CAF - Automation and the Margins of Firm and Worker Adjustment in Argentina de 2026

First version: July, 2026

This version: September, 2026

Small sections of text that are less than two paragraphs may be quoted without explicit permission as long as this document is acknowledged. Findings, interpretations and conclusions expressed in this publication are the sole responsibility of its author(s) and cannot be, in any way, attributed to CAF , its Executive Directors or the countries they represent. CAF does not guarantee the accuracy of the data included in this publication and is not, in any way, responsible for any consequences resulting from its use.

C A F - W O R K I N G P A P E R # 2 0 2 6 / 11

Automation and the Margins of Firm and Worker Adjustment in Argentina

Irene Brambilla | Octavio Bertín| Jessica Bracco| Guillermo Falcone | Azul Menduiña

Azul Menduiña: azulmenduina@gmail.com

CEDLAS-IIE-UNLP.

CONICET.

Irene Brambilla: irene.brambilla@gmail.com

Octavio Bertín: octaviobertin2001@gmail.com

Jessica Bracco: braccojessica@gmail.com

Guillermo Falcone: guillermofalcone@gmail.comC A F - W O R K I N G PA P E R # 2 0 2 6 / 1 1

T h i s v e r s i o n : S e p t e m b e r 1 1 , 2 0 2 6

Automation and the Margins of Firm and Worker Adjustment in Argentina Irene Brambilla1,2 Octavio Bertín1 Jessica Bracco1 Guillermo Falcone1 Azul Menduiña1 Abstract We examine how automation reshapes firm labor demand and incumbent-worker trajectories in Argentina over 2007–2022. We link customs records on automation-capital imports to administrative employer–employee and simplified self-employment tax data, identify adoption as discrete import spikes, and estimate stacked event studies comparing adopters and their incumbents with not-yet-adopting firms. Five years after adoption, employment is 19 percent lower and the wage bill 20 percent lower, while firm-level wages barely move; the contraction runs through hiring rather than dismissals. Incumbents’ reported earnings stay flat, but their formal attachment weakens—fewer months in registered wage employment, more exits from the registry, and more transitions into simplified self-employment. Automation costs here appear less as payroll wage cuts than as separations, reallocation, and movement out of registered wage employment.

JEL codes: J23, J31, J46, J63, O33

Keywords: automation, worker displacement, firm labor demand, formal employment, self-employment, Argentina

1CEDLAS-IIE-UNLP . 2CONICET.

Small sections of text that are less than two paragraphs may be quoted without explicit permission as long as this document is acknowledged. Findings, interpretations and conclusions expressed in this publication are the sole responsibility of its author(s) and cannot be, in any way, attributed to CAF, its Executive Directors or the countries they represent. CAF does not guarantee the accuracy of the data included in this publication and is not, in any way, responsible for any consequences resulting from its use. ©2022 Corporación Andina de Fomento

We thank Pablo Brassiolo and seminar participants at the CAF Workshop for helpful comments and suggestions. We gratefully acknowledge financial support from CAF – Development Bank of Latin America and the Caribbean – for this project. The views expressed herein are those of the authors and do not necessarily reflect those of CAF. All remaining errors are our own.C A F - D O C U M E N T O D E T R A B A J O # 2 0 2 6 / 1 1

E s t a v e r s i ó n : 1 1 d e s e p t i e m b r e d e 2 0 2 6

Automatización y los márgenes de ajuste de empresas y trabajadores en Argentina Irene Brambilla1,2 Octavio Bertín1 Jessica Bracco1 Guillermo Falcone1 Azul Menduiña1 Abstract Examinamos cómo la automatización transforma la demanda laboral de las empresas y las trayectorias de los trabajadores incumbentes en Argentina entre 2007 y 2022. Vinculamos registros aduaneros de importaciones de bienes de capital de automatización con datos administrativos de empleo registrado y del régimen simplificado de trabajo independiente (monotributo), identificamos la adopción como picos discretos de importación y estimamos estudios de eventos apilados. Cinco años después de la adopción, el empleo es 19 por ciento menor y la masa salarial 20 por ciento menor, mientras que los salarios apenas se mueven; la contracción opera vía contrataciones y no despidos. Los ingresos declarados de los incumbentes permanecen estables, pero su vínculo formal se debilita: menos meses en

el empleo registrado, más salidas del registro y más transiciones al trabajo independiente simplificado. Los costos de la automatización aparecen menos como recortes salariales y más como separaciones y reasignación.

