0 )
SPACE NONCONTACT
text description
/* the conftrmatioh request was retorned 3y the postal service as cndelivsra3le and
the client for several Months has been unable to communicate with the debtor */
inference
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 4.5 NW 0.0 )
SPACE NCNT&NOTACT
TEXT DESCRIPTION - -
/* NONCONTACT AND NOTACTIVE */
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION AND ( NONCONTACT NOTACTIVE )
RULES CONSEQUENTS ( RESERVE PW 5.0 NW 0.0 )
SPACE NCTPAY
TEXT DESCRIPTION
/* THE CUSTOMER HAS STATED HIS INTENT TO PAY LITTLE OR NOTHING OF THE DELINQUENT
BALANCES */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 3.0 NW -1.0 )
SPACE BANKRUPT
TEXT DESCRIPTION ■
/* THIS DE3T0R IS IN 3ANXRUPTCY-TYFS PROCEEDINGS */
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 3.0 NW -1.0 )
SPACE LAWYER
TEXT DESCRIPTION
/* LEGAL COUNSEL GIVES POOR PROSPECTS OF ANY SIGNIFICANT RECOVERY FROM THIS DE3T0K
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 5.0 NW -1.0 )
SPACE BANXSWRSLAW
TEXT DESCRIPTION
/* 3ANXROPT AND WRITEOFF AND LAWYER •/
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION AND ( 3ANKSCPT WRITEOFF LAWYER )
RULES CONSEQUENTS ( RESERVE PW 9.0 NW -2.0 )
-37-
S?ACS ECONOMIC
TEXT DESCRIPTION
/* ECONOMIC FACTORS ARE CAUSING PARTICULARLY BAD EFFECTS ON THIS CUSTOMER'S ABILITY
TO PAY */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE ?W 3.0 NW 0.0 )
SPACE BAN&ECOSLAW*
TEXT DESCRIPTION
/* BANKRUPT AND ECONOMIC AND LAWYER */
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION AND ( 3ANKRUPT ECONOMIC LAWYER )
RULES CONSEQUENTS ( RESERVE PW 7.0 NW -1.0 )
SPACE GOCDRECORD
TEXT DESCRIPTION
/* THIS DE3T0R HAS A GOOD PAST RECORD OF ULTIMATELY PAYING SUBSTANTIALLY ALL
DELINQUENT BALANCES */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 2.0 NW 2.0 )
SPACE LEGAL
TEXT DESCRIPTION -
/* THE MERITS OP THIS DEBTOR'S LIKELY COUNTERCLAIM AGAINST YOUR CLIENT'S SUIT
INDICATE THAT A LEGAL ACTION WOULD 3E FRUITLESS */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 3.0 NW -1.0 )
SPACE LEGALSNPAY
TEXT DESCRIPTION
/* LEGAL AND NOT? AY */
INFERENCS
PRIOR 0.0
LOGICAL DEFINITION AND ( LEGAL NOT? AY )
RULES CONSEQUENTS ( RESERVE PW 6.0 NW -1.0 )
*
SPACE NPAYSGDREC
TEXT DESCRIPTION
/* NOT? AY AND GOCDRECORD */
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION AND { NOTPAY GOODRECORD )
RULES CONSEQUENTS ( RESERVE PW -2.0 NW 0.0 )
STOP
-38--
SPACE NCTACTIVS
TEXT DESCRIPTION
/* NOT OF ACTIVE V
INFERENCE
PRIOR 0.0
logical definition not active
s?^cz cr&np3slact
text description
/* creditmgr and ( not problems ( :;ot active */
i:jference
PRIOR 0.0
logical definition and ( creditmgr nctprc3lems active )
rules consequents ( reserve ?w -3.0 sw 3.0 )
space correspond
text description
/* recent correspondence and other data in the customer's credit file supports
3elief in the c0llecta3ility of this account */
i:;ferencs
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE ?W -2.0 Nil 4.0 )
SPACE CCRRES&ACT
TEXT DESCRIPTION
/* CORRESPOND AND ACTIVE */'
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION AND ( CORRESPOND ACTIVE )
RULES CONSEQUENTS ( RESERVE P» -3.0 KW 3.0 )
SPACE CCRRESSWORJC
TEXT DESCRIPTION
/* CORRESPOND AND WORKOUT */
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION AND ( CORRESPOND WORKOUT )
RULES CONSEQUENTS ( RESERVE PW -3.0 NW 3.0 )
S??JZZ WRITEOFF
TEXT DESCRIPTION
/* TOTAL WRITS OFF OF THIS ACCOUNT, IF REQOIRED, WILL REPRESENT A MATERIAL ADJUST-
MENT •/
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE FW 2.0 NW 0.0 )
SPACE NC RESPONSE
TEXT DESCRIPTION
/* THERE WAS NO RESPONSE TO YOUR CONFIRMATION REQUEST NCR TO A FOLLCW-UP REQUEST */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RE^ZKTt PW 1.0 NW -1.0 )
-39-
SPACE COLAGENCY
TEXT DESCRIPTION
/* THIS ACCOUNT HAS 3EEN ASSIGNED TO A COLLECTION AGENCY OR LAWYER */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 2 NW -1 )
SPACE PORMEREMP
TEXT DESCRIPTION
/♦THIS CELHJ2UENT ACCOUNT IS FROM A FORMER EMPLOYEE */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 2 NW 0 )
SPACE NEW? AID
TEXT DESCRIPTION
/* DESPITE THE PRESENCE OF THIS DELINQUENT ITEM, NEWER ITEMS HAVE BEEN FULLY PAID
INFLUENCE . .
