ACCOUNTING INFORMATION SYSTEMS AND ORGANIZATION LEARNING: A SIMULATION

Pe r gamon
Accting., Mgmt. & Info. Tech., Vol. 7, No. 1, pp. 1-19, 1997
© 1997 Elsevier Science Ltd. All rights reserved
Printed in Great Britain
0959-8022/97 $17.00 + 0.00
Pl h S0959-8022(97)00001-5
ACCOUNT I NG I NF ORMAT I ON S YS T E MS AND
ORGANI Z AT I ON L E ARNI NG: A S I MUL AT I ON
Ar i s M. Ou k s e l
University of Illinois at Chicago
Ke n Mi h a v i c s
Roosevelt University
Pe t e r Ch a l o s
University of Illinois at Chicago and City University of Hong Kong
Ab s t r a c t - - Ac c o u n t i n g Information Syst ems may facilitate or i mpede organizational learning. Critical
attributes of account i ng syst ems that have the potential to affect organizational learning include: (1)
characteristics of the i nformat i on envi ronment , whet her uniform, dispersed or clustered importance weights;
(2) i nformat i on distribution, whet her overlapping or segregated information; and (3) information coordination
mechani sms, whet her expert t eams, majority vot i ng t eams or hierarchies. Organizational learning and
performance was si mul at ed in t he following manner: (i) the organization was faced with a cont i nuous sequence
of repetitive but not identical problems; (ii) t he organizational task was subdivided between analysts; and (iii)
anal yst s learned by basi ng their deci si ons on t he relationship found bet ween i nformat i on available to t hem and
organizational out comes. Simulation resul t s indicated that learning in flatter (team) organizations is generally
more accurate t han in hierarchical organizations. Learni ng is also faster with majority t eams t han hierarchies,
but slower with expert teams. Overlapping account i ng i nformat i on t ransmi ssi on bet ween agent s was found to
offer onl y limited benefits. These fi ndi ngs have implications for the desi gn of accounting information syst ems
in organizations. © 1997 Elsevier Science Ltd
INTRODUCTION
Accounting Information Systems (A.I.S.) play a central role in organizational learning,
prompting claims that "the aim of the design of A.I.S. is quite simply to improve organizational
learning" (Emmanuel, Otley & Merchant, 1990, p. 371) and "to activate learning and
experimentation" (Senge, 1990, p. 253). Our interpretation of firm learning builds on three
classical observations of behavior (Levitt & March, 1995). The first is that a great deal of
organizational learning is based on routines (Cyert & March, 1963; Nelson & Winter, 1982)
which are environmentally conditioned (Fiol & Lyles, 1985). The second observation is that
organizations are goal orientated. Their behavior depends on the relationship between targeted
and achieved outcomes. The third observation is that organizational actions are historical.
Output goals are tempered by interpretations of the past, as organizations adapt incrementally
and learn from goal attainment in response to feedback about outcomes (Den Hertog &
Wielinga, 1992). The emphasis on routines and the ecology of learning distinguishes this
2 A.M. OUKSEL et al.
approach from alternative theories and closely resembles paradigms of individual learning found
in the accounting literature (Hammond &Deane, 1973; Harrell, 1977).
An interactive learning process of periodic planning, measurement, feedback and perform-
ance appraisal characterizes many A.I.S.: (1) goals are determined; (2) environmental
information is coded, stored and subsequently retrieved; (3) deviations of actual outcomes from
predetermined goals are recorded for future corrective action; and (4) goal attainment is
rewarded (Flamholtz, Das & Tsui, 1985, p. 39). Recent evidence from the accounting literature
supports this view of organizational learning. The most commonly cited applications include
project management reports such as path analysis for production scheduling; profit planning
systems for lines of business which report actual to forecast revenue and expense comparisons;
revenue budgets which analyze market share, volume, price, etc.; reports used to monitor
competition; and human resource planning budgets (Simons, 1995, p. 108).
A number of factors are critical in A.I.S. design. Organizational memory, whether tacit or
formalized, may be systematically coded but information may not always be readily available.
Organizations vary in the emphasis they place on formal routines. Goals and feedback, by
definition, are strongly conditioned by the environment in which the firm operates. Firms
operating in uncertain environments face challenges in implementing A.I.S. designed to promote
routinized organizational learning. As environmental uncertainty increases, organizations need
to adapt their A.I.S. in order to promote learning (Ouchi, 1977).
The impact of organizational complexity on learning is also critical in A.I.S. design
(Galbraith, 1977; Weick & Roberts, 1993). Organizational hierarchies and multi-divisional
responsibility centres shape information routines. The complexity of organizational design
usually precludes complete dissemination of accounting information to all divisional sub-units.
Typically, information is aggregated or compressed as it flows vertically up the organizational
hierarchy. Laterally, information is often selectively distributed across sub-units. The ecological
structure of the organization complicates modeling of the learning process. Senge (1990, p. 289)
argues that hierarchies often impede organizational learning, while teams more effectively
promote learning. Recent initiatives to flatten organizational structures, establish cross-
functional teams, and increase electronic information sharing are in part motivated by a desire
to improve organizational learning.
In this study, we replicate and extend previous computational organization theory research
(Carley, 1992) within an A.I.S. framework. Specifically, we measure previously unexamined
effects of environmental information characteristics on team and hierarchical information
processing across segregated and overlapping A.I.S. Outcome effectiveness is measured in terms
of both accuracy and speed of organizational learning. As this is extremely difficult, if not
impossible, to capture in an organizational setting, a simulation is used. While the simulated task
is admittedly stylized, the methodology lays the groundwork for a set of empirically testable
propositions for future experimental and field research.
THEORETICAL DEVELOPMENT
Organi zat i onal l earni ng
Current literature on organizational learning tends to be theoretically fragmented, drawing on
analogies of individual learning or simply using organizational learning as an explanation for
many different kinds of observed organizational change or adaptation (Argote, Beckman &
Epple, 1990; Carley, 1992; Fiol & Lyles, 1985; Levitt & March, 1988). Some authors (Argyris
& Schon, 1978; March & Olson, 1976) deal explicitly with individual learning, while others
ACCOUNTING INFORMATION SYSTEMS AND ORGANIZATION LEARNING 3
(Lant & Mezias, 1992) propose a learning model which accounts for patterns of changes at the
level of the organization. Argyris and Schon (1978) distinguish between single- and double-loop
learning. This distinction can be thought of in terms of operational (know-how) and strategic
(know-what) learning. Single-loop learning occurs at the operational level because it concerns
processes, while double-loop learning occurs at the strategic level because it concerns the
definition of goals.
The effect of individual learning on organizational learning can be seen by analyzing the
impact of persistence of shared models on the performance of organizations. Others emphasize
the importance of individual learning only in so far as personal skills, insights and knowledge
become "embodi ed in organizational learning routines, practices, and beliefs that outlast the
presence of the original individual" (Attewell, 1992). This persistence of conceptual knowledge
is consistent with the notion of organizational learning adopted in this paper, namely "the
encoding of inferences from history into routines that guide future behavior" (Levitt & March,
1988).
Organizational learning however is more than the sum of individual learning of constituent
members. Rather, it is the ability to store data (in some type of memory), to recognize patterns
and rules from this data, and to build shared mental models across individuals and divisions of
an organization. A.I.S. play a critical role in shaping these routines. The notion of shared mental
models is central to work on continuous organizational i mprovement (Senge, 1990).