JEL codes: J23, J31, J46, J63, O33

Keywords: automatización, desplazamiento de trabajadores, demanda laboral, empleo formal, trabajo independiente, Argentina

1CEDLAS-IIE-UNLP . 2CONICET.

Pequeñas secciones del texto, menores a dos párrafos, pueden ser citadas sin autorización explícita siempre que se cite el presente documento. Los resultados, interpretaciones y conclusiones expresados en esta publicación son de exclusiva responsabilidad de su(s) autor(es), y de ninguna manera pueden ser atribuidos a CAF, a los miembros de su Directorio Ejecutivo o a los países que ellos representan. CAF no garantiza la exactitud de los datos incluidos en esta publicación y no se hace responsable en ningún aspecto de las consecuencias que resulten de su utilización. ©2022 Corporación Andina de FomentoBRAMBILLA ET AL. 1 1|INTRODUCTION Automation can make firms more productive, more competitive, and in some cases larger. But it also reopens a fundamental question in labor economics: do workers at automating firms share in those gains, or do they bear the costs of adjustment? A large literature studies robots and automation at the industry, region, and firm levels (Acemoglu and Restrepo, 2020; Graetz and Michaels, 2018; Koch et al., 2021; Bonfiglioli et al., 2024). Yet the central welfare question is ultimately about workers. Even when automation improves firm performance, incumbent workers may not benefit proportionally; they may instead bear the costs through

separations, wage changes, or weaker labor-market attachment. The effect of automation on workers presents an additional layer of concern in developing economies. In those settings, unemployment and wage reductions are not the only margins of adjustment; instead, workers may also move into informality, low-productivity self-employment, or intermittent participation, often with limited income support during the transition. Existing evidence for Latin America suggests that adverse shocks are often absorbed through transitions across employment states, with informal employment acting as a buffer for workers leaving formal jobs (Brambilla et al., 2023; Dix-Carneiro and Kovak, 2019; Dix-Carneiro et al., 2021). But that margin of adjustment need not be benign: it may come with lower earnings, weaker protection, and a more fragile attachment to stable formal jobs. A separation induced by automation may therefore trigger not just a temporary earnings loss, but a persistent transition to low-quality work. In that sense, automation may have much sharper distributional consequences in economies where informality is pervasive and a central margin of adjustment. In this paper, we study how automation reshapes firm labor demand and worker trajectories in Argentina over 2007–2022. We combine three administrative sources: transactionlevel customs data on imports of automation-related capital goods, linked employer– employee records from the Sistema Integrado Previsional Argentino (SIPA), the country’s formal social-security registry, and records from AFIP , Argentina’s tax authority, on monotributo, the country’s simplified self-employment tax regime. This allows us to identify firms experiencing directly observed, import-based automation events, measure their pre-event incumbent workforce, and then follow those workers across all subsequent formal employers. Rather than relying on regional or local labor market exposure measures, this matched employer-employee approach enables us to study the dynamics at the firm

