Profit from the Learning Curve

Winfred B. Hirschmann

From the Magazine (January 1964)

Practice makes perfect. A thing can always be done better not only the second time but each succeeding time by trying. This everybody knows. But how many know that the pattern of improvement can be sufficiently regular to be predictive? How many realize that such patterns can characterize, not only individual performance, but also the composite performance of many individuals organized to accomplish a common task?

The industrial learning curve quantifies such performance. It has evolved from experience in airframe manufacture, which found that the number of man-hours spent in building a plane declined at a regular rate over a wide range of production. Such continuing improvement was so common in the aircraft industry that it became the normal expectation in the war time mass production of aircraft; thus, production and other types of performance were customarily scheduled on some basis of progressive betterment.

But this is not the practice in industry generally. Although learning curves have been recognized in industries other than aircraft, they have not been as widely accepted. Instead, predictions are usually based on assumptions of level performance and constant costs.

In this article I will argue that this practice is not only incorrect but costly. People do learn, and they learn according to a generally predictable pattern. The learning curve, I believe, is an underlying natural characteristic of organized activity, just as the bell-shaped curve is an accurate depiction of normal, random distribution of anything, from human I.Q.’s to the size of tomatoes. Wherever people strive to do better, improvements result; otherwise, how would progress take place?

By failing to capitalize on this natural phenomenon, managers will not encourage continued efforts once they become convinced that “further improvements are not possible.” Further improvements are always possible over time, so long as people are encouraged, or even ordered, to seek them. Thus, an understanding of the learning curve becomes of crucial importance to the business manager.

Characteristic Curve

Before I demonstrate its application, however, it is necessary to clarify just what I mean by the learning curve.

Learning patterns were reportedly first observed for manufacturing operations in 1925 by the commander of the Wright-Patterson Air Force Base in Ohio.¹ During subsequent years, definitive studies of aircraft assembly showed the following pattern: the fourth plane required only 80% as much direct labor as the second; the eighth plane, only 80% as much as the fourth; the one hundredth, only 80% as much as the fiftieth; and so on. Thus, the rate of learning to assemble aircraft was concluded to be 80% between doubled quantities. On an arithmetic chart, with linear coordinates, the relationship is a curve, showing a rapid initial decline that later trails off (see Exhibit I-A). On a double logarithmic chart, however, it is a straight declining line, which reflects a constant rate of reduction (see Exhibit I-B). Such a straight-line relationship is easier to draw and use for prediction purposes.

Exhibit I. The 80% Learning Curve

While the learning curve is a universal phenomenon, it has many variations in form; for example, there are great variations in the level at which a curve starts (i.e., the cost of the first unit). This is simply because of the different ranges of complexity of items. Nevertheless, the slope of the curve is common to a wide variety of experience. In fact, it was the regular finding of the common slope of about 80% for fighter, bomber, and transport planes that started speculation about a general theory of learning curves.² Further investigation showed that, although operations having essentially the same proportions of labor content have fairly common slopes, other operations differ in characteristics and corresponding slopes of curve.

Thus, operations paced by people have steeper slopes than those paced by machines. For example:

In airframe manufacture three-fourths of the direct labor input is assembly; the balance is represented by men engaged in machine work. In such a largely man-paced operation, an 80% curve is commonly found.

But when the proportion of assembly work is lower, the downward slope of the curve is not so steep. If the ratio of assembly to machine work is 50/50, the slope is about 85%. If the ratio is one-fourth assembly and three-fourths machine work, the operation is largely machine-paced, and the slope is around 90%.

These results might be expected since learning is related to people—the fewer the people, the less the capacity for learning.

These various percentages provide a point of departure on which to base predictions. Such predictions, in general, have proved more valid than those based on an assumption of level performance, i.e., a condition of no learning.

As Exhibit II, depicting the experience of several aircraft companies, shows, it would be possible to draw a line to fit the successive points of a plot; they would not fall directly on it, but most of them could be covered by a band centered on the line. Both define a declining trend. The learning curve is the line that fits the points; and the boundaries of the band are the upper and lower limits within which random variations cause the points to fall.

Exhibit II. Data Illustrating Normal Learning Curve and Some Deviations Source: Miguel A. Reguero, “An Economic Study of the Military Airframe Industry,” Wright-Patterson Air Force Base, Ohio, Department of the Air Force, October 1957, pp. 231–235.

Whenever points fall outside these limits, there is usually an assignable cause, just as there is for any other operation which “gets out of control.” For example, Exhibit II indicates:

A “leveling-off,” or even a “toe-up,” can occur, such as at the end of a contract when workers are transferred to other production lines and closing-out operations become inefficient.

