Policy Research Working Paper 11051 Technology Sophistication across Establishments Diego A. Comin Xavier Cirera Marcio Cruz Finance, Competitiveness and Innovation Global Department January 2025 Policy Research Working Paper 11051 Abstract This paper examines technology sophistication in estab- function. There is significant variation in technology sophis- lishments. To comprehensively measure technology tication across and within countries, explaining 31% of sophistication, a grid is created that covers key business productivity dispersion and over half of the agricultural functions and the technologies used to conduct them. Ana- productivity gap. The sophistication of widely used tech- lyzing data from over 21,000 establishments in 15 countries, nologies is more relevant for productivity than the most the authors find that the most widely used technology is advanced technologies. More sophisticated technologies are usually not the most sophisticated available in the business appropriate for both developed and developing countries. This paper is a product of the Finance, Competitiveness and Innovation Global Department. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at xcirera@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team TECHNOLOGY SOPHISTICATION ACROSS ESTABLISHMENTS Diego A. Comin Xavier Cirera Marcio Cruz We thank David Baqaee, Jan DeLoecker, Mary Hallward-Driemeier, Apoorv Gupta, John Haltiwanger, Maurice Kugler, Bill Maloney, Martha Martinez Licetti, Paolo Mauro, Denis Medvedev, Dani Rodrik, Giacomo Ponzetto, Tommaso Porzio, Doug Staiger, Edouard Schaal, Chris Snyder, Eric Verhoogen, and seminar participants for insightful comments. Kyung Min Lee actively participated in earlier drafts of this paper and we are indebted to him. Harneet Singh provided outstanding research assistance. We thank Aman Mahajan, Magda Malec, Mariana Pereira, Santiago Reyes, Antonio Martins Neto, Soumya Agrawal, and Adrian Forest for inputs to sections of the paper. We also thank the industry experts and partner institutions for their support in designing and implementing the survey. An extended list of acknowledgements is provided in the appendix. Financial support from the infoDev Multi-Donor Trust Fund, the Korea World Bank Group Partnership Facility (KWPF), and the Competitive Industries and Innovation Program (CIIP) is gratefully acknowledged. This manuscript supersedes Cirera et al.(2020), Cirera et al. (2021) and Cirera, Comin and Cruz (2024). The views expressed in this paper are solely those of the authors and do not necessarily reflect those of the World Bank Group, its Board, or the National Bureau of Economic Research. © 2025 by Diego A. Comin, Xavier Cirera, and Marcio Cruz. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source. JEL No. O1, O3, O4 NBER Working Paper No. 33358. Diego A. Comin Marcio Cruz Dartmouth College IFC, World Bank Group Economics Department 2121 Pennsylvania Avenue, NW 6106 Rockefeller Hall, Room 327 Hanover, NH 03755 Washington, DC 20433 and CEPR marciocruz@ifc.org and also NBER diego.comin@dartmouth.edu Xavier Cirera The World Bank 1818 H ST NW Washington, DC 20433 United States xcirera@worldbank.org 1 Introduction Technology is central to some of the most fundamental economic questions. Yet, our un- derstanding of these issues often depends on indirect and limited measures. A long tradition, dating back to Ryan and Gross (1943) and Griliches (1957), has characterized technology in establishments by the presence of a few (typically one) advanced technologies. This ap- proach faces several limitations. First, the number of technologies is very small compared to those used in an establishment. Additionally, the tasks where establishments use these technologies are neither comprehensive or representative of the business functions conducted in an establishment. Second, measures based on the presence of advanced technologies do not provide information on how establishments without them produce. In particular, we do not know how sophisticated the technologies used are relative to the frontier. This concern is particularly relevant in developing countries where advanced technologies are less widely diffused. Third, traditional measures do not capture how intensively a technology is used, which is crucial to explain income divergence across countries (Comin and Mestieri, 2018). This omission limits our understanding of whether establishments predominantly use the most sophisticated technologies they have adopted and the importance for productivity of these technologies relative to the most widely used technologies.1 In this paper, we develop a new approach to directly and comprehensively measure the sophistication of technologies used in establishments. Our first step is to create a two- dimensional grid structure, which we refer to as ‘the grid’. Following the task-based pro- duction function approach (Zeira, 1998; Grossman and Rossi-Hansberg, 2008; Acemoglu and Autor, 2011; Acemoglu and Restrepo, 2018), the horizontal dimension of the grid covers the key tasks an establishment conducts, grouped into broader categories that we call business functions (BF). To this, we add a vertical dimension that represents the range of technologies that can be used to perform the key tasks in each business function. The grid encompasses 1 Since the classic work of Ryan and Gross (1943) and Griliches (1957) on hybrid corn, many have applied this approach to measuring technology in establishments in other sectors. For example, Davies (1979) studies the diffusion of 26 different manufacturing technologies, each typically relevant in only a single narrow sector, Trajtenberg (1990) measures the presence of CAT-scanners in hospitals, Brynjolfsson and Hitt (2000); Stiroh (2002); Bresnahan, Brynjolfsson and Hitt (2002); Akerman, Gaarder and Mogstad (2015) measure the presence of some ICTs such as computers or access to the internet. Other efforts include the Survey of Manufacturing Technology by the Census Bureau discontinued after 1993, which covers 17 specific technologies, including numerically-controlled machines, computer-aided design or engineering technologies, programmable controllers and local area networks (see Dunne (1994)); or the Canadian Survey of Advanced Technologies with 41 and 50 technologies, depending on the round (see for example Boothby, Dufour and Tang (2010)). More recently, the Advanced Business survey, Acemoglu et al. (2022), also administered by the US Census Bureau and that focused on five generic, frontier technologies: AI, robotics, dedicated equipment, specialized software and cloud computing. Unlike the previous studies, the Advanced Business Survey asks for the intensity with which the firm uses these advanced technologies. 1 63 business functions: seven general business functions (GBF) relevant to all sectors, and 56 sector-specific business functions (SSBF) across 12 sectors (agriculture, livestock, food pro- cessing, apparel, leather goods, automotive, pharmaceutical, other manufacturing, wholesale and retail, financial services, land transport services, and health services). In total, the grid spans 305 technologies. The grid has three properties. First, it is comprehensive both in terms of the business functions and of the technologies considered in each business function. Second, it is rele- vant for any establishment and country, regardless of its level of development. Third, the technologies in each business function are ranked according to their sophistication, from the simplest to the most complex which represents the world technology frontier. We implement the grid in the Firm Adoption of Technology (FAT) survey, administered to over 21,000 establishments that constitute representative samples in 15 countries: South Korea, Poland, Croatia, Chile, the Brazilian state of Ceará, Georgia, Vietnam, the Indian states of Uttar Pradesh, Tamil Nadu, Gujarat and Maharashtra, Ghana, Bangladesh, Kenya, Cambodia, Senegal, Ethiopia, and Burkina Faso. FAT collects three types of information. First, it gathers establishment-level data on sales, inputs, education of the workers and the managers, management practices, etc. Second, it records whether each sector-specific business function is conducted in-house. Third, and most relevant, FAT documents the technologies from the grid used by each establishment in each business function and, of these, which one is the most widely used technology. Using the information from FAT, we develop two measures of technology sophistication at the business function-establishment level: ‘MOST’ for the most widely used technology, and ‘MAX’ for the most advanced technology available. We use these measures to study three topics: the use of technology at the business function level, the cross-establishment variation in technology sophistication, and the relationship between technology sophistication and productivity across establishments. At the business function level, our analysis reveals that the most widely used technol- ogy (MOST) is usually not the most sophisticated one available (MAX). The gap between MAX and MOST is persistent, indicating that MAX and MOST represent two distinct and relatively independent processes of technology upgrading within establishments. By aggregating across all functions of an establishment, we derive establishment-level measures of technology sophistication that inherit the comprehensiveness of the grid. We observe significant variation in technology sophistication across establishments, both across and within countries. The dispersion in technology sophistication across establishments varies considerably across countries, increasing with per-capita income. Technology sophis- tication is positively associated with establishment size, the human capital of its workers, 2 the quality of management practices, exporter status, multi-national and multi-establishment status, and shows an inverted U-shaped relationship with establishment age. Examining the relationship between technology sophistication and productivity across establishments, we show that technology sophistication is strongly associated with produc- tivity. Differences in technology sophistication account for 31% of the variation in productiv- ity across establishments. This result holds even when controlling for management quality, human and physical capital, and markups. The productivity regressions yield three, additional insights. First, MOST is more rele- vant than MAX for establishment productivity, highlighting the importance of focusing on MOST in technology measurement and modeling. Second, there is significant variation across sectors in the share of productivity dispersion accounted for by technology sophistication. For example, it accounts for 50%, in agriculture but only for 28% in services. Consequently, differences in technology sophistication account for more than half of the agricultural pro- ductivity gap between high- vs. low-income economies (Caselli, 2005). Third, we examine whether technology is equally beneficial in both high- and low-income economies (Basu and Weil, 1998; Acemoglu and Zilibotti, 2001) by comparing the elasticity of productivity with respect to technology sophistication. This elasticity is not smaller in low-income countries, suggesting that the productivity gains from using more sophisticated technologies are not limited to advanced economies. Our study of technology sophistication is closely related to various lines of research. The patterns documented about the use of technology at the business function level provide an empirical counterpart to the two frameworks used to model technology in production: quality ladders (Aghion and Howitt, 1992) and love-for-variety models (Romer, 1990). These canonical paradigms predict that the best available technology at the business function level (MAX) suffices to characterize the technology sophistication of the business function and drives the establishment’s productivity. However, our findings highlight that MOST is not only key to characterizing the sophistication of technologies used in a business function, but is also more strongly associated with productivity across establishments than MAX. The notions of inter- and intra-firm diffusion are connected to MAX and MOST. Mansfield (1963) introduced the concept of intra-firm diffusion to describe the gradual increase in the use of a technology (e.g., diesel locomotives) within a firm after its adoption. The small body of literature following Mansfield has studied the intra-firm diffusion of a few technologies, in a few countries, such as numerically controlled machines in UK metalworking (Battisti and Stoneman, 2003) and e-commerce in the UK and Switzerland (Battisti et al., 2007; Hollenstein and Woerter, 2008). These studies have produced mixed results on whether intra- and inter-firm diffusion follow similar processes and whether earlier (less-sophisticated) 3 technologies remain the most widely used for a long time. Our broader study establishes general patterns about the gap between MAX and MOST. Additionally, with the comprehensive coverage of technologies and productivity measurement at the establishment level, we reveal a novel finding: the different associations between productivity and the MAX and MOST measures of sophistication. There is a long tradition studying the relationship between the adoption of advanced tech- nologies by an establishment and productivity.2 Studies in this literature typically consider a limited number of technologies. For example, Acemoglu et al. (2022) find an association be- tween productivity and the number of advanced technologies a firm adopts (see footnote 1 for a list of the five technologies considered). Our analysis of the relationship between technology sophistication and productivity builds on this literature. In addition to the greater number of technologies considered, our exploration extends previous work in three directions. First, the use of a grid that provides a dense coverage of business function and ensures that the technologies considered are representative of the main tasks conducted in the establishment. In particular, they do not just reflect generic tasks that apply to establishments in a wide range of sectors but they also include a large number of tasks that are specific to individual sectors. Second, the estimates found in the literature (e,g,. Acemoglu et al. (2022)) are in line with the coefficient for MAX in the productivity regression. However, the literature has not included measures of MOST in the productivity regressions. A key novel finding in our analysis is that the association of productivity and technology sophistication is much stronger with MOST than with MAX. This cross-establishment finding is consistent with the findings in Comin and Mestieri (2018) who study the relationship between productivity growth and the intensity of use of technologies across countries and over time. Third, most studies of productivity and technology across establishments are limited to a single country. Even though our 15 country-sample is far from representative of the 200+ countries in the world, it suffices to demonstrate that the strong association between technology sophistica- tion and productivity holds both within and between countries and that the within-country association does not differ between developed and developing economies. There are clear methodological and conceptual parallels between our effort to measure and study technology sophistication and the seminal work by Bloom and Van Reenen (2007) on management practices. As with technology, there is a long tradition of documenting specific management practices in a limited number of companies. The groundbreaking studies by Bloom and Van Reenen (2007) and Bloom et al. (2019) have greatly extended this scope by 2 For example, Hubbard (2003) focuses on on-board computers in trucks, Bartel, Ichniowski and Shaw (2007) on computer numerically controlled (CNC) machines and computer-aided design (CAD) software, Hjort and Poulsen (2019) on high-speed internet, Gupta, Ponticelli and Tesei (2020) on cellphones. 4 measuring the quality of management practices across 18 dimensions related to operations, planning, monitoring, and human resources, covering thousands of firms in many countries. Similar to FAT, data on management practices is collected via firm surveys. Experts rank practices based on their quality, and an establishment-level score is constructed to study the drivers of management practices and their association with productivity. We perform a similar analysis for technology sophistication. Beyond the similarities in measurement meth- ods, Bloom, Sadun and Reenen (2012) have hypothesized that technology sophistication and management practices are complementary. We explore the complementarity of technology and management in the context of productivity regression across establishments, finding supporting evidence. The rest of this paper is organized as follows. Section 2 introduces the FAT survey, and describes various validation exercises of the sophistication rankings, and the data collected. Section 3 presents the technology sophistication measures and illustrates key insights with examples from specific establishments, and sectors in FAT. Section 4 explores the use of technology at the business function level. Section 5 studies technology sophistication across establishments. Section 6 investigates the relationship between technology sophistication and productivity across establishments. Section 7 concludes. 2 The Survey The FAT survey (henceforth, “the survey") collects detailed information for nationally rep- resentative samples of establishments in agriculture, manufacturing, and services about the technologies each establishment uses to perform key business functions necessary to operate in its respective sector. In the following sub-sections, we describe the survey design and implementation, relegating further details to section A in the Appendix. 2.1 Structure The survey is composed of five modules. Module A collects information on the general characteristics of the establishment.3 Modules B and C cover the technologies used. Mod- ule D focuses on barriers to, and drivers of, technology adoption, while Module E gathers information about the establishment’s financial statements and employment. 3 The survey is designed, implemented, and weighted at the establishment level. For multi-establishment firms, the survey targets the establishment randomly selected in the sample. The survey can be downloaded at the following address (https://dcomin.host.dartmouth.edu/files/FAT_Survey_complete.pdf). The implementation manual which includes all instructions for interviewers, training materials and a full descrip- tion of the technologies in the grid can be downloaded at (https://dcomin.host.dartmouth.edu/files/ Implementation_Manual_TAS_29112023.pdf). 5 The survey differentiates between general business functions (Module B), which comprise tasks that all establishments conduct, regardless of the sector where they operate, and sector- specific business functions (Module C), which are potentially relevant only for establishments in a given sector. All establishments in our sample respond to Module B, but only those belonging to the sectors for which we have developed a sector-specific module respond to C. To attain a wide coverage that allows a meaningful study of sector-specific technologies, we develop sector-specific modules for 12 significant sectors in the economy, including agriculture (crops and fruits), livestock, food processing, wearing apparel, leather goods and footwear, automotive, pharmaceutical, other manufacturing, wholesale and retail, financial services, land transport services, and health services.4 These sectors have been selected based on their share in aggregate value-added, employment and number of establishments and they cover all three industries (agriculture, manufacturing, and services). 2.2 The Grid To design Modules B and C, we determined the business functions covered and the list of technologies, from most basic to most sophisticated, that can be used to implement the key tasks in each function. We call the resulting structure "the grid". To construct the grid, we followed three steps. First, we conducted desk research review- ing the specialized literature. Second, we held meetings with World Bank Group experts on each of the sectors covered. Third, we reached out to external consultants with significant experience (at least 15 years) in a given sector. For example, the external experts in agri- culture and livestock were agricultural engineers and researchers from Embrapa-Brazil. For food processing, apparel, automotive, pharmaceuticals, transportation, finance, and retail, as well as for the GBFs, we relied on senior external consultants selected by a large man- agement consulting organization. For health, our team relied on consultants and physicians with practical experience in both developing countries and advanced economies. In total, more than 50 experts participated in the construction of the technology grid. The resulting grid is composed of 7 general and 56 sector-specific business functions and contains a total of 305 technologies (See Section A.1.1 of the appendix for details on the procedures followed to define the grid). All technologies in the Grid are precisely described so that respondents and enumera- tors can objectively establish their use. Figure 1 presents the general business functions considered in the survey and the possible technologies that can be used to conduct each of 4 The granular information that can be obtained with the FAT survey allows us to explore central questions on technology policy in developing countries. One example, itself a product of this paper, is the World Bank policy report "Bridging the Technological Divide" (Cirera, Comin and Cruz, 2022). 6 them. The grid contains 7 GBFs: Business administration, production planning, sourcing and procurement, marketing, sales, payment methods, and quality control. Each function considers between 4 and 7 technologies. For example, an establishment can gather and an- alyze customer information for marketing purposes using face-to-face conversations, online chats via WhatsApp or the internet, structured customer surveys, customer relationship management (CRM) software to store contact information, interaction history, and commu- nication preferences, or big data analytics and/or artificial intelligence to uncover trends and make informed marketing decisions. Figure 2 presents the grid for one sector-specific module, agriculture. The grid considers six SSBFs for agriculture which include land preparation, irrigation, weeding and pest management, harvesting, storage and packaging. For example, to prepare the land for cultivation, a farm can use manual labor with simple tools such as hand-held hoes, or rakes, animal-aided instruments such as ploughs, equipment manually operated such as tractors, motor tillers, or rotators, or equipment supported by digital tech- nologies such as GPS, software or precision agriculture tools. Section A.1.1 of the appendix reports the grids for all other SSBFs and the implementation manual precisely defines each of the technologies in the grid. 2.3 Ranking of Technology Sophistication In addition to identifying key business functions and relevant technologies, industry experts ranked the technologies in each function based on their sophistication. More sophisticated technologies can perform a wider variety of tasks, more complex tasks, or perform tasks with greater accuracy and speed. The experts’ deliberations and resulting sophistication rankings, shown on the grid, were produced before the survey administration. This approach to ranking technologies resembles the World Management Survey (Bloom and Van Reenen, 2007), which relies on experts to rank management practices according to their quality. Given the importance of the ranking for our analysis, we evaluated the coherence of the expert rankings through a three-stage validation process implemented in 14 of the 63 business functions on the grid including most of the GBFs and SSBFs in agriculture, food processing apparel, and retail. The three stages are as follows: 1. Comparison of Key Features: We compared the technologies in each business function along three dimensions invoked by the experts: functionality, integration, and au- tomation. Functionality refers to the capabilities a technology offers to handle more complex tasks, in a faster way, on a larger scale, with greater accuracy and reliability. Integration reflects a technology’s ability to connect and interact seamlessly with other systems by ex- changing data and coordinating processes. Automation enables the technology to execute 7 processes, make decisions, and generate outcomes independently, without human interven- tion. 2. Novelty and Cost: We documented the year of invention and the cost of each technology and studied their correlation with the experts’ rankings. Although novelty and cost do not define sophistication, more sophisticated technologies tend to be newer and more expensive. 3. Large language models (LLMs): We conducted two exercises to validate our rank- ing through LLMs. First, we asked ChatGPT to rank the technologies based on their levels of sophistication. We replicate this exercise following specific definitions of sophistication based on functionality, integration, and automation. Second, we asked ChatGPT to identify a specific task for each of the 14 business functions and estimate the time required to perform the task with each technology. To collect the information in the first two stages, we relied on multiple sources, including the official description of specific leading brands supplying these technologies. For GBFs we collected information from multiple companies websites, including Microsoft, Google, SAP, Oracle, QuickBooks, IBM, Sage, NetSuite, BambooHR, Trello, Salesforce, Workday, Meta, Qualtrics, Survey Monkey, Amazon, Shopify, LinkedIn, among others. These companies have more than 80% of the global market share for technologies used in business administration, such as standard software (e.g, spreadsheet) and enterprise resource planning (ERP) systems, and a large share of the market across GBFs.5 In addition, we consulted specialized websites (e.g, tech.co; erpresearch.com; getapp.com) that provide comparisons across these products, totaling more than 50 original sources of information. Similar exercises with multiple sources of information were replicated for SSBFs. We illustrate the validation methodology using the example of business administration, a GBF that includes finance, accounting, and human resources processes. Table 1 summarizes the three-step validation procedure for each technology in business administration. The least sophisticated technology, handwritten processes, can only perform basic manual administra- tion tasks such as transaction entry, bookkeeping, or employee records handling, without any integration or automation features. Standard software like Microsoft Excel or Google Sheets helps with basic functionality to perform mathematical and statistical operations, including charts, and handle financial account, and HR records. However, it requires manual inputs and knowledge to build specific applications, with limited integration and automation. Mobile apps, such as QuickBook online, are pre-designed to perform these tasks with some 5 Estimates based on Enlyft dataset (Cirera, Comin and Cruz, 2022). These companies are recognized as key players by various specialized sources estimating market potential for ERP (e.g., Research and Market, Fortune Business Insight), even if there are variations in their market share estimations. 8 integration and automation features, but with limited scale and customization. Specialized software, such as Oracle Financial, have high functionality, integration, and automation ca- pabilities, within specialized domains. Finally, enterprise resource planning (ERP systems), such as SAP and Oracle NetSuite provide comprehensive functionality with full integration within and across business functions, with a high level of automation. Comparing the fea- tures of business administration technologies results in a ranking that matches the grid’s sophistication ranking. More sophisticated technologies in business administration are embodied in more ex- pensive software. For example, standard software such as Microsoft 365 (which includes Microsoft Excel) and Google Sheets costs between $12 and $18 per user/month; apps such as Quickbooks Online and BambooHR cost between $30 and $200 per user/month; special- ized software such as Oracle Financials, Intuit Quickbooks and Workday HCM cost between $120 and$600 per user/month; ERP systems such as SAP ERP or Oracle Netsuite cost over $1700 per user/month. In business administration, there is no clear relationship between technology novelty and our sophistication ranking (e.g., the SAP ERP system was intro- duced in 1981, while Microsoft Excel was first available in 1985).As shown in section A.2 of the appendix, for most SSBFs, especially in agriculture and manufacturing, we observe a positive and strong association between technology novelty, cost, and our sophistication rankings. ChatGPT’s ranking of the technologies in business administration based on functionality, automation, and integration coincides with the expert ranking. Furthermore, the ranking is robust to variations of the prompts provided to ChatGPT focusing on specific dimensions of technology sophistication (e.g, exclusively based on functionality, integration, or automa- tion). Finally, the estimated time to perform a task (e.g., managing payroll and financial reports) provided by ChatGPT is consistent with the experts’ sophistication ranking. For example, it takes roughly 5 hours to conduct the task with handwritten processes, 2 hours with computer and standard software, 1.5 hours with an app, 1 hour with specialized software and 30 minutes with an ERP system. We replicated these exercises for business administra- tion in ChatGPT over 100 iterations, to account for its probabilistic features and potential variation in a typical activity, and the patterns are consistent. The ChatGPT rankings are strongly and positively associated with the experts’ rankings, following a similar ranking order of sophistication in all iterations. Overall, the validation exercise across all 14 business functions supports the experts’ ranking of sophistication.6 6 In section A.2 of the appendix, we provide the results for the other 13 business functions where we implement the validation of the expert rankings. 