level, determining how automation affects employment relationships, reported earnings, worker reallocation, formal labor-market attachment, and subsequent transitions into selfemployment, non-employment, or informality. We define automation events as discrete spikes in firm imports of automation-related capital goods. In Argentina, where advanced automation equipment is primarily sourced from abroad, such spikes provide a useful observable proxy for large episodes of automation adoption. The capital goods that we consider are industrial robots, machining centers, numerically controlled machinery, automatic data-processing equipment, and industrial control instruments. We use a baseline definition of automation imports that includes all those categories and also a narrow definition based on industrial robots alone. Regardless of the particular definition used, an important caveat is that it captures investment in automation made by firms that import capital goods directly while it excludes firms that purchase automation equipment through integrators, distributors, or related firms (Bessen et al., 2025). Our event-study design, which we describe below, partly addresses this concern by comparing firms with different timing of observed direct-import adoption, rather than comparing direct importers with arbitrary non-importing firms. Indirect adoption before the observed spike remains possible, but if future direct importers have already begun adjustingBRAMBILLA ET AL. 2 through third-party adoption, their outcomes would tend to move before treatment and make our estimates more conservative. In the next sections, we discuss these measurement issues in detail, including the restrictions we impose to limit the risk of capturing imports intended for resale. Against this background, we interpret our estimates as the effects of directly observed, import-based automation events, not as effects for the full universe of automation investments. Our empirical strategy builds on a simple insight, also emphasized by Bessen et al. (2025): automation investment is lumpy. Firms do not adopt automation smoothly over time; they make large, discrete purchases of capital at identifiable moments. We therefore define an automation event as the first year in which a firm’s automation imports, measured relative

to its average wage bill, exceed three times their own average in other years, and we use the staggered timing of those first spikes in a stacked difference-in-differences event-study design (Cengiz et al., 2019; Wing et al., 2024). This shifts identification away from crosssectional differences between firms and toward the timing of discrete adoption episodes within a common event-time framework. Around each event, we observe outcomes for three years before the event and follow firms and incumbent workers for five years afterward. Given our 2007–2022 data coverage, this implies that the estimation sample includes events occurring between 2010 and 2016. At the worker level, we focus on workers with stable pre-event attachment to the firm and trace their trajectories across all subsequent formalsector employers. Because the design compares adopters with firms that adopt later, what identification requires is not that adoption be random but that itstimingbe unpredictable among eventual adopters. We show that it largely is: recent changes in employment, wages, exports, hiring, and separations have essentially no power to predict whether a firm’s first spike occurs this year rather than later. The results point to five facts. First, registered labor demand at automating firms contracts sharply. Employment and the wage bill move in parallel with those of later adopters before the event and begin to fall after the first post-event year; by event year +5, employment is 19 percent lower and the wage bill 20 percent lower than at comparable firms that adopt only later. Second, that contraction runs through a reduction in hiring rather than through dismissals. Separating the two gross flows, we find that hires fall by 21 points of pre-event employment by event year +5, while separations never rise significantly above the control path at any horizon. The gap opens because adopters stop replacing the workers who leave, not because they shed the workers they have—so a substantial part of

the incidence falls on the workers the firm never hires, who are invisible in any analysis based on incumbents alone. Third, firm-level wage outcomes barely change: average and median wages end the window less than one percent below the control path, are statistically indistinguishable from zero, and are much smaller than the 19 percent employment decline. The wage-bill contraction is thus a quantity adjustment, not a price adjustment. Fourth, for incumbent workers the reported-wage margin likewise shows no robust response: total annual earnings and average monthly wages, both across all registered wage jobs and at the original employer, remain flat and imprecisely estimated. Fifth, worker adjustment appears instead on the attachment margin. By the end of the window, incumbent workers spend 0.14 fewer months per year in registered wage employment and are 1.0 percentage point more likely to have left the registry—both distinguishable from zero only at the 10 percent level—while the sharpest and most precisely estimated margin is a 0.49 percentage-point rise in the probability of being observed only in the simplified self-employment regime. Scaled to a cohort of 1,000 incumbents, the estimates imply 25 additional departures from the original employer by event year +5, of which 15 move to another registered wage job and 10 leave formal wage employment altogether—half of them reappearing in simplified self-employment and half no longer observed in any of our records. These effects are modestBRAMBILLA ET AL. 3 in levels, but persistent: automation does not generate mass exit from the formal system, but it moves a small non-negligible share of incumbents away from standard registered wage employment. Three further findings qualify this average picture. First, the near-zero average earnings response is not uniform: incumbents of small adopters see annual earnings fall by 4 percent— the only group in our data with a clear earnings decline—while at large adopters adjustment

