Managing Learning

Curves in Factories
by Creating and
Transferring Knowledge

Michael A. Lapré
Luk N. Van Wassenhove

T
he learning curve phenomenon is widely known. As organizations
gain experience with production, productivity and quality improve
at a decreasing rate. Early empirical studies showed that the loga-
rithm of unit cost decreased with the logarithm of cumulative num-
ber of units produced at a uniform rate—the learning rate.1 The disappointing
managerial implication of most learning curve studies is that the only way to
accelerate cost reduction is to accelerate cumulative production volume, which
may not always be desirable or feasible. Yet, experts such as CEO Ray Stata of
Analog Devices have argued that “the rate at which individuals and organiza-
tions learn may become the only sustainable competitive advantage.”2 The
importance of managing learning rates becomes all the more apparent when
we consider variation in learning rates. Learning rates exhibit considerable vari-
ation across industries, within industries, within firms, and even across workers
within a factory. Some factories reduce unit cost by 50% after doubling cumula-
tive production volume, others do not learn at all.3 Consequently, there is an
opportunity to manage learning rates.
Managing learning rates is by no means an easy task. Barriers to learning
in factories include the following:
▪ Separation of cause-and-effect in time and space (dynamic complexity).
▪ Presence of too many variables, which makes it difficult to comprehend
a system in its entirety (detail complexity).

We gratefully acknowledge thoughtful comments provided by two anonymous reviewers. We thank

the management and employees of N.V. Bekaert S.A. for their unstinting cooperation. The work by
Michael Lapré was supported by a summer research grant from the Dean’s Fund for Faculty
Research at the Owen Graduate School of Management, Vanderbilt University.

CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003 53

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

FIGURE 1. Simplified, Schematic Process Flow in a Bekaert Factory

Dry Heat Wet Wire Bunching

Drawing Treatment Drawing
Brass
Coating
Process

Factory
Floor

Note: Rectangles and squares represent machines. Actual numbers of machines per step are disguised.Whereas upstream departments consist
of a handful to a few dozen machines, downstream departments have hundreds. In reality several steps are repeated.

▪ Simultaneous existence of equally plausible but mutually contradicting

explanations of a situation (ambiguity).
▪ Lack of a full understanding of the effects of input variables of a process
on the output (incomplete technological knowledge).
As a result, individuals in an organization can create their own potentially inac-
curate beliefs, or “myths,” based on subjective interpretations of events. Over-
turning false myths is a difficult challenge in managing factory learning curves.
Misperception of feedback along with poor inquiry and scientific skills can make
the challenge even more difficult.4
As part of a multiyear research project, we studied learning under such
conditions in factories at N.V. Bekaert, S.A., a Belgian multinational corpora-
tion.5 Bekaert is the world’s largest independent producer of steel wire. In partic-
ular, Bekaert’s steel cord division produces about one third of the world’s output
of steel wire (called “tire cord”) used in the production of steel belted radial tires.
Bekaert’s process flow is depicted in Figure 1. Thick wires are drawn through
dies that progressively reduce their diameters. At intermediate points, wires are
heat treated, and a chemical process coats the wires with brass. Coated wires are
drawn even thinner through dies submerged in soap solution. Very thin wires
(called “filaments”) are bunched around each other to form tire cord. The sim-
plest cord has two filaments; the most complex, hundreds.
In the early 1980s, Bekaert’s production personnel faced many barriers
to learning about its production process:

54 CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

▪ Detail complexity arose from the need to coordinate hundreds of

machines, hundreds of process variables to control, different customer
specifications for wires that sometimes had to be processed together,
and the different technical skills required at the different stages.
▪ Dynamic complexity arose from the ease with which effects of problems
experienced at any machine or production stage could be transmitted to
other machines and stages.
▪ Ambiguity arose as production personnel often disagreed on causes for a
particular production problem such as fractures. One person could blame
raw materials, another person could blame intermediate process settings,
and there would be no way to tell who was right.
▪ Incomplete technological knowledge arose as nobody could specify the
combination of the hundreds of process settings that would prevent the
occurrence of defects.
We studied Bekaert’s learning efforts in face of these barriers at three
levels of analysis. First, we studied learning processes in quality improvement
projects.6 Second, we studied the impact of such learning processes on a factory’s
learning curve.7 Third, we studied production lines created specifically to learn
about production processes in three factories.8

Learning Processes are Multidimensional

The first step in opening up the black box of the learning curve is to
examine processes through which organizational members learn. In its simplest
form, organizational members experience events and observe what is happening.
They assess their observations and design a conceptual response to the assess-
ment. Finally, they test designs via implementation, and the learning cycle of
Observe-Assess-Design-Implement begins again.9 Bekaert calls learning cycles
“Symptom-Cause-Remedy-Action,” which readily maps onto Observe-Assess-
Design-Implement.
In the early 1980s, technological knowledge about tire cord production
was very underdeveloped. As a result, production crises in the form of more
than 100 times the normal amount of
process interruptions could happen at any Michael A. Lapré is an Assistant Professor of
time at any point of the production process. Operations Management at the Owen Graduate
School of Management at Vanderbilt University.
Bekaert responded by embarking on a Total <[email protected]>
Quality Management program at its flag-
Luk N. Van Wassenhove is the Henry Ford Chaired
ship Aalter plant. The firm was among the Professor of Manufacturing at INSEAD. <luk.van-
first to do so in Europe. Bekaert’s efforts [email protected]>
seem to have paid off as the firm won two
European Foundation for Quality Management Awards in the early 1990s.
Aalter trained its factory personnel in structured problem solving. The Symp-
tom-Cause-Remedy-Action (SCRA) approach became a standard tool. Many
teams tackled specific quality problems with the SCRA approach in quality

CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003 55

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

improvement projects. The teams documented their progress in final project

reports, which, under a management edict, contained detailed information in
a standardized format.
We had unlimited access to all of Aalter’s records on quality improvement
projects undertaken between 1982 and 1991. We selected all projects that dealt
with: product attributes or process control; had progressed to at least past the
testing stage; and had been adequately documented. This selection left us with
62 projects. We found that learning processes in these projects exhibited a lot of
variation along two dimensions: conceptual learning and operational learning.
Conceptual learning is the process of acquiring a better understanding of cause-
and-effect relationships. Faced with a particular quality problem, say wire frac-
tures, a project team could use scientific models and statistical experiments to
assess underlying causes for the problem at hand. The team could then use its
assessment to design a concept or theory to explain the occurrence of wire frac-
tures. The more a team uses science and statistical experimentation and the
more detail of the theory developed is incorporated in the final design; the
higher is the team’s level of conceptual learning. Operational learning, on the
other hand, is the process of obtaining validation of action-outcome links. Oper-
ational learning is high if a team modifies action variables and observes positive
follow-up of experimental results. In short, a high level of conceptual learning
yields know-why, a high level of operational learning yields know-how.10
One project team aimed to reduce high levels of fractures in two spiral
wrapped cords. In the final production step, a thin filament was wrapped around
a cord. The high levels of fractures occurred near the core of a spool supplying
the filament. The first response was to not use the final 500 meters of filament
on the spool. As this could not be a tenable solution, the team introduced spools
with larger core diameters. The number of fractures dropped by an order of
magnitude. Operational learning in this project was very high. The introduction
of spools with larger cores was a clear modification of an action variable, which
led to clear positive follow-up of results, namely, a sharp drop in fractures.
Conceptual learning in this project was low. Years later, Bekaert found
out what happened. As the spool was emptying out, the pay-off tension was
increasing, which in turn increased the stress on the filament. As a result, the
probability of filament fractures rose sharply. The conceptual learning that did
not occur in the project took place later when Bekaert personnel created and
tested a physical model that quantitatively specified how breaking load changes
as a function of variables such as filament diameter, machine speed and pay-off
tension.
Therefore, learning processes are multidimensional. Projects can exhibit
significant variation along conceptual learning and operational learning.

Global Impact of Conceptual and Operational Learning

So far, we have identified learning processes in local projects. Typically,
local projects do not even affect one percent of total plant waste. What is the

56 CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

impact of locally acquired knowledge on a factory’s global learning curve? Put

differently, did any of the knowledge created through local quality improvement
projects allow the Aalter plant to speed up global waste reduction?
Bekaert managers observed that many local projects had created knowl-
edge applicable to the entire plant. Dissemination of local knowledge was a far
cry from routine practice. Typically, project results did not even spread in the
vicinity of the local project context. Often personnel would argue that local
knowledge obtained elsewhere would not be transferable because of slight dif-
ferences in machine types, product specifications, raw materials, and so on.
To investigate the impact of conceptual and operational learning on the
factory’s learning curve, we classified our quality improvement projects into four
categories. We compared scores for conceptual and operational learning with the
median scores in the sample of all projects and classified each project as high or
low conceptual learning and high or low operational learning. For each category,
we defined a knowledge stock variable—the cumulative number of projects
completed to date. Finally, we investigated the impact of the four knowledge
stocks on the global learning curve. Specifically, how did the knowledge stocks
affect the global rate of learning? Figure 2 shows the results.11
▪ Firefighting—In projects with low levels of conceptual and operational
learning, teams hardly reflect on cause-and-effect relationships, imple-
ment (at most) minor changes, and obtain little validation. We call such
projects “firefighting.” Firefighting did not affect the global rate of learn-
ing. In the absence of root-cause analysis and operational results, factory
personnel ignored these efforts.
▪ Artisan Skills—Projects with low conceptual learning but high operational
learning create know-how. Operational learning yields solutions that
work. However, because conceptual learning is low, the reasons why
these solutions work are ill-understood. These projects are closer to an
art than a science. Hence, we label them “artisan skills.” Surprisingly,
artisan skills did not affect the global rate of learning either. Lack of con-
ceptual learning leaves a lot of room for ambiguity. Personnel in other
areas of the factory would argue that locally obtained artisan skills would
not be applicable to their part of the factory because of slight differences
in product specifications, machine settings, raw materials, and so on. Lack
of conceptual learning makes artisan skills difficult to transfer.
▪ Non-Validated Theories—In projects with high conceptual learning and
low operational learning, teams use science and statistical experiments
to assess cause-and-effect relationships and come up with a high level of
detail in the designed solution. Yet, they fail to include a verification step
to ensure that operational results are indeed obtained when implement-
ing the new design. Consequently, we label these projects “non-validated
theories.” Quite surprisingly, non-validated theories slowed down the
global rate of learning. Why did local non-validated theories cause prob-
lems globally? Many of the non-validated theories used insights from
R&D and engineering staff. The R&D and engineering staff were typically

CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003 57

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

FIGURE 2. The Impact of Four Knowledge Stocks on the Learning Rate

Conceptual Learning Non-validated Theories: Operationally Validated Theories:

Disrupt Learning Curve Accelerate Learning Curve
High 3.2% – 3.6%

Firefighting: No Impact Artisan Skills: No Impact

Low

Low High
Operational Learning

Note: Percentages denote the impact of one additional quality improvement project on the initial learning rate. For example, one
additional operationally validated theory would speed up the rate of waste reduction by 3.6%. On the other hand, one additional
non-validated theory would slow down the rate of waste reduction by 3.2%.

