RESEARCH DOSSIER

Learning and Dynamic Firm Productivity

Abstract. This research dossier examines how firms acquire, retain, transfer, and lose productive capability through production experience and deliberate changes to business processes. It integrates traditional experience curves, dynamic producer optimization, productivity-state transitions, organizational knowledge, and empirical identification. Historical and international evidence from manufacturing, contracting, training, exports, management, and digital assistance is compared according to the outcome measured and the design used. Conditional derivations distinguish scale effects from cumulative learning, technical efficiency from revenue productivity, and within-firm improvement from survival selection. Original analytical synthesis identifies the conditions under which experience has lasting economic value. The study is an evidence-based narrative monograph with illustrative calculations; it introduces no new named economic theory and estimates no new microdata model.

Keywords: Learning by doing; firm productivity; experience curves; organizational learning; knowledge depreciation; process innovation; production functions; dynamic optimization; causal identification; firm selection; productivity persistence; JEL D21; D22

The scientific object of dynamic firm productivity

The productive capacity of a firm changes as its members carry out tasks, encounter errors, modify methods, acquire information and reorganize the resources they control. The economic question is how those changes alter the feasible transformation of inputs into outputs over time, how much of the resulting change can be attributed to experience, and which mechanisms preserve or destroy it. Firm productivity is dynamic when today's production decisions influence tomorrow's productive opportunities. A static cost function, evaluated at a fixed technology, is one component of this analysis; the evolution of that technology supplies the other.

Learning by doing designates productive improvement caused by experience in performing economically relevant activities. Productivity persistence designates serial dependence in measured or latent efficiency; it describes a pattern rather than identifying a cause. Process innovation designates implemented changes in methods or business functions. All three may interact, although each admits independent observation. A firm may learn a better sequence without introducing a survey-classified innovation; an innovation may come from purchased machinery without substantial internal learning; measured productivity may persist because of permanent location advantages.

The unit of analysis is principally the operating firm, with plants, tasks, teams, products and contracting relationships examined when the economic mechanism requires finer resolution. This dossier explains internal producer optimization and productivity evolution. Industry equilibrium appears only where selection, demand and appropriation affect inference about the firm. The object is narrower than a general history of technical progress and broader than fitting a declining unit-cost curve.

Historical origins of the experience curve

Industrial measurement preceded a general economic account of experiential technical change. Wright's (1936) aircraft cost study examined production quantity and cost using observations accumulated in engineering practice. His charts distinguish several notions of cost and recognize weaknesses in accounting information for individual aircraft. This is already a warning against treating every reported experience curve as the same statistical object.

The historical significance lies in the change of explanatory variable: the position of a unit within a series carries information beyond contemporaneous inputs. The tenth and hundredth units may employ the same formal blueprint and nevertheless require different labor effort. Design continuity provides a potential comparison, while the industrial context introduces batch size, investment, supplier changes and scheduling.

Arrow (1962) brought this idea into economic theory through learning by doing, relating accumulated economic activity to the development of technical knowledge. His treatment introduces a divergence between private and social returns to investment under the specified spillover mechanism. The transition from engineering cost forecasting to economic analysis is therefore a transition from describing a production regularity to examining its source, consequences and allocation. A firm-level analysis must decide which experience is appropriable, which is transferable, and which is already reflected in externally available technologies.

Subsequent empirical work increasingly exploits production records and designs that separate knowledge from selection and input choice. This historical movement has not produced a single universal learning coefficient. It has produced an increasingly precise set of questions about the organization of production.

Productivity, cost, quality, and the production boundary

Let physical output in period t be Q_t, labor hours H_t, productive capital services K_t, and intermediate inputs M_t. Labor productivity is Q_t/H_t. A conventional gross-output production representation is Q_t = A_t F(K_t,H_t,M_t). If F is correctly specified and quantities are measured consistently, A_t represents output relative to the specified input bundle. It is a residual of a defined representation, not a direct instrument reading of knowledge.

A firm may increase output per labor hour because it replaces labor with capital. Under a Cobb-Douglas example, Q = A K^alpha H^(1-alpha), log(Q/H) = log A + alpha log(K/H). Capital deepening raises labor productivity even with A fixed. If quality improves, the relevant Q may also change when physical unit counts remain constant. A cost decline can reflect input price changes, lower defects, scale or substitution as well as learning.

Output and input boundaries must be matched. Gross output includes materials consumed in production; a value-added specification aims to exclude intermediate consumption and requires additional assumptions for a valid production representation. Outsourcing may lower recorded labor hours while moving work outside the firm. Comparing two accounting periods without adjusting the boundary can attribute such a transfer to internal efficiency.

Foster, Haltiwanger, and Syverson (2008) use separate quantities and prices to illuminate the distinction between physical and revenue measures. Their evidence concerns particular industries with usable data. Economically, an increase in expected demand may increase revenues and survival without the production frontier shifting. A defensible productivity claim identifies output, inputs, deflator, quality convention, unit, and time interval before interpreting a residual.

Learning as a change in feasible production

Consider a firm that can perform a recurring task using a fixed physical installation. In early periods, the operation requires experimentation, checking and correction. Later periods may be able to reproduce the same output with fewer errors and less effort. The feasible set has expanded if the firm can produce at least the former output with weakly fewer inputs while retaining its alternatives, conditional on the relevant knowledge and constraints.

Learning can modify the production function in several ways. A neutral improvement scales achievable output for all input combinations in a specified domain. A labor-augmenting change makes some labor tasks more effective. A materials-saving change reduces scrap or rework. A reliability improvement changes the distribution of good output rather than the engineering maximum. These changes imply different substitution and cost responses, so a scalar efficiency factor is best understood as an approximation.