Organizational learning in this case is viewed as "i mprovi ng actions through better knowledge
and understanding" (Fiol & Lyles, 1985) or as "increasing an organization' s capacity to take
effective actions" (Kim, 1993) by using discovered patterns and rules to guide future decision
making. This requires constant monitoring of information at the individual as well as at the
organizational level, a role for which A.I.S. are ideally suited.
Envi ronment al uncert ai n~ and A.I.S. coordination
The environment facing the firm has been shown to have a significant impact on A.I.S.
design. Khandwalla (1972) was one of the first researchers to emphasize the importance of
environmental competition on the fi rm' s information system. A similar conclusion was reached
by Otley (1980) who determined that budgetary systems were strongly influenced by the fi rm' s
environment. Gordon and Miller (1976) found that heterogeneity of product lines was a primary
determinant of A.I.S. reporting frequency and sophistication. Another study, pertaining to
environmental uncertainty (Govindarajan, 1984) found no connection between budgetary
accounting systems and organizational learning effectiveness until the mediating effect of
uncertainty was considered. This considerable body of work relating environmental predictabil-
ity to A.I.S. has been summarized in such works as Galbraith (1977), Mintzberg (1979) and
Pfeffer (1982) but "has yet to be fully incorporated into accounting systems research"
(Emmanuel et al., 1990, p. 62).
Early and Hopwood (1981) distinguished environmental uncertainty in diagnostic caus e-
effect relationships from uncertainty in strategic objectives. When objectives are clear and
cause- ef f ect relationships are well understood, decisions are programmable. Exampl es of this
might include an optimization function when constraints are known, or a discounted cash flow
analysis under relative environmental certainty. Rarely however do A.I.S. provide perfect
decision predictability. As the environment becomes less certain, Earl and Hopwood (1981)
argue that the accounting system "serves as a learning machine" from which probabilistic
inferences are made based upon syst em inputs. This would be the case for example when
budgetary revenues are compared to actual revenues for a new product, or when a post audit of
an investment in an uncertain environment is performed.
4 A.M. OUKSEL et al.
This probabilistic view of organizational learning bears a striking similarity to individual
probability learning models (Hammond &Deane, 1973). Compensatory linear additive models
measure subject learning in terms of environmental predictability, consistency, and matching of
subject cue weights to actual environmental weights. Feedback allows the individual to better
match his/her decision weights to ecologically valid environmental weights. Individual studies
of learning in auditing and managerial accounting environments have found that cue feedback
improves multiple cue probability learning and the matching of subject decision weights to
environmental weights (Ashton & Brown, 1980; Harrell, 1977). Similar studies have not
however been performed at the organizational level. Our study methodologically borrows in
spirit from individual learning experiments in operationalizing learning at the organizational
level.
Contingency theorists argue that A.I.S. should be designed to match environmental
uncertainty (Gordon & Miller, 1976; Govindarajan, 1984). Lawrence and Lorsch (1967) have
shown that environmental uncertainty affects divisional structure and the A.I.S. design used to
integrate these divisions. Multi-divisional organizational structure recognizes the unique
environmental uncertainty facing different product market environments. Under such a
structure, it is highly improbable that top management possesses the best predictive information
model for each divisional environment. Sub-unit managers not only have a better understanding
of local environments, but are also in a position to react more quickly to this information.
Hierarchical information transmission functions well in centralized and relatively predictable
environments. It encourages specialization, improves monitoring and reduces uncertainty
(Williamson, 1975). However, vertical information flows also lead to information compression
and time delays as information travels up the hierarchy. A.I.S. are often designed to aggregate
information in summary reports at each hierarchical level. For example, sales goals are often
determined at the top level. These goals are then translated into divisional and product line
profitability objectives. Production is determined by plant operations and goals are set for each
plant. Once sales and production occur, periodic divisional profit and plant performance are
reported to upper levels but detailed production reports and marketing variances remain at lower
levels.
The organization learns by comparing reported results to budgeted goals at each reporting
level. Actual performance feedback is initially gathered at the local level. Divisional production
and sales reports are examples of such feedback. Periodically, A.I.S. issue summary financial
control reports. Examples include production variances for cost centers and divisional
contribution margins for profit centers. The organization learns by comparing reported results to
budgeted goals at each reporting level. This vertical cybernetic loop constitutes organizational
learning in its simplest form.
Accounting forecast errors of individual departments however become magnified when
aggregated at each successive level (Otley & Berry, 1979). Galbraith (1977) has suggested that
vertical information distortion be reduced by improving lateral communication through team
processes and matrix information structures. Recent trends towards cross-functional teams and
flatter organizational structures typify these information communication changes. Flatter A.I.S.
are replacing vertical A.I.S. (Den Hertog & Wielinga, 1992). Cross-functional teams that now
perform previously distinct value activities require greater lateral information flows in order to
be effective. Outsourcing and electronic data interchange between internal and external sub-
units also demand greater horizontal A.I.S.
The trend towards flatter organizational structures suggests that as less information
compression occurs, organizational learning will improve. This ignores, however, environmental
contingencies facing the organization. When accounting information is very differentiated or
ACCOUNTING INFORMATION SYSTEMS AND ORGANIZATION LEARNING 5
clustered across sub-units, learning how to recognize and integrate this diverse information
becomes problematic. Horizontal information transmission is much more vulnerable than vertical
transmission to the degree of local sub-unit expertise. The impact of a single front line analyst is
not diluted by the filtering action of layers in the hierarchy. Operating under much narrower spans
of control, hierarchies have the potential to recognize highly diagnostic sub-unit information to a
greater degree than flatter organizational structures with much wider spans of control.
Hierarchical information transmission in uncertain environments also directly affects the
speed of organizational learning. In flat organizations with relatively uniform sub-unit
environments, simple decision rules are presumably much faster than hierarchical decisions.
While decision models such as simple majority rule of sub-units are simple and quick, learning
how to distinguish where more important sub-unit information resides is much more
complicated. In such cases, under a flat information structure, the speed of organizational
learning will decrease in proportion to the number of sub-units. With different sub-unit
information, the wider the span of control, the slower the organizational learning. A hierarchy
with more levels, but progressively smaller spans of control at each level, should learn more
quickly albeit, as noted, less accurately in many cases.
Our study examined the relative speed and efficiency of organizational learning under flat and
hierarchical A.I.S. Within the flatter organizational structure, two decision rules were employed,
majority and expert teams (defined later). The information system was conditioned by its
environment. Three types of probabilistic information were examined: uniform, dispersed and
clustered information weights. In the uniform case, each organizational sub-unit' s information
had the same importance relative to overall problem resolution. In the clustered case,
information varied in importance and information bits of similar weights were grouped together
by sub-unit. In the dispersed case, information bits also had varying weights of importance but
were randomly distributed across sub-units.
The success or failure of organizational learning when combining different information cues
to determine organizational outcomes depends critically, as argued, upon the coordination of
environmental information across sub-units. Organizational learning under flat information
coordination was posited to differ significantly from hierarchical learning as a function of the
environment. While previous research (Carley, 1992) has examined accuracy and speed of
learning under uniform binary bit information, the effect of clustered and dispersed probabilistic
environmental information upon organizational outcomes has not been examined. Accordingly,
the following general propositions were examined:
Proposition 1: Learning in flat organizations (majority and expert teams) under
uniform and dispersed information is more accurate than learning in hierarchical
organizations. With clustered information, only expert terms outperform
hierarchies.
