Working PaperPDF Available

Combining humans and machines: A new frontier in supply chain management in the retail sector

  • Ops Mend
1 of 18
Combining Humans and Machines: A New Frontier in Supply
Chain Management in the Retail Sector
Luis Herrero
University of Zaragoza
Santiago Kraiselburd
MIT-Zaragoza International Logistics Program
INCAE Business School
Rogelio Oliva
Texas A&M University
MIT-Zaragoza International Logistics Program
Noel Watson
MIT-Zaragoza International Logistics Program
Among the fundamental aspects to consider when optimizing overall supply chain
performance is the financial impact of supply-related decisions at the points closest to
the end consumer. To achieve better performance in this area, the most advanced
retailers typically model their processes using complex automatic systems that
encompass large amounts of data, operational parameters, and specialized software.
But because no automatic system can fully capture the incredible complexity of the real
world, introducing appropriate uses of human intelligence to the operation of these
systems can be the key to superior performance.
The “mom and pop” stores that traditionally provided food, clothing, and other must-have
products have, since the 1940s, been giving way to chains operated with store support
systems duplicated or shared across hundreds of storefronts. The new moms and pops
are professional store managers, and the relationship between business and store
managers (and of course customers and store managers) has become one of the key
determinants of retail performance.
A challenging, almost philosophical question concerning the role of store managers and
their relationship to the business has persistently perplexed retailers. Is the store
manager a partner or an employee, an executor or a trustee? These questions raised
themselves as we examined one of the more mundane activities a store manager can
2 of 18
be called upon to perform managing inventory replenishment decisions. Whereas
technological progress in recent years would seem to suggest that store managers
should assume more of an executor role with respect to such mundane activities, our
analysis suggests otherwise. Specifically, we believe that one key to superior
performance is a collaborative partnership between store managers and the
management information systems (machines) that support them. Notwithstanding the
apparent simplicity of such decisions understood by many practitioners and academics
alike to be deceiving we believe the lessons learned from our study do have broad
Technology has been a key enabler of performance since the 1940s, facilitating the
automation of many processes that used to be performed by human beings. What can
only be characterized as an explosion in the use of technology in recent years, in the
retail sector in general and mass consumption sector in particular, is coincident with the
advent of enterprise planning information systems and quick-to-follow supply chain
planning systems. Developed initially for manufacturing, these systems’ potential value
to other industries quickly became apparent (attendant implementation challenges were
less quickly perceived).
Interestingly, these supply chain planning systems automated the use of decision-
making algorithms and heuristics already in use in the very organizations in which they
were being introduced. In retail, inventory replenishment decision-making systems
offered the promise of improving return on investment in inventory and radically
transforming the management of the business and the role of people charged with
inventory decisions.
The promise of algorithm-supported decision-making was expounded in the seminal
HBR article, “Rocket Science Retailing is Almost Here Are You Ready?i The reality
emphasized in the article was that automation does not eliminate, or exhaust
opportunities to improve, the human factor. Their reference to “the marriage between art
and science” spoke to the authors’ recognition that someone must guarantee the quality
of the information processed by such systems and adapt system parameters to
changing realities, and that for processing certain patterns and information the human
mind is superior to a computer.
But confidence in this marriage as relates to store managers and inventory
replenishment decision-making is not necessarily high within the retail community,
among practitioners or academics. We encountered essentially two philosophies. One
can be summed up as “What the system asks for is sacred,” the other as “The system
suggests, but the manager decides.” Asked what they would expect to see from our
study, three quarters of retail practitioners expected store managers to make matters
worse when they modified recommendations from an automated replenishment system.
Can this marriage work then? Our research indicates that some of the best companies in
the world are successfully exploiting the relationship between human and machine. In
3 of 18
examining some of the dynamics that could influence the outcomes of partnerships
between store managers and in-store information systems, we find that if the work
method described here is not applied, the first philosophy, the so-called sacred system,
normally performs better than the second.
But we also found clear and convincing evidence that significant performance
improvements are possible when the second, “manager is the final decision maker,”
philosophy is correctly applied. Here, we propose as the second stage of Rocket
Science Retailing” judicious, methodical, harmonious decision-making by managers co-
existing with technology.
To better understand the conditions that favor one philosophy over the other, we first
consider the specific advantages and disadvantages of automation versus reliance on
human intervention in retail operations.
The Logic of Automation in Retail
A modern supermarket might sell anywhere from a few thousand to several thousand
items. Consider a chain with some 100 storefronts (major world chains have ten times
this number) and a store format with only 5,000 SKUs per store, and assume that the
stores process restocking orders, on average,
weekly (a rather typical frequency). A quick
calculation reveals that the chain must make, on
average, 500,000 restocking decisions per week!
Bearing in mind that the net profit margin of a
supermarket chain seldom exceeds 5-10% of sales,
and considering typical sales per year, number of
stores, and SKUs, the conclusion is that each
restocking decision affects, on average, only a few
cents of net profits. It is the large number of
decisions, which individually have limited impact on
overall retail performance but collectively determine
a chain's profitability, that has motivated the sector’s
increased use of information technology (IT). This is
so, in part, because the increasing speed of
computers can accommodate the scale of such
decision-making in response times reasonable for
such operations.