9 2.4 Information Collected in FAT The survey collects information in three broad areas: the business functions conducted by an establishment, the use of technologies in each business function, and information on the establishment’s financial statements, workers, and management. Business functions. The business functions that comprise the horizontal dimension of the grid cover the key tasks involved in production. Explorations conducted at the piloting stage of the survey as well as the responses to the questions on the use of technologies in GBFs demonstrate that these functions are conducted in-house and that respondents are aware about the technologies their establishments use in the GBFs.7 We formally explore the relevance of each sector-specific business function in each establishment through a screener question that asks whether a sector-specific function is conducted in that establishment. This information helps us assess the relevance of establishment-level measures of technology sophistication based only on the technologies used in functions conducted in-house. Technology questions. The survey has two types of questions about the technologies used to conduct each business function. First, it asks whether the establishment uses each of the technologies listed in the grid. After identifying the technologies that are used by the establishment in a business function, the survey asks which technology is the most widely used in that function. The answers to these questions permit us to differentiate between the range of technologies present in the business function vs. the intensity with which they are used. FAT also asks whether the establishment uses “other technologies” in the business function in addition to those contained in the grid. Only in 3.6% of the business functions establish- ments declare that “other” technologies are used in the business function, and only in 0.8% of the business functions "other" is the most widely used technology. The low frequency of “other” demonstrates the comprehensiveness of the technologies in the grid. Other variables. The survey also includes other standard questions about financial state- ments’ information, employment, education of the employees, and education and experience of the manager. The survey collects information on four management practices from MOPS, including the presence of formal incentives, number of key performance indicators (KPIs), 7 Due to space constraints in the survey and the information revealed during the pre-pilot, we decided to not directly ask about whether establishments conduct each GBF in FAT. Proxying the fraction of GBFs that are not conducted in house by the share of GBFs for which the establishment responds that either "does not use" or "does not know if it uses" to all the technologies in the grid for the BF, we find that only 3.9% of GBFs are not conducted in-house. 10 frequency of KPI review, and time frame of production targets. The answers to these ques- tions are used to construct a management z-score following the methodology in Bloom and Van Reenen (2007). Despite covering only four of the 16 variables collected in MOPS, the FAT z-score based on this subset of questions accounts for 90.5% of the cross-establishment variance of the original MOPS z-score for Mexican establishments collected by ENAPROCE. 2.5 The Data Our analysis is based on primary data collected from establishments in 15 countries: South Korea, Poland, Croatia, Chile, Brazil (Ceará), Georgia, Vietnam, India (Uttar Pradesh, Tamil Nadu, Gujarat and Maharashtra), Ghana, Bangladesh, Kenya, Cambodia, Senegal, Ethiopia, and Burkina Faso. Several factors were considered in deciding where to implement the FAT survey. We targeted countries on different continents (Asia, Africa, South America, and Europe), with different levels of income, for which there was access to a high-quality sampling frames. In these countries, we collected data from 21,055 randomly selected es- tablishments from the sampling frames. Table 2 shows the distributions of the sample by country, sector, and size groups and Table C.1 provides descriptive statistics. The median establishment in our sample has 9 workers, with an average of 34 workers. 20% of workers have a college degree, 19% of firms were multi-establishments, 18% are part of a multina- tional firm, 17% are exporters, 18% are 5 years old or younger, and 76% have electricity, computers, and internet access. 2.5.1 Sampling Our data is representative for a universe of about 2.1 million establishments. The samples are nationally representative for establishments with 5 or more workers. For each country, the sampling frame is based on the most comprehensive and up-to-date establishment-level census data available from the respective National Statistical Office (NSOs) or similar au- thority. The survey is stratified on three dimensions - sector, firm size, and region - so that we can construct representative measures of technology for aggregates along these dimensions. Sampling weights are based on the inverse probability of selecting establishments within each stratum.8 8 Table A.15 provides information about the distribution of firms by country, sector, and size groups within the universe covered by the FAT survey. For the state of Ceará in Brazil and the Indian states of Tamil Nadu, Uttar Pradesh, Gujarat, and Maharashtra, it is representative at the state level. Section A of the Appendix provides more details on the sampling frame, survey implementation and data collection, and sampling weight. 11 2.5.2 Measures to minimize bias and measurement error The literature on survey design has identified three types of potential bias and measure- ment errors based on whether they originate from non-responses, the enumerator, or the respondent (Collins, 2003). In what follows, we briefly describe the steps taken in designing and implementing the FAT survey to minimize these errors. Appendix A.5 provides a more detailed description of the measures implemented to minimize potential bias. Non-response bias. To maximize response rates and minimize potential biases associated with non-response (Gary, 2007), we followed best practice procedures. First, we partnered with national statistical offices and industry associations to use the most comprehensive and updated sampling frame available. Second, we hired data collection companies or agen- cies which were supported by endorsement letters from local institutions and which had demonstrable experience in nationally representative firm-level surveys. Third, we followed a standard protocol in which each firm was contacted several times to schedule an interview. Fourth, we mostly used face-to-face or phone interviews, which usually have higher response rates than web-based interviews.9 Enumerator bias and error counts. The survey, training, and data collection processes were designed to minimize enumerator biases and data collection errors. First, we used closed-ended questions to make coding the answers a mechanical task, thereby eliminating the need for the enumerator to interpret the answers or exercise subjective judgement when coding them. Second, the same standardized training was implemented in each country in the local language, with enumerators, supervisors, and managers leading the data imple- mentation. Third, we conducted a pre-test pilot of the questionnaire in each country using establishments not included in the sample. Fourth, to attain greater quality control dur- ing the data collection process, enumerators recorded the answers via Computer-Assisted Personal Interviews (CAPI) or Computer-Assisted Telephone Interviewing (CATI) software, and we regularly monitored the data collection process using standard algorithms to analyze the consistency of the data.10 9 These procedures are in line with suggestions of good practice for implementation by (Bloom et al., 2016). We use online surveys only for Georgia and Croatia. In Georgia, we partnered with the National Statistical Office, which resulted in exceptionally high response rate. Face-to-face interviews were not possible during the pandemic. See Table A.16 for the mode and date of data collection in each country. 10 Randomized survey experiments with household surveys have demonstrated that a large number of errors observed in Pen-and-Paper Personal Interview (PAPI) data can be avoided with CAPI or CATI (Caeyers, Chalmers and De Weerdt, 2012). For Georgia and Croatia, we used Computer Assisted Web Interviewing (CAWI). 12 Respondent bias. We took several steps to minimize respondent bias. First, we ensured that the interview was arranged with the appropriate person or persons; main managers (and other managers, such as plant managers and HR managers, in larger firms). Second, we used a closed-ended design in the questionnaire such that the respondent was questioned about specific technologies one at a time and was not told beforehand all the technologies that were associated with each business function. This design reduced measurement error in respondent’s answers. Third, we pre-tested the questionnaire in each country to ensure that our questions were clearly worded within the specific geographical and cultural con- texts of each country, reducing the need for subjective judgement in responses (Bertrand and Mullainathan, 2001). Fourth, to avoid social desirability bias, which may cause respon- dents to overstate the use of more sophisticated technologies, the survey avoided the words "technology" and "sophistication", employing more neutral terms such as "methods" and "processes" instead. 2.5.3 Ex-post checks and validation exercises We conducted several ex-post checks to assess the quality of the collected data. Non-response bias. The average (unit) response rate on the survey varies by country and ranges between 15% and 86%. For example, the response rate was 80% in Vietnam, 57% in Senegal, 39% in Ceará, Brazil, 24% in Korea, and 15% in Croatia. These response rates are high relative to typical response rates in establishment-level surveys, which are around 5 to 10% and are consistent with response rates observed for WMS which are around 40% (Bloom et al., 2016). To minimize potential non-response bias, we adjusted the sampling weights for unit non-response. The adjustment was calculated at the strata level, so that the weighted distribution of our respondent sample across strata (sector, size, region) exactly matches the distribution of establishments in the sampling frame.11 We conducted three tests to assess potential biases from unit non-response-rates.12 In each of these exercises, presented in Section A.6 of the Appendix, we find no statistical difference in the number of employees, technological sophistication, wages, and share of workers by skill and education 11 Table A.17 in the Appendix A provides the response rate by country, defined as the ratio between establishments that responded to the survey and the total number of eligible establishments in the sample for which we attempted to conduct an interview. The response rates were higher when national statistical agencies implemented the survey. Section A.4 of the appendix provides more details on sampling weights. 12 First, using the information from the sampling frame, we check if there are differences in the aver- age number of workers per establishment between respondents and non-respondents within stratum. Sec- ond,using information on the number of contact attempts, we compare the establishment-level technology sophistication in GBFs, described in the next section, between establishments with above and below the average number of attempts. Third, in a similar vein, we compare establishments in the first list of contacts provided to interviewers, versus those provided subsequently. See Table A.18 to A.24 in Appendix A. 13 between establishments in the group that proxies for the response sample and the group of establishments that proxies for the non-response sample. Response bias. To assess the relevance of response bias, we conducted a parallel pilot in Kenya where we re-interviewed 100 randomly selected establishments with a short version of the questionnaire. For those establishments, we randomly selected three business functions and asked about the presence of the relevant technologies. We estimated a probit model to assess the likelihood of consistent answers between the original and the back-check interviews, controlling for establishment-level fixed-effects. Reporting the use of a technology in the back-check interview is associated with 80.6% of the likelihood of reporting the use of the same technology in the original interview. Conversely, reporting that a technology is not used in the back-check interview, is associated with a 70.7% likelihood of not being reported in the original survey. These estimates do not differ between establishments of different sizes.13 Validation using external sources. We evaluate the quality and reliability of the data collected by comparing it to external sources in Korea (KED) and Brazil (RAIS). We focus on variables related to establishment size, productivity and technology. Table A.24 shows that the weighted sample averages of the labor variables in the FAT data (number of workers, average wages, share of college workers, share of low- and high-skill workers) are not statisti- cally different from the averages in the universe of firms from the RAIS dataset. In the Brazil matched establishments, we find a strong correlation between FAT measures of log value- added per worker and the log of average wages from RAIS (See Table A.23). In the Korean matched establishments, we find very high cross-establishment correlations (above 0.93) in the log levels and growth rates of sales and employment, as well as in log labor productivity (0.73).14 Additionally, the average adoption rate of ERP systems in Korean manufacturing establishments in FAT is similar to Chung and Kim (2021), who used a similar sampling frame (32% vs. 40% in Chung and Kim, 2021), and there is a strong cross-establishment 13 The re-interviews produced 1,661 answers, 106 interviews times 3 business functions times an average of 5.2 technologies per function. Both the original and back-end interviews in the pilot are conducted by phone by different interviewers. The correlation between the binary responses in survey and pilot is 73% ranging from 65% in business administration to 77% in sales across business functions, and from 85% among the most basic technologies to around 61% in intermediate, and 77% at the most advanced technologies across functions. 14 In Korea we merge FAT with the Korea Enterprise Data (KED), a leading supplier of business credit reports on Korean businesses. In Brazil, we merge the data with the Relação Anual de Informações Sociais (RAIS), which is an administrative database maintained by the Ministry of Labor providing information on salaries for all formal workers in Brazil. The FAT survey asks about sales and the number of employees for two periods. The most recent year for which the information is available (i.e. the year before the implementation of the survey) and two years before that. For Korea, these reference years are 2019 and 2017. 14 association between the book value of machinery and equipment in KED and the estab- lishment technology sophistication measures (MOST and MAX) from FAT, which will be explained in the next section. Internal validation. We conduct an additional validation exercise of the technology mea- sures, by studying whether establishments with larger sales, employment and sales per worker are more likely to use top-tier technologies, which are the more sophisticated technologies in each BF and are marked in bold in Appendix A.1. Specifically, we estimate a linear probabil- ity model for each business function, where the dependent variable is binary and equal to 1 if the establishment uses one of the technologies classified as top-tier for the business function and 0 otherwise. The model includes a full set of country- and, for the GBFs, 2-digit sector fixed effects. The independent variables are either (log) sales, (log) employment or (log) sales per worker. We find that the coefficients for these variables are positive and significant in a large majority of business functions.15 These ex-post checks further reassure us about the soundness of the survey design, the data collection process, and the accuracy of responses. 3 Measures of Technology Sophistication We next introduce measures of technology at the business function and establishment levels, constructed using information collected by the FAT survey. Before analyzing these mea- sures, we illustrate the granularity of the grid and how these measures can characterize the sophistication of technology used by establishments, with examples from FAT. 3.1 Technology Measures We denote by AN U Mf,j the number of different technologies from the grid used in business function f in establishment j . When more than one technology is used in a business function, we explore whether the technologies used are contiguous in the sophistication ranking of the grid or, instead, there are sophistication gaps in the vector of technologies used. Formally, we define the sophistication gap of establishment j in business function f (SGf,j ) as a binary variable that takes the value of 1 if the establishment uses technologies with sophistication rank τ and τ + k for k ≥ 2 in function f but does not use the technology with sophistication 15 For sales we find a positive coefficient in 100% of BFs (85% significant at 5% level); for employment 98% are positive (93% significant); and for productivity 80% are positive (52% significant, and never negative and significant). 15 rank τ + p, for 1 ≤ p < k . SGf,j is 0 when there are no gaps and at least two technologies are used in the function.16 We study the sophistication of the technologies used in a business function with two variables. M AXf,j measures the sophistication of the most sophisticated technology used in the given business function, while M OSTf,j reflects the sophistication of the most widely used technology in the business function. The starting point to construct these measures is the experts’ rankings of the technologies, from least to most advanced, rf ∈ 1, 2, ..., Rf .17 r −1 We define the relative rank of a technology as r ˆf = Rff −1 . Note that r ˆf ∈ [0, 1]. We follow the standard approach of constructing cardinal measures of the sophistication of a technology by applying an affine transformation to the relative rank, r ˆf . In Section 6, we show that affine transformations are a reasonable cardinalization of ordinal technology measures because establishment (log) productivity is approximately linear in the cardinalized measures of technology sophistication. Specifically, we define M OSTf,j and M AXf,j as M OSTf,j = 1 + 4 ∗ r M OST ˆf,j . (1) M AXf,j = 1 + 4 ∗ r M AX ˆf,j , (2) where r M OST ˆf,j and r M AX ˆf,j are the relative sophistication rankings of the two technologies. By construction, M OSTf,j , M AXf,j ∈ [1, 5], and M AXf,j ≥ M OSTf,j . We also use a similar transformation to define a scaled measure of the number of technologies used in a business function (N U Mf,j ).18 Since the most sophisticated technologies in the grid define the current (world) tech- nology frontier, M AXf,j and M OSTf,j represent the closeness of an establishment to the technological frontier in a business function. M AXf,j and M OSTf,j are of independent importance as they capture different aspects of the technology upgrading processes in the business function. M AXf,j increases when a firm implements a new technology that is more sophisticated than those currently used in a given business function. This technology may not be new to the establishment, but it is new to the business function of the establishment. 16 SGf,j is not defined when less than two technologies are used in the function (i.e. AN U Mf,j < 2). 17 Because several technologies may be assigned the same sophistication, the highest rank in a function Rf may be smaller than the number of possible technologies Nf . In a small number of business functions, the technologies covered are used in various subgroups of tasks. For example, in the body-pressing and welding functions of the automotive sector, the survey differentiates between technologies used for pressing skin panels, pressing structural components and welding the main body. In cases like this, we construct ranks of technologies for each subgroup of tasks within the business function, and then aggregate the resulting indices by taking simple averages across the tasks groups. See Appendix B.1.1 for more details. 18 AN U Mf,j −1 Formally, we define N U Mf,j as N U Mf,j = 1 + 4 ∗ N f −1 , where Nf is the number of different technologies in the grid for the business function f . 16 Therefore, increases in M AXf,j capture technology improvements as those in quality ladder (e.g., Aghion and Howitt, 1992) or horizontal variety (e.g., Romer, 1990) conceptualizations of technology in production. Increases in M OSTf,j occur when a new establishment’s most widely used technology in the business function is more sophisticated than the previous one. This new technology may be entirely new to the business function or an existing technology whose use has been expanded. Therefore, M OSTf,j is more related to Mansfield (1963)’s concept of technology diffusion within the firm, specifically within the business function, rather than to innovation. Relevant outcomes and observable characteristics are often reported at the establishment level. We construct establishment-level technology measures as simple averages of N U Mf,j , M AXf,j and M OSTf,j across the business functions of an establishment. Specifically, we define N U Mj , M OSTj and M AXj as: Nj Sf,j Sj = (3) f =1 Nj where S = {N U M, M OST, M AX }, and Nj is the number of business functions covered for establishment j . 3.2 An Illustration Before studying the general patterns of technology use in establishments, it is useful to become familiar with the grid and the measures of technology sophistication by exploring some examples from FAT. To begin appreciating the level of detail in the grid, we examine two medium-sized es- tablishments in apparel retail: one in India (establishment 1) and the other in Vietnam (establishment 2). Figure 3 plots the M AXf,j index in each business function for both establishments. For instance, establishment 1 uses a dynamic pricing system that automat- ically adjusts prices based on demand conditions, while establishment 2 uses an automated markup technology that collects information on costs and applies a uniform markup. The pricing business function includes five technologies (Figure A.7). From least to most so- phisticated, these are: manual pricing (prices set without a formal account of the costs), automated markup, automated promotional pricing (prices adjusted based on seasonal fac- tors), dynamic pricing, and personalized pricing (prices adjusted at the individual customer level using data analytics such as data mining and machine learning). The value of the MAX index for establishment 1 (4) indicates that it is one notch below the technology frontier in pricing, while the value for establishment 2 (2) shows that it is three notches below the 17 frontier. Establishment 1 also has a higher MAX level in other business functions such as mer- chandising and inventory. In these functions, it uses digital merchandising systems (DMS) and automated inventory controls, respectively. In contrast, establishment 2 selects products to display on shelves manually and uses a warehouse management system with specialized software.19 Nevertheless, establishment 2 has a higher MAX index than establishment 1 in other functions such as customer service, quality control, and sales. Specifically, establishment 2 attends to customer requests made online, checks product quality using statistical process control with software monitoring, and sells its products online using an external digital platform. In contrast, establishment 1 attends to customer requests over the phone, checks product quality manually with the support of digital technologies, and sells products directly at the establishment. The variation in the relative technology sophistication rankings of establishments 1 and 2 across different business functions underscores the importance of a comprehensive coverage when characterizing the technological sophistication of an establishment. Focusing on one or a few functions or technologies provides an imprecise, and possibly biased, characterization of the sophistication of the technologies used in an establishment. Next, we move from the establishment to the sector level and explore the cross-establishment distribution of technology sophistication in the food processing sector. We focus on the fab- rication business function, which is relevant for all manufacturing establishments in FAT. Some of the technologies it covers, such as numerically controlled machines and robots, have been widely studied in automation research. The grid considers six classes of technologies. In increasing order of sophistication, these are (1) manual processes, (2) machines controlled by operators, (3) machines controlled by computers, (4) robots, (5) additive manufacturing in- cluding rapid prototyping and 3D printing, and (6) other advanced manufacturing processes such as laser, plasma sputtering, high-speed machine, E-beam and micro-machining). The top right panel of Figure 4 plots the distribution of M AXf,j in fabrication across food processing establishments in South Korea. The histogram reveals a significant disper- sion across establishments in the most sophisticated technologies available for production in fabrication. Establishments that process food using the world frontier’s fabrication tech- nologies coexist with others that just use manual processes. It is also worth noting that, in contrast to the popular perception, most establishments do not use robots or other more 19 DMS is used to execute core merchandising activities, including product management, inventory re- plenishment, purchasing, vendor management, and financial tracking. One example of automated inventory controls is Computer Assisted Ordering (CAO), an inventory replenishment system that can use either sales or inventory algorithms to prepare a suggested reorder. 18 sophisticated technologies (i.e., M AXf,j ≥ 3) even in such an advanced economy as South Korea.20 To explore the cross-country differences in technology sophistication, the top panel of Figure 4 also plots the histogram of M AXf,j in Senegal (left) and India (middle). There are stark cross-country differences in the distribution. The mean and variance increase uniformly with the level of development in the country (i.e. Senegal, India, South Korea). Additionally, the cross-establishment distribution of M AXf,j is most skewed to the right in Senegal, and least in Korea. The bottom panel in Figure 4 shifts the focus to the most widely used technology. In particular, it shows the histogram of M OSTf,j in fabrication across food processing es- tablishments for each country. By construction, the distribution of M AXf,j stochastically dominates the distribution of M OSTf,j , as M AXf,j ≥ M OSTf,j . However, the distributions of MAX and MOST differ significantly. For example, in 65% of Indian food processing es- tablishments the most sophisticated technology used in fabrication is ’machines controlled by operators.’ Yet, this technology is the most widely in only 35% of establishments. The gap between MAX and MOST in this example motivates a deeper exploration of whether MAX and MOST are statistically distinct across a broad range of business functions and countries and, if that is the case, their relative importance in shaping the relationship between technology sophistication and productivity across establishments. 