A rise in the curve can occur in the middle of a contract too, owing to a substantial interruption (such as that caused by introducing changes in a model, by moving operations to a new building, or by halting operations for a while so that forgetting occurs). Shortly after operations recommence and skill in handling changes is acquired, the curve declines rapidly to approach the old slope. Such a break in the curve occurs frequently enough to have acquired the descriptive term “scallop.” In fact, if, instead of merely a change being made, a new model is introduced, or a new type of item is put into production, the scallop occurs initially and the curve essentially starts again. Thus, the direct labor input reverts back to what it had been when the first item of the preceding type was put into production (assuming that the two items were of similar type and configuration).

Such learning curve performance was so common in aircraft assembly that it became expected as a matter of course. And it is still used consistently today as a basis for projecting costs, forecasting manpower requirements, rating personnel, scheduling production, and negotiating multimillion dollar contracts.

Barriers to Acceptance

While learning curves have been reported for operations in other industries (particularly by personnel who have migrated from aircraft companies), their occurrence seems to be considered the exception rather than the rule. Why has a technique used so long and validated in one industry not been adopted in commerce generally? Here are some possible reasons:

There is a lack of awareness that improvement patterns can be reasonably well quantified. Although everyone has had the experience of improving his skill by repetition, this experience is generally assessed qualitatively. Improvements can seem irregular and fluctuating at the time they occur. Consequently, it is not obvious that, over the long term, these wobbling improvements can be tracing a definable trend. Studies of estimates based on empirical trends in the aircraft industry indicate that improvement curves actually approach a reliability comparable to that experienced with engineering construction estimates.

Skepticism that improvement can continue may be another factor that has limited more general acceptance of this technique. After arduous hours or months of effort to achieve an increment of advance, it is natural to feel that the last ounce of betterment has been wrung out of an operation. As a result the expectation of still further improvement seems quite unrealistic.

Many companies believe, “Our business is different,” and consequently that such curves do not apply to their operations. Credence is given this conclusion by the observations of some experienced practitioners who expect the learning curve to be inapplicable or have little value in such industries as basic chemicals, plastics, and petroleum refining.³ Of course, each business is different from others. Nevertheless, since it is universal experience that over the long run a method can always be improved in succeeding times by trying, there is a basis for believing that progressively improving performance can be universally brought about by effective trying.

Learning curves which are already occurring may not be recognized by many people. With the 80% learning curve characterizing the mix of three-fourths assembly and one-fourth machine work commonly found in airframe manufacture, improvement was so rapid that the direct labor hours for the tenth plane manufactured were less than half those required for the first one. Such a marked decline virtually forced attention to it. Nevertheless, 11 years passed between 1925, when learning curve performance was first noted in aircraft, and 1936, when it was publicly reported by the manager of the Buffalo plant of Curtiss-Wright Corporation.⁴ Consequently, other industries with much smaller ratios of assembly to machine work can be expected to be less susceptible to improvement, and therefore to have curves with less slope. The smaller amount of improvement could be obscured by larger forces, or be offset by such opposing effects as inflation. Improvement might also escape notice by occurring over a longer time, or be attributed to other causes.

Finally, there may also be a lack of awareness that the learning curve can describe group as well as individual performance, and that any group can comprise not only direct labor but also behind-the-scenes people who deliberately seek improvements in the process. In airframe manufacture, for instance, such groups may include tool engineers who contrive new jigs and fixtures. In other types of manufacture, there may be staff groups which work on different combinations of operating conditions to improve yields or devise instruments to improve control.

The last point is particularly important. Yet most managements have failed to recognize that technological progress is a kind of learning. Assigning specialists to seek technical improvements and to incorporate them in operations obviously helps bring about improvement. The industrial learning curve thus embraces more than the increasing skill of an individual by repetition of a simple operation. Instead, it describes a more complex organism—the collective efforts of many people, some in line and others in staff positions, but all aiming to accomplish a common task progressively more efficiently.

This broader concept may be the reason why the phenomenon has many names: “manufacturing progress function,” “cost-quantity relationship,” “cost curve,” “experience curve,” “efficiency curve,” “production acceleration curve,” “improvement curve,” and “performance curve.” Such terms are more commonly used in writing than in conversation, where the term “learning curve” predominates. It seems preferable to retain the general name, but to remember that it can be broken down into meaningful terms which are more specific.

Testing Applicability

If learning curve performance is a natural characteristic, then such performance should be found not only for more types of activities already recognized as responsive, but also for unlikely operations such as those not previously reported or believed susceptible. Petroleum refining offers a good example of the type of industry to which the learning curve might be thought to be inapplicable. It is characterized by large investments in heavy equipment, and is so highly automated that learning is thought to be either nonexistent or too small to be of value. Let us see how true this belief is.