Proposition 2: Learning in flat expert (majority) organizations under uniform and
dispersed information is slower (faster) than hierarchies.
Environmental uncertainty and A.LS. distribution
Information diffusion is low when it is available to only a subset of organizational sub-units.
The wider spans of control found in flat organizations often required broader dissemination of
information. In a distributed task, such as the one examined in this study, sub-units
independently solve components of an organizational problem. Each unit manager has local
information which may or may not be shared with other managers. Organizational tasks with
distributed A.I.S. are common. Product development depends on localized information from
6 A.M. OUKSEL et al.
marketing. Divisional Income Statements report sub-unit profitability. Divisional cost
information is used in inter-divisional transfer pricing negotiations. In such cases, local
accounting information may or may not be shared across sub-units.
Advances in information technology have improved A.I.S. diffusion in several ways. First,
these systems have the ability to simplify complex accounting data by reporting comprehensible
summary information. Trends in revenue data, quality reports and number of new orders are
examples of this type of information. Second, the proliferation of networked computers allows
information to be shared electronically. Profit planning goals can be broadly disseminated. Point
of sale data collection can be transmitted to production, marketing and finance in real time about
customer buying patterns. Third, database management allows managers to ask "what i f"
questions about alternative scenarios such as investments or pricing. All of this permits more
relevant and timely data to be transmitted about the fi rm' s environment.
Organizations increasingly recognize the value of cross-functional information sharing
networks. Lateral information flow has increased with the advent of recent software information
packages such as Lotus Notes; by the formation of cross-functional teams; and by the wider
spans of control found in flat organizations. These lateral networks define the distribution of
information in terms of number of nodes or links. Using network analysis, shared accounting
information has been found to amplify as well as minimize errors in individual units (Monge,
1987). As managers gain access to more information, theoretically greater organizational
learning is possible, depending on the value of shared information. Sharing information of
diagnostic importance should clearly be more useful than sharing redundant information or
information of low relevance. Vulnerability to local optima is decreased and outcomes have
been found to improve (Hutchins, 1990).
In the extreme, all units may share all information. In such cases, information diffusion is
complete and redundant. Practical considerations, however, limit information distribution.
Coordination costs arise and time delays occur as network nodes increase (Malone, 1987).
Information revelation between units may be subject to misrepresentation. Cognitively, the
ability of individuals to process large amounts of information and to integrate this information
is limited. Decision makers suffer from information overload as the amount and complexity of
information increases. Learning time also increases. While A.I.S. should be designed to activate
learning, information overload may limit the organization' s ability to focus attention. We
speculate that organizational learning differs significantly between shared and segregated A.I.S.
as a function of the environment. The following general propositions were examined:
Proposi t i on 3: Organizational learning under overlapping A.I.S. is more accurate
than learning under segregated A.I.S. with clustered information, but there is no
difference under uniform or dispersed information conditions.
Proposi t i on 4: Overlapping A.I.S. systems decrease the rate of organizational
learning relative to segregated A.I.S. under all information conditions.
SIMULATION
Two aspects of A.I.S. are addressed: management reporting structure, "who reports to whom".
and task decomposition, "who has access to what information". Our approach is an extension of
Carl ey' s (1990, 1992) experiential learning model (ELM) research framework for examining
organizational learning performance under various design constraints and operating conditions.
In previous work, Carley (1992) examined the role of hierarchy and overlapping information
ACCOUNTING INFORMATION SYSTEMS AND ORGANIZATION LEARNING 7
under uni form i nformat i on conditions. We replicate this and ext end the model to envi ronment al
uncertainty. Specifically, we exami ne how A. I. S. design is cont i ngent upon the nature of
differential i nformat i on available t o individual agents.
Before expl ai ni ng the model , the fol l owi ng t er mi nol ogy must be defined: (1) a bit of
e v i d e n c e - - a single bi nary input whi ch the organi zat i on gathers f r om its external envi ronment ,
where each bit represents the presence or absence of some feature of the envi ronment ; (2) a
t a s k - - t h e set of bits of evi dence for whi ch a single organi zat i onal deci si on must be rendered;
(3) a s u b t a s k - - a subset of the input bits of evi dence; (4) an a g e n t - - a deci si on maker within the
organi zat i on hierarchy, who convert s input evi dence from bel ow in the organi zat i on structure
into an out put decision r ecommendat i on, whi ch is passed up the structure; (5) an e v e n t - - t h e
process by whi ch an input task is conver t ed into a single organi zat i onal decision.
The model assumes that : (1) organi zat i onal deci si on maki ng behavi or is historically based;
(2) organi zat i onal learning depends on the bounded rationality of individual deci si on makers
who make up the organi zat i on; (3) individual deci si on makers anal yze subsets of the overall
task; (4) subordinates condense their input data into out put r ecommendat i ons to their superiors,
and this i nformat i on compr essi on occurs at each node in the structure; (5) overall organi zat i onal
deci si ons do not require that a consensus be reached (e.g. a legitimate pol i cy mi ght be to let the
maj ori t y opi ni on rule); (6) the overall organi zat i onal deci si on is t wo val ued (e.g. go/ no go); (7)
each intermediate deci si on is similarly t wo valued; and (8) the organi zat i on faces quasi-
repetitive integrated deci si on maki ng tasks.
Quasi -repet i t i ve tasks are similar but not identical to previ ous tasks. An integrated task means
the task is t oo compl ex for a single deci si on maker to handle. Fi rm exampl es include periodic
divisional budget s and per f or mance reports, and mul t i -product scheduling. Sub-deci si ons of
multiple agents must be combi ned in some fashion, dependi ng on organi zat i onal design, to reach
an overall decision. The tasks of interest here are assumed to be non- decomposabl e, meani ng
that combi ni ng the correct solutions to each sub task may not al ways yi el d the correct solution
to the overall task.
In this model , each time the organi zat i on faces an input task and det ermi nes what its
organi zat i onal response will be, a deci si on maki ng event is said to have occurred. Each of the
input tasks that the organi zat i on faces is represented by a bi nary string of n bits, whi ch we
denot e by b = ( b i . . . . . b n ) . Each bit bi represents one el ement of the input task. We can thus
consi der each bit as a feature of the input task and the string of bits as input evi dence that the
organi zat i on must anal yze t o det ermi ne an appropriate response. Thi s is not uncharacteristic of
frequent l y model ed account i ng deci si on envi ronment s such as audit or manageri al j udgment s in
whi ch deci si on makers are present ed with either sequential of si mul t aneous probabilistic
i nformat i on bits and asked t o reach a bi nary deci si on such as investigate vari ances/ do not
investigate variances, i nvest / do not invest, sampl e further/do not sampl e further, or qual i fy
opi ni on/ do not qual i fy opinion.
The organi zat i on is represent ed by a number of agents (the sub-deci si on makers) each of
whom has access to a subt ask (subset) of the n bit string, bi, bi + 1 . . . . . bj where 1 < i < j < n.
Each agent exami nes his/her local memor y of prior tasks (bit patterns) and the correspondi ng
past deci si on out comes in an at t empt to learn what the deci si on on the current task ought to be.