perceived as experts; consequently their ideas were readily imple-

mented even though the ideas were not validated in the factory. So,
non-validated theories spread. Why were they disruptive? Many of these
ideas were developed at central R&D. Learning experiences in a small
scale R&D laboratory, however, are very different from learning experi-
ences in a full-blown manufacturing environment (low fidelity). Even if
production methods developed at R&D are replicated in a plant, specific
elements in the full-scale manufacturing environment (such as equip-
ment configurations) can cause deterioration in performance. Basically,
fine-tuning of these ideas is necessary before they can be spread on a
wide scale.12
▪ Operationally Validated Theories—In projects with high levels of both con-
ceptual and operational learning, teams use scientific models and statisti-
cal experiments to develop theories that explain the occurrence of process
interruptions. Using these theories, teams implement changes and obtain
validation. Hence, we label these efforts “operationally validated theo-
ries.” The only knowledge stock that enhanced the global rate of waste
reduction was the stock of operationally validated theories. In these pro-
jects, teams not only obtained solutions that worked, the solutions were
derived using scientific principles. This provides a very powerful combina-
tion to overturn subjective “myths.” Operationally validated theories iso-
late the principles from idiosyncratic local conditions. Notwithstanding
differences in machine types, product specifications, or raw materials,
operationally validated theories help to identify the principles that are
transferable. So, for local projects to accelerate a factory’s global learning

58 CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

curve, high levels of both conceptual learning and operational learning

are required.

Designing a Production Line to Accelerate the Learning Curve

Only 25% of Aalter’s quality improvement projects were operationally
validated theories that accelerated the learning curve. The other projects either
slowed down or did not affect the learning curve. Is the implication that there
is only a 25% chance for a quality improvement project to enhance the bottom
line? Fortunately, the answer is no.
In the late 1980s, Bekaert realized that central R&D laboratories lacked
characteristics of the dynamic production environment encountered in factories.
Bekaert therefore relocated process optimization to the factory. In 1988, the
company established a “Model Line” at Aalter (MLA) for an important represen-
tative product and gave a senior R&D manager the responsibility to create fun-
damental process control knowledge without sacrificing the production of
saleable wire.
The MLA was a fundamentally new organizational structure at Bekaert.
Typically, Bekaert tire cord factories organized personnel and machines by pro-
duction departments corresponding to the stages in the production process. The
MLA, on the other hand, was an integrated line corresponding to a single prod-
uct cutting across all production departments. People and machines were dedicated to
a single product and managed by a team led by a senior R&D manager. The MLA
team routinely defined and conducted projects for productivity and quality
improvement. The team set improvement goals based on both R&D and pro-
duction experience. The team collected data on any product/process variable
deemed relevant. Its unique data capture allowed the team to link data across
process stages. The MLA team typically solved problems by building scientific
models from which it derived testable hypotheses. Hypotheses were tested with
natural and controlled experiments. The resulting regression output was used to
modify limits for Statistical Process Control (SPC). The MLA consistently pro-
duced operationally validated theories.13

Model Line A: Higher Stages of Technological Knowledge

Table 1 shows how we measured technological knowledge. The concept
of “ability to specify impact of variability of known causal variables” captures a
project team’s ability to predict how quality/productivity would vary if changes
were made in the levels of a given set of process parameters. The measure
ranges from level 1 (clueless) to level 5 (a scientific model). We found that both
conceptual learning and interdepartmental problem solving allowed project
teams to “climb” the stages of technological knowledge.14
Typical quality improvement projects reached levels 2 or 3 at best. In
other words, project teams would be able to predict a positive or negative effect
on quality (level 2) or predict the magnitude of the effect of a single change
within a small range (level 3). The MLA team, on the other hand, conducted

CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003 59

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

TABLE 1. Stages of Technological Knowledge

Level Ability to Specify Impact of Variability of Known Causal Variables

1 None
2 Specify Direction
3 Specify Impact of Local Changes
4 Empirical Equation
5 Scientific Model

Note: Adapted from R.E. Bohn,“Measuring and Managing Technological Knowledge,” Sloan Management Review, 36/1 (Fall 1994):
61-73.

projects that routinely reached levels 4 and 5. Regression output from natural
and controlled experiments yielded empirical equations that quantified the
impact of several variables at the same time for broad ranges (level 4). Several
projects yielded complete scientific models (level 5).
In 1993, Bekaert transferred production of the MLA product to a new
U.S. plant. The process control knowledge derived from level 4 and level 5 tech-
nological knowledge allowed the new U.S. plant to become the Bekaert plant
with the highest productivity and quality performance.