A useful empirical distinction is between movement toward an existing frontier and a shift of the frontier available to the operating unit. If a worker discovers a procedure already known to another plant, the firm may move toward an industry benchmark while its own feasible techniques expand. Frontier terminology depends on which population defines the benchmark. For a new plant using a mature technology, catch-up may be the major source of dynamic improvement.

Knowledge acquisition costs matter. Time spent experimenting is an input, even when recorded as ordinary labor. A valid intertemporal comparison includes the resources consumed during the unsuccessful trials that made later success possible. A firm cannot infer a net economic return merely from the lower cost of its final successful operation.

The conventional cumulative experience function

The standard power-law representation for a unit cost or labor requirement at experience E > 0 is c(E) = c_0 (E/E_0)^(-b), with c_0 > 0, E_0 > 0 and b nonnegative. Holding the specification and input-price convention fixed, the ratio when experience doubles is c(2E)/c(E) = 2^(-b). A progress ratio of 0.80 corresponds to b = -log(0.80)/log(2), approximately 0.322: costs decline by 20% on doubling, not by 80%.

For an illustrative calculation, set c_0 = 100 labor-hours at E_0 = 1 and posit an 80% progress ratio. At experience levels 2, 4 and 8 the corresponding requirements are 80, 64 and 51.2 hours. These are mathematical examples, not aircraft industry estimates. The assumed relation implies that the absolute savings of subsequent doublings become smaller even though the proportional change remains constant.

An individual-unit specification differs from a cumulative-average specification. The average through N units is the sum of the unit requirements divided by N. If c(n) = c_0 n^(-b) with 0 < b < 1, then for large N the average is approximately c_0 N^(-b)/(1-b) by integral approximation. The asymptotic exponent is the same, but the level differs; an average of the first units cannot be used as the requirement of the next unit.

Power laws are useful summaries for a range of observed experience. An asymptotic prediction of zero cost is inappropriate for production with essential positive input requirements. A positive floor, a bounded technological regime or a piecewise relation may provide a more credible extrapolation. The fitted slope should not be interpreted as a causal parameter until competing explanations receive separate treatment.

Flow, stock, and depreciation of productive experience

Cumulative output gives each past unit permanent weight. Effective experience may instead diminish when routines are unused, workers depart, equipment changes or tacit interfaces deteriorate. A conventional stock representation is E_(t+1) = (1-delta) E_t + q_t, where q_t is current experience-producing activity and delta is a depreciation fraction between zero and one. Iterating yields E_t = (1-delta)^t E_0 plus the discounted sum of previous activity. The expression is a standard accounting representation, not an estimate of a particular industry's forgetting rate.

If q_t is held at q and delta is strictly positive, the stock converges to q/delta. The convergence is monotonic from below when the initial stock lies below the steady state. At zero production, the stock falls at geometric rate. A hypothetical stock of 100 with delta = 0.10 becomes 81 after two inactive periods. If c is strictly decreasing in E, the model predicts a restart cost associated with the interruption, even when lifetime output is greater than before.

Benkard's (2000) aircraft production analysis is an empirical motivation for replacing immutable lifetime experience with a stock capable of declining. The result cannot by itself distinguish every source of loss. Reduced capacity use, changes in workforce, model switching and altered supplier reliability may affect restart costs independently.

Depreciation may consequently be a function of the type of knowledge and the environment. Explicit instructions and reusable fixtures can survive inactive periods; coordination may require repeated joint performance. Researchers need data on shutdown length, workforce continuity, product variants and capital changes to interpret estimated memory decay. A single depreciation parameter compresses many possible mechanisms and is valuable only within the boundaries of its application.

Scale economies and the temporal structure of production

Increasing returns within a production period mean that output responds more than proportionately to an expansion of all variable inputs at a given state of technology. Learning concerns the effect of previous activity on current productive capacity. A firm may display either or both. The distinction is especially important in a rapidly growing plant, because current output and accumulated output are then closely correlated.

Suppose observed average cost is c(q,E) = a + f/q + d E^(-b). The f/q term represents fixed cost spreading; the last term represents a stylized experience effect. A rising order book lowers cost through both channels when q and E grow together. A regression using only log E generally conflates them. A sudden change in orders that moves current capacity utilization while inherited experience stays fixed could help identify the scale component if the demand change has no contemporaneous direct effect on efficiency or quality.

Batch economies introduce another layer. Setups and changeovers are incurred for production runs; a larger batch spreads those costs. Repetition may also teach operators to shorten the setup itself. Observing lower setup time per batch after cumulative setups increases is more informative about task learning than observing lower average production cost per unit.

A sequence of throughput, shifts, staffing, downtime and product-mix records supports a richer separation of these mechanisms than aggregate monthly accounts. The investigator should ask whether two observations occur at comparable load, whether inputs are measured per scheduled or active hour, and whether utilization constraints changed over time. No general monotonicity follows for unit costs when orders fluctuate.

Individual skill and organizational capability

A worker can acquire dexterity, diagnostic knowledge and judgment. The organization can retain separate knowledge in instructions, software, quality-control routines, tooling designs and coordination practices. The two forms of retention are complementary: experienced workers generate improvements, and documented routines can make the improvements available to a successor.

The manufacturing learning review of Argote and Epple (1990) identifies variation in learning rates and describes mechanisms that include employee turnover and knowledge transfer. Empirically, the productivity of a plant that continues to perform well after personnel changes supplies evidence consistent with organizational retention. It does not identify precisely which repository of knowledge is responsible. Selection of equally skilled replacements or changes in capital can produce similar observations.

Task assignment is also endogenous. More experienced workers may be assigned more difficult jobs; a raw comparison of their completion times can understate acquired ability. Conversely, managers may allocate easy tasks to new workers, masking early skill gaps. A study needs standardized task characteristics, preassignment information and, where possible, repeated observations across workers and teams.