In processing units

Petroleum refining comprises such process operations as distillation, cracking, and reforming. From time to time, scattered observations have appeared in the literature to the effect that the capacity of some of the units performing these operations is larger than their design capacity. In 1951, the worldwide, installed design capacity of fluid catalytic cracking units was 1,200,000 barrels per stream day; however, the actual throughput was about one-third greater—as much as 1,600,000 barrels per stream day! This aggregate is composed of the throughputs of the individual cracking units at a point in time. It does not show how they changed over time.

But by plotting the performance of individual units at a point in time against their age at that time, we obtain a clue to this pattern; see Exhibit III. The points are the ratios of the achieved capacity to the design capacity as determined from published tabulations. The selected points of Exhibit III were calculated from these data. For example, the first unit (A) at the time of this research—1958—was one and one-half years old and by that time had achieved approximately 116% of design capacity. The second unit (B) at the time of this research was four years old and had achieved about 125% of design capacity. In general, as the older units show (C, D, E, F, G, and H), performance rapidly improved in the first few years, and continued at a slower rate in later years.

Exhibit III. Capacity Achieved by Different Fluid Catalytic Cracking Units of Various Ages (in 1958)

Exhibit IV shows that successive annual points for an individual cracking unit indicate that growth occurs in a step-wise fashion, so that the points are scattered in a band instead of lying on a smooth curve. The pattern of improvement indicated by the colored line, which is the same curve as the one in Exhibit III, resembles the inverse of a learning curve on an arithmetic chart. If the parameters are changed so that the number of days to process 100,000 barrels is plotted against cumulated throughput on a logarithmic chart, a declining straight line can be drawn through the points (see Exhibit V). The line has a “slope” of about 90%, as might be expected from a machine-paced operation which involves comparatively little direct labor.

Exhibit IV. Performance of an Individual Fluid Catalytic Cracking Unit

Exhibit V. Learning Curve—Plot of Performance Corresponding to Exhibit IV

What accounts for this improvement in performance? Safety margins for critical equipment are included during design stages of a project to ensure getting required (design) performance. Thus, actual performance can and should be higher than the design target. Operators will soon learn to take advantage of built-in safety margins. However, equipment not considered critical in design and without extra safety margins may limit initial performance to the design target. Obviously, removal of such a bottleneck can result in marked improvement. But as time passes, fewer and fewer bottlenecks remain to be uncovered, so progress slows.

These circumstances explain a relatively rapid early growth and a subsequent gradual slowing down as more and more capital and ingenuity are required to eliminate further bottlenecks. In this industry with its heavy dependence on machinery, the improvement curve appears to reflect technological resourcefulness. So it seems reasonable to believe that this technical skill will continue to result in such enhancement patterns so long as increased demand or other incentives occur to prod the search for improvement, and so long as it is backed up by the present level of research and engineering efforts.

In start-up operations

But if learning is greatest where the most people are involved, the learning curve should be revealed most clearly in refinery operations which have a high labor content. One such circumstance would be the start-up of units after shutdown periods for repairs. Specialists are on hand to assure proper functioning of instruments. Extra craft people are assigned to handle emergencies. More supervisors are present to give guidance. The regular crew is particularly busy in routing flows, opening and closing valves, lining out system components, and operating manual controls before the automatic ones are cut in. Shift workers may at times “double over.”

This relatively high human activity content of start-ups suggests that the people involved should be susceptible to a significant degree of learning. And actual experience confirms this. Over a period of ten years the time necessary to put a Whiting refinery fluid cracking unit of the American Oil Company on stream dropped to less than half the time initially required.

In maintenance

In view of the repetitive nature of much maintenance work and the continuing efforts made to organize for better efficiency, the performance of a maintenance department as a whole might be expected to show progressive learning. And, as Exhibit VI shows, this expectation is sound. The points in the exhibit show a declining trend for productive labor man-hours on maintenance and shutdowns during the years 1949–1956. At the end of that period the plot seems to be leveling off. If the management had speculated then about this curve, it might have felt that it had reached a plateau, that maintenance had learned how best to do the jobs required so that a further decline would not occur, at least for a while. But actually, as the 1958 point (plotted as a cross) shows, the trend continued, ending up about where it should have been expected.

Exhibit VI. Productive Refinery Labor Man-Hours for Maintenance and Shutdowns

This is not difficult to explain. Department performance is a composite of many individual tasks. And if it is a reflection of learning, then individual maintenance operations should follow established learning curves. This expectation is confirmed not only in refining but also in other large manufacturing plants. Exhibit VII depicts a record of the performance of a repetitive maintenance operation by General Electric Company at its Richmond-Washington AEC plant. It shows that the time required to replace a group of parts during shutdowns declined along a 76% curve.