4 Technology Sophistication at the Business Function We use the FAT dataset to examine technology at the business function level. We explore two issues: the range of technologies used in each function and the comparison between the most widely used and the most sophisticated technology available in the business function. We then compare our findings with the predictions of state-of-the-art models of technology in production and draw conclusions about relevant extensions. 4.1 The Array of Technologies Used in the Business Function We begin our analysis of the vector of technologies used in a business function by counting the number of technologies from the grid that an establishment uses, denoted as AN U Mf,j . Table C.2 reports, for each function, the average AN U Mf,j across all establishments that conduct the function in-house. On average, establishments use two different technologies 20 This is also true in other sectors with greater penetration of robots such as the automotive sector. Even if we weight establishments by size (e.g., employment or sales) only 56% of establishments in automotive fabrication in South Korea use robots or a more sophisticated technology in fabrication. 19 per function. This average is consistent across both general and sector-specific business functions. The distribution of AN U Mf,j reveals that 62.6% of functions use more than one technology, and 28.3% use at least three. The vector of technologies used in a business function is relevant to two economic liter- atures. Schumpeterian models predict that when adopting a more sophisticated technology, an establishment will abandon the less sophisticated ones it was using. In models of techno- logical leapfrogging, late adopters skip the less sophisticated technologies to directly use more sophisticated ones. Although FAT only provides cross-sectional data, it can be informative about the empirical support for these predictions. We explore the frequency of instances where establishments (i) completely skip or abandon less sophisticated technologies, (ii) use the least sophisticated technology despite having more advanced options, and (iii) create sophistication gaps by skipping some technologies in the business function. Of the 37.4% of functions where only one technology is used, 52.8% use the least sophisti- cated technology. In the remaining 47.2%, the technology used is not the least sophisticated. Therefore, only in 18% (37.4% * 47.2%) of functions, establishments have fully skipped or abandoned all less sophisticated technologies. Conversely, in 70.4% of functions where multiple technologies are used, one of the technologies the establishments use is the least sophisticated technology in the grid. Finally, sophistication gaps are infrequent. Overall, they occur in 25% of business functions: 27% among GBFs and 17% among SSBFs. The GBFs where gaps are more frequent are payments (48%), business administration (34%), and sales (28%). These observations show that establishments continue to use less sophisticated technolo- gies even in business functions where more advanced technologies are available. FAT, there- fore, does not support the predictions of Schumpeterian and leapfrogging models regarding the abandonment and skipping of less sophisticated technologies. Additionally, the infrequency of sophistication gaps, combined with the regular use of the least sophisticated technology, implies that we can approximate the entire vector of technolo- gies used in a business function just by the most sophisticated technology it uses. In other words, M AXf,j approximately captures the entire adoption history of the establishment in a business function. To further explore the role of MAX in the technology upgrading process, we estimate the following regression: M AXf,j = αj + αf + β ∗ N U Mf,j + uf,j (4) where αj and αf are establishment and business function fixed effects. The point estimate of β , presented in column 2 of Table 4, is 0.84. The close to one-to-one movement of MAX and NUM suggests that the technologies that are introduced by an establishment in 20 a business function are typically more sophisticated than existing ones. 4.2 The Intensity of Use of Technology Technology research has mainly focused on the presence of new or advanced technologies in establishments, often neglecting the intensity of use of existing technologies. This focus is based on the belief that the technologies establishments use most intensively are the most sophisticated technologies they have adopted. Therefore, measuring the most widely used technology seems redundant, as M AXf,j is considered a sufficient statistic for M OSTf,j . Departing from this tradition, we study the sophistication of the most widely used tech- nology in business functions, M OSTf,j . Our primary objective is to determine whether, M OSTf,j is indeed redundant, as suggested by the literature, or if it provides distinct in- sights into the sophistication of technologies used in a business function. To explore this, we examine the MAX-MOST gap at the business function level. First, we document the magnitude and frequency of this gap. Second, we analyze the relationship between MOST and NUM, comparing it to the relationship between MAX and NUM. Third, we examine potential drivers of the MAX-MOST gap by studying its association with relevant proxies across establishments. The MAX-MOST Gap. The average difference between M AXf,j and M OSTf,j across establishments and functions is 0.68. This gap is significant economically and statistically given that sophistication measures range from 1 to 5, with standard deviations of 1.23 for M AXf,j and 1.09 for M OSTf,j . To study the frequency of MAX-MOST gaps, we calculate, for each establishment, the fraction of functions with multiple technologies that exhibit a MAX-MOST gap. The average frequency of the MAX-MOST gap across establishments is 0.62. The average is similar for GBFs (0.62) and SSBFs (0.61). However, the average frequency of MAX-MOST gaps in SSBFs varies significantly across sectors, ranging from 28% in health services to 82% in financial services. Except for health services and pharmaceuticals, all SSBFs have an average gap frequency of at least 50% (see Figure 5). MAX-MOST gaps are also the norm across all countries, although there are significant differences, with average frequencies ranging from 51% in South Korea to 83% in Burkina Faso. Figure 6 shows that the average frequency of MAX-MOST gaps decreases with income, exhibiting a correlation of -0.55. To further understand the differences between MAX and MOST, we regress M OSTf,j on M AXf,j and study the fraction of the (within-establishment) variance in M OSTf,j accounted 21 for by M AXf,j . Specifically, we estimate: M OSTf,j = αj + αf + β ∗ M AXf,j + uf,j , (5) where αj and αf are establishment and function effects.The estimates, reported in column 1 of Table 4, reveal that while M AXf,j and M OSTf,j are positively correlated within es- tablishments, M AXf,j explains only 34% of the variance in M OSTf,j . This indicates that M AXf,j is not a sufficient statistic for M OSTf,j . To explore why this is the case, we regress M OSTf,j on N U Mf,j and compare this estimate with that of regressing M AXf,j on N U Mf,j in specification (4). Specifically, we estimate M OSTf,j = αj + αf + β ∗ N U Mf,j + uf,j . (6) The estimates, reported in column 3 of Table 4, imply that a 1-unit increase in N U Mf,j is associated with an increase in M OSTf,j by just 0.25, suggesting minimal impact of new technology adoption on the most widely used technology. This contrasts with the 0.84 estimate for M AXf,j , showing that, the extension of the use of existing technologies in a business function and the adoption of new technologies are distinct technology upgrading processes and they are driven by different forces. Drivers of the MAX-MOST Gap. A first step towards understanding the nature of the MAX-MOST gaps is to study whether they are transitory or permanent. Sluggishness in the extension of the use of new technologies could cause a transitory MAX-MOST gap.21 Alternatively, the gap could reflect persistent factors that induce establishments to under- utilize the most sophisticated technologies available in a function. We consider the subsample of functions where establishments have adopted top-tier technologies, marked in bold in the grids presented in subsection A.1. For these tech- nologies, FAT collects information on the year of adoption in the business function. We then divide this sample into two groups: those where M AXf,j = M OSTf,j and those where M AXf,j > M OSTf,j . For each group, we examine the distribution of the years since adopt- ing the top-tier technology in the function.Figure 7 shows that the distributions are similar, indicating that time is not a significant factor in closing the MAX-MOST gap, and that this gap is persistent. We continue exploring the nature of the MAX-MOST gaps by considering three potential 21 As in vintage capital models where establishments slowly replace obsolete technologies embodied in old capital as it depreciates (e.g., Benhabib and Rustichini, 1991). 22 drivers. First, establishments may struggle to extend the use of sophisticated technologies due to a lack of necessary worker skills, making human capital a limiting factor. Second, sophisticated technologies are often embodied in physical capital, and restricted access to capital may create a gap. Third, the gap may result from managerial mistakes rather than constraints. Worse managers or managers with imprecise or biased perceptions of their establishment’s technological sophistication may tend to underutilize available technologies. We define GAPj as the number of functions with a MAX-MOST gap relative to the number of functions where the establishment uses multiple technologies. The establishment’s human capital is measured by the fraction of college-educated employees, Hj . Restricted access to credit is measured by a dummy variable that reflects whether the establishment has been denied a loan application within the last year, N OLoanj . We study the role of managerial mistakes with two different variables. The first is the management z-score that reflects the quality of management practices. The second is the manager’s bias in technology perception, Biasj , measured by the difference between the manager’s assessment of the technology sophistication in the establishment and the actual technology sophistication.22 We explore the relevance of these factors for the MAX-MOST gap by estimating the following specification: a GAPj = αs +αc +β0 ∗M ultiplej +βa ∗Dj +βh ∗Hj +βl ∗N OLoanj +βm ∗z −scorej +βb ∗Biasj +uj , (7) where αs are two-digit sector fixed effects, αc are country fixed effects, M ultiplej is the fraction of BFs where the establishment uses multiple technologies, and Dj a are a set of dummy variables that capture the establishment age. The estimates, reported in Table 5, support all the proposed drivers of the MAX-MOST gap. Having a loan rejected, a lower fraction of college educated workers, and a positive bias in the perceived sophistication of technology and worse management practices are all positively associated with the fraction of functions with a MAX-MOST gap in the estab- lishment. Additionally, an establishment’s age of at least 11 years is associated with a lower frequency of the gap. These estimates are robust to controlling by the fraction of functions where multiple technologies are used by the establishment. 4.3 Taking Stock The way technology is conceptualized in production is crucial for many economic models, as it influences their predictions about optimal technology choices and the relationship between 22 The specific question in FAT asks managers to rate the technology in their establishment relative to other establishments in the world. We convert this score to a 1-5 scale and compare it to S j which is the simple average of M AXj and M OSTj . 23 technology and productivity in establishments. The two main paradigms for this are Romer’s (1990) love-for-variety model and Aghion and Howitt’s (1992) quality ladder model. The findings in this section provide an empirical benchmark for evaluating these models and offer guidance on necessary extensions to more accurately depict technology sophistication at the business function level. The fact that establishments typically use multiple technologies per business function and do not abandon less sophisticated technologies indicates that the technologies in the grid are not perfect substitutes. The persistent gap between MAX and MOST suggests that establishments face significant costs in extending the use of sophisticated technologies already in place, beyond the costs of adopting new technologies. Our evidence indicates that these costs may be influenced by the establishment’s access to skilled workers and capital. Additionally, managerial errors, reflected in worse management practices and in the bias in the manager’s assessment of his establishment’s sophistication, contribute to this gap. These factors create a wedge between MAX and MOST, as shown by the small fraction of MOST’s variance explained by MAX and their differing associations with NUM. Notably, the MAX-MOST gap is larger in lower- income countries. Existing frameworks of technology in production ignore the variation in the intensity of use as an important determinant of the technology sophistication of the business func- tion. This omission limits their capacity to describe technology sophistication in the es- tablishment and, based on the evidence provided in Section 6, their capacity to study the cross-establishment relationship between technology and productivity. 5 Technology Sophistication Across Establishments We move from the business function to the establishment level to study the sophistication of technology across establishments. We are interested in two issues: the variation in technology sophistication across establishments and the association between technology sophistication and establishment characteristics. Technology Sophistication in the Establishment. We measure technology sophisti- cation in an establishment by averaging the technology sophistication across the business functions conducted in the establishment. This measure omits the sophistication of tech- nologies used in functions that the establishment outsources to other establishments. This omission is not important if establishments outsource a small number of functions. As dis- cussed in section 2, the pre-pilot, together with the answers to the technology questions in 24 FAT, strongly suggest that GBFs are conducted in-house in an overwhelming majority of es- tablishments. Similarly, 87% of the relevant sector-specific business functions are conducted in-house. Therefore, the technology sophistication of in-house functions is a good proxy for the sophistication of the technologies establishments have access to both directly and indirectly via the sourcing of functions. Reassuringly, all the establishment-level findings we present next are robust to controlling for the fraction of functions an establishment conducts in-house (Cirera, Comin and Cruz, 2024). Cross-establishment Variance in Technology Sophistication. Table 6 reports the key statistics of the cross-establishment distribution of technology sophistication. There is a large variation in technology sophistication across establishments. The standard deviation of M AXj is 0.76 and the difference between the sophistication of the establishments in the 80th and 20th percentile of the distribution (p80-p20 gap) is 1.28. The standard deviation of M OSTj is 0.63, and the p80-p20 gap is 1.16. Technology sophistication varies significantly across countries. The difference between the average sophistication in the countries with highest and lowest levels are 1.53 for MAX and 1.01 for MOST. Figure 8 studies the relationship between technology sophistication and per capita income across countries. There is a strong positive correlation between per capita income and both measures of technology sophistication. For MAX the correlation is 0.78 and for MOST it is 0.94. Technology sophistication also varies significantly within sectors. The standard deviation of M AXj within sectors ranges from 0.89 in agriculture to 0.69 in manufacturing, while for M OSTj it ranges from 0.68 in agriculture to 0.62 in manufacturing. Note that for both MAX and MOST, the sector with largest cross-establishment dispersion in technology sophistication is agriculture and the sector with smallest is manufacturing. Technology sophistication also varies significantly within countries. For example, mea- suring the within-country dispersion in technology sophistication by the difference in sophis- tication between the establishments in the 80th and 20th deciles in a country, we find that the average p80-p20 gap across countries is 2.17 for MAX and 1.56 for MOST. However, this gap varies considerably across countries. Figure 9 shows the relationship between within- country dispersion in technology sophistication and per-capita income. The p80-p20 gap in a country is positively associated with per capita income. However, the strength of the association differs significantly between MAX and MOST. While the correlation of income with the p80-p20 gap of MAX is 0.33, for MOST it is 0.95.23 23 If we measure the within-country dispersion in technology sophistication by the standard deviation, the correlations with (log) per capita income are 0.15 for MAX and 0.7 for MOST. 25 Cross-establishment Correlates of Technology Sophistication. We explore the es- tablishments’ characteristics that are associated with technology sophistication by estimating the following specification: Sj = αc + αs + β ∗ Xj + uj (8) where Sj = {M AXj , M OSTj }, αc and αs denote country and 2-digit sector fixed effects, respectively, and Xj reflects the characteristics of the establishment including fraction of employees with college degree, quality of management practices, size, age, exporter, multi- national, and multi-establishment status. Table 7 reports the estimates for M AXj (column 1) and M OSTj (column 2). We find that both measures of technology sophistication are positively associated with employees’ human capital, the quality of management practices, larger establishment size, exporter, multinational, and multi-establishment status, and they have an inverted U-shape relation- ship with establishment age. 6 Technology Sophistication and Productivity The relationship between technology and productivity is central to several important liter- atures. It is crucial to study the drivers of the large differences in productivity we observe across establishments and countries (e.g., Klenow and Rodríguez-Clare, 1997; Bartelsman, Haltiwanger and Scarpetta, 2013; Syverson, 2011). These cross-country differences in pro- ductivity are even more pronounced among agricultural establishments, as highlighted by the literature on the agricultural productivity gap (Caselli, 2005). A natural hypothesis is that the cross-establishment variation in productivity reflects differences in technology sophistication across establishments. This answer prompts the question of why establishments implement technologies that differ so much in sophistication. The literature on appropriate technology suggests that this may be the case because estab- lishments in low-income countries do not benefit from using sophisticated technologies as much as those in high-income economies. One reason for the cross-country heterogeneity in the marginal product of technology sophistication is that more advanced technologies may require complementary inputs that are relatively scarce in low-income countries. Note in any case that technology inappropriateness cannot explain the enormous variation in technology sophistication we observe within countries. In this section, we use standard productivity regressions to explore the relationship be- tween technology sophistication and productivity across establishments, thereby shedding light on these literatures. 26 6.1 Productivity Regressions To explore the relationship between productivity and technology sophistication, we estimate variations of the following productivity regression: Yj = αc,s + βk ∗ Kj + βh ∗ Hj + γ ∗ Sj + θ ∗ Xj + uj (9) where the dependent variable is the log of sales per worker in establishment j , Kj is the log of the book value of capital per worker, Hj is the percentage of workers in the establishment with a college degree, Sj represents measures of technology sophistication in the establishment, Xj is a vector of controls, αc,s reflects various combinations of 2-digit sector and country dummies (typically not interacted), and uj is classical measurement error.24 Columns 1 to 3 of Table 8 report the estimates of specification (9) with technology sophistication measured by the average of M AXj and M OSTj , denoted by S j , and with both country and sector dummies (column 1), only sector dummies (column 2) and country- specific sector dummies (column 3). In all three specifications we find a large and positive coefficient of S j in the productivity regression. In the baseline, with country and sector effects, an increase by one point in technology sophistication is associated with an increase in the (log) productivity of the establishment by roughly 0.5. The point estimate is roughly the same when allowing for country-specific 2-digit sector effect. However, the estimated coefficient of technology sophistication increases to 0.63 when excluding the country-fixed effects, suggesting that differences in technology sophistication are even more relevant to account for cross-country than within-country differences in productivity. Since the establishment’s productivity is measured by its sales per worker, one may wonder whether the association between technology sophistication and productivity is driven by the association with establishments’ prices or with establishments’ output per worker. FAT does not contain information on the prices of the goods and services produced by each establishment. However, it collects information on the markup charged for the main good or service sold by the establishment.25 We ascertain the relevance of markups in the relationship between technology sophistication and productivity by including the markup as a control 24 The estimates are robust to measuring productivity as value added per worker, using the log of sales as dependent variable including the log of employment as a control, or to calibrating the coefficients of capital and labor to the average sectoral share of the compensation to employment and capital in total sales (e.g., De Loecker and Syverson (2021)). 25 This information is collected for Croatia, Chile, Brazil, Georgia, Vietnam, India, Cambodia, Bangladesh, Senegal and Ethiopia. As a result, controlling for markups reduces the sample from 13046 to 8553 estab- lishments. The point estimate of the coefficient of S j in the baseline specification for the subsample of establishments with markup information is 0.52, very similar to the point estimate in column 1. 27 variable. Column 5 of Table 8 reports the estimates. The markup in the main product is positively associated with sales per worker. However, controlling for the markup does not change the point estimate or the significance of the coefficient of S j in the productivity regression. This finding supports the conclusion that the relationship between technology sophistication and productivity across establishment operates through output per worker rather than through prices. Recent studies in productivity have highlighted the role of managerial practices (See Bloom and Van Reenen (2007) and Bloom et al. (2019). We explore the role of managerial practices in the relationship between technology sophistication and productivity by includ- ing the management practices z-score as a control. Consistent with Bloom et al. (2013), we estimate a positive coefficient for management practices (column 5). Its magnitude is relatively modest, as a one standard deviation increase in the management practices score is associated with an increase in establishment productivity by 6.2 percentage points. We further explore the possibility advanced by Bloom, Sadun and Reenen (2012) that technol- ogy sophistication and management practices are complementary. To this end, we introduce in the specification an interaction between S j and a dummy that takes the value of 1 if the management score is above the median (column 6). We find that the coefficient of this interaction variable is positive and significant, suggesting that a key role of managers is the proper implementation of more sophisticated technologies. Importantly, controlling for the quality of management practices has little bearing on the estimated coefficient of tech- nology sophistication, demonstrating the robustness of the relationship between technology sophistication and establishment productivity. Linearity. The productivity regressions can help assess the appropriateness of the cardi- nalization used to construct the technology sophistication measures from the ordinal infor- mation extracted from FAT. An appropriate cardinalization of an ordinal variable is one that accurately captures its projection into a relevant cardinal variable. For technology sophis- tication, the most relevant variable is the (log) of productivity at the establishment level. Therefore, we can assess the appropriateness of the linear cardinalization used to construct M AXf,j and M OSTf,j by exploring whether the relationship between (log) productivity and the measures of technology sophistication across establishments is approximately linear. Table 9 explores the linearity of the relation between S j and productivity. Column 1 reports the estimates of the productivity regression, allowing the coefficient of S j to differ between establishments ranked above or below the median sophistication level. We find that the coefficient of the interaction between S j and the "above median sophistication" dummy is negative but it is quantitatively small, and significant only at the 10% level. 28 Column 2 introduces greater flexibility in the specification by replacing the S j with three dummies that reflect whether the establishment’ sophistication falls in one of three intervals: [1.5-2.5), [2.5,3.5), and [3.5,5], leaving out the interval [1-1.5). These intervals are constructed so that they span the entire range of S j , and each contains a significant portion of the establishments in the sample. The estimated coefficients of these dummies imply that the increments in (log) productivity associated with a unitary increase in average sophistication are .42 (with an standard error of 0.05) when moving from the first to the second interval, .5 (s.e. 0.06) when moving from the second to the third, and .39 (s.e. 0.08) when moving from the third to the fourth.26 These estimates demonstrate that the slope of the relationship between S j and (log) productivity is roughly constant, and therefore, well approximated by a linear relationship. This finding reassures us that the linear cardinalization used to construct the technology sophistication measures accurately represents the mapping from ordinal technology sophis- tication measures to establishment productivity. Dimensions of technology sophistication. The variable S j aggregates different dimen- sions of technology sophistication in the establishment. It includes both the sophistication of technologies in general and in sector-specific functions, as well as the M AXj and M OSTj measures. Next, we unpack S j to explore which dimensions of technology sophistication most significantly impact the relationship between S j and productivity across establishments. We start by decomposing S j into the average sophistication across the GBFs (S GBF,j ) and across the SSBFs (S SSBF,j ). Columns 1 and 2 of Table 10 report the estimates that result from replacing S j with S GBF,j and S SSBF,j in specification (9). To ensure consistency in the coverage, in this particular exercise, we restrict the sample to establishments in sectors where technology information is recorded for both GBFs and SSBFs. We find that establishment productivity is strongly and positively associated with technology sophistication in both GBFs and SSBFs, but the association is stronger for GBFs with the coefficient for (S GBF,j being roughly three times larger than that for S SSBF,j . Next, we separate S j into M AXj and M OSTj . In the specification with country and sector fixed effects (column 3), both coefficients are positive and significant, indicating that both adopting new technologies and expanding the use of existing sophisticated technologies are associated with higher productivity. However, the coefficient for M OSTj is six times larger than the coefficient for M AXj , suggesting that the expansion of the use of existing technologies is much more relevant for productivity than the adoption of new technologies to the business function. This asymmetry is even more pronounced across countries. In the 26 The increments in average S j for each consecutive pair of intervals are 0.742, 0.881, 0.924. 