When routines embody know-how, worker-level and firm-level productivity effects have different lifetimes. A portable individual skill can leave the firm; an efficient procedure can remain in the organization even if its inventor leaves. Economic incentives determine whether a worker shares discoveries, whether the firm invests in documentation and whether knowledge is accessible under time pressure. The productivity effect is a property of the organization of experience as well as the amount of experience.

Learning under uncertainty and learning about ability

Two distinct state variables can change with experience: actual productive capability and beliefs about capability. The distinction is fundamental for firm dynamics. A firm that learns about a previously unknown production cost adjusts output, investment and survival decisions, even when the underlying technology remains unchanged.

In Jovanovic's (1982) selection model, operating firms receive evidence about their efficiency and update their expectations. The resulting variation in growth and exit arises from incomplete information and selection. This mechanism explains why a cross section of surviving older firms may look more productive even if age has no positive causal effect on underlying efficiency.

For a simple textbook signal example, let an unknown fixed cost parameter theta have a normal prior with mean m and variance v, and let a signal x = theta + epsilon have independent normal error variance s. The posterior mean is m + v/(v+s), and the posterior variance is vs/(v+s). Repeated observations reduce uncertainty. The improved decision can lower expected losses without theta itself changing. This illustrative calculation uses a conventional Bayesian update rather than reproducing the complete industry model.

Actual learning by doing can coexist with this discovery process. A producer may become better at its operations while discovering where its comparative advantages lie. A dynamic empirical design should avoid using a single age variable to stand for both mechanisms. Repeated standardized performance tests, observed changes to procedures and the behavior of unsuccessful firms provide possible discriminating evidence.

The intertemporal production decision

The firm chooses current output and resources in the presence of future consequences. Consider the conventional discounted optimization problem V(E,z) = max_q {R(q,z) - C(q,E,z) + beta E_z'[V((1-delta)E+q,z')]}, where beta lies between zero and one and z contains other relevant states. The maximization is subject to a feasible output set, financing and all applicable constraints. The expectation concerns the next-period random state. This is an illustrative application of standard dynamic optimization.

For a differentiable interior optimum with a transition independent of q except through experience, the first-order condition is R_q - C_q + beta E_z'[V_E] = 0. The final term is the marginal continuation value of experience generated by output. If learning lowers future cost and is retained, V_E can be positive. Then current optimal output may exceed the output chosen by a comparable firm that ignores future experience benefits.

The argument assumes that activity generating experience has the same unit as q and that the firm appropriates its continuation value. Learning may instead depend on experiments, setups, defects or task diversity. Financing limitations may prevent the firm from exploiting future gains; expected exit can shorten the relevant horizon. A positive continuation value does not establish that producing at a loss is optimal without comparing the magnitude of current losses with future returns and alternatives.

In the particular example of two dates, let an extra unit produced today cost 5 more than its revenue and yield a known future saving of 8. At beta = 0.9, the discounted saving is 7.2, and the incremental surplus is 2.2, absent other consequences. If the knowledge is lost with probability one half, expected discounted saving falls to 3.6 and the activity is no longer profitable on these assumptions. Persistence thus directly changes the economics of developing capability.

Persistence, shocks, and the distinction between causes

A frequently used empirical representation for log productivity omega_t is omega_(t+1) = a + rho omega_t + gamma x_t + u_(t+1), where x_t represents one or more predetermined productivity-enhancing activities. A stationary autoregression with constant a and |rho| < 1 has a long-run mean a/(1-rho) when x is zero and innovations have mean zero. A one-time positive shock of d to omega has an expected contribution rho^h d after h further periods under the stated assumptions.

Persistent productivity can be generated by durable knowledge, physical investments, management, local infrastructure or persistent measurement error. If the observed variable is a revenue residual, product demand and pricing can also induce persistence. An estimated rho is therefore a persistence coefficient conditional on the specification, not a direct estimate of the physical survival of knowledge. Comparing it with delta from the experience-stock equation requires a mapping from E to omega and an explicit shock process.

For an illustrative rho = 0.8, a one-unit log-state impulse has half its initial expected size after log(0.5)/log(0.8), approximately 3.11 periods. A model with rho close to one has a long impulse horizon, but finite samples may distinguish near-unit-root persistence from permanent differences poorly. Firm fixed effects remove time-invariant levels but can bias short dynamic panels when lagged dependent variables and transformed errors are correlated.

The relevant empirical question is which productivity component remains once temporary inputs, task difficulty, demand, prices and other covariates are considered. Evidence of persistence should be accompanied by a clearly stated time unit, lag structure and uncertainty interval rather than interpreted as a universal property of firm memory.

Process innovation and deliberate investment in knowledge

Experience-driven improvement can arise within normal production. Deliberate process innovation uses resources to create and implement different procedures, control systems or organizational arrangements. The Oslo Manual business definitions (OECD and Eurostat, 2018) specify implementation and a substantial difference from the firm's previous process. The statistical definition covers more than factory technology. An idea recorded in a design file, without application, does not yet satisfy the implementation condition.

Firm-level R&D spending is an input into a discovery and development process. The output of that process is uncertain. In endogenous productivity research, Doraszelski and Jaumandreu (2013) allow R&D to influence the transition of latent productivity. Their Spanish manufacturing application examines the 1990s; the linked January 2013 author version reports more than 1,800 firms in nine industries. The estimates are dependent on the model and sample rather than direct measurements of every firm's discoveries.

A persistent gain can require successive complementary investments: changing equipment, revising instructions, training personnel and collecting diagnostic data. Individual activities may initially appear unproductive when accounting recognizes their costs immediately while benefits occur later. A process may also become obsolete before the investment is recovered.

The economic return to innovation depends on the relevant counterfactual. The benchmark is the feasible path under the best available alternative, including ordinary learning or adoption of outside methods. Evaluating a new technique against a frozen preinvestment baseline exaggerates its incremental return if the old method would have improved anyway. Measuring both implementation and postimplementation performance helps identify when the productivity transition occurred.