Exhibit VII. Maintenance Learning Curve at a General Electric Plant Source: Carl A. Bennett, Application of a Learning Curve to a Maintenance Problem, Second Annual Quality Control Symposium of the Dallas-Fort Worth Section, ASQC, March 16, 1957.

In construction

Building new items of heavy equipment also appears to be characterized by learning. The 245,000 barrel refinery at Fawley, England, which went on stream in the early 1950’s, could have been duplicated 5 years later at 70% of the original cost. This decrease in cost represents a rate of decline of about 7% per year. The per-barrel investment costs of units for some individual processes (thermal cracking, polymerization, catalytic cracking, and catalytic reforming) also decline progressively. In the case of thermal cracking, the decline continued for 33 years.

A learning curve measure of the rate of decline for fluid catalytic cracking units is suggested by the observation that the steel required and the investment cost in 1955 were estimated to be one-third of those required to duplicate the capacity of the original downflow fluid plant, which was built in 1942. During the intervening years, about 3 million barrels of fluid cracking capacity were built. When, as in Exhibit VIII, these two cost points are plotted on a logarithmic scale against the cumulative installed design capacity of fluid catalytic cracking plants and connected by a solid straight line, the slope of that declining line (in black) turns out to be about 80% (in constant dollars).

Exhibit VIII. Learning in Construction of Fluid Catalytic Cracking Units

Note that price rises can distort the picture. If the “Nelson’s Refinery Construction Cost Index” for petroleum refining equipment is applicable, the actual dollar cost in 1955 of building a 1942 unit would have been 2.2 times as much as in 1942.⁵ After taking into account the technology which enabled it to be duplicated in 1955 for one-third this cost, we arrive at a unit cost in 1955 dollars of 2.2 × 33.3% or 73% of the 1942 dollar cost. On this basis, the cost reduction over a 13-year period corresponds to a slope of 94%, and is represented by the black line on Exhibit VIII. The fact remains that the true technological benefit is measured by the 80% curve (the colored line.)

It might seem surprising to find that the construction of multimillion dollar processing units, which are comparatively few and built to order, is characterized by the same 80% slope commonly found in manufacturing many different items in large numbers on production lines. However, since such construction is largely assembly work, this finding is consistent with the general learning curve experience: that is, operations with similar ratios of assembly to machine work have similar learning curve slopes.

This decline reflects learning by construction people in how to build so as to reduce unit costs, and by research, engineering, and operating people in what to build. It provides a clue to the relative contributions of capital and of technology to learning curve performance. The fact that a plant built to duplicate the performance of an original one could be built with one-third the steel indicates that the what-to-build contributions of the progressively improved technology embodied in successive plants greatly outweighs the how-to-build contributions of capital for the better tools and equipment employed in building them.

The great extent to which technology can be dominant over investment has been shown by three recent studies. These cover somewhat different time periods, but each independently concludes that about 90% of the U.S. growth in output per man-hour has been due to technological change, and only about 10% to increased investment in capital equipment.⁶

Technology also contributes to building larger plants with attendant economies, because construction costs do not increase proportionately with size. Doubling the capacity of a plant does not necessarily double the cost. Rather, it increases it by some lesser amount, which can be represented by an exponent of the size. A plant twice as large may cost about 2^0.7 or 1.6 times as much as the base size. If this 0.7 cost exponent applies to fluid cracking plants, then the per-barrel investment costs of various multiples of size decline, as shown by the column of crosses on the right side of Exhibit IX. Since plants have been built two or more times larger than originally planned, the economies of size from progress in research and engineering also contributed to offsetting the effects of inflation.

Exhibit IX. Learning in Construction of Various Size Fluid Catalytic Cracking Units

In this instance, technological progress did decrease costs more than inflation increased them. This circumstance might suggest that depreciation allowances have been more than adequate to provide the capital necessary to replace units when they are retired. Actually, rapid obsolescence is concomitant with rapid technological progress. A competing firm with a new, low-cost unit has the advantage of a smaller capital charge for its products and can thereby profitably sell them at lower prices. To survive in the face of such competition, companies may be forced to replace existing units before they are fully amortized and before depreciation reserves have become large enough to pay for the replacements. And this is particularly true for units depreciated by the straight-line and other methods required for equipment installed before 1954.

Industry performance

An industry is an aggregate of components. We can reason that if learning in components is widespread, it should be reflected in aggregate performance. A logarithmic plot (in Exhibit X) of man-hours per barrel versus cumulated barrels of crude oil refined in the United States since 1860 results in the fairly regular type of decline such reasoning would lead us to expect.