Replicating Model Lines for Knowledge Creation and Transfer

In two years, the MLA had achieved dramatic improvements in quality
and productivity. Moreover, the MLA team had created fundamental process
control knowledge. In 1991, impressed by the MLA track record, Bekaert intro-
duced model lines for similar, important, representative products at two other
major European tire cord plants (B and C). We refer to these model lines as MLB
and MLC1. The first manager of the MLA was appointed corporate model line
manager responsible for coordinating these model line efforts at central R&D. In
1993, Bekaert added a second model line at plant C (MLC2). A model line team
typically codified knowledge from projects in monthly activity reports, which
were shared with central R&D and other model line teams. At bi-annual meet-
ings, model line managers also shared knowledge at central R&D. Additionally,
they shared knowledge directly between them as well as with central R&D
through informal channels such as phone, fax, and visits.
By 1995, the new model lines had not delivered results comparable to the
MLA record. Particularly, creation of technological knowledge on the new model
lines had not progressed to the same extent. Looking back in 1996, the corporate
model line manager commented:
“Initially, we thought we could solve everything in a [model] line with natural
data, just doing the statistical experiments. But this is not true. [Experiments]
need to be guided by scientific models. This, however, is very difficult. How do
you create this? . . . Two things are key: responsibility, which so far has been

60 CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

TABLE 2. Data Collected to Compute Total Factor Productivity

Physical Quantities Fixed Prices/Wages

Output Good Tonnage Produced Sales Price

Input 1 Direct Labor Hours Direct Labor Cost Per Hour
Input 2 Indirect Labor (engineers, technicians) Monthly Indirect Labor Cost
Input 3 Machine hours Machine Cost Per Hour (depreciation + cost of capital)
Input 4 Raw Material Wire Rod Cost Per Ton

largely driven by existing culture in a plant, and the creativity or research skills of
[model line personnel].”

It is worthy to note that Bekaert is not an exception in terms of struggling

with replication of a successful business model (the first model line at Aalter).
Time and again, scholars have observed that failure to replicate best practice
permeates modern industry. From kitchens in a single commercial food firm to
regional operating units within a telephone company, large productivity differ-
ences continue to exist. Local knowledge within firms remains untapped, lead-
ing chairmen of firms such as Texas Instruments and Hewlett-Packard to lament:
“If only we knew what we know.”15 The ubiquity of this problem makes it inter-
esting to analyze the difficulties Bekaert experienced in replicating the success of
the first model line.
To investigate whether projects accelerated learning curves on the model
lines, we needed an overall measure of productivity. Model line teams routinely
monitored partial productivity and quality measures such as labor hours per ton
or rejects on a particular process step. Yet, management lacked an overall mea-
sure for productivity. Combining readily available data on direct and indirect
labor, capital, raw material usage, and yield, we computed a Total Factor Produc-
tivity (TFP) measure—the value of output divided by the sum of costs of inputs:
price93output t
TFPt  ,

j wagej,93input jt
where price93 is the 1993 sales price per ton, output t output produced in month t,
wagej,93 1993 unit cost for input j, and input jt quantity of input j used in month t.
The key idea of TFP is to fix prices and wages, so one can focus on productivity
evolution of a production unit while leaving out inflation effects. In other
words, a production unit can manage the conversion of inputs into outputs,
but has no control over inflation effects. Table 2 summarizes the monthly data
for physical quantities and 1993 prices/wages we collected to compute TFP for
each model line. Managers at Bekaert felt that TFP was a particularly good pro-
ductivity measure for model line performance as it provided a dynamic perspec-
tive, it was not distorted by inflation effects, and it aggregated trade-offs between
the partial measures management typically focused on.16 Figure 3 shows the
TFP evolutions for the four model lines. Table 3 provides additional background
information on the four model lines.

CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003 61

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

FIGURE 3. Logarithm of Total Factor Productivity for the Four Model Lines

MLC2

MLB

MLC1

MLA

Sep-89 Sep-90 Sep-91 Sep-92 Sep-93 Sep-94

TABLE 3. Start-up Dates, Organizational Context, and Productivity Improvements

on Model Lines

Model Line A B C1 C2

Start Production of Product 1988 1990 1988 1972

Start Model Line Concept 1988 1991 1991 1993
Team: # Engineers 2 1 1
# Technicians 2 1 2
# Workers 4 2 2
Prior Experience on the Team R&D, plant plant plant
Year ML Manager Joined Bekaert 1972 1989 1990
ML Manager Reported to CEO plant manager plant manager
Interdepartmental Authority to with the with with department managers,
Commit Departments to Action model line department who had some experience
for Experimentation team managers with interdepartmental teams
Productivity Improvement 100% 4% 25% 7%

Note: Productivity improvement is measured by percentage change in Total Factor Productivity (TFP) compared to the MLA. To protect
Bekaert’s proprietary information, we defined TFP improvement on the original Model Line A (MLA) as 100%. We expressed TFP
improvements on the other model lines as a percentage compared to the TFP improvement on the MLA. For example, 25% productivity
improvement on MLC1 means that TFP improvement on MLC1 was only 25% compared to the TFP improvement on the MLA.

62 CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

Total Factor Productivity Regression Analysis

Learning by doing is typically associated with the idea that repeated practice makes perfect.
We used cumulative production volume to assess learning by doing on a model line.To study
transfer of learning by doing from other model lines, we used cumulative production volume
summed across all other model lines.To assess formal learning through explicitly defined qual-
ity/productivity projects, we used cumulative completed projects.To study transfer of formal
learning from other model lines, we used cumulative projects summed across all other model
lines.We included three control variables in our analysis.Two of our control variables have
been used before in TFP studies in process factories like Bekaert’s: capacity utilization and
reject rate. Capacity utilization captures any non-learning related productivity improvement
due to economies of scale. Reject rate captures any changes in quality policies that were not
learning related. At plant C, we found that low TFP values were associated with changes in raw
material suppliers. Every change in supplier led to readjustments in process settings and some-
times the introduction or deletion of an entire process step.To isolate changes in raw material
suppliers from any learning (in)efficiencies, we used the percentage of supplier change—the
share of raw materials for which a different supplier was used compared to the previous
month.Table 4 shows the TFP regression results.