Example of Psychological
Overconfidence: overestimating
accuracy or performance in
uncertain situations.
Frame Dependence: being
influenced by the way a problem
is “framed” or the manner in
which a situation is described.
Representativeness: thinking that
causes and effects must resemble
each another when this is not
necessarily the case.
Gambler’s Fallacy: seeing
correlation where it does not
Influence of emotions: responding
on the basis of emotion rather
than calculation of benefits and
4 of 18
Nor is it just the scale of decision-making that provides
the logic for automation in retail. Human-based or
judgmental decision-making is known to be
inconsistent. In the 1960s, operations management
scientists observed that production rules determined by
past decisions made by production schedulers could
outperform future decisions by the very same
production schedulers (for examples, see Bowman
1963 ii and Sterman 1989 iii). These observations were
evidence of a variance bias, that is, an inclination for
decision making to be appropriate, on average, but
exhibit significant variance around this average.
This inconsistency is only the tip of the iceberg with
respect to the so-called psychological limitations of
human-based decision-making. Experimental
economists have identified a number of consistent
biases including overconfidence, frame dependence,
representativeness, gambler's fallacy, and the influence
of emotions (see details in insert). These at times
serious limitations in human judgment are the primary
motivation for outsourcing decision-making to rules-
based algorithms.
In Defense of the “Human Factor”
The source of the advantages of computers over
human-based judgment is the existence of
economically programmable solutions for a large
number of restocking situations. Availability of
information, ease of interpretation of data, and ease of
problem recognition are among the factors that render a
solution suitable for automated computation (see details
in insert). But not all solutions represent a good trade-
off in terms of the effort required to implement the
solution versus the quality of outcomes. Economically
programmable solutions are those that generate a
positive return on effort.
Whereas not all optimal solutions (solutions that
produce the best possible outcome for some defined
objective) may be classified as economically
programmable or even programmable, heuristics
(solution methods that involve a trade-off between
Among the factors that contribute to the
non-programmable nature of a solution
Availability of information
Automated demand forecasts can be
based on historical data, but what can
information technology do when historical
data is limited or useless? This is precisely
the case for new products, which, by
definition, lack historical data, an extremely
frequent situation in the retail sector due to
the generalized trend of ever more
promotions and new product launches
(each of which much be given its own SKU
and price).
Ease of interpretation of data
In general, the more short-lived demand,
the greater the dependence on cultural
parameters and codes that are difficult to
quantify but can be estimated by an expert,
and the more valuable such experts’
opinions. An example is fashion, one of the
sectors in which demand changes most
Ease of problem recognition
Another important factor that favors human
intervention is response speed to structural
changes. Because many automatic
systems base forecasts on some type of
weighted average from historical data, it
can take some time to adjust automatically
generated forecasts to shifts in demand
(usually by averaging the new reality using
historical data that is not necessarily valid).
In many cases, only the point of sale
manager is able to react sufficiently quickly
to the thousands of variables in the local
environment that may abruptly affect
demand. These might include, for example,
the competition, circumstances of the
global environment, a meteorological
phenomenon, an important public event
and isolated changes in the local
environment. All of these factors are
impossible to introduce in an automated
decision making model, but they
significantly affect the demand behavior.
5 of 18
quality of solution and time to generate the solution), being almost by definition
programmable, provide options for economically programmable solutions. Heuristics are
thus generally the practical option for many automated inventory replenishment
problems, the other option being, of course, human judgmental decision-making.
Based on our definitions and observations, judgmental decision-making seems to be
appropriate in cases in which its return on effort is greater than that of a programmable
solution (optimal or heuristic). This observation implies that it is not enough that the
solutions generated by humans be better than those generated by programmable
heuristics. Magnitude of savings and number of problem opportunities also contribute to
this calculus. For example, if the number of problem opportunities is small, judgmental
decision-making may not justify its labor cost.
As IT systems encompass more and more programmable heuristics and optimal
solutions, a relevant question becomes “Where are the opportunities for the human
factor?” Three empirical observations can be made.
First, according to Oliva and Watson (2009), any decision support system will have blind
spots, that is, unintentional but systematic biases resulting from lack of information in a
particular domain or limited algorithms or heuristics for processing the information.iv
Informational blind spots result from information or processing lags or information
channel breakdowns in an IT system. Procedural-based blind spots, which usually
derive from algorithms that are inappropriate or inadequate or missing from the IT toolkit,
precipitate mismatches between solution and problem. Blind spots, albeit generally
unintentional (designers of decision support systems do not know what they do not
know), often result from explicit design choices, as when it becomes too expensive to
collect and maintain the requisite information. Human decision-making, to the extent that
it does not exhibit these blind spots for some decisions, might complement, and thereby
improve, an automated decision support system.