Exhibit X. Man-hours per Barrel Refined in the Petroleum Industry

Other industries show similar declines, as illustrated by Exhibit XI for the U.S. electric power industry and by Exhibit XII for the U.S. basic steel industry.

Exhibit XI. U.S. Electric Power Industry—Suggests Learning Curve Decline Source: John E. Ullman, “Economics of Nuclear Power,” Science, April 4, 1958, p. 140.

Exhibit XII. Man-Hours per Unit of Output in U.S. Basic Steel Industry—Suggest Learning Curve Decline 1935–1955 Source: Bulletin 1200, Washington, U.S. Department of Labor Statistics, September 1956, and other sources.

Such increases in aggregate productivity reflect the joint effect of many interrelated influences. Among them are technological advance, increased capital investment, better methods of management, increased health and education of workers, and improved communications. Their over-all aim, however, is progress, and such declines can be regarded as the result of learning how to do things better.

Since construction, maintenance, processing, and start-up operations are common in industry generally, and since learning curves can characterize such activities in the petroleum industry (where they were neither previously reported nor expected to be significant), it seems reasonable to believe that similar component activities can follow learning curve patterns in other areas also, and may in fact already be contributing to the progress occurring. The credibility of this generalization is reinforced by understanding the elements of learning and the practices which promote it.

Important Implications

What does all this seem to add up to? Essentially this. There are two main factors which affect learning: (1) the inherent susceptibility of an operation to improvement, and (2) the degree to which that susceptibility is exploited. In detail:

1. Inherent susceptibility is related to the human content of an operation. It is reflected in the ratio of assembly to machine work. The greater the human content, the greater the susceptibility for improvement.

2. The degree to which susceptibility is exploited is related to the dynamic content of the environment—the drive and resourcefulness of management and its skill in stimulating supervisors and technical people to be creative and workers to be productive. The greater the dynamic content of the environment, the greater the exploitation of inherent susceptibility.

Effect of faith

One of the factors affecting the dynamic content of the environment is faith. If progress is believed possible, it will likely be sought; and if it is looked for, there is some possibility of finding it. Conversely, if improvements are considered unlikely, there will be little urge to seek them. A defeatist philosophy can be engendered which so debilitates an effort that it helps to produce the very condition it assumes. Industrial engineers have long known that once a quantitative objective is imposed on an organization, there are strong forces created to fit the objective. There are a number of studies showing that progress did cease when the actual unit time reached the originally estimated time for the job.

Consider, for instance, the experience of a company with a “cost-plus” subcontract from an aircraft company:

This company expected that costs would begin at a high level, but would drop during an initial period of acquiring familiarity, and then level off. This is what actually occurred, and the subcontractor was satisfied.

The aircraft company, however, pressed for continued reduction in unit costs. To its people, learning curve performance was such a common phenomenon that they did not initially explain it to the subcontractor as the basis of their expectation. When unit costs ceased to decline, however, lower echelons of the aircraft company talked with lower echelons of the subcontractor, who proved to be unfamiliar with learning curve experience and did not believe it. When lack of progress continued, the upper management of the aircraft company talked to its counterpart in the subcontracting firm, and found that level equally unfamiliar and unable to understand why the reduced costs already achieved were not satisfactory. Finally, the aircraft company insisted that unit costs could and must decline, and that if the subcontractor company did not know how to lower them, he would be shown. A team of aircraft people explained the learning curve and showed how to get the results which produce continued progress. Thereafter, the decline occurred along the learning curve originally expected by the prime contractor.

Open-ended expectations

The limiting effect of “ceiling psychology” also shows up in the expected performance of new equipment. If it is designed to operate at a given rate and fails to perform as expected, great efforts will be made to bring it up to target. It is taken for granted that rated or specified performance can be achieved. But once that level is reached, expectations are proved justified, and attention is directed elsewhere on the presumption that rated output is the limit of capacity.

Such were the circumstances in the following example:

A new machine had a given rated capacity, but initial production was less than half that capacity. Some officials in the company had opposed the selection of this equipment on the ground that they did not believe the quality and quantity would be acceptable. Its initial performance supported their belief, and they recommended that the machine be replaced with the one they had favored.

However, those who had selected the original one were still convinced it could perform adequately, and asked for more time. The firm’s management engineering consultant supported them; he had noted the occurrence of some improvement along a line that suggested “learning.” The machine was retained for further trial, and output gradually rose to the rated capacity in about four weeks.

With “normal” output achieved, no further improvement was expected, and concerted efforts were about to cease. But the consultant noted that the improvement pattern had continued to trace a learning curve. He reasoned that progress might continue, and therefore engineers were kept assigned to the machine. The result was a further increase to the level of triple the rated capacity in four months.