That is, t hey try to learn f r om their past experience, for exampl e whet her or not past cash fl ow
proj ect i ons were accurat e or what was the cause of hi gh defect rates. To study organi zat i onal
l earni ng vi a this model , it is assumed that the organi zat i on initially knows not hi ng about the bits
of evi dence that compri se each task ot her than the fact that each bit is t wo-val ued (0 or 1) and
that the overall deci si on to be reached is similarly t wo-val ued (0 or 1). The const ruct s of the
general model and their interrelationships are descri bed in Ouksel and Mi havi cs (1995).
8 A. M, OUKSEL et al.
Three different stylized organizational structures are studied. These structures can all be viewed
as being hierarchical in architecture, but composed of varying numbers of levels (one, two, or
three). The first is a single level structure which Carley referred to as a "majority team". The
second is a two level structure referred to as an "expert team," and the third is a three level structure
known as a "hierarchy" (Carley, 1990). Figure 1 shows the structures used in the simulation.
Twenty-seven bits (n = 27) of evidence are available per input task and each bottom level agent
handles three bits. A 27 bit task is used here since it is the smallest number of bits which allows at
least a 3 level hierarchy where each decision maker has an odd number of subordinates. An odd
number of bits is chosen so that a unique correct decision can always be directly determined (i.e. so
the majority rule decision mechanism would never result in a tie).
Each bottom level agent examines his/her past experience with the 3 bit data pattern
encountered and probabilistically determines which organizational response (0 or 1) is more
likely to be correct. The agent bases this decision upon the historical accuracy of the bit pattern
Hierarchy
(final decision = leader's judgement)
/to , o o , t 7
... bits of evidence ... 0
I 0 1 1 0 I 0 I
Expert Team
(final decision = leader's judgement)
o I o I o 7
t t t b , t s o f e , , i d e n . . . . . t ~ ) t
I o 1 1 o I o I
Agents
Majorily Team
(final decision = majority w.'~te)
, , t , t o , , , t , , t t /
... bits of e'.,idence ...
o 1 I o I o I /
Fig. 1. Organization structures,
ACCOUNTING INFORMATION SYSTEMS AND ORGANIZATION LEARNING 9
encount ered relative t o the actual out c ome - - i . e , the hit rate. In the case of the maj ori t y team,
these predi ct i ons become "vot es" and the organi zat i onal response is det ermi ned by maj ori t y
rule, whereas in the ot her structures the predi ct i ons become r ecommendat i ons to the agent s'
superiors. Each superior in t urn treats his/her subordi nat es' r ecommendat i ons as input dat a t o
his/her own deci si on maki ng process, f ol l owi ng the same probabilistic historical learning
procedure as empl oyed by t he subordi nat es (i.e. pattern mat chi ng). An exampl e of this woul d be
sales forecast s submi t t ed by product line managers whi ch are used to det ermi ne overall
di vi si onal sales estimates. These est i mat es are then revi ewed at a hi gher level and consol i dat ed
into a corporat e sales budget .
The basic model assumes that the actual envi ronment al classification funct i on, whi ch
associates a correct deci si on wi t h each task, is based on a simple maj ori t y rule. That is, i f more
t han n/2 dat a bits have values of 1, t hen the true out come is 1 (also denot ed Hi ). Otherwise, t he
true out come is 0 (also denot ed Ho). We assume n is odd to avoi d "ties" or i nconcl usi ve data,
such as a bit pattern of 0 0 11. Suppose n = 9 and there are 3 bot t om level "agent s" or deci si on
makers report i ng t o a single manager. Aft er a number of task trials, a t ypi cal deci si on maker ' s
memor y mi ght resembl e that shown in Table 1.
Furt her suppose the current input t ask was model ed as 1 1 1 0 10 0 0 1. That is, agent # 1 sees
111 and agent #2 sees 0 10 and agent #3 sees 0 0 1. Aft er mat chi ng on the input pattern 11 1
agent #1' s deci si on woul d be a 1; with pattern 0 10 agent #2' s deci si on is 0; and wi t h 0 0 1 agent
#3 decides to r ecommend a 0 (Table 1). The manager ' s input pattern is thus 1 0 0 and after
mat chi ng this t o local memory, he/she sets the overall organi zat i onal deci si on equal to a 0.
However, he/she is incorrect. The true out come is a 1 since the input t ask has mor e l ' s than O's
(5 vs. 4). Thi s t ype of task, where compl et el y correct sub-deci si ons at each stage do not
necessari l y lead to a correct overal l decision, is what is referred to as a "non- decomposabl e
deci si on task" (Carley, 1992).
"Task decomposi t i on scheme" is defi ned as the set of subtasks assi gned to the agents at the
l owest level of the organi zat i onal structure. The "segregat ed" scheme is one in whi ch each dat a
bit can be anal yzed by one and onl y one bot t om level agent (i.e. the subtasks do not overlap).
The "overl appi ng" task decomposi t i on scheme is one in whi ch a dat a bit can be anal yzed by
mor e t han one agent. Fi gure 2 shows t he t wo di fferent t ask decomposi t i on schemes that were
studied. Thi s overl appi ng scheme was i ncl uded t o study the effect of dat a r edundancy on
organi zat i onal learning.
This si mul at i on captures t he not i on of "wei ght s" of evi dence and accords with envi ronment al
reality. That is, not every bit of evi dence is necessarily equal l y i mport ant in det ermi ni ng the
Table 1. Typical memory for an agent
Input evidence H0 H1
0 0 0 57 7
0 0 1 42 22
0 1 0 42 22
0 1 1 22 42
1 0 0 42 22
1 0 1 22 42
1 1 0 22 42
111 7 57
l0 A.M. OUKSEL et al.
Agents
E1 E2 E3 E4 E5 E6 E7 E8 E9
S e g r e g a t e d
E19 E20 E21
Bits of Evidence
E22 E23 E24 E25 E26 E27
Agents
E6 E7 E8 E9
Overlapping
E22 E23 E24 E25 E26 E27 E1 E2 E3 E4 E5
Bits of Evidence
Fig. 2. Task decomposition schemes.
correct deci si on for an input task. The potential for non- uni f or m wei ght s of evi dence is model ed
via the i ncl usi on of i nt eger coeffi ci ent s for each bit of evi dence. Thus each bi nary variable, bi,
is multiplied by its wei ght coefficient, wi, and the summat i on of these product s is then compar ed
to some t hreshol d value to det ermi ne whet her the correct deci si on is H 0 or Hi .
As ment i oned earlier, it is reasonabl e to suspect that organi zat i onal learning per f or mance
depends on the "goodness of fit" bet ween an organi zat i on' s structure and its environment• In
addition, it seems intuitive, due to the law of averages over a large sampl e of simulation runs,
that non- uni f or m wei ght s distributed across input data bits woul d not yield results different f r om
the uni form wei ght s case.