Second, it is likely that blind spots in an IT approach reflect (increasingly more frequent)
previously uncommon” or “special” events, some of which will have serious
consequences if ignored. Examples include the promotion or introduction of competing
substitute products, weather and traffic incidents, strikes, and so forth.
Third, with current technologies, the operation of information systems still relies on the
support of human participants. With current technologies, humans still play an integral
role in the organizational/operations processes and systems that provide the data that is
used by these information systems and that execute their recommendations. Therefore
even in settings characterized by programmable solutions, humans inevitably act as
intermediaries between the information system that makes recommendations and the
reality that these information systems try to both assess and affectv
6 of 18
Categorizing Retail Operating Systems
Our distinction between information system-based and human decision-making is
instructive for categorizing retail operating systems both past and present. In terms of
human decision-making, a further distinction between centralized human decision-
making, as at a regional or national head office, and decentralized human decision-
making, as in a store by a store manager, is useful. Our categorization scheme relies on
a fundamental distinction between which entity, centralized information technology
(hereafter, “IT”) or centralized (or decentralized) human intelligence (hereafter, “HI”),will
have final versus supportive responsibility for decision-making. By supportive
responsibility is meant the provision by an entity of some information that supports the
final decision maker. Decentralized HI, for the purposes of the present discussion, is
understood to be store managers.
One categorization of retail operating systems based on these distinctions is depicted in
Table A. In this table, the typical modern approaches of centralized large chains (e.g.,
Home Depot) fall within the first column, where the final decision is left to the IT system.
Note that within this context the IT system could be supplemented with either expertise
centralized in the organization headquarters, or with local information from the
decentralized store managers. The second column represents alternative approaches
used by centralized large chains (e.g., Zara), and the third column represents an
approach that yields the final decision rights to the local store managers. Again, any of
these approaches can be supplemented by any of the two other approaches.
Final Decision Maker
Human Intelligence
Table A
Traditional “mom and pop” store managers (or owners) have total final responsibility for
decision-making. In the absence of a centralized HI or IT system, they make decisions
based on personal experience and capabilities. Prior to the adoption of IT by retail, early
retail chains assigned final responsibility for decision-making to store managers, albeit
with support from centralized HI. At the advent of and during the so-called “technological
revolution,” the continuing evolution of retail chains saw, first, store managers’ decision
making subsumed by a centralized HI model, and, second, new centralized IT
developed to support the new model.
The current trend among large retailers has been a transition towards IT centralization.
The justification for this transition is the inconsistency and bias of human decision
7 of 18
makers and the difficulty in assessing the impact of a rigid decision algorithm and the
loss of context-rich local data from store managers. Our proposal is to enrich these
decision support systems with the flexibility and responsiveness of human decision
makers. First by giving final decision rights to a centralized HI, and, eventually, to local
decision makers.
We now briefly illustrate, in a series of case studies, scenarios in which each of these
models was adopted. In the first two cases (A1-A2), variations of the widely accepted
trend of assigning final decision-making responsibility to IT were adopted, not all of them
successful. The next case (B) describes the adoption of a centralized HI model that is
performing quite well, and the concluding cases (C1-C2) scenarios in which final
decision-making responsibility was successfully assigned to store managers. We
maintain that case (B), although extremely interesting, applies primarily to sectors
characterized by high uncertainty and the domination of local by global trends, in other
words, special cases. We believe that the most common retail settings are best served
by the approach described in cases (C1-C2).
Case A1. Final decision making responsibility resides in centralized IT with
centralized HI and store managers in support: Home Depot
The shift to centralized decision-making in a modern retail chain is exemplified by Home
Depot. During the first 20 years of the company’s 38-year history, store managers had a
great deal of autonomy in decision-making related to their stores including management
of inventory replenishment. A radical change to the store’s culture and decision
processes with the adoption of a centralized IT system with extensive HI support did not
produce the expected It was not for want of many good ideas, but owing to the
complex reality of the retail business, that the new operating system (OS) did not work
well on the floor. Positive aspects of a centralized IT model replete with “modern”
technology and processes did not compensate for the loss of important input formerly
contributed by store managers, a consequence, perhaps, of the new organizational
culture that deemphasized the importance of the role of store managers.
Case A2. Final decision-making responsibility resides with centralized IT with local IT
support: Wal-Mart and the use of RFID
Wal-Mart announced in July 2010 that its stores would begin using RFID tags, which
enable the location and movement of inventory to be automatically tracked by
strategically located radio frequency readers. The company had previously used such
tags to track pallets and cases in its supply chain. Until recently, their price had made
item-level tracking infeasible, but the inevitable reduction in the price of such tags has
increased the possibility of operating systems that can truly be categorized as
centralized IT with local IT support.
In this example of how technology can improve and radically change operations, the use
of RFID casts local IT in a supportive role that implies less dependence on the discipline
of store personnel to assure the accuracy of inventory level calculations. If a company
8 of 18
embraces a “what the system asks for is sacred” philosophy (Wal-Mart does), local data
will be more accurate and the system’s global performance will improve. If, on the other
hand, the company philosophy is “the system suggests, but the manager decides,” the
new IT support role could be exploited to enable frontline employees to allocate more
time to other support tasks, thereby multiplying the potential of human intelligence.