29 specification without country effects (column 4), the coefficient for M OSTj increases to 0.74 while M AXj becomes insignificant. Interpretation. In section 4, we showed that MOST and MAX represent distinct dimen- sions of technology upgrading. The estimates in Table 10 indicate that these processes are differently associated with productivity. The stronger association of M OSTj with produc- tivity as compared to M AXj has significant implications. Positively, it suggests the need for theoretical frameworks that better link technology sophistication with productivity, empha- sizing M OSTj . Normatively, it highlights that current innovation and technology policies, which focus on increasing M AXj , should have a broader scope and also aim at expanding M OSTj . Additionally, the differential association between establishment productivity and M AXj and M OSTj can shed light on the interpretation of the regression coefficients. In this paper, we avoid drawing causal interpretations from the associations between variables. Specifi- cally, the estimates of the productivity regressions in Table 8 are consistent both with a productivity enhancing effect of sophisticated technologies and with an effect of establish- ment productivity on the return to implementing more sophisticated technologies. However, under this second interpretation, productivity should have relatively symmetric effects on the returns to adopting new technologies and to extending the use of existing technologies. Therefore, the strong asymmetry in the coefficient estimates for M AXj and M OSTj in Ta- ble 10 is more consistent with a causal effect of the extension of the use of sophisticated technologies on establishment productivity. 6.2 Development Accounting Next, we use the estimates from the productivity regressions to conduct development ac- counting exercises. Specifically, we compute how much of the variation in productivity and revenue-based total factor productivity (TFPR) across establishments can be accounted for by differences in technology sophistication. To calculate the contribution to productivity dispersion, we regress the (log) of sales per worker and technology sophistication (S j ) on the country and sector dummies included in the relevant specification of (9). We then residualize these variables and calculate the gap between the 10th and 90th percentiles. The contribution of factor technology sophistication to cross-establishment differences in productivity results from multiplying the 10-90 gap in residualized S j by its coefficient in the productivity re- gression and dividing by the 10-90 gap in residualized productivity. To compute the contribution to T F P R dispersion, we residualize (log) sales per worker and S j by the appropriate country and sector dummies, as well as by Kj and Hj . After 30 obtaining the 10-90 gaps in the residualized productivity and technology sophistication vari- ables, we follow the same procedure as before to determine the contribution of technology sophistication to the dispersion in TFPR. Table 11 reports the results from the development accounting exercises. The first three rows correspond to the productivity regressions reported in the first three columns of Table 8. In the baseline specification with country and sector fixed effects, differences in technology sophistication account for 23% of the differences in productivity and 24% of the differences in TFPR across establishments. Excluding country effects allows us to study these contri- butions both within and across countries. In this case, differences in productivity account for 26% of productivity differences and 31% of TFPR differences across establishments. 6.3 The Agricultural Productivity Gap Cross-country differences in productivity are roughly twice as large in agriculture than in non-agricultural sectors (Caselli, 2005). The FAT dataset is consistent with this so-called agricultural productivity gap as the gap between the (log) productivity of establishments in the 90th and 10th deciles is 5.91 in agriculture, compared to 4 in services. This implies that the 90-to-10 productivity ratio is 6.75 times (i.e. exp(1.91)) larger in agriculture than in services. To study the role of technology sophistication in the agricultural productivity gap, we re- estimate the productivity regression (9) separately for each one-digit sector. The estimates, reported in Table 12 show a strong positive association between technology sophistication and productivity across all three sectors. However, the coefficient of technology sophistication varies significantly across sectors, being largest in agriculture and smallest in services. Rows 4-9 of Table 11 report the contribution of technology sophistication to the dispersion in productivity and TFPR across establishments in each sector. Technology sophistication accounts for 50% of the dispersion in TFPR in agriculture, 30% in manufacturing and 28% in services. Within countries, it accounts for 33% of TFPR dispersion in agriculture, 26% in manufacturing and 24% in services. We find similar contributions to the cross-establishment dispersion in productivity (see column 1 of Table 11). Differences across sectors in the contribution of technology sophistication to cross-establishment dispersion in productivity have implications for the agricultural productivity gap. In agricul- ture, technology sophistication accounts for a 90-10 log-productivity gap that is 1.05 points larger than in services. This means that over half of the agricultural productivity gap (1.05 out of 1.91 log-points) can be accounted for by differences in technology sophistication across establishments. 31 6.4 Appropriate Technology The appropriate technology hypothesis has conjectured that establishments in poor countries do not extensively use sophisticated technologies because the scarcity of human and physical capital limits the potential productivity gains that sophisticated technologies embody (e.g., Basu and Weil, 1998; Acemoglu and Zilibotti, 2001). To formalize this hypothesis, suppose that the productivity of an establishment is given by Yj = Ac eS j , while the cost of implement- C ing technology with sophistication S j is Cj (S j ) = 2j e2S j . In this formulation, the marginal product of technology sophistication, Ac , may vary across countries reflecting the relative abundance of productive factors that are complementary to more sophisticated technologies. The parameter that captures the marginal cost of implementing more sophisticated tech- nologies, Cj , potentially varies across establishments. Establishment j in country c chooses to implement a sophistication level S j = ln(Ac /Cj ). Note that both Ac and Cj affect the sophistication of technologies implemented. However, the marginal product of technology sophistication only depends on Ac . This insight allows us to explore the inappropriateness of more sophisticated technologies by studying whether the estimate of the marginal product of technology sophistication is larger in high- than in low-income countries. To study this prediction, we split the FAT sample between the high-income countries (South Korea, Poland, and Croatia) and the rest. Th e latter group includes countries clas- sified by the World Bank as low- and middle-income, is referred to low-income for brevity. We examine whether the coefficient of technology sophistication in the productivity regres- sions differs between these two groups. Table 13 presents the results. Column 1 includes a dummy for high income countries interacted with S j . Columns 2 and 3 estimate separate productivity regression for the establishments in high- and low-income countries, allowing all the coefficient and the sector dummies to vary between the two subsamples. Column 4 replaces S j by four dummies based on the average sophistication of the establishment and allows the coefficients to vary between the two groups of countries. In all specifications, we find that the coefficient of technology sophistication in the productivity regressions is not smaller for the sample of low-income countries than for the high income. This suggests that Ac is not lower in low-income countries than in high-income countries. A potential concern with this interpretation is that the lack of a differential association between productivity and technology sophistication in high- vs. low-income country estab- lishments may be due to an omitted variable. This could be the case if the omitted variable is more strongly correlated with either technology sophistication or productivity in low-income economies. One possible such variable is access to finance. Omitting this variable in the productivity regression could result in a larger coefficient for technology sophistication in low-income economies than if it were properly controlled for. This bias could mask a ‘true’ 32 lower value of Ac in low-income economies. To explore whether the estimates in columns 1-4 reflect omitted variable bias or correctly reflect that Ac is not lower in low-income economies, we split the sample of establishments along potential proxies for the omitted variable. We then examine if there is a differential association between S j and productivity within the subsamples for high- and low-income economies. If no differential association is observed, the case for omitted variable bias is weakened. We consider two proxies for the omitted variables: the establishments’ human capital and size, respectively measured by the fraction of college-educated workers and the number of employees. We split establishments between those above and below the median fraction of college-educated workers (columns 5-6 of Table 13) and those above and below the (un- weighted) median number of employees (columns 7-8 of Table 13). In all four subsamples, we find that the coefficient of S j is not higher in high-income economies than in low-income economies. This finding suggests that the failure to find a stronger association between technology sophistication and productivity in high-income countries is unlikely to be due to the omission of variables that affect productivity more in low-income economies. This leads us to conclude that our findings suggest that more sophisticated technologies are generally appropriate for use in all countries, regardless of their development level. 7 Conclusions This paper presents a new approach to comprehensively characterize the technologies used in an establishment. Introduces a tool, the grid, that describes the key business functions involved in production and the possible technologies to perform the main tasks in each func- tion. We have implemented this methodology and assembled a dataset covering over 21,000 establishments in 15 countries at all stages of development. An exploration of the FAT dataset has uncovered three main findings. First, the most widely used technology in a busi- ness function (MOST) typically is not the most sophisticated technology available (MAX). This gap between MAX and MOST is not transitory. It reflects the different nature and dynamics of the two upgrading processes: adoption versus extension of the use of an adopted technology. Second, there are large differences in technology sophistication across establish- ments. Factors associated with greater technology sophistication include the establishment size, the human capital of its employees, the quality of the managerial practices, being an exporter, and being part of a multinational or a multi-establishment firm. Third, there is a strong and robust cross-establishment association between technology sophistication and productivity both within and between countries. This relationship (i) is linear, (ii) is much 33 stronger for MOST than for MAX, (iii) accounts for more than 30% of the differences in productivity between establishments, and (iv) for 50% between agricultural establishments, and (v) the association between establishment productivity and technology sophistication is not weaker in low- than in high-income countries. We plan to build on the methodological and empirical contributions of this paper in various directions. First, we intend to extend FAT to collect data in more countries and also to create grids for tasks in new sectors that allow us to provide a detailed technological account of more establishments. Second, the distinct nature of MAX and MOST documented in this paper deserves scrutiny. On the theoretical side, we plan to develop frameworks that connect produc- tivity to the range of technologies used but also to how intensively they are used and that rationalize the observed gap between MAX and MOST. On the empirical front, it seems of first-order importance to assess the relevance of the possible explanations for the gap. A separate issue that we have overlooked in this paper is the aggregation of technology sophistication across business functions into an establishment-level sophistication index. In this paper, we have taken the reasonable shortcut for a descriptive exercise of construct- ing establishment-level measures of technology sophistication as the simple average of the function-level sophistication measures. However, developing a theory that rationalizes the variation between the functions of an establishment in technology sophistication and that provides a foundation for the construction of measures of technology sophistication at the establishment level would be a major development. In this paper, we have intentionally avoided studying the range of business functions conducted in establishments (i.e. the horizontal dimension of the grid). This topic, along with developing models that rationalize the relationship between technology sophistication and the limits of the establishment and its specialization from a task perspective, warrants separate attention. 34 References Acemoglu, Daron, and David Autor. 2011. “Chapter 12 - Skills, Tasks and Technologies: Implications for Employment and Earnings.” In . Vol. 4 of Handbook of Labor Economics, , ed. David Card and Orley Ashenfelter, 1043 – 1171. Elsevier. Acemoglu, Daron, and Fabrizio Zilibotti. 2001. “Productivity Differences.” The Quar- terly Journal of Economics, 116(2): 563–606. Acemoglu, Daron, and Pascual Restrepo. 2018. “The race between man and machine: Implications of technology for growth, factor shares, and employment.” American economic review, 108(6): 1488–1542. Acemoglu, Daron, Gary W Anderson, David N Beede, Cathy Buffington, Eric E Childress, Emin Dinlersoz, Lucia S Foster, Nathan Goldschlag, John C Halti- wanger, Zachary Kroff, et al. 2022. “Automation and the workforce: A firm-level view from the 2019 Annual Business Survey.” NBER Working Paper No. 30659. Aghion, Philippe, and Peter Howitt. 1992. “A Model of Growth Through Creative Destruction.” Econometrica, 60(2): 323–351. Akerman, Anders, Ingvil Gaarder, and Magne Mogstad. 2015. “The Skill Comple- mentarity of Broadband Internet.” The Quarterly Journal of Economics, 130(4): 1781–182. Autor, David H, and David Dorn. 2013. “The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market.” American Economic Review, 103(5): 1553–97. Bartel, Ann, Casey Ichniowski, and Kathryn Shaw. 2007. “How Does Information Technology Affect Productivity? Plant-level Comparisons of Product Innovation, Process Improvement, and Worker Skills.” The Quarterly Journal of Economics, 122(4): 1721– 1758. Bartelsman, Eric, John Haltiwanger, and Stefano Scarpetta. 2013. “Cross-Country Differences in Productivity: The Role of Allocation and Selection.” American Economic Review, 103(1): 305–334. Basu, Susanto, and David N. Weil. 1998. “Appropriate Technology and Growth.” The Quarterly Journal of Economics, 113(4): 1025–1054. Battisti, Giuliana, and Paul Stoneman. 2003. “Inter-and intra-firm effects in the diffu- sion of new process technology.” Research Policy, 32(9): 1641–1655. Battisti, Giuliana, Heinz Hollenstein, Paul Stoneman, and Martin Woerter. 2007. “Inter And Intra Firm Diffusion Of Ict In The United Kingdom (Uk) And Switzerland (Ch) An Internationally Comparative Study Based On Firm-Level Data.” Economics of Innovation and New Technology, 16(8): 669–687. Behaghel, Luc, Bruno Crépon, Marc Gurgand, and Thomas Le Barbanchon. 35 2015. “Please Call Again: Correcting Nonresponse Bias in Treatment Effect Models.” The Review of Economics and Statistics, 97(5): 1070–1080. Benhabib, Jess, and Aldo Rustichini. 1991. “Vintage capital, investment, and growth.” Journal of Economic Theory, 55: 323–339. Bertrand, Marianne, and Sendhil Mullainathan. 2001. “Do People Mean What They Say? Implications for Subjective Survey Data.” American Economic Review, 91(2): 67–72. Bloom, Nicholas, and John Van Reenen. 2007. “Measuring and Explaining Man- agement Practices Across Firms and Countries.” The Quarterly Journal of Economics, 122(4): 1351–1408. Bloom, Nicholas, Benn Eifert, Aprajit Mahajan, David McKenzie, and John Roberts. 2013. “Does management matter? Evidence from India.” The Quarterly journal of economics, 128(1): 1–51. Bloom, Nicholas, Erik Brynjolfsson, Lucia Foster, Ron Jarmin, Megha Patnaik, Itay Saporta-Eksten, and John Van Reenen. 2019. “What Drives Differences in Management Practices?” American Economic Review, 109(5): 1648–1683. Bloom, Nicholas, Raffaella Sadun, and John Van Reenen. 2012. “Americans do IT better: US multinationals and the productivity miracle.” American Economic Review, 102(1): 167–201. Bloom, Nicholas, Renata Lemos, Raffella Sadun, Daniela Scur, and John Van Reenen. 2016. “International Data on Measuring Management Practices.” Amer- ican Economic Review, 106(5): 152–56. Boothby, Daniel, Anik Dufour, and Jianmin Tang. 2010. “Technology adoption, train- ing and productivity performance.” Research Policy, 39(5): 650–661. Special Section on Government as Entrepreneur. Bresnahan, Timothy F., Erik Brynjolfsson, and Lorin M. Hitt. 2002. “Informa- tion Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence.” The Quarterly Journal of Economics, 117(1): 339–376. Brynjolfsson, Erik, and Lorin M. Hitt. 2000. “Beyond Computation: Information Tech- nology, Organizational Transformation and Business Performance.” Journal of Economic Perspectives, 14(4): 23–48. Caeyers, Bet, Neil Chalmers, and Joachim De Weerdt. 2012. “Improving Consump- tion Measurement and Other Survey Data through CAPI: Evidence from A Randomized Experiment.” Journal of Development Economics, 98(1): 19–33. Caselli, Francesco. 2005. “Accounting for Cross-country Income Differences.” Handbook of Economic Growth, 1: 679–741. Chung, Sunghoon, and Minho Kim. 2021. “How smart is a ’smart factory’ ?: Drivers 36 and effects of factory smartization.” Cirera, Xavier, Diego A Comin, and Marcio Cruz. 2022. Bridging the technological divide: Technology adoption by firms in developing countries. World Bank Publications. Cirera, Xavier, Diego A Comin, and Marcio Cruz. 2024. “Anatomy of Technology and Tasks in the Establishment.” National Bureau of Economic Research Working Paper 32281. Cirera, Xavier, Diego A Comin, Marcio Cruz, and Kyung Min Lee. 2020. “Tech- nology within and across Firms.” Policy Research Working Paper, n. 9476. The World Bank. Cirera, Xavier, Diego A Comin, Marcio Cruz, and Kyung Min Lee. 2021. “Anatomy of Technology in the Firm.” NBER Working Paper Series, No. w28080. National Bureau of Economic Research. Coase, R. H. 1937. “The Nature of the Firm.” Economica, 4(16): 386–405. Collins, Debbie. 2003. “Pretesting Survey Instruments: An Overview of Cognitive Meth- ods.” Quality of Life Research, 12(3): 229–238. Comin, Diego, and Martí Mestieri. 2018. “If Technology Has Arrived Everywhere, Why Has Income Diverged?” American Economic Journal: Macroeconomics, 10(3): 137–78. Davies, Stephen. 1979. The Diffusion of Process Innovations. Cambridge University Press. De Loecker, Jan, and Chad Syverson. 2021. “An industrial organization perspective on productivity.” In Handbook of industrial organization. Vol. 4, 141–223. Elsevier. Dunne, Timothy. 1994. “Plant Age and Technology use in U.S. Manufacturing Industries.” The RAND Journal of Economics, 25(3): 488–499. Gary, Pike R. 2007. “Adjusting for Nonresponse in Surveys.” In Higher Education: Hand- book of Theory and Research. 411–449. Springer. Griliches, Zvi. 1957. “Hybrid Corn: An Exploration in the Economics of Technological Change.” Econometrica, 25(4): 501–522. Grossman, Gene M, and Esteban Rossi-Hansberg. 2008. “Trading tasks: A simple theory of offshoring.” American Economic Review, 98(5): 1978–1997. Gupta, Apoorv, Jacopo Ponticelli, and Andrea Tesei. 2020. “Information, Technology Adoption and Productivity: The Role of Mobile Phones in Agriculture.” National Bureau of Economic Research. Hjort, Jonas, and Jonas Poulsen. 2019. “The Arrival of Fast Internet and Employment in Africa.” American Economic Review, 109(3): 1032–79. Hollenstein, Heinz, and Martin Woerter. 2008. “Inter- and intra-firm diffusion of tech- nology: The example of E-commerce: An analysis based on Swiss firm-level data.” Research Policy, 37(3): 545–564. 37 Hubbard, Thomas N. 2003. “Information, Decisions, and Productivity: On-board Com- puters and Capacity Utilization in Trucking.” American Economic Review, 93(4): 1328– 1353. Klenow, Pete, and Andrés Rodríguez-Clare. 1997. “The Neoclassical Revival in Growth Economics: Has It Gone Too Far?” NBER Macroeconomics Annual, 12. Mansfield, Edwin. 1963. “Intrafirm Rates of Diffusion of an Innovation.” The Review of Economics and Statistics, 348–359. Romer, Paul M. 1990. “Endogenous Technological Change.” Journal of political Economy, 98(5, Part 2): S71–S102. Ryan, B., and N. Gross. 1943. “The Diffusion of Hybrid Seed Corn in Two Iowa Commu- nities.” Rural Sociology, 8(1): 15–24. Särndal, Carl-Erik, Bengt Swensson, and Jan Wretman. 1992. Model Assisted Survey Sampling. Springer Science & Business Media. Stiroh, Kevin J. 2002. “Information Technology and the U.S. Productivity Revival: What Do the Industry Data Say?” American Economic Review, 92(5): 1559–1576. Syverson, Chad. 2011. “What Determines Productivity?” Journal of Economic Literature, 2(49): 326–365. Trajtenberg, Manuel. 1990. Economic Analysis of Product Innovation: The Case of CT Scanners. Vol. 160, Harvard University Press. Ulyssea, Gabriel. 2018. “Firms, Informality, and Development: Theory and Evidence from Brazil.” American Economic Review, 108(8): 2015–47. Williamson, Oliver E. 1979. “Transaction-Cost Economics: The Governance of Contrac- tual Relations.” The Journal of Law and Economics, 22(2): 233–261. Williamson, Oliver E. 1988. “Technology and transaction cost economics: A reply.” Jour- nal of Economic Behavior Organization, 10(3): 355–363. Wood, Lexi. 2024. “How Much Does ERP Cost?” https://softwareconnect.com/learn/erp- pricing/. World Bank. 2009. “Enterprise Survey and Indicator Surveys: Sampling Methodology.” The World Bank, Manuscript. Zeira, Joseph. 1998. “Workers, machines, and economic growth.” The Quarterly Journal of Economics, 113(4): 1091–1117. 38 Figures and Tables Figure 1: General Business Functions and Their Technologies 1. Administration (HR 2. Production or 4. Marketing/ 3. Sourcing and 6. Payment 7. Quality processes, Finance, service operations Costumer 5. Sales Procurement methods control Accounting) planning information Manual search Handwritten Handwritten of suppliers, Informal chat Direct sales at the Exchange of Manual, processes processes without (face-to-face) establishment Goods visual or centralized written database. processes Online chat without the Computers with Computers with Direct sales by support of standard software Computers with (e.g. Cash standard software phone or e-mail digital (e.g. Excel) standard WhatsApp or Internet) technologies software Sales through Cheque, Mobile Apps or Digital Mobile Apps or Structured voucher or Manual, digital platforms social media visual or platforms Online social costumer platforms or apps bank wire media, surveys written specialized Apps processes Computers with Specialized software or digital Online sales using Prepaid, credit with the Costumer support of specialized installed for demand planning, platforms Relationship external digital or debit card software demand forecast platforms (e.g. digital Managemen technologies Supplier t (CRM) Amazon, eBay, software Alibaba) Online or Relation electronic Enterprise Resource Enterprise Resource Management payment Planning (ERP) or Planning (ERP) or Statistical (SRM) not Big data Online sales (e- equivalent software equivalent software process integrated with analytics / commerce) using integrated with other integrated with other control production Artificial its own website Online back office functions back office functions planning intelligence through platform Supplier Electronic orders Relation integrated to Automated Management specialized supply Virtual or systems for (SRM) chain management cryptocurrency inspection integrated with systems production planning 39 Figure 2: Sector Specific Business Functions and Technologies in Agriculture 1. Land preparation 3. Weeding and pest 2. Irrigation 4. Harvesting 5. Storage 6. Packaging management Manual application of Manual Rain-fed herbicide, fertilizer, or Manual Product partially or totally exposed Manual packing in pesticide bags, crates or boxes Animal Aided Mechanical application Manual irrigation Animal aided Protected, but not Instruments of herbicide, fertilizer, or instruments controlled temperature pesticide Human-operated mechanical equipment Mechanized Surface Flood Irrigation Biological methods Human-operated Cold or dry controlled for packing in bags, Agriculture by Gravity machines environment crates, or boxes Fully-automated Automated Variable Rate Mechanized Irrigation by Small Application (VRA) Automated packing Agriculture combined Controlled atmosphere Pump directly linked to the (including harvester harvesting, training, Precision Drone Application in pruning, or picking Agriculture) combination with process Sprinkler or center Mechanized Constant monitoring of remote sensing pivot combined harvester products supported by digital technologies Precision Agriculture – Modified atmosphere digital tools, drones packing 40 Figure 3: M AXf,j in two establishments in retail services Note: Figure displays the technology index M AXf,j across all business functions for two individual establishments in retail services. 41 Figure 4: Distribution of technology sophistication in Food Processing (Fabrication) MAX - Senegal MAX - India MAX - Korea .8 .6 .4 .6 .3 .4 Fraction Fraction Fraction .4 .2 .2 .2 .1 0 0 0 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 MAX MAX MAX MOST - Senegal MOST - India MOST - Korea 1 .6 .4 .8 .3 .4 Fraction Fraction Fraction .6 .2 .4 .2 .1 .2 0 0 0 1 1.5 2 2.5 3 1 2 3 4 5 1 2 3 4 5 MOST MOST MOST Note: Figure displays the distribution of the technology measures M AXf,j and M OSTf,j for the fabrication function across establishments in the food processing manufacturing sector in Senegal, India, and Korea. Each column in the histograms corresponds to one technology. From least to more sophisticated these are: (i) manual processes, (ii) machines controlled by operators, (iii) machines controlled by computers, (iv) robots, (v) additive manufacturing including rapid prototyping and 3D printers, and (vi) other advanced manufacturing processes such as plasma sputtering, high speed machine, E-beam, and micromachining. 42 Figure 5: Average MAX and MOST GAP by Class of Business Function Note: Figures displays the average across relevant establishments of the share of business functions f in a given class of functions (e.g., GBF or SSBF in the sector) where M AXf,j > M OSTf,j and N U Mf,j > 1. 43 Figure 6: MAX-MOST GAP across countries Note: The country-level MAX-MOST Gap is the average MAX-MOST Gap across the establishments in the country. 44 Figure 7: Distribution of Years since Adoption of Top-tier Technologies Conditional on MAX-MOST Gap Note: Top-tier technologies are listed in Appendix A. Only functions where the establishment uses multiple technologies are considered. MAX-MOST gap is present in business function if M AXf j > M OSTf j . It is absent otherwise. 45 Figure 8: MAX and MOST across countries Note: Country-level MAX and MOST are, respectively, the weighted averages of establishment-level M AXj and M OSTj , where the weights are sampling weights. 46 Figure 9: Within-country dispersion in technology sophistication and per capita income Note: P80-p20 is the difference between the technology sophistication (MAX or MOST) of establishments in the 80th and 20th percentile of technology sophistication in the country. Percentiles are computed using establishment weights. 