Thus, not putting a ceiling on expectations may permit improvements to continue. It may also result in a faster rate of improvement. Some companies report getting more rapid progress when the operating force is not informed of the target rate.

Conversely, if betterment is not believed possible, then the incentive to seek improvement is reduced, and an atmosphere of maintaining the status quo is encouraged. Assembly workers get to frown on increases as rate-breaking, and engineers who have ideas for possible improvement may hesitate to push them if they will be interpreted as “rocking the boat,” or if failure risks censure.

If there is a limit to learning, it could be expected to be reached after quite a period of time, during which skill and technology would have improved performance to the ultimate level. This is what seems to occur frequently in an established plant. In a Cornell University study, however, two separate items which were thought to have reached an achievement plateau were found to show continued learning when transferred to another firm.⁷

There is evidence that improvement can persist over many years, and on many millions of items. In a man-paced operation involving the assembly of candy boxes, for instance, the learning curve was found to have persisted for the preceding 16 years during which 16 million boxes were assembled by one person.⁸ In a machine-paced operation, progress also has continued over the production of tens of millions of units.

Although the percentage of improvement in such cases becomes small after a long period, the benefit can still be substantial if the number of units involved is large.

Progress by serendipity

The assumption that continual progress can occur may also create an atmosphere which encourages uncovering ideas or recognizing them when stumbled upon. Here is an example of this possibility:

In an automatic operation for forging wrenches, a piece of hot metal is forged into the proper shape in a die. With repeated use, the die “wears” or becomes larger, so that eventually the dimensions of the wrench exceed specifications and the die has to be replaced. Experience indicates the minimum number of forgings that can be made before replacement is required, and likewise the maximum number.

On the evening shift at one company, dimensions of the forgings began to vary erratically and thus indicated the need for die replacement much sooner than usual. Investigation disclosed that a new man was on the job, and that he was tinkering with the furnace temperature. When the tinkering was stopped, the die continued to perform satisfactorily. The matter could have ended there with the customary reprimand to the employee to follow instructions.

However, the investigator noted that there had been some good as well as bad results. In studying the matter, he found that the steel used could be forged at temperatures much higher than those specified without burning or destroying the properties of the metal. New dies were subsequently made smaller, and the metal first forged at the minimum practical temperature. As the die wore (got larger), the forging temperatures were increased. The hotter forgings shrank more on cooling, and in this way continued to fall within tolerance. This practice permitted a die to produce triple the normal output.

Thus, even with automatic operations, there may be transient periods when performance is better than normal. If these situations can be recognized and studied with the same zeal and conviction as is applied to below-normal performance, it may be possible to make the improvement persist. An optimistic research climate may in this way encourage progress by serendipity—making happy and unexpected discoveries by accident.

Wishing not enough

The general knowledge of the pervasiveness of learning and the actual experience of having had it occur, sometimes in the most unexpected places, should lead to the conviction that it can, will, and must occur. Yet even companies that have had much experience with learning report occasions when learning has failed to come about as predicted. Why? The explanation is that merely expecting progress is not enough to bring it about.

One aircraft firm states categorically that if a foreman does not get the expected amount of improvement, he is fired. But the foreman does not control all his environment. If he is provided with improper tools, an insufficient budget, or not enough staff assistance—in other words, if the system is not right—it is the management, some say, that should be replaced.

Perhaps in recognition of this circumstance, the general manager of another airframe company took different action when costs of a given model were not declining as expected:

A conference with his staff developed the explanation that the many changes required by the government were responsible for the lack of satisfactory progress. Further investigation indicated, however, that along with each change required by the government, the plant’s own engineering staff was making additional changes to improve the operation on the ground that, as long as the interruption would occur anyway, the extra changes could be incorporated at substantially no extra cost or disadvantage.

The general manager’s action was to order 15% of the design and tool engineers to be transferred to another contract. Even though there were strong protests that such action would make it impossible to do the necessary work, he stuck by his edict. As a result, the work piled up in front of the remaining people so that they did not have enough time for embellishment. After 30 days, the number of extra changes generated by the engineering department dropped to four for each change initiated by the government, and the cost curve began to decline. The next month, the transfer of an additional 15% of the engineers brought about a further decline in unit costs. Three such transfers were made, and the extra changes generated dropped to an average of about two and one-half for each change required by the government. Concurrently, the unit cost curve established a decline along the expected path.

Just why the number of transfers stopped at three was not explained. Perhaps the previous experience of the general manager which convinced him that progress could be made also taught him that, with the operation of his plant and industry, any further gains would introduce offsetting penalties so that there would be no further net benefit.