Case B. Final decision-making responsibility resides in centralized HI with store
managers and centralized IT in support: Zara
Fashion is perhaps most illustrative of an industry in which knowledge is unarticulated,
trends change rapidly, and human intuition can play a fundamental role. This is
particularly true with quick response manufacturing and resupply systems. Of all the
store chains in this sector, Zara (part of the Inditex Group) might possibly be the best
known. The human factor has played, and continues to play, a major and essential role
at Zara. But a group of MIT researchers and Zara employees has successfully designed
and implemented a system that complements human intuition with automatic software.vii
Orders manually generated by store managers on the basis of the intuition generated by
direct contact with customers are combined by the software with aggregated historical
information to generate sales forecasts. These are used by other software that takes into
account as well the warehouse stock (recall that Zara is known for short production runs
and for changing its inventory frequently without necessarily reproducing what the stores
request) to determine what to send to each store. Managers at Zara headquarters can
modify either the software parameters or the orders.
At Zara, final decision-making responsibility remains with centralized HI, but to store
managers’ support has been added support from centralized IT. Although this transition
seems logical for a product category with a high degree of uncertainty like the fashion
market, the company’s previous model had proved highly effective. Changing it risks
long-term effects that might be difficult to measure. If, for example, limiting the role of
human input were to diminish learning by mangers, performance might be degraded
over the long run.
Case C1. Final decision-making responsibility resides with local HI with centralized
IT support: A Dutch supermarket chain
In some chains, store managers are responsible not only for sales and inventory, but
also for personnel and the significant expense it represents. In some store formats and
in certain sectors, much of employees’ time is dedicated to handling store merchandise,
and the remaining time to dealing with customers. But much of the software used to
automatically generate store orders does so in such a way that merchandise arrives
when demand is highest. Although this is when the inventory is most needed, it is also
when it is least desirable to divert store employees’ attention from customers to receiving
orders. To solve this problem, some managers directly intervene in the system
proposals. This was the case in a Dutch retail chain studied.viii Store managers
systematically modified orders suggested by the system to ensure that they would arrive
9 of 18
at the store on low sales days. Recognizing this, the company modified the system to
better balance the workload at the store level before moving to an automated system.
Case C2. Final decision-making responsibility resides with store managers with
centralized IT and HI in support: Our main research scenario
The attraction of the advantages of a centralized OS in retail described in the
Introduction notwithstanding, we were motivated by a desire to revise and update the
rocket science retailing article to study less obvious instances of the marriage of
technology and human intelligence. Our curiosity was piqued, in part, by the fact that
these cases, being less frequently encountered in practice, are hence less studied. Our
case study is of a decentralized final decision-maker with centralized IT and centralized
HI in support. Our findings of significantly improved performance provide some support
for the subject organization shifting final responsibility for decision-making from
centralized IT supported by decentralized store managers to decentralized store
managers supported by centralized IT and HI.
Our subject organization, Kingston (name disguised) is among the largest in the world in
its sector, with sales in excess of $10 billion and more than 400 stores and 40,000
employees. The company boasts an extensive catalog offering, and its marketing
approach is more customer-centered than those of it principal competitors. The supply
chain structure is complex involving centralized sourcing (Asia, Europe, and America)
and distribution centers and cross-docking facilities. Complexity, in terms of both volume
of products and customers and its management (diversity of suppliers, product
categories, commercial offers, customer services, logistics, circuit types, and so forth),
characterized even the single country operation we chose to study (more than $2 billion
in sales and in excess of 10,000 employees).
The company’s IT solutions were, in isolation, strong systems. Subsequent to their
implementation, the company saw an 80% decline in out of stocks and 40% reduction in
inventory levels, and the systems performed well in benchmarking studies conducted by
an internationally renowned, specialized consultancy. This track record is important, as a
study aimed at determining whether human intervention can improve decisions made by
a state-of-the-art, well implemented, automated IT store-ordering solution would be
meaningless if that solution performed poorly.
Specifically, we were interested in studying decisions made under the company’s "the
system suggests, but the manager decides" philosophy, and assessing whether this
philosophy could be superior to the "what the system asks for is sacred” philosophy. We
analyzed the decisions taken by the system and subsequent interventions by more than
60 managers who modified those decisions over a period of several months. These
managers review each morning, on average, approximately 600 SKUs and receive
approximately 100 order proposals from the automated IT system. The more than
300,000 real and valid decisions we compiled from the 10-15 weeks of data we
examined represented, on average, more than 5,000 decisions per manager.
10 of 18
Kingston’s automatic ordering system, which reviews all planning and supply needs at
most once per week, is composed of a demand planning model-based set of software
modules that trigger order proposals and a set of parameters used in the calculations.