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 1.5 NW 0 )
SPACE NO PAY EVER
TEXT DESCRIPTION
/* NO PAYMENTS HAVE EVER 3EEN RECEIVED FROM THIS CUSTOMER */
PRIOR 0.0
ROLES CONSEQUENTS ( RESERVE PW 2 NW -0.5 )
SPACE ISSUENOTS
.TEXT DESCRIPTION
/» THIS DE3T0R HAS ISSUED NOTES FOR THE UNPAID PORTIONS OF HIS ACCOUNT */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 0.0 NW -0.5 )
SPACE OCTSAVSN?
TEXT DESCRIPTION
/* OUTSTAND AND AVEAGE AND NOT? AY */
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION AND ( OUTSTAND AVEAGE NOT? AY )
RULES CONSEQUENTS ( RESERVE PW 1 NW 0 )
SPACE CSTSCOSNPE
TEXT DESCRIPTION
/* CREDITSTC? AND COLAGENCY AND NCPAYEVER */
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION AND ( CREDITSTCP COLAGENCY NCPAYEVER )
^ULZS CONSEQUENTS ( RESERVE PW S NW 0 )
SPACE CST&NPS&WO
PEXT DESCRIPTION
/* CREDITSTC? AND NCPAYEVER AND WRITEOFF */
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION AND ( CREDITSTCP NOPAYEVER WRITEOFF )
RULES CONSEQUENTS ( RESERVE- PW 3 NW 0 )
-40-
SPACE COLLECTED
TEXT DESCRIPTION
/* YOUR TESTS OR CASH RECEIPTS INDICATE THAT THE DELINQUENT PORTIONS OF THE ACCOUNT
IN QUESTION HAS 3EEN COLLECTED SUBSTANTIALLY IN THEIR ENTIRETY SUBSEQUENT TO THE
DATE OF THE BALANCE SHEET */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW -30.0. NW 1.0 )
SPACE PROBLEMS
TEXT DESCRIPTION
/* SERIOUS PROBLEMS WITH THE DELINQUENT PORTIONS OF THIS ACCOUNT VERS REVEALED
THROUGH CONFIRMATION CAUSING YOU TO 3ELLEVE THEY ARE INVALID */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 8.0 NW 0.0 )
SPACE NOTPRCBLEMS - '
TEXT DESCRIPTION
/* NOT OF PROBLEMS */
INFERENCE
PRIOR 0.0
LOGICAL DEFINITION NOT PROBLEMS
SPACE RIGOROUS
TEXT DESCRIPTION
/* COLLECTION EFFORT 3EING APPLIED BY YOUR CLIENT TO COLLECT THIS ACCOUNT IS LESS
RIGOROUS THAN IS DESIRA3LE */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW 1.0 NW 0.0 )
SPACE WORKOUT
TEXT DESCRIPTION
/* RECENT COLLECTIONS TCr'ARD THE DELINQUENT PORTIONS OF THIS ACCOUNT ARE PFOCEEDINC
SATISFACTORILY */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW -3.0 NW 3.0 )
SPACE CREDITMGR
TEXT DESCRIPTION
/* THE CREDIT MANAGER, OR OTHER COMPANY OFFICIAL, EXPRESSES A STRONG BELIEF IN THE
ULTIMATE COLLECTION OF SUBSTANTIALLY ALL THIS ACCOUNT */ f
INFERENCE *
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW -1.0 NW 2.0 )
SPACE ACTIVE
TEXT DESCRIPTION
/* ALTHOUGH A PORTION OF THIS CUSTOMER'S TOTAL BALANCE IS STILL DELINQUENT, HE
CONTINUES TO BE AN ACTIVE CUSTOMER */
INFERENCE
PRIOR 0.0
RULES CONSEQUENTS ( RESERVE PW -2.0 NW 3.0 )
-41-
REFERENCES
Abdel-Khalik, A. R. , and El-Sheshai , K., "Information Choice and Utili-
zation in an Experiment on Default Prediction," Journal of Accounting
Research (Autumn 1980), pp. 325-342.
Ashton, R. H., "A Descriptive Study of Information Evaluation," Journal
of Accounting Research (Spring 1981) 19.
, Human Information Processing in Accounting (American Accounting
Association, 1982).
Biggs, S. F., and Mock, T. J., "An Investigation of Auditor Decision
Processes in the Evaluation of Internal Controls and Audit Scope
Decisions," ROADS Paper No. 80-7, (April 28, 1980).