Result of need

This circumstance illustrates that merely expecting progress does not bring it about. It is not ordained by fate to arrive on schedule, but must be continuously, vigorously, and resourcefully sought. Such drive is usually the result of need. In the case of animals, the need is usually hunger—the desire to survive. It has been found to lead to learning curves for a whole range of organism complexity: amoebas, ants, snails, rats, and monkeys.

This same need underlies national and industrial progress. It has been said that civilization rides on the gun carriage—that the threat to survival in a war is both so large and so apparent that masses of people strive with efforts they would otherwise not make to get progress which would otherwise not result.

Survival is also the drive in industry. The proprietor of a new business will work long hours to get his company on its feet, and threats to profit will spur efforts in an established firm. If a company coasts, it loses ground; and if it continues to slip, it faces bankruptcy. The threat of economic extinction is like any other threat of death. It prods the firm to strive for improvements and for the progress necessary to continue. When improvements must occur for survival, the question of whether they can occur becomes trivial. The problem becomes one of uncovering what has to be there.

Concepts not new

These concepts are well known. They are reflected in the motto, “Progress Is Our Most Important Product,” and the slogan, “Petroleum Promotes Progress.” They underlie the familiar phrases, “necessity is the mother of invention” and, “practice makes perfect.” It is the universal experience epitomized by such proverbs that makes it reasonable to expect that any activity involving organic life is susceptible to learning curve improvement.

The contribution of learning curve studies has been to validate this expectation by finding learning curves for both simple and complex organisms, and for various levels of organized activity. These studies have brought the concept into sharper focus and put it into a more meaningful and useful framework by introducing a measure of quantification where previously there was only qualitative expectation.

These results can be summarized as learning curve doctrine:

1. Where there is life, there can be learning.

2. The more complex the life, the greater the rate of learning. Man-paced operations are more susceptible to learning or can give greater rates of progress than machine-paced operations.

3. The rate of learning can be sufficiently regular to be predictive. Operations can develop trends which are characteristic of themselves. Projecting such established trends is more valid than assuming level performance or no learning.

4. Learning is related to the dynamic context of the environment. Faith and incentive stimulate progress, and provide the drive to exert the energy, resourcefulness, skill, and persistence needed to bring it about. Conversely, “ceiling psychology” and the tendency to maintain the status quo, or not to rock the boat, inhibit learning.

Practical Applications

Enunciating these fundamentals to a manager who finds he has been experiencing learning in some of his operations might seem akin to expecting him to be astonished at being informed that he has been writing prose all his life. But there is more here than a mere labeling of what is happening. Suppose we itemize some of these applications.

1. If learning has been occurring, it is reasonable to believe that it can continue, if the dynamic content of the environment remains the same. It is prudent to include this potential of existing operations for continued improvement in expansion plans and other longer term forecasts. Reflecting it in capital expenditure plans could influence the size of the next increment of capacity, or defer the increment, and thereby enable more of evolving technology to be eventually incorporated as well as provide other earnings on the capital expenditure in the meantime.

2. Another application lies in choosing between modernizing existing plants and replacing them. It seems reasonable to expect more potential for improvement to be in new units of equipment than to remain in existing ones which have had most of their bottlenecks removed. This difference in potential is an advantage to the replacement alternative which has generally been considered intangible. It can now be evaluated.

3. There is also an incentive for a company to refine its practices so as to capitalize more fully on the potential inherent in its daily operations. If historical data are on hand for plotting, they can provide an experience base for predicting further improvement. Even if available—not in terms of unit labor required versus cumulative output, but as costs versus time or other parameters—historical data may still be useful if they define a trend because, once established, a trend is inclined to persist.

Since past records reflect past practices, they should only be used as a point of departure. Increased management attention, more effort, greater incentives, the application of a higher caliber of resourcefulness and creativeness, or other changes in the dynamic content of the environment may lead to a faster rate of progress. If the aim of better performance is set in a framework of goals which contain a realistic allowance for betterment, ways of achieving improvement may be sought which would otherwise not be sought. Such goals can constitute a real challenge to a group’s capacity to meet them. And in this way the expectation may help bring about the improvement. Evidence that this approach can be fruitful is provided by two examples. They were not originally reported as learning, but they reflect the application of learning curve doctrine. Thus:

In Du Pont’s 1,000-acre, 1,800-man Sabine River petrochemical works in Orange, Texas, production at the end of 1955 had reached the highest level in the plant’s history, with the average for the year being 25% above that of 1954. Plant size had increased by 13% over the preceding two years. Nevertheless, in the face of the increased maintenance burden, “goals set with due allowance for increased productivity” helped reduce over-all maintenance costs despite higher wages, higher material prices, and higher employee benefits.