47 Table 1: Comparison of Technology Categories: Business Administration Handwritten pro- Standard Software Mobile apps Specialized Software Enterprise Resource cess Planning (ERP) Functionality Basic manual tasks Handles financial, Pre-designed to han- Extensive specialized Comprehensive man- (e.g., simple book- accounting, and HR dle financial, account- tools for complex finan- agement of all finance, keeping, and employee record, with manual ing, and HR record. cial, accounting, and accounting, and HR records). inputs or built in Limited scale and cus- HR management. processes. functions. tomization. Integration No. Limited. It requires Good integration ca- Integration with sys- Full integration with a manual processing or pabilities with limited tems and customizable wide range of functions additional templates. customization. reporting tools. and customization. Automation No. Limited. It requires Good automation for High-level of automa- High-level of automa- manual scripting. specific processes with tion within their spe- tion across all func- limited scale. cialized domains. tions. Experts ranking 1 2 3 4 5 Reason for Ranking Manual processes with Basic functionality and Good functionality, in- High functionality, in- Comprehensive func- minimal functionality, some level of automa- tegration, and automa- tegration, and automa- tionality, full integra- 48 no automation, and no tion. Limited integra- tion, but limited scale tion capabilities within tion, and advanced integration capabilities. tion capabilities. and customization. specialized domains. automation. Technology Example Paper Ledger Microsoft Excel, QuickBooks Online, Oracle Financials, In- SAP ERP, Oracle Net- Google Sheets BambooHR tuit QuickBooks (Desk- Suite top), Workday HCM Cost for acquiring the Negligible Microsoft Excel: QuickBooks Online: Oracle Financials: SAP ERP; Oracle technology $159.99; Google Sheets $30-$200/month; Bam- $600+/month; In- NetSuite price varies. (Free-$18/user/month) booHR: $108/month tuit QuickBooks: Average ERP $1,740- for 20 employees $1,481+/year/user; $9,330/month. Launch year Pre- 1900 Microsoft Excel: 1985; QuickBooks Online: Oracle Financials: SAP ERP: 1981; Oracle Google Sheets: 2006 2001; BambooHR: 2008 1989; Intuit Quick- NetSuite: 1998 Books: 1998; Workday HCM: 2006 ChatGPT ranking 1 2 3 4 5 ChatGPT time (task)* 5 hours 2 hours 1.5 hours 1 hour 30 min Sources: Product description on the websites of various companies, including Microsoft, Google, QuickBooks, Bamboo HR, Oracle, SAP, and Workday. Wood (2024) provides estimates of average costs for ERP software. The prompts for ChatGPT ranking and estimated time to perform a typical task – manage payroll and prepare financial statements – are available in the appendix. Table 2: Number of establishments in FAT by country, sector and size Sector Size Total Agri. Manu. Serv. Small Medium Large Bangladesh 903 - 744 159 361 232 310 Brazil* 1531 96 726 709 690 563 278 BurkinaFaso 600 80 142 378 335 187 78 Cambodia 794 - 333 461 583 142 68 Chile 1095 44 321 730 545 390 160 Croatia 710 46 272 392 472 183 55 Ethiopia 1476 149 747 580 999 330 147 Georgia 1800 196 768 836 741 632 427 Ghana 1262 85 350 827 774 382 106 India** 3242 101 1841 1300 1822 912 508 Kenya 1305 155 438 712 499 421 385 Korea 1551 128 658 765 656 569 326 Poland 1500 90 624 786 779 394 327 Senegal 1786 204 679 903 1219 395 172 Vietnam 1499 110 806 583 774 426 299 Total 21055 1485 9449 10121 11249 6158 3646 Note : * Brazil refers to state of Ceará; ** States of Tamil Nadu, Uttar Pradesh, Gujarat, and Maharashtra in India. The survey does not cover agriculture or services in Bangladesh, nor agriculture in Cambodia. In India, only the states of Gujarat and Maharashtra have agriculture included in the survey. 49 Table 3: Average level of technology measures AN U Mf,j N U Mf,j M AXf,j M OSTf,j Nf ABFs 2.1 2.1 2.7 2.0 4.8 GBFs 2.1 2.0 2.6 2.0 5.3 SSBFs 2.1 2.1 2.7 2.0 4.8 Notes : See Section 3.1 for definitions of variables. The table reports the average across the specific class of business functions, after averaging across establishments using sampling weights. Table 4: Relationship between technology measures M OSTf,j M AXf,j M OSTf,j M AXf,j 0.55∗∗∗ (0.01) N U Mf,j 0.84∗∗∗ 0.25∗∗∗ (0.01) (0.01) N 187497 187497 187497 R-squared 0.66 0.75 0.50 BF FE Y Y Y Firm FE Y Y Y Variation Explained 0.34 0.47 0.05 Notes : This table reports the regression estimates of specifications 5, 6, and 4. To compute the last row, we first residualize the dependent and indepen- dent variables by regressing them on the fixed effects, and then we regress the residuals of the dependent on those of the independent. The reported figure is the corresponding R2 . Regressions are estimated using establishment-level sampling weights. Standard errors are clustered at the establishment level. *, ** and *** denote 10%, 5% and 1% significance respectively. *, ** and *** denote 10%, 5% and 1% significance respectively. 50 Table 5: Cross-establishment Drivers of MAX-MOST Gap GAPj (1) (2) Hj -0.07∗∗∗ -0.06∗∗∗ (0.01) (0.01) N OLoan 0.04∗∗∗ 0.05∗∗∗ (0.01) (0.01) Bias 0.01∗∗∗ 0.01∗∗∗ (0.00) (0.00) Management (Z-Score) -0.00∗∗∗ -0.00∗∗∗ (0.00) (0.00) Age: 6 - 10 Years -0.01 -0.00 (0.01) (0.01) Age : 11 - 15 Years -0.04∗∗∗ -0.04∗∗∗ (0.01) (0.01) Age : 16+ years -0.02∗∗∗ -0.02∗∗∗ (0.01) (0.01) M ultiple -0.08∗∗∗ (0.01) N 16605 16605 R-squared 0.10 0.11 2-Dig. Sector FE Yes Yes Country FE Yes Yes Notes : MAX-MOST Gap, GAPj , is defined at establishment level as number of BFs with M AXf j > M OSTf j , over number of BFs with N U Mf,j > 1. The base categories are Size: Small, and Age ≤ 5 Years. Establishments weighted by sampling weights. N OLoan is a binary variable taking value 1 if the establishment has any loan rejected in the last year. Bias is defined as [(1 + 4/9 ∗ Self P erceptionScore) − Sj ], where Self P erceptionScore asks the establishments their perceived technology ranking (scale from 1 to 10) as compared to the whole world. M ultiple is the fraction of BFs in the establishment with N U Mf,j > 1. *, ** and *** denote 10%, 5% and 1% significance respectively. 51 Table 6: Cross-establishment distribution of M AXj and M OSTj M AXj M OSTj Sector Mean SD p20 p50 p80 Mean SD p20 p50 p80 Overall 2.61 0.76 1.99 2.52 3.27 2.02 0.63 1.47 1.93 2.53 Agriculture 2.63 0.89 1.81 2.51 3.73 2.07 0.68 1.37 2.03 2.68 Manufacturing 2.61 0.69 2.02 2.53 3.18 2.01 0.62 1.47 1.92 2.52 Services 2.61 0.79 1.97 2.51 3.30 2.02 0.64 1.47 1.93 2.52 Notes - Statistics are calculated using establishment-level sampling weights. 52 Table 7: Technological sophistication and establishment characteristics (1) (2) M AXj M OSTj Hj 0.40∗∗∗ 0.24∗∗∗ (0.02) (0.01) Management (Z-Score) 0.13∗∗∗ 0.11∗∗∗ (0.00) (0.00) Size: Medium 0.31∗∗∗ 0.22∗∗∗ (0.01) (0.01) Size: Large 0.66∗∗∗ 0.46∗∗∗ (0.02) (0.02) Age: 6 to 10 0.09∗∗∗ 0.13∗∗∗ (0.01) (0.01) Age: 11 to 15 0.03∗∗ 0.14∗∗∗ (0.01) (0.01) Age: 16+ 0.00 0.03∗∗∗ (0.01) (0.01) Foreign owned 0.25∗∗∗ 0.25∗∗∗ (0.02) (0.01) Exporter 0.19∗∗∗ 0.15∗∗∗ (0.02) (0.01) Multi-establishment 0.27∗∗∗ 0.18∗∗∗ (0.01) (0.01) N 17161 17161 R-squared 0.46 0.38 2-Dig. Sector FE Yes Yes Country FE Yes Yes Notes : Estimates of M AXj and M OSTj on establishment characteristics using establishment-level sampling weights. The base categories are Size: Small, and Age: ≤ 5 Years. *, ** and *** denote 10%, 5% and 1% significance respectively. 53 Table 8: Productivity and Technology Sophistication ln(Sales per worker) (1) (2) (3) (4) (5) (6) (7) Kj 0.234∗∗∗ 0.266∗∗∗ 0.223∗∗∗ 0.300∗∗∗ 0.304∗∗∗ 0.233∗∗∗ 0.231∗∗∗ (0.007) (0.007) (0.007) (0.009) (0.009) (0.007) (0.007) Hj 0.191∗∗∗ 0.585∗∗∗ 0.150∗∗∗ 0.084∗ 0.109∗∗ 0.189∗∗∗ 0.202∗∗∗ (0.040) (0.042) (0.040) (0.048) (0.049) (0.040) (0.040) Sj 0.493∗∗∗ 0.631∗∗∗ 0.504∗∗∗ 0.527∗∗∗ 0.530∗∗∗ 0.460∗∗∗ 0.422∗∗∗ (0.019) (0.020) (0.019) (0.024) (0.024) (0.020) (0.022) Markup 0.189∗∗∗ (0.052) Management (Z-Score) 0.062∗∗∗ -0.003 (0.011) (0.018) S j * D(High Management) 0.068∗∗∗ (0.015) Constant 6.118∗∗∗ 5.951∗∗∗ 7.948∗∗∗ 5.484∗∗∗ 5.208∗∗∗ 6.206∗∗∗ 6.217∗∗∗ (0.175) (0.147) (0.310) (0.203) (0.217) (0.175) (0.175) N 13046 13046 13046 8553 8553 13046 13046 R-squared 0.407 0.234 0.435 0.341 0.342 0.409 0.410 Sector FE Yes Yes Yes Yes Yes Yes Country FE Yes No Yes Yes Yes Yes Sector x Country FE Yes Notes : Estimates of specification (9). D(High Management) is a dummy that takes the value 1 if the management z-score of the establishment is above the median, and 0 otherwise. Markup is the gross markup (1+markup%)for the main product or service produced in this establishment. Columns (4) and (5) are calculated only for the sample where the markup data is collected. All regressions estimated using establishment-level sampling weights. *, ** and *** denote 10%, 5% and 1% significance respectively. 54 Table 9: Linearity of Relationship Between Productivity and Technology Sophistication ln(Sales per worker) (1) (2) Kj 0.23∗∗∗ 0.24∗∗∗ (0.01) (0.01) Hj 0.19∗∗∗ 0.24∗∗∗ (0.04) (0.04) Sj 0.56∗∗∗ (0.04) S j * D(High Sophistication) -0.03∗ (0.02) D(1.5 ≤ S j ≤ 2.5) 0.31∗∗∗ (0.04) D(2.5 ≤ S j ≤ 3.5) 0.75∗∗∗ (0.05) D(S j ≥ 3.5) 1.11∗∗∗ (0.07) Constant 6.02∗∗∗ 6.77∗∗∗ (0.18) (0.17) N 13046 13046 R-squared 0.41 0.40 Sector FE Yes Yes Country FE Yes Yes Notes : D(.) are binary variables that take the value 1 if the establishment/country satisfies the condition in parenthesis and 0 otherwise. High Sophistication represents that the establishment has above-median S j ; the different intervals for S j represent that the establishment’s S j is in the given interval. All regressions are estimated using establishment-level sampling weights. *, ** and *** denote 10%, 5% and 1% significance respectively. 55 Table 10: Productivity and Dimensions of Technology Sophistication ln(Sales per worker) (1) (2) (3) (4) Kj 0.254∗∗∗ 0.287∗∗∗ 0.235∗∗∗ 0.267∗∗∗ (0.008) (0.009) (0.007) (0.007) Hj 0.119∗∗ 0.6482∗∗∗ 0.233∗∗∗ 0.633∗∗∗ (0.053) (0.054) (0.041) (0.042) S GBF,j 0.304∗∗∗ 0.452∗∗∗ (0.028) (0.029) S SSBF,j 0.081∗∗∗ 0.170∗∗∗ (0.024) (0.026) M AXj 0.084∗∗∗ -0.037 (0.023) (0.023) M OSTj 0.433∗∗∗ 0.740∗∗∗ (0.025) (0.027) Constant 5.958∗∗∗ 5.594∗∗∗ 6.103∗∗∗ 5.971∗∗∗ (0.184) (0.156) (0.174) (0.145) N 8877 8877 13046 13046 R-squared 0.407 0.254 0.410 0.250 Sector FE Yes Yes Yes Yes Country FE Yes No Yes No Notes : All regressors are establishment-level measures. All regressions estimated using establishment-level sampling weights. *, ** and *** denote 10%, 5% and 1% significance respectively. 56 Table 11: Development Accounting Sector Country FE Contribution to log of Sales per TFPR worker Overall Y 0.23 0.24 N 0.28 0.31 Sector X Country 0.23 0.25 Agriculture Y 0.30 0.33 N 0.44 0.50 Manufacturing Y 0.27 0.26 N 0.33 0.30 Services Y 0.20 0.24 N 0.26 0.28 Notes : The table reports the contribution of S j to the cross-establishment dispersion in productivity, (log) sales per worker, and TFPR, as discussed in the text. Cross-establishment dispersion is measured by gap between the establishments in 90th and 10th deciles of the distribution of relevant variable. The first three rows correspond to the estimates from the first three columns of Table 8. Rows 4 through 9 report contributions from sectoral regressions reported in Table 12. Table 12: Productivity and Technology Sophistication - Across Sectors ln(Sales per worker) (1) (2) (3) (4) (5) (6) Kj 0.342∗∗∗ 0.442∗∗∗ 0.234∗∗∗ 0.281∗∗∗ 0.218∗∗∗ 0.244∗∗∗ (0.023) (0.024) (0.008) (0.009) (0.011) (0.012) Hj 0.507∗∗ 0.825∗∗∗ 0.164∗∗∗ 0.604∗∗∗ 0.165∗∗∗ 0.586∗∗∗ (0.227) (0.236) (0.061) (0.063) (0.058) (0.060) Sj 0.648∗∗∗ 1.023∗∗∗ 0.584∗∗∗ 0.701∗∗∗ 0.458∗∗∗ 0.587∗∗∗ (0.088) (0.086) (0.023) (0.025) (0.030) (0.030) Constant 5.402∗∗∗ 3.067∗∗∗ 6.471∗∗∗ 6.326∗∗∗ 7.502∗∗∗ 7.410∗∗∗ (0.419) (0.233) (0.124) (0.108) (0.420) (0.139) N 825 825 6032 6032 6189 6189 R-squared 0.716 0.577 0.480 0.327 0.382 0.186 2 Dig. Sector FE Yes Yes Yes Yes Yes Yes Country FE Yes No Yes No Yes No Data Agri. Agri. Manu. Manu. Serv. Serv. Notes : All regressions estimated using establishment-level sampling weights. *, ** and *** denote 10%, 5% and 1% significance respectively. 57 Table 13: Technology Appropriateness ln(Sales per worker) (1) (2) (3) (4) (5) (6) (7) (8) Kj 0.23∗∗∗ 0.11∗∗∗ 0.31∗∗∗ 0.24∗∗∗ 0.21∗∗∗ 0.26∗∗∗ 0.24∗∗∗ 0.23∗∗∗ (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) Hj 0.20∗∗∗ 0.45∗∗∗ 0.09∗∗ 0.25∗∗∗ 0.54∗∗∗ 0.48∗ 0.67∗∗∗ 0.10∗ (0.04) (0.09) (0.05) (0.04) (0.07) (0.29) (0.07) (0.05) Sj 0.51∗∗∗ 0.47∗∗∗ 0.50∗∗∗ 0.37∗∗∗ 0.63∗∗∗ 0.34∗∗∗ 0.54∗∗∗ (0.02) (0.04) (0.02) (0.04) (0.03) (0.03) (0.03) D(1.5 ≤ S j ≤ 2.5) 0.31∗∗∗ (0.04) D(2.5 ≤ S j ≤ 3.5) 0.78∗∗∗ (0.05) D(S j ≥ 3.5) 1.21∗∗∗ (0.08) S j * D(High Income) -0.07 0.08 -0.36∗∗∗ -0.13 ∗∗ -0.00 (0.04) (0.07) (0.07) (0.06) (0.06) D(2.5 ≤ S j ≤ 3.5) * D(High Income) -0.12∗∗ (0.06) D(S j ≥ 3.5) * D(High Income) -0.33∗∗∗ (0.11) Constant 6.08∗∗∗ 11.34∗∗∗ 5.31∗∗∗ 6.77∗∗∗ 6.56∗∗∗ 5.51∗∗∗ 6.55∗∗∗ 6.03∗∗∗ (0.18) (0.36) (0.18) (0.17) (0.41) (0.20) (0.22) (0.31) N 13046 2104 10942 13046 5803 6874 6383 6663 R-squared 0.41 0.30 0.38 0.40 0.39 0.47 0.43 0.42 Sector FE Yes Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Data All High Income Low Income All High Hj Low Hj High Emp. Low Emp. Notes : D(.) are binary variables that take the value 1 if the establishment/country satisfies the condition in parenthesis and 0 otherwise. The different intervals for S j represent that the establishment’s S j is in the given interval; high income is satisfied if the establishment is in one of the three high-income countries, which are - South Korea, Poland and Croatia. High Emp. and Low Emp. are categories defined on the basis of above and below median number of employees. High Hj and low Hj are based on above and below median fraction of college-educated workers. Base category for the high-income countries in column 6 is D(S j < 2.5)*D(High Income). The first two sophistication categories have been merged for high-income countries because only 49 establishments (1% of all high-income estab.) belong to the group D(S j < 1.5). All regressions are estimated using establishment-level sampling weights. *, ** and *** denote 10%, 5% and 1% significance respectively. 58 NOT FOR PUBLICATION Contents (Appendix) A The FAT Survey 60 A.1 The Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 A.1.1 The Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 A.1.2 General Business Functions . . . . . . . . . . . . . . . . . . . . . . . 62 A.1.3 Sector Specific Business Functions . . . . . . . . . . . . . . . . . . . . 63 A.1.4 Barriers and Drivers . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 A.1.5 Financial Statements and Workers . . . . . . . . . . . . . . . . . . . 77 A.2 Validation of the Ranking of Technology Sophistication . . . . . . . . . . . . 77 A.2.1 Comparison between experts’ and ChatGPT’s sophistication rankings 92 A.3 Sampling Frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 A.4 Survey Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 A.5 Measures to Minimize Bias and Measurement Error During Survey Design and Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 A.6 Ex-post Checks and Validation Exercises . . . . . . . . . . . . . . . . . . . . 105 B Construction of Measures 113 B.1 Technology Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 B.1.1 Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 C Additional Figures and Tables 118 C.1 Technological Sophistication . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 D Detailed Acknowledgments 122 59 A The FAT Survey This section provides more details on the Firm Adoption of Technologies (FAT) survey and its implementation. We start with a description of the grid of technologies in FAT. Then we describe the sampling frames used and the construction of sampling weights. We finalize describing all the tests conducted to minimize potential biases, including validation exercises ex-post implemented with external data sources. A.1 The Survey The FAT survey is a multi-country, multi-sector, representative firm-level survey. It collects information on the technologies used by firms in specific business functions that encompass the key activities that each firm conducts. Compared to existing firm-level surveys, the FAT survey covers a significantly larger number of technologies and business functions (Table A.1), and a wider range of sectors; for example, it covers agriculture distinguishing between crops and livestock. Table A.1: Coverage of Firm-Level Technology Surveys # of # of Includes Firms Surveys Technologies Business Functions in Agriculture Firm-level Adoption of Technology (FAT) Survey 305 63 Yes Manufacturing Technology Survey (MTS) 17 0 No Survey of Advanced Technology (SAT) 57 3 No Community Survey on ICT Usage and E-Commerce in Enterprises 9 0 No Information & Communication Technology Survey (ICTS) 4 0 No Annual Business Survey (ABS) 2018 Technology module 10 0 No Annual Business Survey (ABS) 2019 Technology module 5 0 No Note: The Number of technologies and business functions are computed by authors. MTS, ICTS, and ABS were conducted by the United States Census Bureau. SAT was conducted by the Statistics Canada. ICT Usage in Enterprise is conducted by EUROSTAT. The FAT survey addresses important knowledge gaps compared to other surveys mea- suring technology at the firm or establishment level. To start, the number of technologies covered is rather limited when compared to how many technologies are involved in pro- duction processes. Second, their focus on the presence of advanced technologies makes it impossible to understand how production takes place in companies without such advanced technologies. This concern is most relevant in developing countries where advanced tech- nologies have diffused less. Third, since their unit of analysis is the firm, existing studies are not designed to analyze what business functions benefit from each technology. This draw- back is particularly problematic for general technologies that can be relevant for multiple business functions. Finally, existing surveys largely omit questions about how intensively a 60 technology is employed in the firm, and therefore, they do not reveal whether a technology that is present is widely utilized or just marginally. Specifically, the FAT survey comprises five sections: • Module A– Collects general information about the characteristics of the establishment; such as sector, multi-establishment and ownership. • Module B – Covers the technologies used in seven general business functions. • Module C – Covers the use of technologies for functions that are specific to each of 11 agriculture, manufacturing, and services sectors. • Module D – Includes questions about the drivers and barriers for technology adoption. • Module E – Collects information on employment, financial statements and perfor- mance, which allow us to compute labor productivity and other measures at the es- tablishment level. A.1.1 The Grid We construct a technology grid that identifies the main business functions and the key technologies used to carry out the tasks of each business function. To design modules B and C, the survey draws upon the knowledge of experts in production and technology in various fields and sectors. These experts provided their knowledge on: i) what are the key general and sector-specific business functions, ii) what are the different technologies used to conduct the main tasks in each function, and iii) how are the different technologies related both in terms of their sophistication and the degree of substitutability between them. First, we started with desk research revising the specialized literature identifying business functions and technologies across the value chain.27 Second, for each sector, as well as for the general business functions, we hold meetings with private sector specialists at the World Bank Group to validate the initial findings and start to define the key business functions and technologies. Third, we hold meetings with Lead and Senior Economists across the World Bank Group, including the International Finance Corporation (IFC), from different fields of specialization and wide experience with sectoral projects in several countries (e.g. agriculture, manufacturing, retail, transport, health, etc.). Fourth, we hold meetings and validation exercises with external senior consultants, with wide experience on the field (e.g. 27 This process involved the revision of peer-review journals and reports from international organizations and industry associations. 61 at least 15 years), including experience with firms in developing countries as well as advanced economies. The source of external senior consultants in the last layer of quality control varied across sector. For agriculture and livestock, the validation exercise was conducted with agricultural engineers and researchers from Embrapa, an agricultural research institution from Brazil. For food processing, wearing apparel, pharmaceutical, transport, and retail, as well as for the general business functions, the team hired external consultants through a large manage- ment consultant organization. For automotive sector, the team has hired a senior consultant directly. For health, the team invited directly five physicians with different field of special- izations and practical experience in hospitals in clinics in the United States and low income countries in Saharan Africa. The validation exercise with sector specialists were organized as follows. First, the team would explain the purpose of the project, present the initial findings, and share a draft with identified business functions and technologies associated with them. The sector specialists would have between one and two weeks to reflect on the material to validate them or propose a new combination of business functions and technologies associated with them. After receiving the revised material, a second meeting with sector specialists would be organized with the FAT survey team to discuss the proposal and converge towards an updated combination of business’s functions and technologies. In what follows we describe the grids for both types of business functions. A.1.2 General Business Functions Figure 1 in Section 2 shows the 7 general business functions in FAT and the possible technolo- gies used to conduct them. The business functions identified are: business administration (HR processes, finance,accounting), production planning, procurement and supply chain management, marketing and product development, sales, payment methods, and quality control. These are business functions that in addition to being central in the functioning of the firm, are also retained in some capacity (or some tasks) within the firm. The technologies used for these business functions tend to be more available and off-the-shelf technologies, often ICT technologies. For example, for administrative processes, these range from hand- written processes (the least sophisticated) to the use of enterprise resource planning which are software that allow for real time, integrated management of the main business processes. With the help of management consultants, we identify the technologies feasible for each business function and develop similar rankings of sophistication based on the consultants understanding of the number of tasks and complexity that the technologies can handle. One important characteristic of the grid is that the sophistication rankings are not fully 62 hierarchical for all business functions. In the case of sales, for example, firms can use var- ious technologies, and while online sales are more sophisticated technologies that sales on the phone or email, there is no clear sophistication ranking between sales made in the com- pany’s website or using online platforms; both are complementary. A similar example occurs with payment methods; firms may use a variety of them, often depending on the financial infrastructure in the country. A key advantage of the grid structure is that it allows to accommodate the use of more than one technology by business function. The survey questionnaire is implemented so respondents are asked first about the use of each of the technologies in the grid. Then, for those technologies selected in each business function, the respondent is asked to identify the one that is more intensively used in implementing the tasks of the business function. Finally, when using one of the most advanced technologies, the respondent is also asked to provide the year of adoption. This allows to uncover new facts about technology adoption and use by allowing to build new measures of technology sophistication at the business function level based on extensive measures, the most sophisticated technology, and intensive measures, the technology used more intensively. It, also allows to calculate measures of diffusion lags for advanced technologies. A.1.3 Sector Specific Business Functions For the sector-specific technologies, a similar approach was used to identify key business functions and associated technologies in 12 sectors of activity across agriculture, manufac- turing, and services (including agriculture-crops; livestock; food processing; wearing apparel; leather and footwear; automotive; pharmaceuticals; wholesale and retail; transportation; fi- nancial services; health services; other manufacturing). One business function, fabrication, was also included for all manufacturing sectors. Identifying key business functions and the frontier in each sector required significant interaction with several sector specialists. These functions tend to be associated with sector-specific production processes. Here, we present all sector-specific business functions and associated technologies cov- ered by the FAT survey in the first and second phases of data collection. These figures complement the information provided in Section 2, particularly Figure 2, which describes the functions and associated technologies for SSBFs in agriculture, among SSBFs. The com- plementary information is provided for all SSBFs, including Livestock (Figure A.1), Food Processing (Figure A.2), Wearing Apparel (Figure A.3), Leather and Footwear (Figure A.4), Automotive (Figure A.5), Pharmaceutical (Figure A.6), Wholesale and Retail (Figure A.7), Transportation (Figure A.8), Financial Services (Figure A.9), Health Services (Figure A.10), 63 and Other Manufacturing (Figure A.11)). 28 3. Animal healthcare 4. Herd Management and 5. Transport 1. Breeding 2. Feeding Monitoring Manual Rapid diagnostic tests transport Breed substitution Household waste Human monitoring Non-motorized Pest sprays vehicles Natural grasslands and Inbreeding or pastures Animal-aided monitoring Crossbreeding Motorized vehicles Live-attenuated, inactivated or subunit vaccines Feedlots or grazing Artificial insemination Integrated systems: crop- system1.5" (AI) pasture; tree-crop-pasture; Specialized / tree-pasture; tree-crop climate- controlled DNA or RNA-based vaccine vehicles Molecular genetics Automated cameras and Forage Crops video Disease medication Supplementary feed to Drones grazing pastures: hay, silage, grains feed Analog tracking devices attached to animals Manufactured or mixed feed Digital tracking device attached to animal Genetically-modified feed Figure A.1: Agriculture - Livestock: Business Functions and Technologies 28 As the survey is rolled out in other countries, the number of additional sectors included in the survey is also increasing. 