Transfer within and between firms

Knowledge may be linked to a product, a particular site, a team, a supplier relationship or a common pool of technical information. Transfer changes the initial conditions of another activity. A receiving unit that adopts a tested procedure can avoid a portion of the experiments originally required to discover it, although adaptation and validation may still impose costs.

Internal transfer can be examined by comparing adoption across otherwise similar plants using the same equipment and specifications. If a later-starting plant begins with a lower defect rate than an earlier one, this is consistent with transfer, but new equipment, worker selection or improved suppliers may also explain the gap. Evidence becomes more persuasive when an identifiable procedural innovation moves between locations, the relevant performance measure changes afterward, and alternative changes are accounted for.

External knowledge can be acquired through labor mobility, suppliers, consulting, licensing and observation. The recipient benefits only when the imported information applies to its feasible methods. A sophisticated technique may lower theoretical unit requirements yet prove costly to implement with unavailable complementary resources.

Arrow's private and social returns (1962) highlight a condition under which knowledge generated by investment benefits other producers. If the innovating firm cannot appropriate the benefits, privately chosen activity may fall below a social optimum given the model's assumptions. In a firm-specific setting, retention and contracts determine appropriation. The existence of learning therefore does not establish the existence or magnitude of an externality; spillovers require a separate evidential argument.

Contracting relationships and joint experience

A production chain can possess capabilities that arise from repeated interaction between legally independent firms. A supplier may learn a buyer's tolerances; a contractor may learn the client's scheduling priorities; both can improve coordination without changing their general technical competence. The relevant experience stock belongs partly to their match.

Kellogg (2011) studies this question through learning by drilling in the Texas oil industry. The paper constructs detailed drilling records and examines experience accumulated jointly by producers and drilling contractors. Its reported results support a relationship-specific component that is imperfectly transferable to new partners. The setting supplies unusually concrete measures of task completion, while geology, match formation and assignment require empirical controls.

Consider a firm's total experience across all jobs and its pair-specific experience with a given contractor. If costs fall more with the latter after accounting for the former, the pattern is consistent with match-specific knowledge. Causality depends on ruling out selective rematching after shocks and on whether persistent match quality has been accounted for. Pair fixed effects absorb stable pair attributes but do not automatically solve endogenous retention.

Contractual implications follow conditionally. Stable relationships may preserve valuable joint know-how. A policy of frequent competitive reallocation may reduce prices but impose learning losses, while long commitments may weaken incentives or restrict better alternatives. Switching costs include loss of relational productivity as well as transactional costs. This is a conventional comparative-static implication of costly transfer, not a general prescription for contractual duration.

Measurement of tasks, defects, and quality

Counts of finished units are insufficient when the quality distribution changes. A production system produces accepted output, rejected output and partially completed work. Let q_g be accepted units and h total hours including rework. If the initial process makes 100 units with 10 rejects in 100 hours and a later process makes 100 units with 2 rejects in the same hours, accepted-unit productivity rises from 0.90 to 0.98 per hour, about 8.89%. Gross counts do not change. The example is hypothetical and assumes that accepted units have common quality.

More generally, define quality-adjusted output as the sum of product quantities weighted by a specification-based index. Choice of weights matters: transaction prices combine quality, cost, bargaining and demand. Engineering tolerances, reliability and standardized adjudication may provide alternatives. For service activities, customer satisfaction, repeat contacts, resolution rates and error severity can supplement speed.

When quality and quantity trade off, faster processing can worsen expected economic outcomes. A higher rejection rate creates future customer costs and may shift liability outside the measured production boundary. The Egyptian export experiment (Atkin et al., 2017) is a concrete reminder that raw physical throughput and productive capability can move differently when specifications change.

Researchers should document whether defective output is counted before or after rework, whether quality checks changed, and whether monitoring intensity itself altered recorded failure rates. An improved inspection system can temporarily increase recorded defects by detecting previously unnoticed problems. The interpretation requires a measurement model for the observed indicator as well as a theory of the production process.

Revenue productivity, prices, and distortions

Revenue productivity based on deflated sales is often necessary in administrative data. If R_i = p_i Q_i and all sales are deflated by a common industry price P, the measured output is p_i Q_i/P. Relative prices remain in this measure. Under a known production function, the log residual equals log physical productivity plus log(p_i/P), conditional on quantities and the simplifying assumptions.

For an example, take two firms with identical physical input bundles and outputs of 100 units. One sells at 10 per unit and the other at 12. Their revenue residuals differ by log(1.2) even though their physical productivity is equal. Conversely, a technically efficient producer may charge a lower price to capture volume. A higher revenue residual need not correspond to lower resource use.

Foster et al. (2008) provide evidence that productivity and profitability selection differ when producer prices and quantities can be separately observed. This finding motivates a distinction between the performance state relevant to technical efficiency and the state relevant to survival. Firm-level financial accounts remain valuable, provided that the interpretation acknowledges their content.

A revenue-based learning curve can reflect customer relationships, premium quality, market expansion and changing markups. Comparing prices within narrowly comparable product specifications, exploiting physical quantities where available and estimating demand jointly with production can reduce ambiguity. Demand outcomes are economically real, yet their value is distinct from a pure gain in technical productivity. An analytical dossier must avoid replacing the whole economic performance of a firm with a single technical metric.

Production-function estimation and simultaneity

Inputs respond to information held by the firm. Suppose log output is y_it = alpha k_it + beta l_it + gamma m_it + omega_it + epsilon_it. If managers observe omega_it when selecting flexible inputs and the investigator does not, the input choices correlate with the composite regression disturbance. Ordinary least squares then generally fails to recover the production elasticities under the maintained structural interpretation.