We used regression analysis to study the impact of four learning variables

on TFP. For each model line, we assessed learning by doing (“practice makes
perfect”), formal learning (through projects), transfer of learning by doing, and
transfer of formal learning (see Sidebar “Total Factor Productivity Regression
Analysis” and Table 4). The MLA improved TFP via formal learning. The MLB
did not use any form of learning or transfer of learning to change TFP. Model
lines C1 and C2 improved TFP via learning by doing, and MLC1 also used trans-
fer of learning by doing to improve TFP. To understand why the MLA was the
only line to successfully use formal learning and no transfer of formal learning
occurred between the model lines, we need to look at stability of production
environments and organizational context.

Pitfall 1 in Replicating Model Lines:

Stable Environment
Table 4 showed a significant disruptive effect of changing suppliers for
MLC1. Figure 4 further illustrates the impact of changing suppliers on the TFP
evolution for MLC1: The biggest disruptions in TFP were all associated with
changes in raw material suppliers. While the MLC1 had one preferred, principal
supplier (“S”), shortages and/or quality problems at supplier S forced the MLC1
to sometimes use back-up suppliers. Two types of events were particularly dis-
ruptive: having to use at least 50% of raw materials provided by back-up sup-
pliers, or having to use at least 15% of raw materials provided by a different
back-up supplier compared to the previous month. Changes in raw materials
required readjustments in process settings, and sometimes the introduction or
deletion of an entire process step.

CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003 63

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

TABLE 4. Total Factor Productivity Learning-Curve Estimations

MLA MLB MLC1 MLC2

Learning by Doing n.s. n.s. 0.202* 1.887**

Log(cumulative volume)
Transfer of Learning by Doing n.s. n.s. 0.010* n.s.
Log(cumulative volume from others)
Formal Learning 0.093** n.s. n.s. n.s.
Log(cumulative projects)
Transfer of Formal Learning n.s. n.s. n.s. n.s.
Log(cumulative projects from others)
Economies of Scale 0.156** 0.043** 0.105** n.s.
Log(capacity utilization)
Quality Policies 1.619** 1.150** 1.206** 1.615**
Reject Rate
Continuity in Suppliers n.a. n.a. 0.059** n.s.
%Supplier Change

Notes: The dependent variable is Log(TFP). R2 of the model was 0.98.We allowed for autocorrelation.The autocorrelation coefficient was not
significant.Total number of observations: 149. * indicates statistically significant at 0.05. ** indicates statistically significant at 0.01. n.s. means not
statistically significant. n.a. means not applicable (for MLA and MLB there were no back-and-forth changes in raw material suppliers, so for MLA
and MLB no coefficient was estimated for %supplier change).

FIGURE 4. The Impact of Changing Suppliers on Total Factor Productivity: Model Line C1

Jan-92 Jan-93 Jan-94 Jan-95

Note: Open squares (▫) indicate months during which (i) more than 50% of raw materials was provided by back-up suppliers, and/or (ii) more
than 15% of raw materials was provided by a different back-up supplier compared to the previous month.

64 CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

Dynamic production environments are characterized by a lot of variabil-

ity. In conducting experiments to create technological knowledge about certain
variables, one should try to keep as many possibly confounding factors as con-
stant as possible. Continuity in resources (such as workers, equipment, and raw
material suppliers) can enhance a stable environment for experimentation (as
was the case for the original model line A). In the MLC1 case, a partnership with
supplier S might have prevented significant disruptions in productivity.17

Pitfall 2 in Replicating Model Lines:

Organizational Context
Figure 5 provides insight as to why projects on model lines B, C1, and
C2 were different from the MLA. Figure 5a shows “management buy-in”—the
extent to which plant management gave model line teams authority to define
their own projects to work on. Plants B and C were organized hierarchically by
production departments as opposed to the MLA which crossed all departmental
barriers. Figure 5b shows “interdepartmental problem solving”—the extent to
which projects spanned multiple production departments.

The Original Model Line A

The MLA had full authority to define its own projects (Figure 5a). The
MLA team typically pooled fragmented, scientific, engineering, and experience-
based knowledge to build a scientific model and to formulate testable hypoth-
eses. Hypotheses were tested with natural and controlled experiments. The
resulting regression output gave the team formulae to derive SPC rules to con-
trol process variables. A significant part of the MLA projects was truly inter-
departmental spanning at least three production departments (Figure 5b).
Interdepartmental problem solving is critical in a production environment char-
acterized by dynamic complexity—separation of cause-and-effect in time and
space.
Consider an MLA project set out to improve an important driver of pro-
ductivity in the wet wire drawing stage: drawability—the number of dies that
wear out during the production of one metric ton of wire.18 (Replacing worn
out dies is labor intensive and causes significant machine downtime.) The MLA
manager constructed a chemical model based on knowledge he had acquired
during his 16 years at R&D. At R&D, he had worked on many projects involving
all the production stages of making tire cord. For example, in the 1970s, he had
participated in an R&D project on the ability of tire cord to withstand corrosion.
From this R&D project, he remembered that some copper-related variables
determined in the brass-coating stage (a previous production stage, see Figure
1) affected drawability. Based on prior production experience, the model was
expanded to include additional variables, such as wire roughness and breaking
load—all determined in production stages preceding wet wire drawing. The MLA
team derived testable hypotheses from the model and tested these hypotheses
with controlled experiments on the factory floor (i.e., conceptual learning was

CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003 65

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

FIGURE 5A. Management Buy-in for Projects Conducted on the Model Lines

100%
Model-line projects
80% defined by model-line
team
60% Model-line projects
defined by plant
40% management
Department-wide
20% projects defined by
plant management
0%
MLA MLB MLC1 MLC2