64 1. Input testing 2. Mixing/blending/cooking 3. Anti-bacterial 4. Packaging 5. Food storage 6. Fabrication Minimal- Manual Minimal Manual Sensory (visual, Processes processing packing in protection, smell, color, Manual process preservation bags, bottles or some exposure etc.) methods boxes to outside Machines elements controlled by Mechanical Anti-bacterial operators Review of supplier Human operated equipment requiring wash or mechanical Ambient testing on Certificate operation by humans soaking of Analysis equipment for conditions in Machines packaging in bags, closed building controlled by bottles or boxes computers Partially automated Thermal process with minimal Processing Some climate human interaction Technologies control in Robots Non-computer- Automated secured controlled testing kits process with building minimal human Fully automated interaction Additive manufacturing Other advanced process controlled methods including Fully including rapid solely by computers High-pressure automated prototyping and 3D Automated testing such or robotics processing (HPP) Fully Automated climate and printers as chromatography or or Pulsed electric with Robotics security- spectroscopy field(PEF) controlled building Other advanced manufacturing processes (e.g. laser, plasma sputtering, high speed machine, E-beam, micromachining) Figure A.2: Food Processing: Business Functions and Technologies 65 1. Design 2. Cutting 3. Sewing 4. Finishing 5. Fabrication Manual Basic Processes Manual design Manually manual and hand drawing Manual cutting sewing ironing Machines controlled by Digital or semi- Sewing Electric operators digital design machine high- using specialized Cutting machine pressure manually 2D drawing manually operated steam iron operated software Machines controlled by computers Computer Aided Semi-Automatic Semi- Design (CAD), 3D cutting machine automated Tunnel design, virtual (straight knife, sewing finisher prototyping round knife, die machines cutting machine) Robots Automatic or Automated Form Computerized sewing finishing cutting machine machines machine (no Laser: water jet, knife, other) Additive manufacturing High tech including rapid prototyping pressing and 3D printers Automatic or 3D Knitting machine Computerized cutting machine Other advanced manufacturing (Laser) processes (e.g. laser, plasma sputtering, high speed machine, E-beam, micromachining) Figure A.3: Wearing Apparel: Business Functions and Technologies 66 1. Design 2. Cutting 3. Sewing 4. Finishing 5. Fabrication Manual Basic Processes Manual design Manually manual and hand drawing Manual cutting sewing ironing Machines controlled by Digital or semi- Sewing Electric operators digital design machine high- using specialized Cutting machine pressure manually 2D drawing manually operated steam iron operated software Machines controlled by computers Computer Aided Semi-Automatic Semi- Design (CAD), 3D cutting machine automated Tunnel design, virtual (straight knife, sewing finisher prototyping round knife, die machines cutting machine) Robots Automatic or Automated Form Computerized sewing finishing cutting machine machines machine (no Laser: water jet, knife, other) Additive manufacturing High tech including rapid prototyping pressing and 3D printers Automatic or 3D Knitting machine Computerized cutting machine Other advanced manufacturing (Laser) processes (e.g. laser, plasma sputtering, high speed machine, E-beam, micromachining) Figure A.4: Leather and Footwear: Business Functions and Technologies 67 4. Plastic Injection 5. Productive Assets 1. Assembly 2. Body pressing 3. Painting 6. Fabrication Molding Management Molding of non- Machines controlled Pressing of skin panels Water-based painting visible interior plastic Breakdown Manual by operators using operators using operators components using maintenance system Processes operators Flexible Machines Pressing of structural Molding of plastic Preventative or controlled by Manufacturing Cells components using Solvent-based painting exterior body parts predictive operators (FMC) or Flexible operators using operators using operators maintenance system Manufacturing Systems (FMS) Machines Welding of main body Molding of non- controlled by using operators Water-based painting visible interior plastic Model Based computers automated using Lasers components Condition Monitoring robotics automated using robotics Computer Robots Pressing of skin numerically Solvent-based panels using robotics Molding of plastic controlled (CNC) painting automated machinery using robotics exterior body parts Additive automated using manufacturing robotics including rapid Pressing of structural prototyping and Robot(s) without components using sensing 3D printers robotics 4-9 axis computer Other advanced numerically Welding of main body using robotics manufacturing controlled processes (e.g. laser, plasma sputtering, high speed machine, E-beam, micromachining) Figure A.5: Automotive: Business Functions and Technologies 68 4. Compression, 2. Raw Material 3. Mixing & 5. Quality 1. Facilities Encapsulation (NOT 6. Packaging 7. Fabrication Weighing & Compounding Control FOR SYRUPS OR Dispensing DRY POWDERS) Unfiltered air Manual, Manual filling of Manual Manual in filling Manual titrimetric/ pills in bottles Mixing Processes space Beam Scales Compression, chromatographic analyses OR Encapsulation with dosing dies placement of Planetary syrups, powders Machines Basic air Mixers in bottles or controlled by filtration Analog Electronic pouches operators Scales OR Motorized chromatography homogenizers Compression, Encapsulation Machines HEPA air Electronic filtration Electronic For pills: Slat controlled by chromatography Counters, computers Scales High Speed, Automated with data Compression, Cottoners, High Shear acquisition Ultra HEPA Encapsulation Cappers, Granulators Robots air, Automated Labelers pressurization Weighing OR Systems Integrated Additive manufacturing control Compression, machine filling Fluid Bed including rapid Processors Encapsulation of syrups, prototyping and 3D (not with powders in printers syrups) bottles or pouches Other advanced manufacturing Automated Automated, processes (e.g. laser, Compounding Integrated plasma sputtering, high Packaging Lines speed machine, E-beam, micromachining) Figure A.6: Pharmaceutical: Business Functions and Technologies 69 1. Customer Service 2. Pricing 3. Merchandising 4. Inventory 5. Advertisement Manually selecting Handwritten record Paper based At the Store Manual Cost products keeping communication Category Management Computer databases with Call Help Desk Automated markup Radio, Billboards, TV tools manual updates Social Media (e.g. Retail Merchandising Warehouse Management Facebook, WhatsApp, or Automated promotional Systems or Digital Email or mobile phone System & bar codes similar) Merchandising Automated inventory Online requests Dynamic pricing systems Product trend analytics control (CAO) or Vendor Search Engine Marketing managed inventory or Radio-frequency identification Social Media (youTube, Chatbots Personalized pricing Fb, Twitter, Instagram) Automated Storage and Retrieval systems Big data or Artificial Intelligence Figure A.7: Wholesale and Retail: Business Functions and Technologies 70 1. Planning 2. Plan execution 3. Monitoring 4. Performance 5. Maintenance measurement Event driven at All manual paper driven Handwritten Manual process with predetermined check system (e.g. repair, the support of fax, points of load Manually monitored regulatory, licensing, information to create and reported load plans text, or phone calls. transactions insurance, warranty, performance, and parts Event driven at management) Manual process with predetermined intervals Non-specialized the support of digital with the support of Information collected software, MS platforms or mobile digital platforms or Applications: Excel, by electronic file share mobile apps Information collected by apps Word, Power Point, electronic file and shared (e.g. Email or FAX to create load plans) etc. through Email or FAX. Information Paper documentation exchanged via web- exchange on daily, weekly or monthly Computer or apps with based specialized Batch information communication intervals (e.g. invoice, Batch information collected bill of lading, transportation by software installed on collected by software protocol (e.g. Email reporting applications installed ERP to or WhatsApp) regulatory, etc.) transportation equipment by service and cost integrated with ERP to create ERP generated performance metrics load plans Information collected report fleet management Specialized software by software installed documentation interface via Internet, on the transportation Specialized software including GPS, equipment (e.g. GPS, installed on the dynamic routing E-log – Driver status, transportation Real time information by Real time information (weather, traffic), E- load monitoring) equipment. (e.g. GPS, online software interface by online software log, driver status and E-log – Driver status, with ERP to manage, interface with ERP to safety, load monitoring load monitoring) document, and report create load plans File exchange fleet asset status between ERP File exchange File exchange between ERP between ERP integrated applications and integrated integrated applications and applications and delivery equipment software applications delivery equipment delivery equipment software applications Figure A.8: Land Transportation: Business Functions and Technologies 71 2. Client 5. Operational 1. Customer services 3. Loan applications 4. Approval process identification support Teller (Face-to- Teller with Paper-based Analysts based on Writing records from face) documentation applications paper applications employees ATM Machines Online passwords Mobile/Phone Analysts based on Digital accounting application digital information On-line on company websites Online passwords and Channel partners, token devices Loan officer, paper Automated decision Digital network Mobile banking based mechanisms Digital authentication provided by Internet applications specialized and mobile apps firms Artificial intelligence or big-data analytics Biometric identity verification Blockchain Figure A.9: Financial Services: Business Functions and Technologies 72 1. Health 3. Equipment 2. Scheduling Management 4. Medication 5. Procedures appointments of patient management records Emergency Dept. Functioning Personal visit Handwritten Diagnostic Myocardial Ultrasound and paper Manual/ Paper monitoring and treatment Childbirth Trauma infarction/Stroke Post-operative Process administration of sepsis care area of medicine Functioning CT Phone call, SMS, e-mail Digital Defibrillation Information Barcode Traction ICU Treatment Cesarean System identification (closed for medicine with Section fracture) Functioning administration antibiotics MRI Specialized software or mobile to patients Coronary Blood bank app for Electronic Open Angiography appointment Health High-risk Treatment or Multivessel without automated Records with labor of Fracture coronary specialized Resuscitation, revascularizati Lab to test Pulse reminders mechanical blood and oximetry software on ventilation, urine glucose Specialized software or mobile control, and Cloud-based renal control ECG app for Electronic Health Functioning appointment with Xray Records (EHR) automated reminders and confirmation Figure A.10: Health Services: Business Functions and Technologies 73 Fabrication Manual Processes Machines controlled by operators Machines controlled by computers Robots Additive manufacturing including rapid prototyping and 3D printers Other advanced manufacturing processes (e.g. laser, plasma sputtering, high speed machine, E-beam, micromachining) Figure A.11: Other Manufacturing: Business Functions and Technologies 74 For sector-specific business functions, digital technologies tend to be embedded in other technologies that are usually at the frontier. This is a common feature, particularly in agriculture and manufacturing, and has important implications in terms of the costs of adoption and the importance of network effects. For example, among methods commonly used by agricultural firms to perform harvesting, the most basic option is to harvest manually, followed by animal-aided instruments, human-operated machines, or a single tractor with one specific function (such as a single-axle tractor), a combined harvester (machines or tractors that combine multiple functions fully operated by the worker), and combined harvester using the support of digital technologies (such as global positioning systems [GPS] or computing systems integrated with the tractor). Unlike GBFs, the application of digital technologies for harvesting requires other sophisticated equipment or machines. In addition to the possibility of computing different measures of technology sophistication for sector specific business functions, an important feature of the sector specific grid is the fact that it includes screening questions that allow for the fact that not all the business functions are carried out within the establishment. In other words, not all entries in the grid need to be implemented at the establishment or at the firm. While the tasks of most general business functions related to management and organization are usually carried out within the boundaries of the firm - either in the same establishment or in another establishment of the firm if multi establishment - some sector specific business functions can be carried out in another establishment within the same firm (insourcing), or they can be (outsourced) to a different firm. Our approach is, therefore, rooted in a view of the firm similar to Coase (1937), where firms are agents coordinating and implementing tasks. The advantage of this approach is twofold. In addition to the fact that this approach allows a better identification of technology and its use as described above, it allows to study critical questions such as the organization of the firm and tasks (Williamson, 1979), and more importantly the relationship between organizational modes, transaction costs and technological choice (Williamson, 1988). After finalizing the FAT questionnaire, we pre-piloted it in Brazil and Senegal. We personally conducted the face-to-face interviews, in collaboration with enumerators and su- pervisors trained to conduct data collection with firms from different sectors and size groups. In the pre-pilot stage, we tested if the business functions and technologies covered by the questionnaire were comprehensive and clearly understood by respondents, through detailed discussions and follow up questions with representative of firms, which led us to make the necessary adjustments to the survey. For example, we experimented with survey designs that asked about the fraction of time/output/processes that were conducted with each of the technologies in the business function. We decided against using this approach to reflect the intensity of use of technologies because it was harder for respondents to answer precisely, 75 and as a result led to a more subjective interpretation, which made the comparability of answers across business functions and companies harder to interpret. A.1.4 Barriers and Drivers In addition to the information on the technologies used by firms, the survey also collects information on potential drivers of and barriers to technology adoption. First, the survey asks whether the firm acquired new machines, equipment, and software in the last three years; and in the case of machines or equipment, whether these were leased, purchased as new or secondhand. The survey also asks questions on links to larger firms and multinationals, either via value chain linkages as a supplier or buyer, or via the CEO previous experience working in a MNE or a large firm. The survey also asks questions about access to finance and trade status. The first ques- tion is about having secured a loan in the previous three years for purchasing equipment, machinery or software. On more general access to finance, the survey asks how many times the establishment needed to borrow money to expand production but could not obtain fi- nance. On trading status, the survey asks whether the firm is an importer, an exporter; and if an exporter, what is the share of sales that is exported. A key complementary factor for technology adoption is the quality of management. The survey pays special attention to management by collecting information on the top manager’s background and on management practices. Specifically, FAT asks about the level of educa- tion attainment of the top manager in the establishment, whether she has studied abroad, and whether she has experience in multinationals. In addition, the survey contains four ques- tions about management practices. These include four questions from MOPS (Bloom et al., 2016) on the number of KPIs, the frequency with which they are monitored, the horizon of production targets and a question on the use of formal incentives. Though the information we collect on management practices is more restricted than the sixteen questions in MOPS, we have used information from the Mexico ENAPROCE survey and show that the index that emerges from the small number of variables collected is highly correlated with the full MOPS index and it captures a large fraction of the cross-firm variance in the quality of management practices.29 To investigate also the potential role of policies on technology adoption, the survey asks questions about awareness about existing public programs to support technology upgrading; 29 Specifically, we use data from Mexico ENAPROCE survey and calculate the correlation between a management quality index with the 4 questions in FAT and the overall index using all questions of MOPS that are in ENAPROCE. The correlations are 0.74 for 2015 survey and 0.73 for 2018; which suggests that with less questions we are still able to capture most of the variation in management quality. 76 and whether the firm is a beneficiary of such programs, and if so, what type of support the firm received. While the approach of the survey is as much as possible to ask factual questions, it is also important to understand the perceptions that entrepreneurs and managers have about what are the main barriers and drivers of the decisions to adopt new technologies. To this end, the questionnaire asks the respondent to select the most important obstacle and driver for adoption from a closed list of options. As barriers we include: lack of information and tech- nical skills, uncertainty about demand, cost, lack of finance, government regulations or lack of infrastructure. As drivers, we include competition, adoption by other firms, production of new products, accessing new markets, cost reductions or adjusting to regulations. The survey also asks managers to benchmark their business technology sophistication level in relation to other firms in the country, and also vis-a-vis more advanced firms internationally. This helps to understand the role of beliefs of the main managers in technology adoption decisions and potential behavioural biases and overconfidence. A.1.5 Financial Statements and Workers In addition to the information on the technologies used by firms, the survey also collects financial statements information, information on the business owners, employees, and on potential drivers of and barriers to technology adoption. Financial statements. The survey asks the establishment about its total sales, material inputs, replacement value of capital stock, energy consumption, wages and employment. This allows to construct measures of nominal value added per worker, and capital per worker. Workers. Beyond the number of employees, the survey asks questions that provide information on the education of the workers (share of workers with primary, secondary and tertiary education), and about the occupation composition of the labor force (share of Managers, Professionals, and Technicians; Clerical support workers and sales workers; Production workers and Service workers). A.2 Validation of the Ranking of Technology Sophistication To evaluate the coherence of the rankings of technology sophisticated from industry experts, we implement a multi-stage validation exercise. First, we select 14 business functions for which we compare the technologies in the grid along three dimensions that define their so- phistication and that summarize some of the arguments used by experts to justify their rankings: functionality, integration, and automation. Functionality refers to the capabilities a technology offers to handle more complex tasks, in a faster way, on a larger scale, with 77 greater accuracy and reliability. Integration reflects a technology’s ability to connect and interact seamlessly with other systems by exchanging data and coordinating processes. Au- tomation enables the technology to execute processes, make decisions, and generate outcomes independently, without human intervention. For these functions, we also document the year of the launch and the cost of each of the technologies in the grid. Though these variables do not define the sophistication of a technology per se, more sophisticated technologies tend to have been developed more recently and be more expensive. This is particularly the case for sector specific technologies used in the production of goods (e.g., agriculture and manufacturing). We relied on multiple source collect information on the functionality, integration, and automation features, as well as the launch and the cost of the technologies in the grid. In addition to broad desk research, we consult official websites with description of specific lead- ing brands supplying these technologies (e.g., Microsoft website provides detailed information on the features of Microsoft Excel, including prices, which is used as a proxy for standard spreadsheet software, one of the technologies in the grid for business administration). In addition, we validate the rankings by using AI-powered large language models. First, we asked ChatGPT to rank the technologies in the grid in terms of technology sophistication. Second, we ask ChatGPT to rank technology sophistication based on specific definitions of functionality, integration, and automation and compare those with the industry experts rankings resulted in the technology grid, based on the methodology we described in section 2. Third, we ask ChatGPT to identify a specific task for each of the 14 business functions and estimate the time required to perform the task with each technology. Tables A.1 - A.13 summarize these exercises for the following business functions: sourc- ing, marketing, sales, payment, and quality control, among GBFs; Weeding and harvesting for agriculture; Design and sewing for apparel; Anti-bacterial and packaging for food pro- cessing; Pricing and inventory for retail services; in addition to the example for businesses administration, provided in section 2. We also added for each business function a summary of two of the exercises with ChatGPT: i) The overall ranking of technology sophistication; ii) The estimated time to perform a typical task in the business function. Next sub-section pro- vides more details on these and additional exercises conducted with ChatGPT. Subsection A.2.1 describes the specific ChatGPT prompts used for these exercises and displays scatter plots showing the positive association between ChatGPT and industry expert rankings of technology sophistication. Overall, the validation exercise across 14 businesses functions strongly supports the ex- perts original ranking of sophistication. To start, within each business function, the specific features of technologies are associated with functionality, integration, and automation. 78 Table A.1: Technology Sophistication in Sourcing Manual Search Computers with Mobile Apps or Dig- SRM (not inte- SRM (integrated Standard Software ital Platforms grated) with production planning) Functionality Basic functions (e.g., Moderate features for Good features for Extensive features for Comprehensive features manually finding sup- tracking supplier infor- discovering suppliers, supplier management, for supplier manage- pliers in the yellow list) mation and managing comparing options, procurement, contract ment, procurement, procurement database. and managing basic management, and per- contract management, procurement tasks. formance monitoring. and performance moni- toring. Integration No. Limited. It requires Integrates well with Good integration with Seamless integration manual import/export other digital tools, but other procurement with production plan- functions and data en- lacks customization. tools, but not with ning and other systems. try. production planning. Automation No. Basic automation Moderate level of au- Advanced automation High level of automa- through formulas tomation for tasks such for supplier selection, tion for tasks such as and macros; exten- as supplier searches, order processing, and supplier selection, or- sive manual setup basic order processing, performance tracking. der processing, perfor- and maintenance are and communication. mance tracking, and de- required. mand forecasting. Rank 1 2 3 4 5 Reason for the Rank Least sophisticated due Low sophistication due Moderate sophisti- Sophisticated in sup- High level of sophisti- to complete lack of in- to limited integration cation with decent plier management and cation with specialized tegration and automa- and basic automation, integration and au- procurement with ad- integration for pro- 79 tion, relying entirely on mainly suitable for tomation, but mainly vanced automation, but duction planning and manual processes. small-scale or manual focused on task man- lacks integration with robust automation, procurement processes. agement and basic production planning, though focused on procurement functions. limiting its overall supplier management functionality. and production-related tasks. Technology Exam- Phone Calls, Personal Microsoft Excel (1985), LinkedIn (2003), SAP Ariba (1996), Ora- SAP Integrated Busi- ples Visits, Supplier Cata- Google Sheets (2006), ThomasNet (1898), cle Procurement Cloud ness Planning (2014), logs LibreOffice Calc (2011) Alibaba (1999) (2010), Coupa (2006) Oracle SCM Cloud (2016), Infor Nexus (2019) Cost Cost of phone calls, Microsoft Excel: LinkedIn: $59/month, SAP Ariba: Pricing SAP IBP: Starts at travel, accommodation, $150/year, Google ThomasNet: Free or depends on enterprise $30,600/year; Oracle and catalogs varies Sheets: Free or Custom Pricing, Al- size; Oracle Procure- SCM Cloud: Starts $6-$18/user/month, ibaba: Transaction fees ment Cloud: Starts at $350/month; In- LibreOffice Calc: Free (varies) at $625/user/month, for Nexus: Starts at Coupa: $2,500/month $10,000 Launch Year N/A 1985, 2006, 2011 2003, 1898, 1999 1996, 2010, 2006 2014, 2016, 2019 ChatGPT ranking 1 2 3.5 4 5 Estimated time in a 4 hours 2.5 hours 1.5 hour 1 hour 30 minutes task Source: The table draws upon various sources of information including desk research, website of various companies including Microsoft, Google, LinkedIn, ThomasNet, Alibaba, Oracle, Coupa, InforNexus, and SAP. *CHATGPT estimates the time for selecting suppliers and managing orders, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.2: Technology Sophistication in Marketing Face-to-face chat Online chat Structured surveys Customer Relation- Big Data Analytics ship Management or Machine Learning software (CRM) algorithms (AI) Functionality Collecting qualitative Real-time communica- Designing surveys, Managing customer Advanced analytics, data, gaining in-depth tion with customers, collecting structured data, tracking inter- machine learning, insights, building rap- collecting feedback, feedback, analyzing actions, automating predictive modeling, port with customers. handling customer responses. marketing campaigns, real-time data process- queries. analyzing customer ing. behavior. Integration No integration Basic integration with Moderate integration Good integration with Seamless integration CRM systems and data with data analysis tools other business systems, with multiple data analysis tools. and CRM systems. including email market- sources, CRM sys- ing, social media, and tems, and marketing sales platforms. platforms. Automation No automation Basic automation for re- Moderate level of au- High level of automa- High level of automa- sponses and data collec- tomation for survey dis- tion for tracking in- tion for data collection, tion; advanced features tribution and response teractions, segmenting analysis, and generating available in some tools. collection, but manual customers, and running insights. analysis is often re- marketing campaigns. quired. Rank 5 4 3 2 1 Reason for Ranking Least sophisticated due Effective for immedi- Useful for structured Strong integration and Highest sophistica- 80 to the absence of in- ate customer interac- feedback, but lower au- automation, making tion due to advanced tegration and automa- tion but lacks the so- tomation and integra- it highly effective for analytics, seamless tion, although valuable phistication in automa- tion limit its sophisti- comprehensive cus- integration, and high for qualitative insights. tion and integration of cation compared to top tomer management and automation for action- higher ranks. ranks. marketing. able insights. Technology Exam- In-person Interviews, WhatsApp Business, In- SurveyMonkey, Salesforce, HubSpot Google Analytics ples Focus Groups, Mystery tercom, LiveChat Qualtrics, Google CRM, Zoho CRM Shopping Forms Launch Year N/A 2018 (WhatsApp Busi- 1999 (SurveyMonkey), 1999 (Salesforce), 2014 2005 ness), 2011 (Intercom), 2002 (Qualtrics), 2008 (HubSpot CRM), 2005 2002 (LiveChat) (Google Forms) (Zoho CRM) Cost Cost of time and travel Free or $5/month Free or $25- $1,000-$10,000/month Free or $150,000/year (varies), $2,000-$5,000 (WhatsApp Business), $75/user/month (Salesforce), $1,170- per session (Focus $39-$139/seat/month (SurveyMonkey), $4,300/month (Hub- Groups), $20-$100 (Intercom), $20- Based on invitation Spot CRM), Free or per shop (Mystery $59/user/month (Qualtrics), Free or $20-$50/user/month Shopping) (LiveChat) up to $18/user/month (Zoho CRM) (Google Forms) ChatGPT ranking 1 2 3 4 5 Estimated time in a 4 hours 2 hours 1.5 hours 1 hour 30 minutes task Source: The table draws upon various sources of information including desk research, and websites of various companies including Google, Salesforce, Hubspot, Zoho, SurveyMonkey, Qualtrics, WhatsApp, Intercom, and LiveChat. *CHATGPT estimates the time for collecting customer feedback, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.3: Technology Sophistication in Sales Sales at the Sales by phone, Sales through so- Online sales us- Online sales Electronic or- establishment’s email orders, or cial platforms ing external dig- (own website) ders integrated premises sales representa- ital platforms to SCM systems tives Functionality Basic features for Features for man- Basic features Features for listing Extensive features Comprehensive fea- point of sale, inven- aging customer for creating on- products, manag- for creating and tures for managing tory management, data, tracking sales line storefronts, ing inventory, pro- managing online e-commerce, order and sales tracking. activities, and managing prod- cessing orders, and stores, processing processing, inven- automating sales uct listings, and handling shipping payments, han- tory management, processes. processing sales and returns. dling shipping, and supply chain orders. and inventory integration. management. Integration Limited integration No integration with Limited integration Good integration Good integration Seamless integra- with other systems. CRM and email with inventory with inventory with payment tion with supply systems. management and management and gateways, shipping chain, inventory, processing systems. shipping services. carriers, and inven- and logistics sys- tory management tems. systems. Automation Basic automation Moderate automa- Moderate level of Moderate level of High level of au- High level of au- for sales tracking tion for tracking automation for or- automation for or- tomation for order tomation for order 81 and inventory interactions and der processing and der processing and processing, pay- processing, inven- updates. managing sales inventory updates. inventory updates. ment