runs through separations and transitions into simplified self-employment instead. Reported earnings are nowhere the principal margin of adjustment; where they respond at all, it is in one corner of the firm-size distribution. Second, displacement is concentrated at the top of the pre-event wage distribution rather than the bottom: the probability of leaving the original employer rises monotonically across wage quartiles, reaching 6 percentage points in the highest, while incumbents in the lowest quartile show earnings gains. Third, the results are insensitive to the spike threshold that defines an event, but not to the technology that does: when we narrow the basket of automation goods to industrial robots alone, the worker-level displacement margins disappear and incumbent earnings drift upward instead. The robots-only estimates rest on far fewer adopters and are correspondingly imprecise, but the contrast suggests that automation is a bundle of technologies whose consequences for incumbents differ, and that the displacement we document is a property of broad automation adoption rather than of robot adoption narrowly defined. Our paper is most closely related to Bonfiglioli et al. (2024) and Bessen et al. (2025). Bonfiglioli et al. (2024) study the firm-level effects of robot imports and emphasize that demand shocks can make robot imports and employment positively correlated: expanding firms are more likely to automate, and they are also more likely to hire. This creates a selection problem in which raw adoption-employment correlations can understate displacement or even have the wrong sign. Our setting differs in a way that is useful for this concern. During much of our sample period, Argentine firms faced substantial frictions in importing capital goods and accessing credit, including foreign-exchange and import-authorization constraints. Because the automation capital we study is largely imported in Argentina, the exact timing of observed adoption is likely shaped not only by firms’ desired investment date but also by when they can finance and clear the import transaction. This does not

make adoption randomly assigned, but it weakens the simple story in which observed robot imports merely mark the peak of a firm-specific demand boom. Consistent with that interpretation, we observe no systematic employment growth at adopting firms before the automation event, and the post-event decline is not mechanically driven by mean reversion from an unusually high pre-period. Moreover, any remaining positive demand shock around adoption would work against finding an employment decline: it would raise employment at treated firms and partly offset the labor-saving effect of automation. The fact that we still estimate a large post-event contraction in registered labor demand is therefore hard to reconcile with a purely positive demand-driven adoption story. Bessen et al. (2025) provide the closest worker-level comparison. They link automating firms to individual workers in the Netherlands and show that automation increases separations and lowers cumulative wage income, mainly through fewer days worked. We build on their firm–worker perspective, but the Argentine setting changes both the observed margins and the interpretation of worker losses. Like them, we find weaker attachment among incumbents, both to the original employer and to registered employment more broadly, though the margins differ in precision: our estimate of the separation margin is sizeable but imprecise, and the sharpest signal is the move into simplified self-employment. Unlike in their Dutch evidence, however, reported wage earnings among workers who remain observed in registered wage employment do not show a comparable negative path. The adjustment margin we observe is instead a change in employment status: workers spendBRAMBILLA ET AL. 4 fewer months in registered wage employment, are more likely to exit the formal registry, are more likely to appear only in the simplified self-employment tax regime, 1 consistent with a shift toward weaker and potentially lower-quality formal attachment. This difference is informative. Argentina’s registered wage-employment sector in our sample period was characterized by collective bargaining, union wage schedules, wage floors, payroll taxes,