SI MULATI ON RESULTS
Design of virtual experiment
The design was a 3 × 2 × 2 . Three different organi zat i onal structures were studied: maj ori t y
teams, expert teams, and hierarchies (see Fig. 1). Usi ng a Mont e- Car l o simulation, t wo different
task decomposi t i on schemes were contrasted, segregat ed and overl appi ng (see Fig. 2), as were
three types of wei ght distributions, clustered (similar weights adjacent to each other), di spersed
(dissimilar adjacent weights), and uni form (all wei ght s were equal). For exampl e, 9 9 9 5 5 5 1 1 1
is an exampl e of clustered weights, whereas 9 1 5 9 1 5 9 1 5 is dispersed. Thus, in total 18
different organi zat i onal case mani pul at i ons were studied. The operat i onal i zat i on of these
mani pul at i ons was descri bed in the Simulation section above. The model out put provi ded t wo
AC C OUNT I NG I NF ORMAT I ON S YS TEMS AND ORGANI Z AT I ON L E AR NI NG 11
dependent variables: asymptomatic performance level (learning accuracy) and speed of
learning.
Each case was replicated 50 times. That is, the simulation was run 50 times. Each run
consisted of 2500 sample decision tasks (except for Expert Teams, where 5000 tasks were
needed before organizational performance became asymptotic) 1. Each task involved 27 bits of
probabilistic information, 3 bits each across 9 agents, except the overlapping treatment which
involved 5 information bits. Tasks were repeated each time with information bits randomly
chosen with replacement (see Simulation section). "Performance level" was operationalized by
calculating the percentage of correct decision made in a given set of trials. I f 65 out of the last
100 trials were correct, the performance level was 65% at that point in the learning process.
Speed of learning was measured by the number of decision trials required until the
organization attained at a least a 69% performance level, significantly above a 50% random
accuracy hit rate. The 69% figure was chosen because it lay just below the level at which the
worst performing organizational design (majority team/segregated task decomposition/clustered
weights) had its performance cease to improve over time.
Performance scores were calculated via a movi ng average of the number of correct decisions
out of the most recent 100 trials. In addition, the performance scores were subjected to an
exponential smoothing technique to attenuate much of the variance which they would otherwise
exhibit. The weighting factor was chosen, so that for any time window the estimate of
performance was calculated as: estimate = 0.3* (current estimate) + 0.7* (previous estimate).
The smoothing factor was chosen to minimize the variance in the performance estimates,
without unduly lengthening the number of time periods required to reach the 69% performance
threshold previously discussed.
Simulation results
Results on learning performance and speed are summarized in Tables 2 and 3. To better
visualize the simulation resuilts, Figs 3- 5 are included to show graphs of learning performance
i mprovement (under a segregated task decomposition scheme).
Figure 3 shows that under conditions of uniform weights of evidence the flattest structure, the
majority team, learned both faster and ultimately better than the hierarchy. The i mprovement in
speed was due to decreased cognitive load. The majority team only had to learn 23 bits per agent
whereas the hierarchy had two levels of 23 learning. The majority t eam' s advantage in terms of
learning speed is even more pronounced when compared to the intermediate structure, the expert
team. The cognitive load of the expert t eam was much greater, involving 2 9 bits of learning.
However, given a long enough time horizon, the expert team eventually catches up to a majority
t eam' s learning proficiency level.
Figure 4, with clustered evidence weights, shows that under these conditions the majority
team structure suffers a large penalty in its organizational learning potential, as a one vote/one
person decision rule does not discriminate between analysts with different predictive
information. Figure 5, on the other hand, reveals similar trends with dispersed weights among
the three organizational structures studied to those exhibited under conditions of uniform
weights (see Fig. 3).
In the case of uniform weights, majority teams achieved a final performance level (84.8% for
segregated decomposition and 82.8% for overlapping) significantly above that of hierarchies,
A s a mp l e si ze of 50 wa s c h o s e n s i nce t he bi nomi a l pr oba bi l i t y di s t r i but i on a p p r o a c h e s a n o r ma l di s t r i but i on whe n bot h
n*p a n d n*q ar e >5. Si nce our pr opor t i ons of s uc c e s s (p) r a nge f r o m 0. 70 t o 0. 85 an n = 50 wa s d e e me d
s uf f i ci ent l y l ar ge.
12 A. M. OUKSEL e t al .
Table 2. Simulation results: final performance
Segregated task decomposi t i on Overl apped task decomposi t i on
Majority Expert Hierarchy Majority Expert Hierarchy
team team team team
Uni form weights /a = 84.8 ~ = 82.3 p = 80.0 ~ = 82.8 p = 81.9 p = 78.8
~r = 1.7 cr = 1.9 ~r = 1.8 ~r = 2.0 cr = 1.9 cr = 2.0
Clustered wei ght s ~ = 70.9 la = 83.5 la = 79.4 ~ = 76.2 ja = 83.5 IJ = 80.9
cr = 4.2 cr = 1.8 ~r = 2.0 ~y = 1.9 ~r = 1.5 cr = 1.8
Di spersed weights la = 82.6 ~ = 80.9 la = 78.2 ~t = 83.4 ~ = 82.1 la = 78.2
o- = 2.2 ~ = 1.7 ~r = 2.3 cr = 1.8 o" = 1.7 o- = 2.1
Notes: Each cell has n = 50 simulation runs. The final performance figures reported are the percentage of correct
organizational decisions achieved during the last 100 tasks (i.e. tasks 2401 through 2500).
whi c h s uf f er ed i nf or ma t i on c ompr e s s i on ( 80. 0% f or s egr egat ed de c ompos i t i on and 78. 8% f or
over l appi ng) . I n addi t i on, wi t h uni f or m wei ght s , maj or i t y t eams l ear ned s i gni f i cant l y f ast er t han
hi er ar chi es ( 18. 6 vs. 43. 9 deci s i on per i ods r equi r ed f or s egr egat ed de c ompos i t i on and 21. 2 vs.
50. 9 f or over l appi ng) . The s e r esul t s c onf i r m t hos e f ound by Car l ey ( 1992) f or uni f or m
wei ght s .
Si nce t he t wo dependent var i abl es wer e not s i gni f i cant l y cor r el at ed, t wo s epar at e uni var i at e
anal ys i s of var i ance t est s ( ANOVA) wer e us ed r at her t han a si ngl e mul t i var i at e anal ys i s of
var i ance ( MANOVA) 2. Us i ng Coc hr a n' s C and Bar l et t ' s Box F t est s, homoge ne i t y of var i ances
a s s umpt i ons wer e f ound not t o be vi ol at ed f or pe r f or ma nc e dat a but wer e vi ol at ed f or s peed
dat a. A Li l l i ef or s t est f ur t her i ndi cat ed t hat Tabl e 2 wa s f r om a nor ma l di st r i but i on. I n Tabl e 3,
Table 3. Simulation results: speed of learning
Segregated task decomposi t i on Overl apped task decomposi t i on
Majority Expert Hierarchy Majority Expert Hierarchy
team team team team
Uni form wei ght s la = 18.6 ~t = 88.1 ~ = 43.9 la = 21.2 /a = 77.2 p = 50.9
o- = 5.6 er = 19.2 cr - - 9.2 ~r = 6.2 o" = 16.8 cr = 14.7
Clustered weights ~t = 36.0 ~ = 54.7 la = 27.4 la = 24.4 p = 62.0 p = 30.0
o- = 25.5 o- = 14.5 cr = 8.2 cr = 9.4 ~r = 13.6 o" = 6.2
Di spersed weights ~ = 19.7 ~a = 87.2 ~ = 41.6 ~ = 20.3 la = 72.1 ~ = 45.3
o- = 5.5 o- = 19.3 o- = 10.2 o" = 5.9 o" = 15.2 o- = 10.1
Notes: Each cell has n = 50 simulation runs. The speed of learning figures reported were calculated by dividing the # of
decision tasks encountered (until a 69% exponentially smoothed performance level was attained) by 10. See Figs 3-5.