The system includes:
an information capture module that gathers the information needed, including
logistics and price data and data on supplier services and company operations
(e.g., the logistics network, circuits, delivery schedules, inventory levels, and so
forth), to support calculations by the company’s ERP;
an analytics engine based on multiple statistical algorithm models selected on
the basis of characteristics of the historical series to be analyzed that takes into
account in making item and point of sale calculations real demand and individual
conditions at each point (the module accounts for the in-store logistics workload,
as described earlier in the Dutch supermarket chain’s modified IT solution);
a local adaptation parameterization module that considers individual
characteristics of each item and supplier at each point of sale (e.g., the physical
space reserved for each item at each point of sale, delivery days for item or
supplier, and delivery conditions).
Kingston permits category managers to intervene and modify the system in two ways:
a. by adjusting local system data and parameters, primarily inventory level,
merchandising space requirements, and lead times;
b. by modifying, as by reducing or increasing amounts, or even voiding, the system's
order proposals, or generating orders that the system would not have proposed.
Given this background, and keeping in mind that specialized benchmarks had verified
the system’s superior performance, we observed the rate of proposal modification to be,
on average, 32% overall, which broke down as follows: 53% order cancelations, 28%
newly initiated orders, 9.5% expansions, and 9.5% reductions. The overall percentage of
proposal modification seems high, but it was whether the modifications were appropriate
and their impact on the business that were important to analyze.
Do humans hurt or help with management of inventory replenishment?
We evaluated our results by calculating the financial results of all decisions including:
decisions proposed by the system that were not modified,
decisions that were modified by the managers (e.g., voids, reductions and
increases, and manually generated orders).
The unit of analysis for each decision was performance over the time period during
which a decision was relevant. We estimated the net impact of these decisions by
simulating the results had the changes had not been made.
11 of 18
We quantified performance by assigning to each decision a monetary value estimated
via an economic function that takes into account out-of-stock situations and the cost
associated with excess inventory. The economic value of out-of-stock situations was
obtained from the lost profit margin by estimating sales not made, taking into account
the percentage of customers that substitutes another for a product not found, and the
percentage of customers who do not buy anything. The average value used to estimate
conversion of an out-of-stock situation into a sale is 50%, which is more exigent than the
59% reported in other large international studies.ix
Inventory maintenance cost was calculated only for excess inventory, that is, inventory
in the store during the decision period beyond the minimum demand, merchandising, or
logistics requirements. The cost of capital for inventory maintenance was calculated
using the purchase price of the excess inventory at an annual interest rate of 24%, the
company’s estimated inventory holding cost.
Sensitivity analyses conducted for a wide range of values of the previously indicated
cost parameters did not alter our study conclusions significantly.
To compare results across the product categories analyzed, which have diverse costs
and margins, we used as a universal indicator the total cost of each decision divided by
the item’s sales during the period relevant to the decision. Performance is thus
measured as a percentage of sales for each decision.
The results obtained were interesting indeed. Had managers had not modified the
proposed decisions, the average cost of items for which managers had modified
decisions would have been 3.02% of sales. But the actual cost of items for which
managers modified the system’s decisions was 1.11% of sales. This means that, thanks
to human intervention, the cost of affected items was reduced by 1.91% of sales. By
comparison, the cost of items for which the managers did not intervene was 1.07% of
sales (Figure 1). These results indicate that managers’ intervention in the automatic
system clearly added value.
Keeping in mind the mix of unmodified and modified decisions, the overall result was a
favorable impact on the business. The costs (as percent of sales) went from an
estimated 1.95%, if the proposals had been left unmodified, to 1.09% after the human
intervention (Figure 2). To our surprise, we found positive average savings by every
single manager (Figure 3). These results clearly corroborate our hypothesis that
combining human and machine intelligence can improve business performance overall.
Sources of observed improvement
Given that the cost for unmodified proposals was 1.07% of sales , and for the modified
proposals, had they not been modified, 3.02%, it is clear that managers generally
intervened in cases in which the system, for whatever reason, would have generated a
much higher than normal cost. That their interventions resulted in a cost of 1.11% of
sales (4% less efficient than that generated by the system’s unmodified proposals)
12 of 18
indicates that, on average, managers were almost as effective as the automatic system
once they identified the need to intervene. The good judgment evidenced by these
managers’ choices of when to intervene suggests that they recognize when the system
is performing well. Further analysis reveals some asymmetries in the managers’
interventions. That decisions to expand orders significantly increases both fill rate and
excess stock, whereas decisions to contract orders can significantly reduce excess stock
with little effect on the realized fill rate, suggests that managers are highly effective when
contracting, but less precise when expanding, orders.
Figure 3, however, also shows the different mechanisms through which Kingston’s
managers achieve these results. While some managers seem to go though heroic
efforts to obtain savings form a system that would otherwise have very high costs, others
seem to have a system that is creating very good recommendations and consequently
do not need much managerial intervention afterwards.