FIGURE 5B. Interdepartmental Problem Solving on the Model Lines

100%
Projects spanning
80% at least three
departments
60% Projects between
two neighboring
40% departments
Intra-departmental
20% projects

0%
MLA MLB MLC1 MLC2

high). The MLA team found that wire-related variables determined in preceding
production stages (such as copper content) did affect drawability. In addition, the
team learned in an experiment that the temperature of the soap solution was
significant—dipping the wire into a warm soap solution improved drawability by
73%. The installation of a “hot lubricant rinse” in the production stage preced-
ing wet wire drawing achieved the same drawability improvements in full-scale
manufacturing (i.e., operational learning was high). The hot lubricant rinse and
controlling specific wire properties in preceding production stages led to a sharp
increase in productivity. Moreover, faced with the combination of conceptual
and operational learning, the Aalter plant adopted the findings for all products.
Products are primarily defined by the number of filaments and the diame-
ters of the filaments. In addition, products can be customized in terms of proper-
ties such as adhesion (ability to adhere to rubber) and tensile strength of the
wire. The MLA focus on a single product for a single customer kept characteris-

66 CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

tics such as adhesion and tensile strength constant, thus creating a more stable
environment for conducting controlled experiments. Once new technological
knowledge was created on the MLA such as “a hot lubricant rinse improves
drawability,” the Aalter plant could assess how such principles could be trans-
ferred to other products. The Aalter plant found that installing a hot lubricant
rinse before the wet wire drawing stage improved drawability for all products.
The products produced in plants B and C fell in the same range as the Aalter
products in terms of number of filaments, filament diameters, adhesion, and
tensile strength. Therefore, new technological knowledge created on any model
line with high conceptual and high operational learning could be applicable to
all three plants. Bekaert’s replication efforts of the MLA demonstrate how the
organizational context limited formal learning and transfer of formal learning.

Replication Efforts

MLB: Lack of Management Buy-In

Figure 5a shows that management buy-in for the model line concept was
absent in plant B. Plant management never allowed the MLB team to define its
projects. Winter notes: “Management must behave so as to provide a supportive
structure within which people at lower levels can act effectively to improve per-
formance. It cannot achieve significant performance improvements by its own
unilateral action, for it lacks the detailed knowledge required to do so.”19 Fur-
thermore, the majority of all projects were department-wide projects, i.e., pro-
jects conducted for all products within a single process step from Figure 1.
Department-wide projects are the opposite of true model line projects. The
model line concept recognizes interdependencies between process steps, and
controls for as much variation as possible by focusing on a single product.
Department-wide projects, on the other hand, ignore other process steps (and
thereby dynamic complexity) and introduce unnecessary variation by including
all products. Therefore, the MLB could never create insights such as “wire prop-
erty X determined in dry drawing will affect drawability in wet wire drawing
by a factor Y.” Therefore, the old tradition of departmental focus was not aban-
doned, hence the model line concept was not really given a chance. Why?
The MLA manager had full authority over production in all departments
(Table 3). He reported directly to the CEO, allowing the CEO to develop a deep
understanding of the model line concept. The MLA was an example of what
Mintzberg calls “thick management”—getting deeply inside a business, coming
to know its needs and its processes and its people well enough to weld them all
together into a smoothly functioning entity that serves its markets with care and
understanding.20 After the MLA manager was promoted to central R&D, model
line managers reported to plant managers, who in turn reported to the Steel
Cord division. Neither the plant managers nor the Steel Cord division developed
deep, intimate knowledge about the model line concept—a case of “thin man-
agement.” According to Mintzberg, “thin management remains distant from the
subject of its efforts, acting as if it moved pieces on a chessboard, making little

CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003 67

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

effort to influence what those pieces really do, even how they relate to each
other in any but the most superficial ways.”21 Consequently, plant managers did
not have an incentive to transfer authority over production from department
managers to model line managers.

MLC: Lack of Interdepartmental Problem Solving

After three years, management at plant C allowed the MLC team to
define all its projects. After establishing management buy-in for the model line
concept, it was still hard, however, to create scientific knowledge—witnessed
by the absence of formal learning and reliance on learning by doing (Table 4).
MLC1 projects were incremental productivity improvement projects compared
to the breakthrough projects on the MLA. Furthermore, as can be seen in Figure
5b, the MLC team did not engage in interdepartmental problem solving across
multiple production departments. Comparison of MLA and MLC in Table 3 sheds
light on these observations. The MLA possessed a rich, diverse knowledge base.
The manager had been with Bekaert for 16 years, and he could draw from both
R&D as well as production experience. Note in the MLA example on drawability
(discussed above) how the MLA manager utilized his knowledge of an R&D
project conducted in the 1970s to enhance the chemical model that led to
improved drawability. The MLA manager routinely combined such knowledge
covering all stages of the production process. The MLC manager, on the other
hand, possessed only one year of experience with Bekaert, only in plant C. Con-
sequently, the MLC team lacked the knowledge diversity necessary to engage in
scientific experimentation across multiple production departments. Several years
later, the corporate model line manager summarized:
“What we didn’t understand when we started model lines in plants B and C is
that a model line manager needs to have authority over production and projects,
and a young engineer is not a good choice to run a model line. Young engineers
lack experience in formal problem solving. Experiments failed.”