2 Only one of the 18 cells showed a significant correlation coefficient at the a = 0.05 level, and no cells showed
significance at the ~ < 0.01 level.
ACCOUNTING INFORMATION SYSTEMS AND ORGANIZATION LEARNING 13
9 0 ~ M a j o r i t y T e a m ~
!:°o
5 0 ~ . . . . - -
0 0 0 0 0 0 0 0 0 0 0 0 0
OJ ~" tO O0 0 04 'q" ~0 ~ 00q I~"
OJ O4 O4
T i m e ( i n p e r i o d s )
Fig. 3. Learni ng s peed- - uni f or m weights.
- M a j o r i t y 1
• - - H i e r a r c h y
however, the test cast some doubt as to the assumption of normality. Non-lJarametric tests for
differences in medians were therefore used for speed of learning.
More formally, our proposition 1 (P1) posited that majority and expert teams would
outperform hierarchies when information weights were uniform or dispersed. As can be seen in
Figs 3 and 5, mean final performance of both types of teams was better than for hierarchies
across overlapping and segregated information distribution schemes for both uniform and
dispersed conditions (F = 390; P = 0.00) 3, confirming P1. Hierarchies suffered performance
degradation because they encountered information compression losses across two levels of
decision making,
Proposition 1 also predicted that hierarchies would not be the best performer when weights
were clustered. This conjecture was also corroborated. Hierarchies performed significantly
better than majority teams (F = 197; P = 0.00) but significantly worse than expert teams across
overlapping and segregated information distribution schemes under the clustered weight
condition (F = 166; P = 0.00). Looking at the figures shown in Table 2, we see that when
weights are clustered, the majority teams' final performance is significantly worse than either
hierarchies or expert teams. This is also seen in Fig. 4.
A simple majority rule, where each analyst's opinion counted equally towards the
organizational decision, was ineffective when some decision makers had access to more (less)
important information than others. The reason for this is that having each agent's vote count
equally towards the determination of the overall organizational decision is simply not
appropriate when some agents have access to more important evidence than do others. The real
9 0 - E x p e r t T e a m
8 5 ! . . ~ _ 4 ~ _ ~ . ~ - ~ w ( ~
o ~ 8 0 4 H= e r a r c hy _ ~ . . _ ~ . - ~ e - 4 ~ - - ~ . ~ - ~ ~ ¢ ~
7 5 i 4 - - 9 ~ T - ~ - ~ ~ - m
. . - ' ~ ~ n ~ - ~ - - • ~q- = H / = - ~ - a - e ~ - =
~ 7 0 I .~'~;~-.m-IC~ C~=-•-~ i -m-B" I M a j o r i t y I e a m
6 5 i . - ' ' . . c U ] .~
6 0 .~.~
-'i
50
0 0 0 0 0 O
Owl ~ ~ aO 0
0 ~ 0 0
0 4 , . 8 8 8
04 O4 O4
T i m e ( i n p e r i o d s )
Fig. 4. Learning speed- - cl ust er ed weights.
= M a j o r i t y
E x p e r t
• H i e r a r c h y
• ~ In each F test, the degrees of freedom for the numerator (between groups variance) was 1 while the d.f. for the
denominator (within groups variance) was 98.
14 A. M. OUKSEL e t al .
Majority Team
8O ~ , ~
70 I ~ H i ~ y . . ~ j ~ - - -
~ 6 5 t j ~ ~ ' ~ ~ .=
60 ~ , ~ ~ ! ._.
50 ~l~ ' ~"
O O O O Q
¢M ~1" (.O G0, O
r--
(~1 ~1" r i d ~ "
~-- Od O,J
T i m e ( i n p e r i o d s )
Fig. 5. Learning speed--dispersed weights.
-- Majority
~ Expert
* Hierarchy
worl d anal og for this situation is that, with maj ori t y teams, the agents who really "know what
t hey are talking about " have no more influence on the final deci si on than those who may be
merel y guessing. Bot h expert teams and hierarchies had some means by whi ch to learn whi ch
analysts were more reliable, and thus came to wei gh their r ecommendat i ons mor e heavily. The
hi erarchy was still subject to great er i nformat i on compr essi on loss than the expert teams,
however, and thus did not perform as well.
Proposi t i on 2 was also confi rmed. As i ndi cat ed in Table 3, due to the i ncreased cogni t i ve
learning of t wo levels of processing, it t ook hierarchies consi derabl y mor e time to learn than
maj ori t y t eams across overl appi ng and segregat ed i nformat i on distribution schemes, when
i nformat i on wei ght s were uni form or dispersed (X 2 = 283; P = 0.00). This occur r ed because the
maj ori t y t eams' learning rate was dependent onl y on each anal yst ' s speed of learning 23
different bit patterns, whereas hierarchies also had to wait for manager s to learn (2(23)).
However, hierarchies needed si gni fi cant l y less time to learn than did expert teams, again under
uni form or di spersed i nformat i on (X 2 = 237; P = 0.00). The reason for this was that the leader
within an expert t eam had to learn 2 9 different bit patterns. In ot her words, expert teams were
si gni fi cant l y sl ower in learning than either hierarchies or maj ori t y teams because the expert had
512 different possi bl e input dat a patterns to learn. This reflects the c ommon not i on of cogni t i ve
overload.
Proposi t i on 3 was conf i r med since a significant i mpr ovement in organi zat i onal per f or mance
was found with overl appi ng i nformat i on syst ems relative t o segregat ed syst ems across t eams
and hierarchies under clustered i nformat i on wei ght s ( F = 67; P = 0.00), but not under the other
i nformat i on conditions. Here we see that the expansi on of each anal yst ' s data access from 3 bits
of evi dence to 5 bits hel ped reduce the number of analysts whose votes were based on
uni mport ant i nformat i on. In the uni form and di spersed i nformat i on conditions, overl appi ng
i nformat i on had little effect, as anticipated.
The onl y proposi t i on not borne out by the results was P4. No significant differences were
f ound when compar i ng learning speeds for overl appi ng versus segregat ed distribution schemes
for maj ori t y t eams faci ng uni f or m or dispersed i nformat i on weights (X 2 = 0. 048; P = 0.83). The
increase in task compl exi t y of learning 25 patterns v e r s u s 23 was not significant. Had the
overl appi ng distribution scheme been wi der (say 7 bits per analyst), significant differences may
have begun to appear. On the other hand, perhaps the enhanced quality of the anal yst s'
predictions (since each coul d vi ew a larger proport i on of the total probl em) served to mitigate
the increasing compl exi t y of learning more bit patterns.
Additionally, Table 3 seems to indicate a possible increase in learning speed when
i nformat i on is clustered, bot h for hierarchies and f or expert teams.. This is due to the fact that
with clustered weights, some anal yst s' recommendat i ons (the ones with little useful
ACCOUNTING INFORMATION SYSTEMS AND ORGANIZATION LEARNING 15
information) can be safely ignored. The reduction in complexity which results once such
analysts can be identified seems to more than offset the time needed to learn which analysts
these are.