Interviews with managers (see sidebar) revealed that they considered themselves to be
responsible for maintaining the integrity of the data fed to, and adjusting some of the
decision parameters utilized by, the system. We also found evidence that it was for the
managers who adjusted parameter values and corrected data inaccuracies that the
system generated the lowest cost, leading us to hypothesize that most managers exhibit
either proactive (i.e., spend time manipulating system parameters and assuring data
quality), or reactive (i.e., make frequent adjustments to offset high costs occasioned by
poor calibration) behavior.
To test this hypothesis, we split the sample at the median along two dimensions: initial
system cost, that is, the cost had managers not been authorized to modify the system’s
recommendations, and intensity of modifications, measured as the fraction of decisions
modified by a manager. This classification yielded four manager types according to
preferred intervention strategy (see Figure 4). As expected, more than two thirds of
managers fell into the predicted categories, low (high) system cost and low (high)
number of modifications; one-third fell off the main diagonal.
Figure 4 reports the average final cost (as a percentage of sales) achieved by the
managers in each category. Although, as mentioned earlier, manager intervention
improves system performance in all cases (compare the average system cost for each
column with the reported final cost for each category), a quick analysis yields some
interesting insights. First, final costs are clearly much lower for managers who started
with good system recommendations (low system cost). Second, although greater
modification activity seems to suggest lower cost, the difference is not statistically
different when we control for initial system cost. Not surprisingly, the savings differential
due to order modification is lower among those that parameterize the system well, thus
clearly suggesting that managers’ top priority should be to adjusting the system
13 of 18
The difference in leverage between the two possible strategies suggests a clear path for
improvement that is further supported by the fact that data quality and system
parameterization are more easily systematized that the more tacit judgment of when to
modify a system recommendation. Moreover, time and effort spent improving operational
discipline vis-à-vis running the system (e.g., making parameter adjustments and
ensuring data quality) will improve the quality of the decisions generated, thereby
reducing the need for adjustment and increasing the time available to further improve
operational discipline. This reinforcing mechanism (see Figure 5) is equivalent to the
productivity chain” argument made by Deming (1986), which states that resources freed
up by productivity gains should be reinvested in the search for still greater
Managers’ realization that they are responsible for data integrity, or, alternatively, that
they are best supported by the system when it has accurate information, provides an
elegant solution to one of the most vexing issues in retail operations: inventory record
inaccuracy (IRI). That IRI is endemic and has a detrimental effect on the return obtained
from IT investments is well documented.xii Moreover, initiatives aimed at improving
inventory control through inspection and correction yield only temporary improvements
in data quality.xiii. Managers authorized to modify system generated restocking decisions
have an incentive to assess the quality of those recommendations, as opposed to being
unaware of what those restocking decisions even are, and being able to fully assess the
impact of improved data quality and parameter adjustments often willingly allocate time
to further improving operational discipline (see the Data Improvement loop depicted in
Figure 5). The average inventory record inaccuracy (i.e., fraction of SKUs for which the
record inventory does not correspond to the physical inventory) of approximately 35%
found for stores with low system cost compares with the average IRI found in previous
studies of around 65%.
14 of 18
As pressure to increase the speed and
effectiveness of decision-making builds under the
growing volume of decisions to be made in
operational environments, the use of technology
becomes essential. Thus, a global semi-automated
system that improves the decision making process
is essential. However, no matter how good
software, its implementation or use are, no IT
solution can be perfect or complete. In fact, all such
solutions have “blind spots”. Therefore, appropriate
intervention from a layer of human intelligence can
improve results. The Kingston case related above
presents conclusive evidence that it is possible to
consistently improve the performance of a state-of-
the-art automated store replenishment system by
empowering managers to modify, void, and even
substitute their own for its recommendations. It is
possible, for example, by modifying
recommendations at the store level, to correct with
context-rich local information systematic blind spots
in a system. The case of the Dutch retailer
described above represents a vivid example of this
type of learning. Once a blind spot is recognized by
local managers, in this case the congestion that
occurred in the warehouse during the busy times, it
is possible to modify the algorithms to address the
blind spot, thus freeing up more managerial time to
address other blind spots or non-programmable
As well as improving the overall quality of decisions,
this simple adjustment of decision rights has two
additional benefits. One, vesting replenishment
accountability in store managers creates a set of
incentives that assures better information quality in
an environment in which bad data is a persistent
problem. Two, continual review of the system’s, and
managers’, decisions can inform adjustments to the
system and its decision rules. This not only fosters
continuous learning and improvement, but also
suggests refinements to decision rules whereby
day-to-day programmable decisions are taken by
What the managers think about the replenishment
global system
In addition to the numerical data, interviews were
conducted with the category managers to capture
qualitative aspects and the people point of view.
Excerpts from those interviews follow.
What is your assessment of the replenishment system?
“The system is comprehensive, it considers a lot of
“You have control. You can review and decide
whether to modify proposals or not. You control
the system.”
“The restocking decisions under normal
circumstances are very good.”
[B]ecause the system is so comprehensive, and
considers so much information, you have to spend
time getting to understand it. You need training.”
“The system is good, perhaps a bit complicated. I
do not know all the technical details.”