handling, and tory updates, and activities. inventory updates. supply chain man- agement. Rank 6 5 4 3 2 1 Reason Basic functionality Limited sophisti- Moderately sophis- High sophistication High sophistication Most sophisticated and minimal inte- cation due to lack ticated with strong with extensive e- with e-commerce due to its compre- gration or automa- of integration, with integration and commerce function- integration and hensive functional- tion, suitable for moderate automa- moderate automa- ality, good integra- automation ity, seamless inte- traditional retail. tion focusing on tion, suitable for tion, and signifi- gration with supply sales tracking. third-party selling. cant automation. chains, and high automation. Technology Ex- Offline stores WhatsApp/ Facebook Shops Amazon Seller Cen- Shopify SAP Commerce ample Email/ tral Cloud Salesmen Cost Cost of setting up Cost of mobile, in- Facebook Shops $39.99/month + $29-$299/month Starts at stores ternet, travel selling fees $100,000/year Launch Year n/a n/a 2020 2000 2006 2013 ChatGPT rank- 1 2 3 4 4.5 5 ing Estimated time 3 hours 2 hours 1.5 hours 1 hour 30 minutes 30 minutes in a task Source: The table draws upon various sources of information including desk research, and websites of various companies including SAP, Shopify, Amazon, Facebook, and WhatsApp. *CHATGPT estimates the time for processing sales orders, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.4: Technology Sophistication in Payment Criteria Exchange of Cash(Physical Cheque, Prepaid card, Online or elec- Online payment Virtual or Cryp- goods (Barter Currency) voucher, or Debit card, or tronic payment through plat- tocurrency (Bit- Exchange Plat- bank wire at the Credit card through a bank form (PayPal) coin) forms) branch (Bank (Visa) wire (SWIFT) Cheque) Functionality Facilitating the Direct payment us- Payment through Electronic payment Secure interna- Online payment Decentralized dig- direct exchange ing coins and ban- written orders to method allowing tional financial system for send- ital currency for of goods and ser- knotes a bank to pay a payment directly messaging for ing and receiving peer-to-peer trans- vices without using specific amount from a linked bank transferring funds money, purchasing actions money from a person’s account or credit between banks goods and services account line Integration No integration No integration Basic integration Wide integration Extensive integra- Seamless inte- Good integration with banking sys- with merchants, tion with global gration with with digital wallets, tems for processing banks, and pay- banking systems e-commerce plat- exchanges, and cheques ment processors forms, banking some e-commerce systems, and other platforms financial tools Automation No automation No automation Limited automa- High level of au- High level of au- High level of High level of au- tion; manual tomation for trans- tomation for fund automation for tomation for trans- processing required action approval and transfers and trans- transaction pro- action validation record-keeping action tracking cessing, payment and record-keeping 82 notifications, and via blockchain record-keeping technology Rank 7 6 5 4 3 2 1 Reason for Basic functionality Simple functional- Limited functional- Widely accepted Strong integration High functionality, High automation, Ranking with no integration ity with no integra- ity and integration, with robust in- within global bank- broad integration but integration is or automation, the tion or automation, with minimal au- tegration and ing, with reliable across multiple less widespread least sophisticated making it the least tomation, making automation, al- automation, but platforms, and high and functionality is method in the con- sophisticated it less sophisticated though slightly less more suited for automation in pro- limited to certain text of payments than newer meth- sophisticated than large transactions cessing payments platforms ods online platforms Average Costs Transaction fees, No cost Bank fees, varies Merchant fees Bank fees, varies 2.9% + $0.30 per Transaction fees varies (1.5%-3%), annual transaction (varies), network fees ($0-$100) fees Launch Year N/A N/A N/A 1958 1977 1998 2009 ChatGPT rank- 1 2 2.5 3.5 4 4.5 5 ing Estimate time in 2.5 hours 2 hours 1 hour 30 minutes 30 minutes 12 minutes a task 4 hours Source: The table draws upon various sources of information including desk research, and websites of various companies including Bitcoin, PayPal, SWIFT, and VISA. *CHATGPT estimates the time for processing payments, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.5: Technology sophistication in Quality Control Category Manual, Visual or Human Inspection Statistical Process Automated Systems Written Processes with Computers Control for Inspection Functionality Documenting inspection Digitizing inspection pro- Analyzing data, statisti- Automated visual inspec- results, visual checks, cesses, capturing photos, cal process control, gen- tion, defect detection, manual recording of real-time data entry. erating control charts and measurement, sorting. defects. reports. Integration No integration Moderate integration Good integration with High integration with with data management data collection systems manufacturing systems, systems and reporting and reporting tools. data collection, and tools. reporting tools. Automation No automation Moderate level of au- High level of automation High level of automa- tomation for data en- for data analysis, control tion for real-time inspec- try and reporting; inspec- chart generation, and re- tion, defect detection, tions still manual. porting. and data logging. Rank 4 3 2 1 Reason for Ranking Lowest sophistication as Moderately sophisticated Highly sophisticated with Highest sophistication 83 it lacks automation and with digitized processes automated data analysis, due to full automation integration, relying en- but less automated in- good integration with and advanced integra- tirely on manual pro- spection; integration is data systems, but less tion with manufacturing cesses. less robust. real-time functionality systems for real-time compared to automated inspection and data systems. logging. Technology Example Checklists and Written Inspectorio Minitab Cognex In-Sight Reports Average Costs Cost of paper and print- Varies $1,500-$4,000/license $5,000-$20,000/system ing, varies Launch Year N/A 2016 1972 2016 ChatGPT ranking 1 2.5 3 4 Estimated time in a 5 hours 2.5 hours 1 hour 30 minutes task Source: The table draws upon various sources of information including desk research, and websites of various companies including Cognex In-Sight, and Minitab, Inspectorio. *CHATGPT estimates the time for inspecting product quality, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.6: Technology sophistication in Agriculture - Weeding Manual application Biological methods Mechanical applica- Fully-automated Drone Application of herbicide, fertil- of fertilizing, weed, tion of herbicide, fer- VRA tools (Preci- (Advanced Precision izer, or pesticide or pest control tilizer, or pesticide sion Agriculture) Agriculture) Functionality Simple tools for apply- Utilizes natural preda- Mechanized sprayers VRA tools adjust the Drones can perform a ing inputs by hand. tors and organic materi- can cover large ar- number of inputs based wide range of tasks als to enhance soil fertil- eas and apply inputs on real-time soil and such as applying inputs, ity and control pests. consistently. plant conditions. monitoring fields, and collecting data. Integration No integration with dig- Minimal integration Limited integration Integrated with soil and High integration with ital or mechanical sys- with other systems; with digital tech- plant sensors to provide remote sensing technol- tems. relies on natural pro- nologies; primarily data-driven application ogy and on-site sensors, cesses. mechanical. rates. providing real-time data and analytics. Automation Entirely manual; re- No automation; requires Requires human opera- Automated application Highly automated with quires significant labor human oversight and tion but significantly re- based on sensor data, minimal human inter- and time. management. duces the physical effort ensuring optimal use of vention, allowing for and time. resources. precise application and 84 monitoring. Rank 5 4 3 2 1 Reason for Rank Least sophisticated with Low sophistication with Moderate sophistication High sophistication Most sophisticated with basic functionality, no natural functionality, with effective function- with precise functional- high functionality, full integration, and no au- minimal integration, ality, but limited inte- ity and good integration integration with digital tomation. and no automation. gration with digital sys- with digital systems, systems, and high au- tems and requires hu- but slightly less au- tomation. man operation. tomated and versatile than drones. Technology Example Hand-held Sprayer Composting and Natu- Tractor-mounted VRA Sprayer with Sen- Agricultural Drone with ral Predators Sprayer sors Sensors Average Costs $20-$100 $50-$500 $1,000-$10,000 $20,000-$50,000 $10,000-$30,000 Launch Year N/A (Ancient Tool) N/A (Ancient Practice) 1950s 2000s 2010s ChatGPT ranking 1 2.5 3 4.5 5 Estimated time in a 8-10 hours 4-5 hours 6-7 hours 2-3 hours 1-2 hours task Source: The table draws upon various sources of information including desk research, and websites of various online marketplaces for agricultural tools like Home Depot and Rogers Sprayers. *CHATGPT estimates the time for weeding and pest control in a medium-sized field, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.7: Technology sophistication in Agriculture - Harvesting Manual Harvesting, Animal Aided In- Human-Operated Mechanized Process Automated Process Training, Pruning, struments Machines or Single- (Fully Operated by (Supported by Digi- or Picking Axle Tractor (One Worker) tal Technologies) Specific Function) Functionality Basic manual tasks (col- Performs basic func- Performs a specific func- Combines cutting, Combines multiple lecting, cutting, prun- tions with animal tion (e.g., soil prepara- threshing, and cleaning functions (cutting, ing). assistance (e.g., tilling, tion, grass cutting). into one operation. threshing, cleaning) transporting). with high precision. Integration No integration, entirely Limited integration, re- Limited to single- Mechanically integrates Integrates GPS and manual. lies on animal power. function tasks. multiple harvesting computing systems for functions. optimized paths and resource use. Automation No automation, fully re- No automation, depen- Manually operated, no Operated manually by High level of automa- liant on human labor. dent on human and ani- automation. the worker, but mecha- tion, requiring minimal mal labor. nized. human intervention. Rank 5 4 3 2 1 Reason for Rank Least sophisticated with Low sophistication, re- Moderate sophistication High functionality with Most sophisticated with 85 basic functionality, no lying on animal labor with effective function- integration of multiple advanced functionality, integration, and entirely with no digital integra- ality, but limited to sin- mechanical functions, full integration with dig- manual labor. tion or automation. gle functions and re- but requires manual ital systems, and high quires full manual oper- operation. automation. ation. Technology Exam- Harvesting Bags, Felco Horse-Drawn Reaper, Walk-Behind Sickle Tractor-Mounted Har- GPS-Enabled Tractor ples Pruning Shears, Hori Ox-Drawn Mower, Bar Mower, Single-Axle vester, Combine Har- with Harvesting At- Hori Knife Animal-Powered Binder Mower, Hand-Pushed vester, Self-Propelled tachment, Auto-Steer Seeder Forage Harvester Combine Harvester, Smart Forage Harvester Average Costs $30, $50, $35 $300, $250, $400 $600, $600, $450 $7,000, $300,000, $200,000, $400,000, $150,000 $350,000 Launch Years 1960, 1945, 1950 1800s, 1800s, 1850s 1960s, 1970s, 1960s 1920s, 1930s, 1940s 1990s, 2000s, 2010s ChatGPT ranking 1 2 3.5 4 5 Estimated time in a 200 hours 50 hours 30 hours 2 hours 1.5 hours task Source: The table draws upon various sources of information including desk research, and websites of various online sellers of agricultural equipment including John Deere, CLAAS, and Amazon Marketplace. *CHATGPT estimates the time for harvesting a 1-acre field of wheat, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.8: Technology sophistication in Apparel - Design Manual design and hand Digital or semi-digital de- Computer Aided Design drawing sign using specialized 2D (CAD), 3D design, virtual drawing software prototyping Functionality Fundamental capabilities for Robust capabilities for vector Comprehensive 2D and 3D de- drawing and drafting, relying graphics, layout design, and sign features, including ad- on user skill and precision technical drawing vanced modeling, simulation, and rendering Integration No digital integration; stan- Limited integration; can im- High integration with CAM, dalone and physical port/export various file types CAE, and other tools for seam- less workflow Automation No automation; all aspects are Moderate automation with High level of automation with manual snap-to-grid, alignment tools, automatic dimensioning, con- and reusable components straint management, and gener- ative design 86 Rank 3 2 1 Reason for Ranking Lowest sophistication due to re- Moderate sophistication with Highest sophistication due to liance on manual input, lack of strong 2D design capabilities, advanced functionality, seam- integration, and absence of au- but lower integration and au- less integration, and high au- tomation tomation compared to CAD tomation, ideal for complex ap- tools parel design Technology Examples Drafting Table, Mechanical Adobe Illustrator, Corel- AutoCAD, SolidWorks, Au- Pencil, Drawing Paper DRAW, AutoCAD LT todesk Fusion 360 Cost $2 - $300 $20.99/month - $499 (one-time) $60/month - $3,995 (one-time) or $198/year + $1,295/year maintenance Launch Year N/A 1987 - 1993 1982 - 2013 ChatGPT ranking 1 3 4.5 Estimated time in a task 8-10 hours 4-6 hours 2-3 hours Source: The table draws upon various sources of information including desk research, and websites of companies like Autodesk, SolidWorks, CorelDraw, and Adobe. *CHATGPT estimates the time for designing a simple dress, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.9: Technology sophistication in Apparel - Sewing Manually Sewing Machine Sewing Semi-automated Automated 3D Knitting Manually Oper- Sewing Machines Sewing Machines ated Functionality Basic hand stitching, Basic stitching with Automatic tension High-speed sewing, Knits entire gar- mending, and precise manual adjustments adjustment, multiple automated thread ments in 3D with cutting of fabric. for tension and stitch stitch patterns, and trimming, and pro- no seams, enabling length. built-in embroidery grammable stitch complex designs and designs. patterns. high customization. Integration No integration Limited to no in- Integrated with Often integrated Integrated with dig- tegration with digi- pre-programmed with manufacturing ital design software, tal or automated sys- patterns and digital systems for effi- allowing for seamless tems. interfaces for ease of cient production transition from de- use. workflows. sign to production. Automation No automation Minimal automation; Moderate level of au- High level of automa- Fully automated relies on user skill tomation with fea- tion, reducing the knitting process, and manual adjust- tures like automatic need for manual ad- from pattern cre- ments for operation. needle threading and justments and inter- ation to finished thread cutting. ventions. garment. Rank 5 4 3 2 1 87 Reason for Rank- Lowest sophisti- Lower sophistication Moderate sophisti- High sophistication Highest sophisti- ing cation, entirely due to minimal au- cation with useful with advanced func- cation due to full manual with no au- tomation and lack of automation fea- tionality and integra- automation, ad- tomation or digital integration with dig- tures, but limited tion, but less special- vanced functionality, integration. ital systems. compared to fully ized than 3D knit- and seamless inte- automated systems. ting. gration with digital design tools. Technology Ex- Needle and Thread, Mechanical Sewing Electronic Sewing Industrial Sewing Shima Seiki Whole- amples Thimble, Fabric Scis- Machine, Hand- Machine, Com- Machine, Automated Garment, Stoll ADF sors Crank Sewing puterized Sewing Quilting Machine, 3, Kniterate Machine, Foot-Pedal Machine, Embroi- Automated Serger Sewing Machine dery Machine Cost $1 - $30 $100 - $500 $300 - $3,000 $1,000 - $20,000 $7,500 - $250,000 Launch Year Ancient - N/A 1800s - 1900s 1970s - 1990s 1960s - 2000s 1970s - 2018 ChatGPT ranking 1 2.5 3.5 4.5 5 Estimated time in 8-10 hours 3-4 hours 2-3 hours 1-2 hours 30 minutes-1 hour a task Source: The table draws upon various sources of information including desk research, and websites of companies like Kniterate as well as online sellers of sewing equipment like Amazon Marketplace. *CHATGPT estimates the time for sewing a Simple T-shirt, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.10: Technology sophistication in Food Processing - Anti-Bacterial Minimal-processing Anti-bacterial Wash Thermal Processing Other Advanced Preservation or Soaking Technologies Methods (HPP, PEF) Functionality Slows bacterial Reduces bacterial load on Effective in killing bac- HPP and PEF offer ad- growth without killing food surfaces, effective for teria and pathogens vanced bacterial inacti- pathogens, preserving surface decontamination. through heat, ensuring vation with minimal im- food quality. food safety. pact on food quality. UV treatment provides surface-level decontami- nation. Integration Easy to integrate, re- Simple to integrate into Well-integrated in the Moderate to high integra- quires basic modifications existing systems with food industry, adaptable tion into existing produc- to packaging/storage pro- easy-to-implement tanks to different production tion lines with minimal cesses. or spray systems. scales. adjustments required. Automation Limited automation po- Can be automated but re- Highly automated, offer- Fully automated pro- tential, requiring manual quires some manual inter- ing precise control over cesses requiring minimal checks and adjustments. vention for effective appli- processing conditions. human intervention. cation. 88 Rank 4 3 2 1 Explanation for Rank- Lower sophistication due Moderate sophistication, High sophistication due Highest sophistica- ing to minimal bacterial in- as it is effective and easy to reliable pathogen inac- tion due to advanced, activation and reliance to integrate, but with tivation, well-established non-thermal bacterial in- on manual processes for lower automation levels integration in the indus- activation methods, high maintaining effectiveness. and more manual inter- try, and high levels of au- integration potential, and vention needed. tomation. full automation. Technology Examples Modified Atmosphere Chlorine Dioxide Wash, Steam Pasteurization, High-Pressure Processing Packaging (MAP), Vac- Organic Acid Wash, Infrared Heating, Mi- (HPP), Pulsed Electric uum Packaging, Edible Ozone Treatment crowave Pasteurization Field (PEF), Ultraviolet Coatings (UV) Treatment Cost $10,000 - $100,000 $500 - $20,000/year $50,000 - $500,000 $10,000 - $2,500,000 Launch Year 1980s - 2000s 1990s - 2000s 1980s - 2000s 1990s - 2000s ChatGPT ranking 2 2.5 4 4.5 Estimated time in a 1-2 hours 15-30 minutes 30 minutes - 1 hour 1-2 hours task Source: The table draws upon various sources of information including desk research, and some lab reports to provide cost estimates of various anti-bacterial processes. *CHATGPT estimates the time for reducing the microbial load on fresh produce, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.11: Technology sophistication in Food Processing - Packaging Manual Packing in Human Operated Me- Power Equipment Re- Power Equipment Bags, Bottles, or Boxes chanical Equipment quiring Routine Inter- Controlled by Comput- action ers/Robotics Functionality Basic functionality involv- Functionality includes High functionality with au- Highest functionality with ing manual sealing of bags, weighing and filling of tomated processes for bag- advanced capabilities such wrapping of products, and packages, sealing of bags, ging, filling, and wrapping, as precise handling, stack- application of labels to and capping of bottles reducing the need for man- ing, and packaging with packages. with human-operated ual effort. minimal errors. mechanical equipment. Integration No integration, manual Moderate integration, rel- Moderate to high inte- High integration complex- processes not typically in- atively easy to incorporate gration, capable of being ity but can seamlessly inte- tegrated with automated into existing workflows but incorporated into existing grate into existing produc- systems. still reliant on human oper- systems but requiring hu- tion lines to enhance effi- ation. man oversight and routine ciency and reduce manual interaction. labor. Automation No automation, entirely Low to moderate automa- High level of automation, Highest level of automa- dependent on human labor. tion, with significant hu- but still requires routine tion, requiring minimal hu- man involvement required human interaction for load- man interaction once pro- for operation and control. ing, monitoring, and occa- grammed and set up. 89 sional adjustments. Rank 4 3 2 1 Reason for Ranking Lowest sophistication due Moderate sophistication High sophistication due to Highest sophistication due to full reliance on manual due to reliance on human automated processes with to advanced functionality, labor with no automation operation with limited au- significant human oversight seamless integration, and or integration capabilities. tomation, improving speed and interaction needed. full automation with min- and accuracy compared to imal human interaction. manual methods. Technology Examples Handheld Heat Sealer, Mechanical Weighing and Automatic Bagging Ma- Robotic Palletizer, Au- Manual Wrapping Ma- Filling Machine, Foot chine, Automated Filling tomated Guided Vehicles chine, Handheld Label Pedal Operated Heat Machine, Automated (AGVs), Computer- Applicator Sealer, Semi-Automatic Wrapping Machine Controlled Packaging Bottle Capping Machine Line Cost $30 - $300 $1,000 - $10,000 $10,000 - $100,000 $50,000 - $1,000,000 Launch Year 1970s - 1990s 1980s - 2000s 1990s - 2010s 2000s - 2015 ChatGPT ranking 1 2.5 3.5 4.5 Estimated time in a 8-12 hours 4-6 hours 1-2 hours 0.5-1 hour task Source: The table draws upon various sources of information including desk research, and websites of companies that provide packaging solutions like ZoneSun Auto Pack. *CHATGPT estimates the time for sealing 1,000 Bags of food products, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.12: Technology sophistication in Retail - Pricing Manual Cost Automated Markup Automated Promo- Dynamic Pricing Personalized Pricing (Excel or similar) tional Systems driven by Predictive Analytics Functionality Minimal features, rely- Provides basic au- Focuses on automat- Utilizes AI to optimize Offers the most ad- ing entirely on manual tomation for pricing ing promotional pric- prices based on demand vanced features, includ- calculations and record- calculations, applying ing based on predefined prediction, competitive ing personalized pricing keeping. markups, and tracking rules, events, and sea- analysis, and real-time strategies based on cus- data using formulas and sonal factors. data. tomer behavior, pref- templates. erences, and predictive analytics. Integration No integration Limited integration ca- Good integration with Integrates well with Integrates seamlessly pabilities, mainly re- sales, marketing, and e- market data sources with various data quiring manual data im- commerce platforms to and internal business sources, CRM sys- port/export. manage promotions effi- systems to provide tems, and e-commerce ciently. comprehensive pricing platforms to gather strategies. comprehensive cus- tomer data. Automation No automation, entirely Some automation Automates promotional Automated price ad- Highly automated, us- dependent on human la- through formulas, but pricing adjustments, justments based on real- ing machine learning al- bor. significant manual ef- but typically based on time data and algo- gorithms to adjust pric- fort is required for data predefined schedules rithms, requiring min- ing dynamically in real- entry and maintenance. and rules rather than imal manual interven- time based on predic- 90 real-time data. tion. tive models and cus- tomer data. Rank 5 4 3 2 1 Reason for Ranking Lowest sophistication Lower sophistication Moderate sophisti- High sophistication Highest sophistication due to the absence of with basic automation cation with strong with strong automation due to advanced per- automation and inte- and limited integration, promotional features and integration, opti- sonalization, seamless gration, relying entirely relying heavily on man- but limited real-time mizing prices based on integration with multi- on manual calculations. ual processes for data automation and in- real-time data and AI ple platforms, and full management. tegration compared algorithms. automation using AI. to dynamic pricing systems. Technology Exam- Pen and Paper Microsoft Excel, Google Salesforce CPQ, SAP PROS Pricing, Dy- Dynamic Yield, Zilliant, ples Sheets, QuickBooks Hybris, Shopify Plus namic Pricing by Adobe Target Prisync, RepricerEx- press Cost Minimal to None $160/year (Office), $75/user/month, Custom pricing, Custom pricing Free to $12/month, Custom pricing, $59/month, $55/month $25/month $2000/month Launch Year Pre-digital 1983 - 2006 1997 - 2006 1985 - 2014 1999 - 2011 ChatGPT ranking 1 2.5 3 4 5 Estimated time in a 8 hours 4 hours 2 hours 1 hour 2 hours task Source: The table draws upon various sources of information including desk research, and websites of companies like Microsoft, Google, Salesforce, SAP, Adobe, PROS, and RepricerExpress. *CHATGPT estimates the time for setting seasonal discount pricing for a product line, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. Table A.13: Technology sophistication in Retail - Inventory Handwritten Record Computer Databases Warehouse Manage- Automated Inven- Automated Storage Keeping with Manual Up- ment System tory Control and Retrieval Sys- dates tems (AS/RS) Functionality Basic manual entry and Basic inventory tracking Comprehensive inven- Advanced inventory Fully automated sys- tracking of inventory and data management tory management, order control using RFID tems for high-density data. with manual data entry fulfillment, warehouse for real-time tracking, storage, retrieval, and and updates. operations, and supply CAO for automated inventory management. chain optimization. order replenishment, and VMI for vendor- managed inventory solutions. Integration No integration Limited integration ca- Integrates with ERP, Integrates with supply Seamlessly integrates pabilities, primarily re- TMS, and other busi- chain systems, WMS, with WMS, ERP, and quiring manual data im- ness systems for end-to- and ERP systems for other business systems port/export. end visibility and con- seamless data flow and to provide real-time trol. inventory optimization. data and analytics. Automation No automation Minimal automation, Offers significant au- High level of automa- Provides the highest mainly through for- tomation for inventory tion with real-time level of automation, mulas and templates, tracking, order process- tracking, automated handling storage with significant manual ing, and warehouse op- updates, and reduced and retrieval tasks effort needed. erations but requires manual intervention. autonomously with 91 some manual oversight. minimal human inter- vention. Rank 5 4 3 2 1 Reason for Ranking Lowest sophistication Lower sophistication Moderate sophistication High sophistication due Highest sophistication due to the absence of due to limited automa- with strong functional- to advanced real-time due to full automation, automation and inte- tion and integration, ity and integration, but tracking, automated in- seamless integration, gration, relying entirely with significant manual with some manual over- ventory control, and and advanced function- on manual processes. data entry required. sight required for opera- strong integration with ality in high-density tions. supply chain systems. storage and retrieval. Technology Exam- Ledger Books, Paper Microsoft Access, SAP EWM, Oracle Zebra Technologies, Swisslog, Dematic, ples Forms, Manual Logs Google Sheets, Quick- WMS, Manhattan Blue Yonder, IBM Honeywell Intelligrated Books WMS Sterling Cost Minimal $160/year (Office), Custom pricing Custom pricing Custom pricing Free to $12/month, $25/month Launch Year n/a 1983 - 2006 1990 - 2005 1969 - 2000 1900 - 2001 ChatGPT ranking 1 2 3.5 4 5 Estimated time in a 40 hours 20 hours 10 hours 4 hours 2 hours task Source: The table draws upon various sources of information including desk research, and websites of companies like Microsoft, Google, Salesforce, SAP, Oracle, Zebra Tech, IBM, SwissLog, Dematic, and Honeywell Integrated. *CHATGPT estimates the time for conducting a full inventory count, following similar characteristics of a given firm. The prompt for these estimates and CHATGPT ranking is available in the appendix. A.2.1 Comparison between experts’ and ChatGPT’s sophistication rankings We validate the industry experts’ technology sophistication rankings using AI-powered large language models. To start, we prepared files with the list of business functions and associated technologies in the grid, keeping the way they are described in the FAT survey. We generate separate files for GBFs and each specific sector functions. We then uploaded these files and asked ChatGPT to rank the sophistication of these technologies, providing the following prompts: 1. Based on the survey questions from the document, please create a report ranking the level of technological sophistication for each technology within business function on a scale of 1 to 5, where 1 represents the most basic technology and 5 represents the most sophisticated technology. You can use decimal points. Generate a report file with this information. Provide the data in an Excel sheet. 