It is also interesting to note that the final performance levels for hierarchies are not
significantly different across the information conditions or between the task decomposition
schemes. The real world analog here seems to be that hierarchies are fairly stable, consistent
performers. They underperform majority teams when weights of evidence are evenly distributed
among decision making agents, but outperform them when important evidence is concentrated
in certain areas. Similar results were found when comparing expert teams to majority teams.
Simulation parameters and cognitive load
It is important to appreciate the extent to which the results are attributable to cognitive load.
Expert teams learn more slowly due to an increased number of patterns they must recognize. As
this is a function of the number of agents in the team, a natural question to be considered is
whether the results are invariant to team size. In our simulation, decreasing team size from nine
to six decreases the number of patterns to be learned by the team from 2 9 t o 2 6, a 4-fold
decrease. A similar agent decrease does not affect the hierarchy as significantly. The number of
middle level managers would decrease from three to two, resulting in one less 2 3 middle
manager pattern to be learned and one less manager pattern for the leader to learn. Hierarchies
will continue to have less complex patterns to learn because of narrower spans of control. The
speed differential between hierarchies and expert teams however will narrow as team size
decreases result in proportionately greater reduced cognitive load. This is an important
consideration when trading off expert team accuracy against speed.
Second, in our study only one hierarchical level was simulated. As the number of hierarchical
levels increases, hierarchical accuracy further deteriorates due to information compression.
Speed would also decrease as more managerial patterns need to be learned. This would make
teams comparatively more advantageous. Third, overlapping information was simulated using
2 5 ver s us 2 3 agent bits in the segregated condition. As overlapping information only improved
majority team accuracy with clustered information, increasing the overlapping information bits
would improve performance in this condition, but at a cost, speed. Cognitive load would
increase by 2 n for n additional overlapping bits of information.
DISCUSSION
A.I.S. are designed to activate learning across a range of management accounting activities.
Current corporate concerns include, among others, operating budgets, capital investments,
productivity measures, new product development, value added activity analysis, target costing,
outsourcing and process re-engineering. In each of these areas, financial and non-financial goals
are determined. Actual performance is measured relative to these goals as managers learn
through feedback. An empirical question is the extent to which A.I.S. design and performance
in these operational areas have been affected by recent trends in flatter informational structures,
team decision making and lateral information sharing networks. Our study provides preliminary
theoretical evidence with respect to these issues.
Forecast accuracy is particularly important for certain accounting decisions such as budgetary
profit projections, capital budgeting cash flow estimates and target cost projections. Consistent
with Carley (1992) and Williamson (1975), our results suggest that as vertical information
compression is decreased, lateral A.I.S. perform well relative to hierarchical A.I.S., provided the
16 A.M. OUKSEL et al.
informational environment is relatively uniform across agents. This suggests that for fairly
routinized accounting projections such as project management scheduling or operating budgets
for mature product lines, flatter A.I.S. with only local agent feedback should be adopted.
Majority teams follow a one person/one vote decision rule and so have no way of learning to
discriminate between analysts with good predictive information. Assuming relatively uniform
weights of evidence for operating budgets and project management, there is no need for such
discrimination since all analysts have access to equally important information. Thus the majority
team structure is well suited to this type of environment.
When environmental information of differential importance is clustered within agents, lateral
A.I.S. no longer indiscriminately outperform hierarchies. With clustered information, majority
teams are more vulnerable to uninformed analysts' errors than hierarchies. In hierarchies,
multiple levels of information processing buffer the organization from analyst error by reducing
the number of cases in which a single analyst can affect the overall organizational decision
outcome (Carley, 1990). These hierarchical levels on the other hand also introduce information
compression and loss of accuracy. In clustered environments, expert teams outperform both
majority teams and hierarchies. Managers learn to differentially weight their subordinates
recommendations. As found empirically (Gordon & Miller, 1976; Govindarajan, 1984;
Khandwalla, 1972), environmental contingencies affect accounting reporting. In more uncertain
environments, where the relative importance of local information differs widely between agents,
expert learning is necessary. A.I.S. should be designed to track agents and differentially weight
multi-agent information.
Capital investments for example demand multi-agent expertise. A.I.S. should be designed to
monitor cash flow components of project forecast accuracy over repeated capital budgeting
decisions. This reinforces learning of the relative importance of subsets of accounting
information such as technology transfers, quality assurance and taxation issues which are jointly
vital to the success of the investment. Institutional memory of these events can be captured and
transferred to newly formed capital budgeting teams in order to mitigate the risks associated
with environmental uncertainties.
Overlapping agent A.I.S. reporting only improves performance when information is clustered
within majority teams. This accords with intuition. Sharing uniform or randomly distributed
accounting information does not improve performance. Nor is sharing clustered agent
information necessary for an expert team that learns where diagnostic information resides. For
example, performance of process re-engineering does not depend on overlapping production or
human resource information between team members. Since individual expertise is recognized
by the team, there is no point in duplicating this expertise between agents.
The findings related to the value of overlapping A.I.S. in majority teams have implications for
ad hoc cross-functional teams. For example, teams constituted solely for the purpose of
decreasing targeted production costs, developing product prototypes and launching new
products are increasingly common. In order to attain a target cost at Toyota, marketing provides
sales projections for alternative engineering designs under various cost scenarios associated with
these designs. In newly constituted teams, with no possible history of environmental prediction
expertise, a majority rule decision may well prevail. Our results suggest that A.I.S. designed
with overlapping clustered information between agents significantly improves performance.
In today' s rapidly changing technological environment, not only accuracy but speed of
organizational learning is also a vital consideration in A.I.S. design. A.I.S. must reflect the fact
that the firm' s value chain from product development through customer delivery is very time
dependent. Product development times and life cycles are much shorter; supplier linkages are
Just-in-Time; and production throughput and product delivery are faster. When information is
ACCOUNTING INFORMATION SYSTEMS AND ORGANIZATION LEARNING 17
not clustered between agents, our results indicate that majority teams are not only more accurate
than hierarchies but perform significantly faster as well. This corroborates the intuition of
hierarchical A.I.S. having a slower response time than flatter information structures. Routinized
decisions such as raw material deliveries, production scheduling and local performance
reporting feedback are well suited to flatter non-overlapping A.I.S. design. In such
environments, majority teams minimize information compression and thus tend to learn better
than hierarchies, and also more quickly than expert teams.
More compl ex management accounting decisions such as a value chain analysis of a product
line or outsourcing decisions may require greater agent expertise. Information of differential
importance exists between t eam members. While expert teams can lead to i mproved
performance, they do so at a cost, time. Expert t eam leaders need time to learn the relative
abilities of their subordinates. Managers within hierarchies also face this problem, but the
reduced span of control afforded by increased levels of management limit the complexity of
learning (i.e. 23 possible input patterns per manager in the hierarchy versus 29 for the expert
t eam leader). As a result, hierarchies also incur speed of learning delays, but to a much lesser
extent than the expert team.