“[The system] is a bit complicated because you
have to attend to many things at the same time
and we don’t always have things under control.”
How do you ensure the quality of the replenishment
systems recommendations?
“You have to work on [inventory record accuracy]
constantly. The [merchandising minimum
requirements] are reviewed every other month.”
“We have a daily schedule to review inventory
record accuracy, although some times it is not
possible to execute … Pending orders are
reviewed every week. The [merchandising
minimum requirements] every five to six months.
Lead times once a year.”
“When I first joined [the company] I was not used
to working with a system like this. I relied heavily
on modifying orders and did not work the system
parameters. Now we have changed our work
strategy and do things the other way around. We
dont modify orders as much and work heavily on
the input data.”
[T]hree things are necessary: data quality is
fundamental, system parameters are very
important, and order modifications do add value.”
Why do you modify?
“You have to keep a close eye on the system
when it comes to new product lines or seasonal
“The only problem is with seasonal products where
the end of season might vary because of, say,
weather conditions, and you have [to adjust]
orders. The system has no way of knowing this, so
it is a good thing that we can modify orders.”
[A]lthough the [system] works well, now and then
you have to modify the proposals, either because
a promotion is doing well, or [because of] a
competitors response, or there is a problem with
the seasonality.”
“The system needs time to react, but we can
respond more quickly.”
“The program and the system are good, but when
we modify orders we improve things.”
“We try to minimize the number of orders that we
modify. Parameters and data quality we try to
adjust continuously.”
15 of 18
the system and human intervention is invoked for complex or unique situations.
That our findings run contrary to common knowledge and prevailing opinions of
practitioners and academics alike we believe is because the implementation (and
measurement of the performance) of such systems as are described here is particularly
difficult, and well and reliably functioning systems are rare. But so great is the potential
for performance improvement when these systems are supported by discretionary
human intelligence that organizations that adopt this model stand to realize a significant
advantage over their competitors.
Figure 1. Performance according to the type of intervention in the automated
Modified pre modification Modified Not-modified
1.11% 1.07%
(as % of sales)
Performance by Decision Group
16 of 18
w/o Modifications w Modifications
(as % of sales)
Overall Performance
Figure 2. Total performance before and after modifications to the system.
Figure 3. Savings by manager.
17 of 18
Percentage of
Hyperactive (16%)
Average Final Cost
Reactive (35%)
Average Final Cost
Proactive (33%)
Average Final Cost
Passive (16%)
Average Final Cost
Average cost of the system (no modification)
Figure 4. Types of managers according to their actions.
The number in parentheses is the percentage of this type of manager relative to the total number of
managers analyzed. Average savings for modified decisions refers to the average of the costs out of sales
obtained by each type of manager. Final costs are significantly different across columns, but not across rows.
When comparing individual cells, only passive managers’ performance is statistically different from the
performance of the other three types of managers.
Figure 5: A systemic view of the case.
Arrows indicate the direction of causality. Signs (“+” and “–“) indicate the polarity of relationships. A”+” means
that, all else equal, if the cause increases (decreases), the effect increases above (decreases below) what
would otherwise have been. Similarly, a “–“ sign indicates that the relationship is inverse, i.e., if the cause
increases (decreases) the effect decreases below (increases above) what it would otherwise have been. The
R in the loop identifier indicates a self-Reinforcing (positive) feedback process.
18 of 18
i Fisher, M.L., A. Raman, A.S. McCelland, 2000. “Rocket Science Retailing Is Almost Here, Are
You Ready?” Harvard Business Review, (07/08), 115-124.
ii Bowman, E.H., 1963. “Consistency and Optimality in Managerial Decision Making.”
Management Science, 9 (2), 310-321.
iii Sterman, J.D., 1989. “Modeling Managerial Behavior. Misperception of Feedback in a
Dynamic Decision Making Environment.” Management Science, 35 (3), 321-339.
iv Managing Functional Biases in Organizational Forecasts: A Case Study of Consensus
Forecasting in Supply Chain Planning.”
v Raman, A., DeHoratius, N., Ton, Z., 2001. “Execution: The Missing Link in Retail Operations."
California Management Review 43, 3.
vi Ton, Z., C. Ross, 2008. Home Depot. Harvard Business School.
vii Caro, F., J. Gallien, 2008. “Inventory Management of a Fast-Fashion Retail Network.” MIT
Sloan Research Paper No 4656-07.
viii Van Donselaar, K.H. et al., 2009. “Ordering Behavior in Retail Stores and Implications for
Automated Replenishment.” Cornell University, Johnson School, Research Paper #19/09.
ix Corsten, D., T. Gruen, 2003. “Desperately Seeking Shelf Availability: An Examination of the
Extent, the Causes, and the Efforts to Address Retail Out-of-Stocks.” International Journal of
Retail & Retail Management, 31(11/12), 605-617.
x Deming, W.E., 1986. Out of the Crisis. Massachusetts Institute of Technology, Center for
Advanced Engineering Study, Cambridge, Mass.
xi Keating, E. et al., 1999. “Overcoming the Improvement Paradox.” European Management
Science, 17 (2) 120-134.
xii DeHoratius, N., A. Raman, 2008. "Inventory Record Inaccuracy: An Empirical Analysis."