and regulated employment relationships. These institutions may make large worker-specific changes in reported wages less common or less visible in administrative annual earnings data. The point applies in both directions. For workers who remain at the automating firm, one might expect productivity gains, survivor selection, or retention motives to raise compensation. In a highly regulated payroll environment, however, such increases may not appear fully as registered wages if firms rely on nonreported bonuses or off-payroll payments. Evidence from Brazil is consistent with this margin: Feinmann (2024) survey formal employees and document widespread payments under the table, with 24 percent of formal employees reporting that they receive part of their pay off the books.2 We therefore interpret the small reported-wage response cautiously. It does not imply that compensation is irrelevant; it means that the margins we observe most clearly are employment quantities and labor-market states: separations, movement across employers, exits from registered wage employment, and transitions into the simplified self-employment regime. The paper also contributes to the literature on worker adjustment in developing economies and Latin America (Brambilla et al., 2023; Faber, 2020; Carbonero et al., 2023; Dix-Carneiro and Kovak, 2019; Dix-Carneiro et al., 2021). Existing evidence for the region has mostly relied on regional exposure measures and local labor-market outcomes. We instead observe firm-level adoption episodes directly and trace incumbent workers through the formal employer–employee registry and self-employment-tax records before and after those events. This provides a different lens on automation in Latin America: rather than inferring exposure from local labor-market composition, we can measure how workers attached to adopting firms fare inside and outside the standard registered wage-employment regime. It also links our results to evidence that informality and self-employment act as margins of adjustment in developing economies. In our setting, formal wage cuts are not the main

observed response; displacement appears through transitions out of the original employer, out of the formal registry, into the simplified self-employment regime, and possibly into unregistered work or non-employment for workers who disappear from our administrative records. The combination of these registries therefore lets us distinguish continued attachment to registered wage employment from a move toward a more weakly attached, non-wage, or unobserved labor-market state. Finally, the paper reconnects to the broader literature on technological change, robots, and automation. In the task-based tradition, Autor et al. (2003), Acemoglu and Autor (2011), Autor (2015), and Acemoglu and Restrepo (2018) emphasize how new technologies reshape tasks, labor demand, and wage inequality. A growing empirical literature studies these issues using industry, regional, plant, and firm-level data, including Graetz and Michaels (2018), Acemoglu and Restrepo (2020), Dauth et al. (2021), Cheng et al. (2019), Carbonero et al. (2023), Koch et al. (2021), Humlum (2022), Aghion et al. (2020), Bonfiglioli et al. (2024), and Bessen et al. (2025). A central lesson from the firm-level branch of this literature is 1This regime is not informal employment in a legal sense. For our purposes, it is informative because it captures a move away from the standard registered wage-employment relationship, while workers who exit the formal registry and do not appear in monotributo records may be unemployed, out of the labor force, or working in unregistered employment. 2We do not observe off-payroll compensation in the formal social-security records, and we do not use the Brazilian evidence as a direct estimate for Argentina. The point is narrower: in highly regulated Latin

American labor markets, reported payroll wages need not be the only margin of compensation adjustment, either upward or downward.BRAMBILLA ET AL. 5 that firms that automate differ systematically from firms that do not: automating firms are often larger, more productive, and on different trajectories even before adoption (Koch et al., 2021; Humlum, 2022). Measurement also varies across studies, including import records, survey indicators, third-party automation expenditures, and broader proxies based on industrial-equipment investment (Humlum, 2022; Koch et al., 2021; Bonfiglioli et al., 2024; Bessen et al., 2025; Aghion et al., 2020). Our contribution is not the use of trade records per se, but the combination of import-based adoption events with linked employer–employee data in a setting with limited domestic production of automation capital and where we can follow incumbent workers around those events. The remainder of the paper is organized as follows. Section 2 describes the data and the measurement of automation imports. Section 3 defines automation events in the data and presents descriptive evidence on automating firms and workers. Section 4 presents our empirical strategy. Section 5 reports the results, robustness checks, and heterogeneity analyses. Section 6 concludes. 2|DATA AND MEASUREMENT We use three linked administrative sources. Customs transaction records identify firmlevel imports of automation-related capital goods and cover 2007–2022. The Sistema Integrado Previsional Argentino (SIPA), Argentina’s formal social-security registry, provides linked employer–employee records on registered employment relationships and wages, also for 2007–2022. Records frommonotributo, the simplified tax regime for small-scale self-employment, identify workers who leave registered wage employment and register under this regime; these records are available for 2011–2018.