Proxy-variable estimators exploit decisions over investment or intermediates to control for latent productivity under monotonicity and timing assumptions. Ackerberg, Caves, and Frazer (2015) analyze production-function identification and show why apparently informative proxy procedures can fail to identify certain elasticities given the assumed functional dependence. The point is methodological: a more complicated estimator is useful when its identifying variation is credible.

If productivity evolves in response to R&D or exports, imposing an exogenous productivity transition may further distort the analysis. A researcher can allow the transition to depend on such activities, but the resulting coefficient is causal only under restrictions on selection, timing, innovations and the information set. Endogenous learning is part of the model, not a property that estimation automatically proves.

Measurement errors in capital, hours or prices may be persistent. A technically plausible production elasticity does not certify a valid productivity residual. Robustness includes alternative input timing assumptions, multiple functional forms, sector-specific material elasticities where support permits, and comparisons with more direct engineering outcomes. The technical design should be reported sufficiently to distinguish structural assumptions from empirical results.

Research design for causal attribution

A causal learning claim concerns a difference between feasible productivity outcomes under different histories of relevant experience, with the comparison holding other pathways appropriately fixed. The observed association between cumulative production and productivity is an insufficient identification strategy because successful firms ordinarily produce more and may receive better orders. Reverse causality can produce the same sign as learning.

Randomly assigned training, access to orders or deployment timing supplies useful variation if the assignment affects the intended outcome through a documented mechanism. An intention-to-treat estimate compares outcomes by assignment. A treatment-on-treated interpretation requires additional assumptions when participation differs from assignment. When treatment affects several mechanisms, experimental identification of the total treatment effect does not identify each mechanism separately.

Natural experiments may exploit interruptions, production allocations or externally imposed schedules. An instrument for output requires an exclusion restriction: a demand shock that also changes product specification, utilization or investment does not identify pure experience without additional structure. A difference-in-differences approach requires a credible counterfactual trend and attention to anticipation, varying treatment effects, clustered shocks and interference between units.

One practical design combines an event study for productivity before and after an activity change with direct records of procedures, inputs and product specifications. A common-task test after both groups receive identical conditions helps distinguish transferable knowledge from changed production composition. The appropriate comparison period must be chosen before seeing the desired outcome. Reporting uncertainty and all relevant specifications allows readers to assess the strength of the causal link.

Survival, entry, exit, and cohort composition

Firm productivity is often observed only for surviving firms. If low-efficiency producers are more likely to exit, the average productivity of a cohort increases through selection. Hopenhayn's (1992) dynamic industry model treats entry, exit and growth within competitive equilibrium. Jovanovic's information mechanism offers a complementary account of attrition. Neither requires every incumbent to acquire new production knowledge.

For a numerical example, two equally weighted firms begin with fixed productivity levels of 1 and 2. The initial mean is 1.5. If the first exits and the second remains unchanged, the surviving mean becomes 2, a one-third increase over the initial mean. No individual firm has improved. Alternatively, if both survive and improve by 0.5, the mean also becomes 2. The observed mean is compatible with different economic histories.

For a balanced panel, within-firm changes provide a direct descriptive comparison for survivors. However, conditioning on survival can itself be a selection mechanism if treatment affects exit. A complete cohort analysis should seek data on failures, mergers, acquisitions and inactivity, distinguishing genuine closure from changes in identifiers. Bounds may be preferable to an assumption that missing firms evolved like survivors.

At the industry level, aggregate productivity depends on both changing firm-level capability and redistribution of output or employment. The weights used to compute an aggregate must correspond to the productivity measure. A decomposition into within, between, entry and exit terms is descriptive unless linked to a credible causal design. The firm's own learning remains one component of a wider evolutionary process.

Aircraft production and the limits of monotonic learning

Aircraft manufacture offers long production sequences, extensive task specialization and exposure to interruptions and model modifications. Wright's historical observations motivated cost forecasting. Benkard's evidence demonstrates why cumulative unit counts alone may not explain the entire subsequent trajectory.

An aircraft plant's productive state contains tooling, design-specific knowledge, supplier maturity, worker proficiency and sequencing routines. A period of reduced output may erode proficiency or change the composition of the team. When activity resumes, a model in which all previous units have equal permanent weight forecasts a much smoother trajectory than a depreciation-based account.

The appropriate empirical comparison includes aircraft type, variant, production rate, installed tooling, workforce and purchased components. Major redesigns alter the meaning of comparable units. An apparent loss of experience can arise because a new variant has a higher inherent task requirement. Conversely, transfer from earlier variants may give the new product an initial advantage that does not appear in its own cumulative output.

The historical evidence supports modeling experience as a state with possible loss and transfer. It does not make a particular learning or forgetting rate invariant across programs. For long-run planning, uncertainty in delivery schedules matters because it changes the path of experience as well as revenue. This is a direct consequence of intertemporal dependence in production costs.

Automobile assembly and distributed process learning

An automobile assembly line is composed of many sequential and partly interdependent operations. Defects at one station may create rework or disrupt downstream processes. The relationship between a plant's total experience and its aggregate efficiency can therefore be mediated by local discoveries and their propagation rather than a single worker-level learning function.

Levitt, List, and Syverson (2013) examine automobile assembly learning using detailed process information. Their reported evidence associates much of the retained knowledge with plant-level arrangements rather than solely with workers. The production setting permits analysis at a finer level than annual financial accounts and gives operational content to the idea of organizational learning.

Suppose a station has a processing time of 8 minutes while the next station requires 10. Reducing the first to 6 minutes will not increase steady-state line throughput if the second remains the bottleneck. It may nevertheless reduce local labor requirements or create buffers. The value of a local learning improvement depends on synchronization, bottlenecks and the ability to reallocate the saved resources.