Crosby, M. A., "Bayesian Statistics in Auditing: A Comparison of Prob-
ability Elicitation Techniques," Report 80-008, Center for Audit
Research, University of Georgia (1980).
Dawes, R. M. , "The Mind, the Model, and the Task," in F. Restle, R. M.
Shiffrin, N. J. Castellan, J. R. Lindman and D. B. Pisoni, eds.,
Cognitive Theory, Vol. 1 (Erlbaum, 1975), pp. 119-29.
Duda, R., Gaschnig, J., and Hart, P., "Model Design in the Prospector
Consultant System for Mineral Exploration," Expert Systems in the
Micro-Electronic Age, D. Michie, Ed. (Edinburgh University Press,
1979).
Dungan, C. W. , "A Model of an Audit Judgment in the Form of An Expert
System," Unpubished Ph.D. Dissertation, Department of Accountancy,
University of Illinois (1983).
Einhorn, H. J., "Expert Judgment: Some Necessary Conditions and an
Example," Journal of Applied Psychology, 59 (1974), pp. 562-571.
, Kleinmuntz, D. N. , and Kleinmuntz, B., "Linear Regression and
Process-tracing Models of Judgment," Psychological Review, 86 (1979),
pp. 465-485.
Ericsson, K. A., and Simon, H. A., "Verbal Reports as Data," Psycholo-
gical Review, Vol. 87, No. 3, (May 1980).
Ernst, G. W. , and Newell, A., GPS: A Case Study in Generality and
Problem Solving (Academic Press, 1969).
Feigenbaum, E. A., "The Art of Artificial Intelligence: Themes and Case
Studies of Knowledge Engineering," Proceedings of the 5th Inter-
national Joint Conference on Artificial Intelligence, (1979).
-42-
Harrell, A. M. , "The Decision-Making Behavior of Air Force Officers and
the Management Control Process," The Accounting Review, (October
1977), pp. 833-841.
Joyce, E. J., "Expert Judgment in Audit Program Planning," Studies on
Human Information Processing in Accounting, Supplement to the
Journal of Accounting Research, (1976).
Joyce, E. J., and Biddle, G. C. , "Are Auditors' Judgments Sufficiently
Regressive," Research Opportunities in Auditing Distribution Service,
Peat, Marwick, Mitchell Foundation, N.Y. #80-4, (1980).
Libby, R. , Accounting and Hunan Information Processing (Prentice-Hall,
1981).
Lichtenstein, S., Fischhoff, B., and Phillips, L. D., "Calibration of
Probabilities: The State of the Art," in H. Jungermann and G.
de Zeeuw (eds.), Decision Making and Change in Human Affairs,
(Dordrecht-Holland: Riedel, 1977), pp. 275-324.
Michaelsen, R. H., "A Knowledge-Based System for Individual Income and
Transfer Tax Planning," Unpublished Dissertation, Department of
Accountancy, University of Illinois, (1982).
Michie, D., "Knowledge-based Systems," Working Paper UIUCDCRS-R-80-1001 ,
Department of Computer Science, University of Illinois at Urbana-
Champaign, (1980).
Newton, L. K. , "The Risk Factor in Materiality Decisions," Accounting
Review, (January 1977) _52_, pp. 97-108.
Nisbett, R. E., and Wilson, T. D., "Telling More Than We Can Know:
Verbal Reports on Mental Processes," Psychological Review, Vol. 84,
No. 3, (May 1977), pp. 231-259.
Payne, J. W. , Braustein, M. L., and Carroll, J. S. , "Exploring Pre-
Decisional Behavior: An Alternative Approach to Decision Research,"
Organizational Behavior and Human Performance, (February 1978) 22,
pp. 17-44.
Reiter, J., AL/X: An Expert System Using Plausible Inference, (Intel-
ligent Terminals Ltd., 1980).
Reitman, W. R. , Cognition and Thought, (Wiley, 1965).
Shields, M. D., "Some Effects of Information Load on Search Patterns
Used to Analyze Performance Reports," Unpublished manuscript,
University of North Carolina, Chapel Hill, (May 1980).
Shortliffe, E. H., And Buchanan, B. G. , "A Model of Inexact Reasoning
in Medicine," Mathematical Bioscience, Vol. 23, (1975), pp. 351-379.
-43-
Simon, H. A., "On How to Decide What To Do," Bell Journal of Economics,
(Autumn 1978), pp. 494-507.
Turing, A. M. , "Computing Machinery and Intelligence," Mind, (October
1950), reprint in Creative Computing, Vol. 6, No. 1, (January 1980),
pp. 44-53.
Yu, V. L., Fagan, L. M. , Wraith, S. et . al., "Anti-microbial Selection
by a Computer: A Blinded Evaluation by Infectious Diseases Experts,
Journal of the American Medical Association, Vol. 242, No. 12,
(September 21, 1979), pp. 1279-1282.
Zadeh, L., "A Theory of Approximate Reasoning," in Machine Intelligence,
_9_, J. R. Hayes, D. Michie, and L. Mikulich (Eds.), (Ellis Norwood
Ltd., and John Wiley and Sons, 1979).
D/174