Management Science 54, (4 ).
xiii Fleisch, E., Tellkamp, C. ,2005. “Inventory Inaccuracy and Supply Chain Performance: A
Simulation Study of a Retail Supply Chain.” International Journal of Production Economics, 95
(3) 373-385.
Non official bibliography:
This Web site lists a number of cognitive biases and corresponding academic sources:
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
Inventory inaccuracy is a main issue in businesses dealing with physical assets. The aim of this paper is to examine the relationship between inventory inaccuracy and performance in a retail supply chain. We simulate a three echelon supply chain with one product in which end-customer demand is exchanged between the echelons. In the base model, without alignment of physical inventory and information system inventory, inventory information becomes inaccurate due to low process quality, theft, and items becoming unsaleable. In a modified model, these factors that cause inventory inaccuracy are still present, but physical inventory and information system inventory are aligned at the end of each period. The results indicate that an elimination of inventory inaccuracy can reduce supply chain costs as well as the out-of-stock level. Automatic identification technology that is becoming available offers the potential to achieve inventory accuracy.
Despite all the data that retailers and e-tailers can now gather about point-of-purchase information, buying patterns, and customers' tastes, they still haven't figured out how to offer the right product, in the right place, at the right time, for the right price. Most retailers largely ignore the billions of bytes of customer data stored in their databases-or they handle that information incorrectly. As a result, they don"t adequately supply what consumers demand. But some retailers are moving profitably toward what the authors call "rocket science retailing"- a blend of traditional forecasting systems, which are largely based on the gut feel of employees, with the prowess of information technology. Marshall Fisher, Ananth Raman, and Anna Sheen McClelland recently finished surveying 32 retail companies in which they tracked their practices and progress in four areas critical to rocket science retailing: demand forecasting, supply-chain speed, inventory planning, and data gathering and organization. In this article, the authors look at some companies that have excelled in those four areas and offer some valuable advice for other businesses seeking retailing perfection. In particular, the authors emphasize the need to monitor crucial metrics such as forecast accuracy, early sales data, and stockouts-information that will help retailers determine when to tweak their supply-chain processes to get the right products to stores at just the right time. The authors discuss the information technologies now available for tracking that information. They point out the flaws in some reporting and planning systems and suggest alternate methods for measuring stockouts, inventory, and losses.
This paper reports some research, ideas, and theory about managerial decision making. The first research projects dealt with are aggregate production and employment scheduling. From this is developed the idea that management's own (past) decisions can be incorporated into a system of improving their present decisions. Decision rules are developed, with the coefficients in the rules derived from management's past decisions (rather than from a cost or value model). Half a dozen test cases are used to illustrate and test these ideas (theory). Some rationale about decision making in organizations and criteria surfaces is supplied to help interpret the major ideas presented.
Examining nearly 370,000 inventory records from 37 stores of a large retail chain, we find 65% to be inaccurate (i.e., recorded inventory does not equal on-hand inventory), at the time of the physical inventory audit. Using two different measures of inventory record inaccuracy, we demonstrate, empirically, that the variation in these measures is associated with specific inventory management practices and store characteristics. Our analysis highlights where managerial attention is most needed to resolve this critical problem and suggests opportunities for further research.
Working in collaboration with Spain-based retailer Zara, we address the problem of dis- tributing over time a limited amount of inventory across all the stores in a fast-fashion retail network. Challenges speci¯c to that environment include very short product life-cycles, and store policies whereby a reference is removed from display whenever one of its key sizes stocks out. We ¯rst formulate and analyze a stochastic model predicting the sales of a reference in a single store during a replenishment period as a function of demand forecasts, the inventory of each size initially available and the store inventory management policy just stated. Secondly, we formulate a mixed-integer program embedding a piece-wise linear approximation of the ¯rst model applied to every store in the network and allowing to compute store shipment quantities maximizing overall predicted sales, subject to inventory availability and other constraints. We report the implementation of this optimization model by Zara to support its inventory distribu- tion process, and the ensuing controlled ¯eld experiment performed to assess the impact of that model relative to the prior procedure used to determine weekly shipment quantities. The results of that experiment suggest that the new allocation process tested increases sales, reduces tran- shipments, and increases the proportion of time that an important category of Zara's products spends on display.
Desperately Seeking Shelf Availability: An Examination of the Extent, the Causes, and the Efforts to Address Retail Out-of-Stocks
  • K H Van Donselaar
viii Van Donselaar, K.H. et al., 2009. "Ordering Behavior in Retail Stores and Implications for Automated Replenishment." Cornell University, Johnson School, Research Paper #19/09. ix Corsten, D., T. Gruen, 2003. "Desperately Seeking Shelf Availability: An Examination of the Extent, the Causes, and the Efforts to Address Retail Out-of-Stocks." International Journal of Retail & Retail Management, 31(11/12), 605-617.