A firm-level estimate can thus conceal large changes in the internal distribution of tasks. Process-level measurements help determine whether improvement is distributed, concentrated or constrained by a particular operation. A comprehensive evaluation tracks good units, defects, rework, downtime and resource use, including the cost of identifying and implementing improvements.

Management practices and operational discipline

Management can alter the return to existing resources through planning, inventory control, monitoring and problem resolution. A management intervention may introduce methods the firm could technically have adopted before, revealing a constraint in access to usable organizational information or in implementation.

Bloom et al. (2013) investigate management in Indian textile plants through randomized consulting. The final journal abstract reports a 17% first-year productivity improvement for the studied intervention, associated with quality, efficiency and inventory changes. This is the published estimate; a differently reported earlier working-paper estimate should not be substituted without identifying its version.

A practice can improve production only when it changes actual behavior. A checklist can be adopted symbolically while defects remain unchanged; effective monitoring can increase recorded defects before underlying failures decline. The distinction between the recorded adoption of a practice and use of that practice is central. A causal estimate for a consulting package includes the accompanying support and institutional conditions.

The evidence does not imply that every firm would obtain the reported percentage from the same package. Existing practice, managerial capacity, scale, trust, worker incentives and implementation costs influence gains. The research is especially valuable because management changes are documented alongside production outcomes. It supplies a credible example of organizational intervention, while leaving the precise contribution of each component and its long-horizon persistence as separate questions.

Export participation and learning from demanding customers

International orders can change the scale, quality and profitability of production. Firms that enter export markets may already be more productive, so a comparison between exporters and nonexporters typically includes selection. Foreign buyers can additionally deliver technical feedback, specifications or access to methods that change future production capabilities.

Atkin et al. (2017) experimentally varied access to export orders for Egyptian rug producers. The randomized market access study combines quality measurements and a common-product exercise with firm outcomes. The treatment group produced higher-quality goods; unadjusted output per hour could fall, while comparisons at matched specifications and common inputs supported a gain in productive knowledge.

This pattern is economically instructive. A buyer who requests a more difficult rug can increase the time required per square meter even when the producer becomes technically more capable. A measure that treats all square meters as interchangeable understates the change in skill. A common-product exercise reduces the influence of specification differences, although it cannot capture every element of the long-run value of the export relationship.

An export opportunity is a bundle of demand, communication, prices and potentially imported inputs. Identifying the causal effect of the offer is different from identifying the causal effect of the exact amount exported. Knowledge transfer should be traced through observable feedback and the tasks to which it applies. This case supports the learning mechanism in its documented population and motivates, rather than establishes, a general claim for all exporting firms.

Training, coworkers, and service production

Service productivity is often assessed through completion times, case resolution, customer outcomes and individual performance. These settings offer repeated tasks and extensive records, but task difficulty and quality can vary considerably. A new worker's faster resolution rate can reflect learning, increased assistance or changes in case allocation.

de Grip and Sauermann (2012) study training and coworker productivity using randomized assignment combined with worker panel data. Their published abstract reports a performance increase of 10% associated with participation and evidence of positive peer effects. The reported results illustrate both deliberate skill formation and indirect transmission through coworkers. The estimate is specific to the training content and operational setting.

When treated workers help untreated colleagues, the control group may receive benefits. This creates interference for a conventional trial that assumes independent potential outcomes. An experiment can explicitly vary the fraction treated within teams to estimate exposure effects, recognizing that the individual and collective interventions answer different questions.

A firm's return on training includes direct productivity, turnover, wage consequences and knowledge shared with coworkers. Workers may gain general skills that improve their outside opportunities. Training decisions also depend on the duration over which the firm can retain or access the skill. Measuring only the performance during the training period omits the investment cost and the subsequent retention interval.

Digital assistance and the measurement of durable learning

Digital tools can encode and deliver accumulated production knowledge. A worker may improve performance because a system provides a useful answer in real time, because the worker acquires transferable skill, or because the organization changes its task assignment. The economic effects of tool access and durable learning must be distinguished.

Brynjolfsson, Li, and Raymond (2025) study generative AI at work in a staggered deployment across 5,172 customer-support agents. The published article reports an average 15% gain in issues resolved per hour, with heterogeneous effects across prior skill and experience and evidence of learning. The final publication estimate is used here rather than earlier working-paper versions.

The appropriate counterfactual for durable learning is performance under standardized conditions after assistance is unavailable, or evidence that a worker applies the acquired knowledge in a distinct task. A tool outage is not automatically exogenous; workload and system reliability may coincide with other shocks. A design that varies access, measures learning separately and accounts for case mix is more informative.

The firm's economic return also includes software costs, integration, verification, errors and changes in demand. A gain in a task-level measure does not directly equal growth in firm-wide total factor productivity. The economic significance of assistance depends on whether the organization can redeploy saved time, improve quality, expand output or lower the required resource bundle without imposing unmeasured losses. Digital assistance broadens the available pathways for learning while retaining the same need for causal and measurement discipline.

Comparative evidence and international transportability

The empirical cases in this dossier cover aircraft and automobile manufacturing, Spanish manufacturing R&D, Indian textiles, Egyptian handicrafts, Texas drilling and customer-support operations. This is a purposive comparative corpus, not a global probability sample. Its value is mechanistic diversity. Its individual studies estimate different outcomes with different populations and treatment regimes.

Aircraft work highlights temporal loss of capability; automobile assembly illustrates organizational embodiment; drilling reveals the economic value of repeated relationships. The textile intervention highlights adoption of management practices; the export experiment demonstrates the role of quality and communication; training and digital assistance examine individual and shared skill formation. Spanish manufacturing evidence addresses deliberate knowledge investment under a model of dynamic productivity.

The studies cannot be pooled as though a production-hour gain, a management productivity estimate and a change in physical efficiency were estimates of the same parameter. They have different numerators, denominators, observation horizons and designs. A quantitative meta-analysis would require defensible harmonization and sampling rules and access to study-level uncertainty measures. No pooled learning rate is inferred here.