MLC2: Lack of High-Quality Ore to Mine

In 1993, the MLC team was given another model line MLC2. The MLC
team invested only 15% of its time and energy in MLC2. The MLC2 product had
been in production over 20 years (Table 3), and was already “down its learning
curve.” The strong learning by doing effect in Table 4 is based on cumulative
production volume since 1972. The part of the learning curve graphed in Figure
3, from 1993 to 1995, corresponds to the very flat part of the learning curve.
The MLC team continued to focus the bulk of its time and effort on the MLC1—
in part because the MLC team had to deal with the many disruptions in raw
material supply to MLC1 discussed earlier.
Several organizational factors impeded successful replication of the MLA.
Management buy-in for the model line concept was missing in plant B. Knowl-
edge diversity to solve interdepartmental problems was underdeveloped on
the MLC1. The MLC2, with a mature product, was already “down its learning
curve.” We conducted several robustness tests to rule out alternative explana-

68 CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

Ruling out Alternative Explanations

Instead of ongoing transfer of formal learning or learning by doing, knowledge could have
been transferred at the introduction of the model line concept.We tested for initial transfer
of either learning variable and found no evidence of transfer at start-up.
The technology shown in Figure 1 did not fundamentally change over time.We included calen-
dar time in our regression to capture any technological development. Calendar time was not
significant, confirming our assessment that technology did not fundamentally change.
If model lines struggled with different quality problems, knowledge from other lines would have
limited use.We studied reject rates per process step and found a similar distribution across
steps for all model lines. Reject rates were of the same order of magnitude for a particular
step on all lines, yet reject rates from one step to the next could differ by an order of magni-
tude.Therefore, it is unlikely that different quality problems impeded knowledge transfer.
The MLB TFP curve was flat.Was there simply no room for improvement? After the time
horizon studied in Figure 3, the company put more emphasis on cost. Plant B shifted to a
cheaper raw material supplier and immediately boosted TFP by 20%. Hence, there was signifi-
cant room left for improvement on MLB, underscoring our assessment that management of
plant B did not buy in to the model line concept.
We conducted numerous sensitivity analyses on TFP, changing one price or one cost at a time
by 10% (up and down). None of our results changed. It should be noted that our data is much
more precise than this large 10% margin of error.

tions (see Sidebar “Ruling out Alternative Explanations”). These robustness tests
increase our confidence in concluding that organizational factors significantly
impeded successful replication of the original model line concept.

Conclusion
Managing learning curves in factories is not an easy task. Many produc-
tion processes like Bekaert’s are shrouded in dynamic complexity, detail com-
plexity, ambiguity, and incomplete technological knowledge. Faced with such
barriers to learning, production personnel should engage in both conceptual and
operational learning. Knowledge created locally with both conceptual and oper-
ational learning can spread globally, thereby accelerating a factory’s global learn-
ing curve.
Organizational constructs such as Bekaert’s first model line should be run
as a learning laboratory to consistently produce both conceptual and operational
learning. Faced with dynamic complexity, separation of cause-and-effect in time
and space, a model line needs to cut across production departments. To success-
fully create new technological knowledge on a model line cutting across produc-
tion departments, management buy-in is required. Without authority to commit
different production departments to action, experiments are likely to fail.

CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003 69

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

Authority over production is critical, because learning needs to take place on the
factory floor. Note that we are not saying that every model line manager has to
report to the CEO. However, “thick management”—a deep understanding of the
required conditions to run a model line—is needed.
Running a learning laboratory in a factory requires controlling the pro-
duction environment as much as possible. A stable production environment will
reduce noise such that experiments yield better significance results. At Bekaert,
examples of noise reduction included stability in raw material suppliers as well
as focusing on a single product. However, in any specific case, factors that reduce
noise could be different things. Management should figure out which factors
reduce noise in order to maintain laboratory conditions. Running a learning
laboratory in a factory also requires knowledge diversity: different skills from
different production departments and disciplines, as well as a broad and diverse
knowledge base.
Given the managerial opportunity presented by the large variation in
learning rates, firms would be well advised to search for the required conditions
for learning and transfer of learning. Given the resources spent on quality and
productivity improvement efforts in firms and the often disappointing results,
learning to learn requires a concerted and dedicated effort.

Notes
1. For references on early learning curve studies, see Robert H. Hayes, and Steven C. Wheel-
wright, Restoring our Competitive Edge (New York, NY: John Wiley & Sons, 1984), chapter 8;
Linda Argote and Dennis Epple, “Learning Curves in Manufacturing,” Science, 247 (1990):
920-924.
2. Ray Stata, “Organizational Learning: The Key to Management Innovation,” Sloan Manage-
ment Review, 30/3 (Spring 1989): 63-74. For the importance of managing learning rates, see
Ramachandran Jaikumar and Roger E. Bohn, “A Dynamic Approach to Operations Manage-
ment: An Alternative to Static Optimization,” International Journal of Production Economics, 27
(1992): 265-282.
3. For variation in learning rates, see Ferdinand K. Levy, “Adaptation in the Production
Process,” Management Science, 11 (1965): B136-B154; John M. Dutton, and Annie Thomas,
“Treating Progress Functions as a Managerial Opportunity,” Academy of Management Review,
9/2 (April 1984): 245-247; Robert H. Hayes and Kim B. Clark, “Why Some Factories are
More Productive than Others,” Harvard Business Review, 64/5 (September/October 1986): 66-
73; Argote and Epple, op. cit.
4. For dynamic and detail complexity, see Peter M. Senge, “The Leader’s New Work: Building
Learning Organizations,” Sloan Management Review, 32/1 (Fall 1990): 7-23. For ambiguity,
see James G. March and Johan P. Olsen, “The Uncertainty of the Past: Organizational Learn-
ing under Ambiguity,” European Journal of Political Research, 3 (1975): 147-171. For incom-
plete technological knowledge, see Roger E. Bohn, “Measuring and Managing Technological
Knowledge,” Sloan Management Review, 36/1 (Fall 1994): 61-73. For misperception of feed-
back and poor inquiry skills, see John D. Sterman, “Learning in and about Complex Sys-
tems,” System Dynamics Review, 10 (1994): 291-330. For creation of myths, see March and
Olsen, op. cit.; Bo L.T. Hedberg, “How Organizations Learn and Unlearn,” in P.C. Nystrom
and W.H. Starbuck, eds., Handbook of Organizational Design (New York, NY: Oxford University
Press, 1981), pp. 3-27; Daniel H. Kim, “The Link between Individual and Organizational
Learning,” Sloan Management Review, 35/1 (Fall 1993): 37-50.
5. For a scholarly development of the ideas in this article, see Amit S. Mukherjee, Michael A.
Lapré, and Luk N. Van Wassenhove, “Knowledge Driven Quality Improvement,” [Part 2 of
2], Management Science, 44/11 (November 1998): S35-S49; Michael A. Lapré, Amit S.
Mukherjee, and Luk N. Van Wassenhove, “Behind the Learning Curve: Linking Learning