The implication for A.I.S. design is the cost -benefi t tradeoff of speed versus accuracy in
clustered information environments such as capital investments and value chain analysis. When
speed is a predominant concern, hierarchies should be used. When accuracy is essential, expert
teams should be employed. These findings are summarized in Fig. 6, a contingency model for
choosing A.I.S. support based on the parameters identified in this study. When faced with a high
degree of uncertainty with respect to task environmental characteristics, hierarchies seem to
represent a conservative choice since they provide the most stable accuracy and speed of
learning performance across all combinations of environmental parameters in the study. These
findings extend Carl ey' s (1990, 1992) earlier results regarding the robustness of hierarchies to
uncertain environments.
While the task is an abstraction of the organizational learning proces, it does provide a set of
empirically testable propositions for future research. Of interest are issues of information
asymmetries and how problems of agency might impede the learning process. In addition,
organizational learning becomes more complex in dynamic environments. In such cases,
Bayesian learning models might be of interest. The optimal tradeoff between the number of
analysts and breadth of information is particularly important in "rightsized" managerial
environments. Finally, it would be of interest to examine alternative information structures and
Clustered
Wei ght s
Distribution
Dispersed
and Uniform
Hierarchy
Majority
Team
Expert
Team
Majority
Team
T i me Accur acy
Critical Fact or
Fig. 6. A contingency model.
18 A. M. OUKSEL et al.
t h e c o s t o f i n f o r ma t i o n r e l a t i v e t o o r g a n i z a t i o n a l o u t c o me s ( Ma l o n e , 1987) . Al t e r n a t i v e
me t h o d o l o g i c a l a p p r o a c h e s s u c h a s f i e l d s t u d i e s , o r g a n i z a t i o n a l s u r v e y s a n d c o n t r o l l e d
e x p e r i me n t a l t e s t s o f t h e s i mu l a t i o n r e s u l t s n e e d t o b e p e r f o r me d t o e i t h e r s u b s t a n t i a t e o r r e f u t e
t h e p r e l i mi n a r y f i n d i n g s o f t h i s s t u d y .
Acknowledgements--The helpful comment s of the editor, an anonymous reviewer and workshop participants at the
Fifth Biannual Accounting Research Conference at the University of New South Wales and the City University of Hong
Kong are gratefully acknowledged.
R E F E R E NC E S
Argote, L., Beckman, S. & Epple, D. (1990) The persistence and transfer of learning in industrial settings. Management
Science, 36, 140-154.
Argyris, C. & Schon, D. (1978) Organizational learning: A theory_ of action perspective. Reading, MA: Addison-
Wesley.
Ashton, H. & Brown, E R. (1980) Descriptive modeling of auditors' internal control judgments: replication and
extension. Journal of Accounting Research, Spring, 1-15.
Attewell, E (1992) Technology diffusion and organizational learning: The case of business computing. Organization
Science, 3, 1-19.
Carley, K. (1992) Organizational learning and personnel turnover. Organizational Science, 3, 20-46.
Carley, K. (1990) Coordinating for success: trading information redundancy for task simplicity. In Proceedings of the
23rd Annual Hawaii International Conference on Systems Sciences, Kona, Hawaii.
Cyert, R. M. & March, J. G. (1963)A theory of t hef i rm, Englewood Cliffs: Prentice Hall.
Den Hertog, F. & Wielinga, C. (1992) Control systems in dissonance: the computer as an ink blot. Accounting,
Organizations and Society, 17, 103-127.
Earl, M. J. & Hopwood, A. G. (1981) From management information to information management. In H. C. Lucas et al.
(Eds) The information system environment (pp. 123-156). Amsterdam: North-Holland.
Emmanuel, C., Otley, D. & Merchant, K. (1990)Accounting f or management control. London: Chapman and Hall.
Fiol, C. M. & Lyles, M. A. (1985) Organizational learning. Academy of Management Review, 10, 803-813.
Flamholtz, E. G., Das, T. K. & Tsui, A. S. (1985) Toward an integrative framework of organizational control.
Accounting, Organizations and Society, 10, 35-50.
Galbraith, J. R. (1977) Organ&ational design. Workingham: Addison Wesley.
Gordon, L. A. & Miller, D. (1976) A contingency framework for the design of accounting information systems.
Accounting, Organizations and Society, 1, 39-70.
Govindarajan, V. (1984) Appropriateness of accounting data in performance evaluation: An empirical examination of
environmental uncertainty as an intervening variable. Accounting, Organizations and Society, 9, 125-135.
Hammond, K. R. &De a ne , D. H. (1973) Negative effects of outcome feedback in multiple cue probability learning.
Organizational Behavior and Human Performance, Februa~, 30-44.
Harrell, A. M. (1977) The decision making behavior of air force officers and the management control process. The
Accounting Review, October, 833-841.
Hutchins, E. (1990) The technology of team navigation. In I. Galagher, R. E. Kraut and C. Egido (Eds) Intellectual
Teamwork (pp. 191-220).
Khandwalla, P. N. (1972) The effect of different types of competition on the use of management control. Journal of
Accounting Research, 10, 275-285.
Kim, D. H. (1993) The link between individual and organization. Sloan Management Review, 34, 37-50.
Lant, T. K. & Mezias, S. J. (1992) An organizational learning model of convergence and reorientation. Organization
Science, 3, 47-71.
Lawrence, P. R. & Lorsch, J. W. (1967) Differentiation and integration in complex organizations. Administrative Science
Quarterly, 12, 1-48.
Levitt, B. & March, J. G. (1988) Organizational learning. American Review of Sociology, 14, 212-243.
Levitt, B. and March, J. G. (1995) Chester I. Barnard and the intelligence of learning. In O. E. Williamson (Ed.)
Organization Theory (pp. 11-37). Oxford: Oxford University Press.
March, J. G. & Olsen, J. E (1976) Ambiguity and choice in organizations. Oslo Uuiversitets, Forlaget.
Malone, T. W. (1987) Modeling coordination in organizations and markets. Management Science, 33, 1317-1330.
Mintzberg, H. (1979) The structuring of organizations. New Jersey: Prentice Hall.
ACCOUNTING INFORMATION SYSTEMS AND ORGANIZATION LEARNING 19
Monge, E R. (1987) The network level of analysis. In C. R. Berger & S. H. Chaffee (Eds) Handbook of Communication
Science (pp. 61-87). Newbury Park: Sage.
Nelson, R. R. & Winter, S. G. (1982)An evolutionary theory of economic change. Cambridge MA: Harvard University
Press.
Otley, D. T, (1980) The contingency theory of management accounting: Achievement and prognosis. Accounting,
Organizations and Society, 5, 194-208.
Otley, D. T, & Berry, J. (1979) Risk distribution in the budgeting process. Accounting and Business Research, 9,
325-337.
Ouchi, W. G. (1977) The relationship between organizational structure and organizational control. Administrative
Science Quarterly, 22, 95-113.
Ouksel, A. M. and Mihavics, K. (1995) A general model of Organizational Learning (UIC CRIM Technical Report
1995-01) The University of Illinois at Chicago.
Pfeffer, J. (1982) Organizations and organization theory. New York: Pitman.
Senge, P. M. (1990) The fifth discipline: the art and practice of the learning organization. New York: Doubleday.
Simons, R. (1995) Levers of control. Boston: Harvard Business School Press.
Weick, K. E. & Roberts, K. R. (1993) Collective mind in organizations: heedful interrelating on flight decks.
Administrative Science Quarterly, 38, 357-381.
Williarnson, O. E. (1975) Markets and hierarchies: analysis and antitrust implications. New York: Free Press.