Transportability is a conditional claim about whether the relevant mechanism exists in a receiving population and whether constraints alter the magnitude of the response. Changes in labor turnover, infrastructure reliability, access to technical information, input quality, finance and management determine the applicability of a reported result. The strongest international conclusion is that multiple pathways through which activity changes productive capacity have been documented in distinct settings; their quantitative importance requires setting-specific evidence.

A data dictionary and practical empirical protocol

A credible firm panel begins by preserving stable identifiers for firms, establishments, teams and products; recording the date and duration of each observation; and distinguishing ownership changes from plant closures. Core outcomes include physical good output, rejected output, quality characteristics, real sales, hours, productive capital services, material quantities, and input prices. Production quantities should be recorded by homogeneous specification whenever possible.

Historical exposure variables include own task repetitions, cumulative units by product, cumulative setups, experience with each contractor or buyer, training attendance, inactivity duration and records of implemented process changes. Distinguish inherited experience from the interval in which output is measured to prevent a mechanical contemporaneous correlation. Record changes in personnel, shifts, equipment, supplier quality, management and product design.

A defensible primary outcome should be preselected according to the mechanism: good units per hour for homogeneous tasks, a standardized performance test for transferable skills, defects per comparable operation for reliability, or a carefully specified production residual for multidimensional output. Time windows should follow the expected persistence of the knowledge. Quarterly accounts alone often cannot distinguish restart losses from temporary utilization changes.

Statistical controls should address differences in unit definitions, missingness, measurement error, survivorship, sampling weights when required and dependence among observations. Confidence intervals should reflect the level at which exposure or treatment varies. Where records are confidential, a reproducibility plan should supply metadata, code and disclosure-safe summaries, without promising public access to protected microdata.

A prospective research report should separately identify the exposure variation, identifying assumptions, outcome construction, treatment implementation, estimated effect, diagnostics and sensitivity results. The recommended protocol organizes possible future investigations; no original microdata have been downloaded, and no new empirical estimation has been performed for this dossier.

Illustrative sensitivity and counterexamples

Consider a firm observed in two periods. Labor hours per accepted unit fall from 10 to 8 while capital services per accepted unit rise from 2 to 3. The 20% labor saving alone establishes a gain in labor productivity, not necessarily total resource productivity. If an hour of labor costs 10 and a unit of capital service costs 30 in both periods, the labor component falls from 100 to 80, but total measured cost rises from 160 to 170. At a capital-service price of 10, cost instead falls from 120 to 110. These illustrative calculations isolate price weights and substitution.

Another firm experiences a 25% decline in average cost from 20 to 15 as output doubles. If fixed cost is 100 and variable cost is constant at 10, moving from 10 to 20 units reproduces that entire change: 100/10+10 = 20 and 100/20+10 = 15. No learning is necessary. After controlling for current batch size, a decline in variable effort at comparable conditions would provide additional evidence.

Now consider a firm that improves each year but loses the workers responsible for its knowledge. Productivity after turnover can fall despite monotonically increasing lifetime output. An effective stock with depreciation accounts for that possibility, although it is not itself proof of the cause. If the fall is confined to the previous team's tasks, skill loss becomes more plausible than an economy-wide input-price shock.

These examples show why the robustness of a learning finding requires more than an alternative regression specification. The outcome needs an economic interpretation, and the chosen comparison must discriminate plausible mechanisms. Each example is constructed from stated assumptions and is not a report of observational evidence.

Original analytical synthesis

The comparative literature supports four interdependent distinctions that clarify empirical interpretation. The first is between performance and productive capability: recorded output, cost or revenue reflects the task, resource bundle, market and quality as well as the feasible production method. The second is between acquisition and retention: an activity can generate knowledge that later depreciates or remains accessible to the organization. The third is between individual and collective possession: the productivity gain may belong to the worker, plant, firm or relationship. The fourth is between production improvement and the changing composition of observed firms.

These distinctions yield a useful ordering of questions. Begin with the production boundary and economic outcome. Determine what experience plausibly changes within that boundary. Examine how the change is embodied, retained or transferred. Identify exposure variation sufficient to separate learning from improved demand, scale, choice of inputs and survivor composition. Only then evaluate its consequence for current and future resource use.

The central analytical result of this synthesis is that equal cumulative output need not confer equal productive capability. Two firms with identical lifetime production can differ in effective experience because their activities occurred at different times, their tasks differed, their staff and relationships persisted differently, or one received transferable knowledge from outside. Conversely, firms with different output histories may achieve similar capability through external transfer and complementary investment. Cumulative output is thus a potentially informative state proxy, not an invariant measure of productive knowledge.

A second result concerns intertemporal value. The value of experience is determined by its incremental expected future consequences relative to an attainable counterfactual. Identical immediate quality gains may have different value depending on persistence, scope of transfer and the ability to appropriate returns. Evaluating learning requires accounting for learning expenditures, unsuccessful trials and opportunities displaced. This principle follows from standard discounted producer optimization and established empirical distinctions. It is presented as an analytical synthesis; no claim is made to a newly established named theory or priority in the literature.

Finally, an evidence hierarchy is outcome-specific. A randomized training trial strongly identifies the effects of that intervention but need not locate the organization's long-run productivity state. A high-resolution process panel can reveal the location of retained knowledge but may lack randomized exposure. A structural panel estimator can incorporate endogenous productivity transitions but relies on the specification and identification restrictions. Convergence across such approaches is informative when they test complementary aspects of the same mechanism.

Limits of the evidence and directions for investigation

Important limitations concern the observational nature of several production studies, heterogeneity of reported outcomes and the absence of harmonized international firm-level data in this investigation. The study uses selected original research, author manuscripts and official abstracts; bibliographic details and the precise claims used here have been checked against the accessible records. Several journal full texts or visual PDF readbacks were inaccessible, so claims based only on abstracts are deliberately limited to their reported scope. Future empirical replication would require the underlying data and code, adequate permissions and an independent specification audit.