70 CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003

 Managing Learning Curves in Factories by Creating and Transferring Knowledge

Activities to Waste Reduction,” Management Science, 46/5 (May 2000): 597-611; Michael A.
Lapré and Luk N. Van Wassenhove, “Creating and Transferring Knowledge for Productivity
Improvement in Factories,” Management Science, 47/10 (October 2001): 1311-1325. For
further information on Bekaert, see Amit S. Mukherjee, Michael A. Lapré, and Luk N. Van
Wassenhove, “Bekaert: Beyond the Quality Prizes,” INSEAD-CEDEP case 01/94-4253.
6. Mukherjee, Lapré, and Van Wassenhove (1998), op. cit.
7. Lapré, Mukherjee, and Van Wassenhove, op. cit.
8. Lapré and Van Wassenhove (2001), op. cit.
9. Kim, op. cit.
10. Ibid., for similar definitions.
11. For the methodology behind the classification, statistical analyses, and robustness tests, see
Lapré, Mukherjee, and Van Wassenhove, op. cit. For examples of projects for each category,
see Michael A. Lapré and Luk N. Van Wassenhove, “Learning across Lines: The Secret to
More Efficient Factories,” Harvard Business Review, 80/10 (October 2002): 107-111.
12. W. Bruce Chew, Dorothy Leonard-Barton, and Roger E. Bohn, “Beating Murphy’s Law,”
Sloan Management Review, 32/3 (Spring 1991): 5-16; Gary P. Pisano, The Development Factory:
Unlocking the Potential of Process Innovation (Boston, MA: Harvard Business School Press,
1997).
13. For natural and controlled experiments, see Roger E. Bohn, “Learning by Experimentation
in Manufacturing,” working paper 88-001, Harvard Business School, Boston, MA, 1987. For
statistical process control, see Harrison M. Wadsworth, Kenneth S. Stephens, and A. Blanton
Godfrey, Modern Methods for Quality Control and Improvement (New York, NY: John Wiley &
Sons, 1986). For similar approaches to learning in factories, see Dorothy Leonard-Barton,
“The Factory as a Learning Laboratory,” Sloan Management Review, 34/1 (Fall 1992): 23-38;
Chew et al., op. cit.
14. See Mukherjee, Lapré, and Van Wassenhove (1998), op. cit.
15. For persistence of productivity differences, see W. Bruce Chew, Timothy F. Bresnahan, and
Kim B. Clark, “Measurement, Coordination, and Learning in a Multiplant Network,” in R.S.
Kaplan, ed., Measures for Manufacturing Excellence (Boston, MA: Harvard Business School
Press, 1990); Michael van Biema and Bruce Greenwald, “Managing Our Way to Higher
Service-Sector Productivity,” Harvard Business Review, 75/4 (July/August 1997): 87-95. For
failure to replicate best practice, see Carla O’Dell and C. Jackson Grayson, “If Only We Knew
What We Know: Identification and Transfer of Internal Best Practices,” California Management
Review, 40/3 (Spring 1998): 154-174; Gabriel Szulanski and Sidney Winter, “Getting It Right
the Second Time,” Harvard Business Review, 80/1 (January 2002): 62-69.
16. For details on our TFP measure, see Lapré and Van Wassenhove (2001), op. cit. For a gen-
eral description of TFP, see Hayes and Clark, op. cit.
17. For improving supplier relationships, see Charles J. Corbett, Joseph D. Blackburn, and Luk
N. Van Wassenhove, “Partnerships to Improve Supply Chains,” Sloan Management Review,
40/4 (Summer 1999): 71-82.
18. Part of this example draws on Amit S. Mukherjee, “The Effective Management of Organiza-
tional Learning and Process Control in the Factory,” unpublished doctoral dissertation,
Harvard Business School, Boston, MA, 1992.
19. Sidney G. Winter, “Organizing for Continuous Improvement: Evolutionary Theory Meets
the Quality Revolution,” in J.A.C. Baum and J.V. Singh, eds., Evolutionary Dynamics of Organi-
zations (New York, NY: Oxford University Press, 1994), pp. 90-108.
20. For thick versus thin management, see Henry Mintzberg, Mintzberg on Management: Inside Our
Strange World of Organizations (New York, NY: The Free Press, 1989).
21. Ibid.

CALIFORNIA MANAGEMENT REVIEW VOL. 46, NO. 1 FALL 2003 71