2.1|Customs Records and Automation Imports Our first source is Argentine customs data, which cover the universe of firm-level import transactions and report detailed product codes (HS 6–10 digits), declared values, quantities, and countries of origin. We match customs records to firms using the tax identifier (CUIT). As in Acemoglu and Restrepo (2022), we identify automation from the product composition of firms’ imports, but we do not adopt their basket unchanged. Their classification was built to compare aggregate technology adoption across countries; ours has to capture the automation-related capital that Argentine firms actually buy abroad. Four of their automation categories carry over directly: industrial robots, automatic machine tools, numerically controlled metalworking machinery, and automatic regulating instruments. We leave aside automatic welding machines, conveyors, and weaving, knitting and other dedicated textile machinery. We add programmable and numerical controllers, and computing and data-processing equipment. We order these goods from the narrowest to the broadest notion of automation, and carry that ordering through the paper. Definition 1 comprises industrial robots alone; Definition 2 adds machining centres; Definition 3 adds numerically controlled metalworking machinery; Definition 4 adds computing and data-processing equipment; and Definition 5, our baseline, adds automatic control instruments and programmable controllers. Table B.1 lists the codes in each group. Because the boundary of the basket is a judgement rather than a fact, Section 5.6.2 re-estimates every result under all five definitions. Figure 1 shows, by year, the number of Argentine firms importing automation-related goods and the total value of those imports. These series help characterize the prevalence and aggregate relevance of automation-related capital inflows in Argentina over 2007–2022.BRAMBILLA ET AL. 6 These automation importers are a small and distinctive slice of Argentine importing activity. Over 2007–2022, 16,016 firms imported at least one automation-related product, out of

65,244 firms that imported any good—roughly one in four importers—and the 8,503 firms we later classify as having an automation spike account for 13 percent of all importers. By value the share is much smaller: automation goods represent about 1.6 percent of total Argentine imports over the period. The margin we study is therefore not a large share of aggregate trade. It is a concentrated and lumpy investment made by a particular set of firms, and it is that firm-level lumpiness, rather than aggregate exposure, that our design exploits. (a) Firms importing automation goods (b) Total value of automation imports F I G U R E 1Automation-Related Imports in Argentina, 2007–2022

Notes: Panel (a) plots the number of formal-registry firms importing at least one automation-related product each year. Panel (b) plots the total real value of automation-related imports (in millions of real USD) by year.Source:Own computations based on Argentine customs records.

While our measure of the technology shock is constructed using imports, the real object of interest is adoption. Three gaps separate them. First, coverage: firms that automate through domestic distributors, integrators, or related-party procurement never appear as direct importers. Second, attribution: the goods a firm imports need not be deployed within that firm. Third, dating: the date of an import need not be the date at which equipment is installed and put to work. Bessen et al. (2025) bring evidence to bear on the first two by comparing an import-based measure with an expenditure-based measure of automation costs in Dutch data. They find that the two are correlated across firms but only weakly correlated within firms over time, and that direct imports are concentrated in a narrower set of firms and sectors. The within-firm result is the one that matters most for a design like ours, which reads adoption off the timing of a firm’s own imports.

The Argentine setting is well suited to an import-based measure. Argentina is a net importer of automation technology: domestic production of advanced automation capital is limited, and the country is not a significant exporter of these goods, so the re-export problem that complicates the Dutch mapping in Bessen et al. (2025) does not arise here. Each of the three gaps is therefore narrower than it would be elsewhere, and the design narrows them further. First, coverage binds least: firms that automate exclusively through domestic distributors or related-party procurement never enter the sample at all, so they do not contaminate the treated–control contrast. The concern is confined to firms that do both, and even for those, an import several times larger than the firm’s normal flow marks an economically meaningful moment in its investment path. The comparison also uses firms whose own first spike arrives later, rather than firms that never import automation goods, so that permanent differences between direct importers and non-importers do not drive theBRAMBILLA ET AL. 7 estimates.3 Second, we address attribution directly, by excluding sectors whose firms import these goods for resale rather than for use in production, and by building the design around large, lumpy changes in a firm’s own imports rather than small continuous purchases: it is routine purchasing, not discrete investme