A particularly important research problem is the decomposition of persistent efficiency into internal knowledge, capital vintage, worker composition, relationships, management and external technology. Longitudinal data linking the same task across workers, locations and changes in tooling could distinguish these mechanisms. Natural changes in assignment are useful only if the resulting comparison has a credible identifying basis.

A second problem is the durability of improvements generated by external assistance. Measurements after temporary withdrawal, alternative assignments or changes in the tool can help distinguish assisted performance from acquired transferable competence. Researchers should also measure negative outcomes, including quality deterioration, unmeasured work, failed experiments and the cost of maintaining technical systems.

A third problem is the aggregation of firm learning in industries with entry and exit. Data on failures and the output shares of continuing producers can reveal the contribution of selection and reallocation, but assigning causes requires additional variation. International comparisons should respect institutional differences and measure the availability and effective uptake of production technologies.

These questions are genuine empirical research opportunities rather than unresolved assertions needed for the dossier's mathematical examples. The intended contribution is a disciplined synthesis of the documented mechanisms, their conditions and their consequences for a microeconomic account of dynamic productivity.

Conclusion

Firm productivity develops through experience, deliberate investment, transfer, changes to organizational routines and the allocation of productive resources. Learning by doing is one causal mechanism within that development. Its durable economic contribution depends on the nature of the acquired capability, the rate at which it is retained, the places and relationships to which it transfers, and the resources consumed in acquiring it.

Historical experience curves provide a valuable description of recurring industrial improvements. Dynamic representations extend them by incorporating depreciation and uncertainty. Empirical studies demonstrate that knowledge may be organizational, relational or individual, and that management, training, exports and digital assistance can alter productive performance under specified conditions. Evidence from each setting must be read in the units, institutional context and identification design of the original investigation.

The economic distinction between current output and inherited productive knowledge gives the firm an intertemporal decision problem. Experience can generate future value, while interruptions, personnel changes or obsolescence can diminish that value. A firm may also benefit from information about its own efficiency without altering its technology. The observed productivity of survivors cannot by itself distinguish these histories.

The strongest defensible conclusion is conditional and substantive: where a relevant activity produces retained or transferable improvements in a feasible production method, the economic effect of that activity must be assessed across the time span of its consequences. Sound measurement and causal attribution distinguish this effect from simultaneous changes in scale, input composition, quality, prices and selection. This provides a coherent foundation for the microeconomic study of learning and dynamic firm productivity.

References

Ackerberg, D. A., Caves, K., & Frazer, G. (2015). Identification properties of recent production function estimators. Econometrica, 83(6), 2411-2451. DOI: 10.3982/ECTA13408. Source record

Argote, L., & Epple, D. (1990). Learning curves in manufacturing. Science, 247(4945), 920-924. DOI: 10.1126/science.247.4945.920. Source record

Arrow, K. J. (1962). The economic implications of learning by doing. The Review of Economic Studies, 29(3), 155-173. DOI: 10.2307/2295952. Source record

Atkin, D., Khandelwal, A. K., & Osman, A. (2017). Exporting and firm performance: Evidence from a randomized experiment. The Quarterly Journal of Economics, 132(2), 551-615. DOI: 10.1093/qje/qjx002. Source record

Benkard, C. L. (2000). Learning and forgetting: The dynamics of aircraft production. American Economic Review, 90(4), 1034-1054. DOI: 10.1257/aer.90.4.1034. Source record

Bloom, N., Eifert, B., Mahajan, A., McKenzie, D., & Roberts, J. (2013). Does management matter? Evidence from India. The Quarterly Journal of Economics, 128(1), 1-51. DOI: 10.1093/qje/qjs044. Source record

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889-942. DOI: 10.1093/qje/qjae044. Source record

de Grip, A., & Sauermann, J. (2012). The effects of training on own and co-worker productivity: Evidence from a field experiment. The Economic Journal, 122(560), 376-399. DOI: 10.1111/j.1468-0297.2012.02500.x. Source record

Doraszelski, U., & Jaumandreu, J. (2013). R&D and productivity: Estimating endogenous productivity. The Review of Economic Studies, 80(4), 1338-1383. DOI: 10.1093/restud/rdt011. Source record

Foster, L., Haltiwanger, J., & Syverson, C. (2008). Reallocation, firm turnover, and efficiency: Selection on productivity or profitability? American Economic Review, 98(1), 394-425. DOI: 10.1257/aer.98.1.394. Source record

Hopenhayn, H. A. (1992). Entry, exit, and firm dynamics in long run equilibrium. Econometrica, 60(5), 1127-1150. Source record

Jovanovic, B. (1982). Selection and the evolution of industry. Econometrica, 50(3), 649-670. DOI: 10.2307/1912606. Source record

Kellogg, R. (2011). Learning by drilling: Interfirm learning and relationship persistence in the Texas oilpatch. The Quarterly Journal of Economics, 126(4), 1961-2004. DOI: 10.1093/qje/qjr039. Source record

Levitt, S. D., List, J. A., & Syverson, C. (2013). Toward an understanding of learning by doing: Evidence from an automobile assembly plant. Journal of Political Economy, 121(4), 643-681. DOI: 10.1086/671137. Source record

OECD & Eurostat. (2018). Oslo Manual 2018: Guidelines for collecting, reporting and using data on innovation (4th ed.). OECD Publishing; Eurostat. DOI: 10.1787/9789264304604-en. Source record

Wright, T. P. (1936). Factors affecting the cost of airplanes. Journal of the Aeronautical Sciences, 3(4), 122-128. DOI: 10.2514/8.155. Source record