The ratio of maintenance costs to plant investment dropped from 5.9 in 1953, to 5.0 in 1954, and to 4.3 in 1955. In the same period, the number of people required for the maintenance operation was reduced by more than 12% (126 men out of about 975); capital tied up in spare parts and extra machinery was reduced by one-third; no sacrifice was made in upkeep or maintenance standards; and, of considerable significance, no injury of serious consequences occurred.⁹

Here further evidence that such continued pacing supported by continued striving can lead to continued improvement is provided by Henry Ford’s results with his Model T. Ford wrote:

“Since we have the firm policy of steady price reduction, there is always pressure…we have never considered any costs as fixed. Therefore, we first reduce the price to a point where we believe more sales will result. Then we go ahead and try to make the price. We do not bother about the costs. The new price forces the costs down… Although one may calculate what a cost is,…no one knows what a cost ought to be. One of the ways of discovering what a cost ought to be is to name a price so low as to force everybody in the place to the highest point of efficiency. The low price makes everyone dig for profits. We make more discoveries concerning manufacturing and selling by this forced method than by any method of leisurely investigation… We have always made a profit at the prices we have fixed.”¹⁰

The results of this policy were reflected in the price of the Model T. In 1910 when 12,292 cars were made, the price was $950. By 1926 when 15,000,000 cars had been manufactured, the price was only $270. In the meantime, the wholesale price index had risen from 70 in 1910 to 100 in 1926, so that the price in constant dollars could be estimated at about $200. A line connecting these two points on a double logarithmic chart has a slope of about 86%, which is consistent with learning curve expectations for an operation which was largely assembly, especially since price is not the customary parameter on a learning curve plot.

The fact that production of the Model T began in 1909, more than a generation before the first learning curve article was published in 1936, underscores an observation once made by the philosopher, George Santayana: “Those who forget the past…are condemned to repeat it.”

Conclusion

Learning is a property of all living organisms. They can trace improvement patterns characteristic of themselves. Since organized groups can be looked upon as living entities, they can be expected to exhibit learning and to trace such patterns. In the aircraft industry, for example, they commonly do.

Such performance does not just happen. It is the result of continued seeking and resourceful striving. Study of a number of operations which are important components of major industries reveals that they have traced improvement patterns with learning curve characteristics.

In a way, such findings should not be surprising. Unremitting competition has provided a continuing incentive for companies to look for new and better ways of doing things, and the resulting progressive improvements are merely consistent with the common experience that a thing can always be done more efficiently each succeeding time by trying.

Nevertheless, discovering such performance for operations previously considered unresponsive does provide additional tangible evidence that learning can be an underlying, natural characteristic of organized activity. It does not merely extend the catalog of learning curves. Instead, it can help to breed the conviction that such performance should be found elsewhere, and thereby lead not only to scrutinizing all operations to see which additional ones are susceptible, but to assuming that all operations have learning curve potential and to devising ways of making this potential a reality. Thus, it is prudent to reflect learning potential in plans and forecasts.

The most important ingredients in learning curve performance are vision and leadership. Continued improvement is a chain of influences which starts with the conviction that progress is possible, continues with the creation of an environment and support of work which promote it, and results in a flexibility and willingness to change established practices for more efficient ones as they continually evolve. Furthering this chain is part of the practice of management. Consequently, the learning curve can be regarded as a primary tool of management.

1. Miguel A. Reguero, An Economic Study of the Military Airframe Industry, Wright-Patterson Air Force Base, Ohio, Department of the Air Force, October 1957, p. 213.

2. See Frank J. Andress, “The Learning Curve as a Production Tool,” HBR January–February 1954, p. 96.

3. Ibid.

4. T. P. Wright, “Factors Affecting the Cost of Airplanes,” Journal of Aeronautical Science, February 1936, pp. 122–128.

5. Oil and Gas Journal, published monthly.

6. For a discussion of these studies; see “The Growth Force That Can’t Be Overlooked,” Business Week, August 6, 1960, p. 68.

7. Richard W. Conway and Andrew Schultz, Jr., “The Manufacturing Progress Function,” Journal of Industrial Engineering, January–February 1959, p. 48.

8. Glen E. Ghormley, “The Learning Curve,” Western Industry, September 1952, p. 34.

9. Philip S. Skaff, “The Maintenance Challenge in a Petrochemicals Plant,” ASME Petroleum Mechanical Engineering Conference, Dallas, Texas, September 1956, Paper 56-PET-2, p. 9.

10. My Life and Work (New York, Doubleday, Page and Company, 1932), pp. 146–147.

A version of this article appeared in the January 1964 issue of Harvard Business Review.