2. Please expand the previous exercise by providing the rank of technology sophistication for each of these definitions: 1) Functionality: The breadth and depth of features and capabilities that technology offers to perform various tasks and processes within a business function. Higher functionality indicates the ability to handle more complex tasks associated with the business function on a larger scale, faster, with more accuracy and reliability. 2) Integration: The ability of a technology to connect and interact seamlessly with other systems, platforms, or technologies within the business ecosystem. Higher integration means the technology can easily exchange data and coordinate processes with other systems. 3) Automation: The extent to which a technology can perform tasks automatically without human intervention. Higher levels of automation imply that the technology can execute processes, make decisions, and generate outcomes on its own. 3. For each business function, please expand the previous exercise by estimating the time each technology would take to perform a typical task in a company with 10 workers. Please describe this typical task in a separate column and explain the reason for these estimates in another column. Generate a report with the estimates and explanations in Excel. Tables A.1 - A.13 summarize the main results from prompts 1 and 3. Figures A.12-A.16 show the strong positive association between ChatGPT-specific rankings of technology so- phistication for functionality, integration, and automation with industry experts’ rankings. The rankings provided by ChatGPT are strongly correlated with our technology sophistica- tion ranking, which is defined by human experts. The overall correlation across all functions and various definitions is above 90%. The results are similar across all business functions. 92 Business Administration Sourcing 5 Enterprise Resource Planning (ERP) 5 SRM integrated with production planning 4 Computers with specialized software 4 Supplier Relation Management (SRM) not integrated with production planning Mobile apps or digital platforms Supplier social media/digital Relation Management (SRM) not integrated with production planning platforms ChatGPT rankings ChatGPT rankings 3 Mobile apps or digital platforms 3 social media/digital platforms 2 Computers with standard software (Excel) 2 PCs manually updated (Excel) Computers with standard software (Excel) PCs manually updated (Excel) 1 Handwritten processes Functionality 1 Manual search of suppliers Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Marketing Payment 5 Big data analytics or machine learning (AI) 5 Virtual or cryptocurrency Customer Relationship Management (CRM) Online payment through platforms (PayPal) 4 Customer Relationship Management (CRM) 4 Online or electronic payment (bank wire) Prepaid, debit, or credit card ChatGPT rankings ChatGPT rankings 3 Structured surveys (phone, online) 3 Prepaid, debit, or credit card Online chat (WhatsApp) Cheque, voucher, or bank wire 2 Online chat (WhatsApp) 2 Cash Cheque, voucher, or bank wire Cash 1 Face-to-face chat Functionality 1 Cash Exchange of goods Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Sales Quality Control 5 5 Electronic orders integrated with supply chain management systems Automated systems for inspection (laser, sensors) Online sales (e-commerce) via own website 4 Online sales using external platforms (Amazon) 4 Statistical process control with software Sales through social media platforms or apps ChatGPT rankings ChatGPT rankings 3 Sales through social media platforms or apps 3 Human inspection supported by computers or phones 2 Sales by phone, email, or sales representatives 2 Human inspection supported by computers or phones 1 Sales at establishment premises Functionality 1 Manual visual or written processes Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Figure A.12: Comparison between expert’s versus ChatGPT’s in General Business Functions Note: This figure compares the experts’ rankings of technology sophistication (horizontal axis) with Chat- GPT’s rankings across the dimensions of functionality, integration, and automation (vertical axis). 93 Agriculture-Harvesting Agriculture-Irrigation 5 Digital-enabled Automation 5 Automated sensor-controlled Irrigation 4 Mechanized Process without Digital Tech 4 Drip/Sprinkler Irrigation Mechanized Human-operated Machines Process without Digital Tech Drip/Sprinkler Irrigation ChatGPT rankings ChatGPT rankings 3 3 Irrigation by Small Pump Human-operated Machines Irrigation Surface Flood by Small Pump Irrigation 2 Animal Aided Instruments 2 Surface Flood Irrigation Animal Aided Instruments Surface Flood Irrigation Manual Irrigation 1 Manual Methods Functionality 1 Rain-fed Manual Irrigation Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Agriculture-Land Preparation Agriculture-Packing 5 5 Modified Atmosphere Packing Digital-enabled Tractors (GPS) Automated Packing Linked to Harvesting 4 Digital-enabled Tractors (GPS) 4 Automated Packing Linked to Harvesting Human-operated Tractors ChatGPT rankings ChatGPT rankings 3 3 Human-operated Tractors Human-operated Mechanical Equipment 2 Animal Aided Instruments 2 Human-operated Mechanical Equipment Animal Aided Instruments 1 Manual plowing Animal Aided Instruments Functionality 1 Manual Packing Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Agriculture-Storage Agriculture-Weeding 5 Continuous Monitoring Devices 5 Drone Application with Sensors High-end Central Storage Fully Automated VRA 4 High-end Central Storage 4 Fully Automated VRA Cold/Dry Controlled Storage ChatGPT rankings ChatGPT rankings 3 Cold/Dry Controlled Storage 3 Biological Methods Biological Methods Mechanical Application 2 Protected Storage without Temperature Control 2 Mechanical Application Protected Storage without Temperature Control 1 Precarious Storage Functionality 1 Manual Application of Herbicide Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Figure A.13: Comparison between expert’s versus ChatGPT’s in Agriculture Note: This figure compares the experts’ ranking of technology sophistication (horizontal axis) with Chat- GPT’s rankings across the dimensions of functionality, integration" "and automation (vertical axis). 94 Food Processing- Packaging Food Processing-Anti-Bacterial 5 5 Power equipment controlled by computers/robotics Advanced methods (HPP, PEF) 4 4 Advanced methods (HPP, PEF) Thermal Processing Technologies Power equipment requiring routine human interaction Thermal Processing Technologies ChatGPT rankings ChatGPT rankings 3 Power equipment requiring routine human interaction 3 Thermal Processing Technologies Power equipment Human-operated mechanical equipmentrequiring routine human interaction Anti-bacterial wash or soaking 2 Human-operated mechanical equipment 2 Anti-bacterial Minimal-processing preservation wash or soaking methods Anti-bacterial Minimal-processing preservation wash or soaking methods 1 Manual packing Functionality 1 Minimal-processing preservation methods Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Food Processing-Food Storage Food Processing-Input Testing 5 5 building Storage - Fully automated climate and security-controlled Computer testing (chromatography, spectroscopy) 4 4 Computer testing (chromatography, spectroscopy) Storage - Some climate control in secured building ChatGPT rankings ChatGPT rankings 3 Storage - Some climate control in secured building 3 Non-computer-controlled testing kits Storagein Storage - Ambient conditions closedclimate - Some control in secured building building Non-computer-controlled testing kits Review of supplier testing 2 Storage - Ambient conditions in closed building 2 Non-computer-controlled testing kits Review of supplier testing Sensory systems Review of supplier testing 1 Storage - Minimal protection with some exposure Functionality 1 Sensory systems Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Food Processing-Mixing 5 Power equipment controlled by computers/robotics 4 Power equipment requiring routine human interaction ChatGPT rankings 3 Power equipment requiring routine human interaction Power equipment requiring routine human interaction 2 Mechanical equipment requiring human force Mechanical equipment requiring human force 1 Manual process Functionality Integration Automation 0 0 1 2 3 4 5 Expert rankings Figure A.14: Comparison between expert’s versus ChatGPT’s in Food Processing Note: The figure compares the experts’ ranking of technology sophistication (horizontal axis) with Chat- GPT’s rankings across the dimensions of functionality, integration" "and automation (vertical axis). 95 Apparel-Cutting Apparel-Design 5 Automatic/computerized cutting machine (Laser) 5 Automatic/computerized cutting machine (Laser) Computer-Aided Design (CAD) with 3D virtual prototyping 4 Automatic/computerized cutting machine (non-laser) 4 Computer-Aided Design (CAD) with 3D virtual prototyping Computer-Aided Design (CAD) with 3D virtual prototyping ChatGPT rankings ChatGPT rankings 3 Semi-automatic cutting machine 3 Digital or semi-digital design (2D drawing software) Semi-automatic Manually operated cutting machine cutting machine Digital or semi-digital design (2D drawing software) 2 Manually operated cutting machine 2 Digital or semi-digital design (2D drawing software) Manually operated cutting machine 1 Manual cutting Functionality 1 Manual design and hand drawing Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Apparel-Finishing Apparel-Sewing 5 High-tech pressing machine 5 Sewing - 3D knitting High-tech pressing machine Form finishing machine Sewing - Automated sewing machines 4 Tunnel finisher Form finishing machine 4 Sewing - Automated sewing machines Tunnel finisher Sewing - Semi-automated sewing machines ChatGPT rankings ChatGPT rankings 3 3 Sewing - Semi-automated sewing machines Electric high-pressure steam iron Sewing - Manually operated sewing machine 2 Electric high-pressure steam iron 2 Sewing - Manually operated sewing machine Sewing - Manually operated sewing machine 1 Basic manual ironing Functionality 1 - Manual sewing Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Figure A.15: Comparison between expert’s versus ChatGPT’s in Apparel Note: The figure compares the experts’ ranking of technology sophistication (horizontal axis) with Chat- GPT’s rankings across the dimensions of functionality, integration" "and automation (vertical axis). 96 Retail-Advertisement Retail-Customer Service 5 Search Engine Big Marketing 5 data or AI to generate personalized advertisements Chatbots (YouTube) Social mediaSearch Engine Marketing Online requests (website, digital platform, app) 4 Social mediaSearch (YouTube) Engine Marketing 4 Online requests (website, digital platform, app) Social media (YouTube) Online Social Media (e.g., requests (website, digital platform, app) Facebook) ChatGPT rankings ChatGPT rankings 3 Email or mobile phone 3 Social Media (e.g., Facebook) EmailTV Radio, Billboards, or mobile phone Social Media (e.g., Facebook) 2 Radio, Billboards, TV 2 Phone call or help desk Radio, Billboards, TV Phone call or help desk 1 Paper-based communication (in-store signage, leaflets) Functionality 1 At the store Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Retail-Inventory Retail-Merchandising 5 Automated Storage and Retrieval Systems (AS/RS) 5 Product trend analytics (Big Data, Machine Learning) Automated inventory control (CAO, VMI, RFID) Retail Merchandising Systems or Digital Merchandising 4 Warehouse Management Automated System inventory control (CAO, VMI, RFID) 4 Retail Merchandising Systems or Digital Merchandising Warehouse Management System Retail Category Management toolsMerchandising Systems or Digital Merchandising ChatGPT rankings ChatGPT rankings 3 3 Category Management tools Computer databases with manual updates Category Management tools 2 Computer databases with manual updates 2 Computer databases with manual updates 1 Handwritten record keeping Functionality 1 Manually selecting products Functionality Integration Integration Automation Automation 0 0 0 1 2 3 4 5 0 1 2 3 4 5 Expert rankings Expert rankings Retail-Pricing 5 Pricing - Personalized pricing (predictive analytics) Pricing - Dynamic pricing systems (software) 4 Pricing - Dynamic pricing systems (software) ChatGPT rankings 3 Automated promotional (seasonality or predictable events) Automated promotional (seasonality or predictable events) Automated markup(eg-excel) 2 Automated markup(eg-excel) 1 Manual cost Functionality Integration Automation 0 0 1 2 3 4 5 Expert rankings Figure A.16: Comparison between expert’s versus ChatGPT’s in Retail Note: The figure compares the experts’ ranking of technology sophistication (horizontal axis) with Chat- GPT’s rankings across the dimensions of functionality, integration" "and automation (vertical axis). 97 A.3 Sampling Frame The sampling frames were based on the most comprehensive and latest establishment census available from national statistical agencies or administrative business register. Table A.14 provides the main data sources used in the sample frame for each country. Table A.14: Sampling frame by country Country Source Sampling frame Year Bangladesh Bangladesh Bureau of Statistics. Est. census, 2013 2019 Brazil Ministry of Labor Employer census, RAIS, 2018 2019 Burkina Faso Business Registry Business Registry 2021 Cambodia Tax Registry Tax Registry 2022 Chile Business Registry Census on Establishments 2022 Croatia Financial Agency (FINA) FINA Data 2023 Ethiopia Ministry of Trade and Industry (MoTI) Business Registry 2022 Georgia National Statistics Office of Georgia Est. census, 2021 2022 Ghana Ghana Statistical Service Est. census, 2013–18 2021 India Central Statistics Office of India Est. census, 2013–17 2020/23* Kenya Kenya National Bureau of Statistics Est. census, 2017 2020 Korea, Rep. Statistics Korea Est. census, 2018 2021 Poland Statistics Poland Est. census, 2020 2021 Senegal National Agency for Statistics (ANSD) Est. census, 2016 2019 Vietnam General Statistics Office of Vietnam Est. census, 2018 2019 Note : * The states of Tamil Nadu and Uttar Pradesh were surveyed in Wave 1 in 2020. The states of Gujarat and Maharashtra were surveyed in 2023. The universe of study includes establishment with 5 or more employees in agriculture, manufacturing and services. The sector classification is based on the International Standard Industrial Classification of All Economic Activities (ISIC), Rev. 4. More specifically, our sample includes firms from the following ISIC rev 4 sectors: Agriculture (ISIC 01, from Group A); All manufacturing sectors (Group C); Construction (Group F), Wholesale and retail trade (Group G), Transportation and storage (Group G), Accommodation and food service activities (Group I), Information and communication (Group J), Financial and insurance activities (Group K), Financial services (ISIC, 64), Travel agency (ISIC 79, from group N), Health services (ISIC 86, from group Q), and Repair services (ISIC 95, from Group S). 98 Table A.15: Total number of firms in the universe covered by the survey Sector Size Country Total Agri. Manu. Serv. Small Medium Large Bangladesh 15,358 15,358 4,164 3,425 7,769 Brazil* 23,364 392 4,758 18,214 12,771 8,955 1,638 Burkina Faso 57,328 4,808 7,493 45,027 40,189 13,284 3,855 Cambodia 8,172 1,890 6,282 5,842 1,287 1,043 Chile 104,854 7,419 11,943 85,492 65,425 30,071 9,358 Croatia 22,350 524 5,387 16,439 17,038 4,381 931 Ethiopia 144,583 3,670 6,553 134,360 105,038 36,798 2,745 Georgia 14,839 313 2,194 12,332 10,815 3,259 765 Ghana 42,165 880 10,284 31,001 30,133 10,070 1,962 India** 616,833 71,464 233,684 311,685 624,452 70,928 13,514 Kenya 74,255 3,680 5,407 65,168 50,584 16,676 6,995 Korea, Rep. 545,515 1520 167,466 376,529 450,264 82,403 12,848 Poland 244,983 3,021 52,340 189,622 198,107 37,799 9,077 Senegal 9,583 1,051 4,069 4,463 7,805 1,414 364 Vietnam 179,713 1,080 45,805 132,828 135,046 33,107 11,560 Total 2,103,895 117,070 567,829 1,432,427 1,756,723 350,188 75,612 Note : * Brazil refers to state of Ceará; ** States of Tamil Nadu, Uttar Pradesh, Gujarat, and Ma- harashtra in India. The survey does not cover agriculture or services in Bangladesh, nor agriculture in Cambodia. In India, only the states of Gujarat and Maharashtra have agriculture included in the survey. Table 2 provides the distribution of the number of firms sampled in each country, by sector and firm size group. We exclude micro-firms with fewer than 5 employees. Micro firms, particularly in devel- oping countries, are more likely to be informal (Ulyssea, 2018), making them less likely to be captured in the sampling frame; and this would require further adjustment in the survey instrument and sampling design.30 This size threshold is aligned with other firm-level stan- dardized surveys with comparability across countries. The World Bank Enterprise Survey (WBES) also uses a threshold of 5 employees. The World Management Survey (WMS) uses a threshold of 50 employees. We stratify the universe of establishments by firm size, sector of activity, and geographic regions. Our sample is representative across these dimensions. In the firm size stratifica- tion, we have three strata: small firms (5-19 employees), medium firms (20-99 employees), and large firms (100 or more employees). Regarding sector, for all countries, we stratified at least for agriculture (ISIC 01), food processing (ISIC 10), Wearing apparel (ISIC 14), Retail and Wholesale (ISIC 45, 46 and 47), other manufacturing (Group C, excluding food 30 In addition, establishments below this threshold often lack the organizational structure to respond to some of the questions. 99 processing and apparel), and other Services (including all other firms, excluding retail). We use this sector structure of the data for most of the analysis in this paper. Additional sector stratification that were country specific included: motor vehicles (ISIC 29); Leather (ISIC 15), Pharmaceutical (ISIC 21), and Motor vehicles (ISIC 29); and Land transport (ISIC 49), Finance (ISIC 64), and Health (ISIC 86).31 In the geographic stratification, we use sub-national regions. To calculate the optimal distribution of the sample, we followed a similar methodology as described by the World Bank (2009). The sample size for each country was aligned with the degree of stratification of the sample. The data used in this paper corresponds to the first and second phase of the survey implementation. The surveys were administered between June 2019 and the end of 2021 by the World Bank in partnership with public or private local agencies across ten countries: Bangladesh, Brazil (the state of Ceará), Senegal, and Vietnam in the first phase until January 2020. In the second phase, conducted during the COVID-19 pandemic, after January 2020, included Burkina Faso, India (the states of Tamil Nadu and Uttar Pradesh), Ghana, Kenya, Poland, and the Republic of Korea. The mode of data collection was face-to-face before the pandemic and mostly on the telephone during the pandemic. Table A.16: Year and mode of data collection Country Year Mode Bangladesh 2019 Face-to-face Brazil 2019 Face-to-face Burkina Faso 2021 Telephone Cambodia 2022 Telephone Chile 2022 Telephone Croatia 2023 Online Ethiopia 2022 Face-to-face Georgia 2022 Online & Telephone Ghana 2021 Telephone India 2020/23* Face-to-face Kenya 2020 Telephone Korea, Rep. 2021 Telephone Poland 2021 Telephone Senegal 2019 Face-to-face Vietnam 2019 Face-to-face 31 These specific stratifications were taken into consideration when determining sampling weights. 100 A.4 Survey Weights We construct the sampling weights of establishments in two steps. First, we compute design weights as reciprocals of inclusion probabilities. Then, to mitigate the risk of non-response bias, we adjust the design weights for non-response. We adopt a stratified one stage element sampling design and randomly select estab- lishments with equal probabilities within strata. Therefore, the inclusion probability of establishment k, within stratum isr (identified by industry i, size s, and region r), is: nisr πisrk = (A.1) Nisr where nisr is the number of establishments targeted by the survey for stratum isr, and Nisr is the number of establishments in the sampling frame for the same stratum. Accord- ingly, the design weights of establishments are: 1 Nisr disrk = = (A.2) πisrk nisr To adjust the design weights in Equation A.2 for non-response we follow a simple Re- sponse Homogeneity Groups (RHG) approach (Särndal, Swensson and Wretman, 1992), with the groups determined by the strata. In other words, we assume that establishment response probabilities are the same within each stratum, but differ across different strata. Under the RHG approach assumptions, response probabilities can be estimated using the observed re- sponse rates within each group, and bias protection is obtained by dividing design weights by group-level response rates. Denoting with the estimated response probability in stratum isr, and with misr the number of respondent establishments in the stratum (so that misr nisr ), the non-response adjusted weights can thus be written as follows: disrk disrk Nisr /nisr Nisr wisrk = = = = (A.3) θˆisr misr /nisr misr /nisr misr Note that the adjusted weights in Equation A.3 are such that the distribution of our respondent sample across strata exactly matches the distribution of establishments in the sampling frame: wisrk = Nisr (A.4) k ∈Risr where Risr denotes the respondent sample for stratum isr. Because of the different number of establishments in each country, when computing global 101 statistics, we re-scale weights so that all countries are equally weighted. A.5 Measures to Minimize Bias and Measurement Error During Survey Design and Implementation During the design of the survey questionnaire a number of good practices were considered in order to minimize different types of potential biases. The literature on survey design has identified three types of potential bias and measurement errors. These depend on whether they originate from the non-response, the enumerator or the respondent (Collins, 2003). In this section we describe all the steps taken in the design and implementation of the FAT survey to minimize these errors. Non-response bias. A critical potential bias is associated with non-response in par- ticular questions or non-participation in the survey (Gary, 2007). When this non-response follows a pattern that can be linked to factors correlated to the measured object, this non- response is associated with biases. For example, if more technology sophisticated firms refuse to participate because of fear to reveal commercial information, this would result in signif- icant downward bias in estimating the level of technology sophistication. To minimize this risk, we try to maximize participation in the survey and follow three steps. First, we partner with national statistical offices and industry associations to use the most comprehensive and updated sampling frame available, as well as their experience on data collection, which are supported by endorsement letters from local institutions.32 Having up to date contact details significantly improves response and minimized contact fatigue. Second, we follow a standard protocol in which each firm is contacted several times to schedule an interview. We split the sample in different batches, following the order of randomization within stratum, and provide contact information of subsequent batches only after interviewers have shown evi- dence that they have exhausted the number of attempts to complete the initial list. Third, we monitor the implementation, validation of skip conditions and outliers (e.g. financial statements’ information) in real time using standard survey software, and request that any missing information are completed through a follow up call, checked by supervisors. This minimizes risks that enumerators skip the order of their randomly assigned list of firms. Enumerator bias and error counts. Minimizing cognitive biases in respondents in face to face and phone interviews starts with making sure that enumerators are able to imple- ment the survey in a clear and consistent manner. To this end, the survey, training, and data collection processes are largely designed to minimize enumerator biases and data collection 32 These procedures are in line with suggestions of good practice for implementation by (Bloom et al., 2016). 102 errors. First, to reduce the likelihood of coding errors, we use closed-ended questions, which make coding the answers a mechanical task, eliminating the reliance on the enumerator’s interpretation of the answer and subjective judgement to code them, as it is the case with open-ended questions (Bloom et al., 2016). Second, to make sure that implementation is consistent across enumerators within and between country surveys, we implement the same standardized training in each country with enumerators, supervisors, and managers leading the data implementation. The training is led by team members directly involved in the elaboration of the questionnaire and implemented in local languages - English, French, Por- tuguese and Vietnamese,33 and they include vignettes to ensure that enumerators understand the specific technologies they are asking about. The two to three days training consists of one general presentation about the project, covering the main motivation, relevance, cover- age, and protocols that should be used to approach the interviewees and the review of the full questionnaire (question by question). The training material includes pictures of each technology mentioned in the survey both in general and sector-specific business functions, which are shared with enumerators. After going over the full questionnaire and clarifying any questions that emerge, the participants of the training conduct a mock interview using CAPI, under the supervision of our team. Third, to guarantee that translations use words that are understood by firms managers, in each country we conduct a pre-test pilot of the questionnaire with firms out of the sample. A pilot of the questionnaire is implemented in each country with firms out of the sample. This allows to fine-tune questions to the local language, finalize the translation and select the most relevant examples in each question. After the pilot, our teams have the opportunity to discuss with the managers implementing the questionnaires and clarify any potential question over the implementation process. Fourth, to attain greater quality control during the data collection process, enumera- tors record the answers via Computer-Assisted Personal Interviews (CAPI) and Computer- Assisted telephone Interviews (CATI) software.34 Using CAPI/CATI has clear advantages. First, it allows the use of logical conditions and skips which prevent data inputting errors and omitting questions, and also reduces the potential for abnormal values or non-response to specific questions. Second, it reduces substantially the time of implementation of the survey, increasing the quality of responses and minimizing survey fatigue. Supervisors are assigned to review all interviews, identifying missing values and abnormal responses. In addition, the CAPI/CATI system can identify when enumerators complete the survey too fast and 33 In the case of Vietnamese, we used translation services support. 34 Randomized survey experiments with household survey has demonstrated that a large number of errors observed in Pen-and-Paper Personal Interview (PAPI) data can be avoided in CAPI (Caeyers, Chalmers and De Weerdt, 2012). 103 other abnormal issues that can raise concerns about the quality of the interview. Finally, CAPI/CATI also allows for the core team to regularly monitor the data collection process and use standard algorithms to analyze the consistency of the data at different stages of data collection and by watches, thus providing continuous feedback and quality control. Respondent bias. Perhaps the most important type of bias relates to cognitive biases from respondents. These biases can be large in surveys with open ended questions or where concepts can be largely subjective. Specifically, two broad groups of factors can trigger response errors: cognitive, which affect the comprehension of the questions, and framing, which may cause biased answers due to the perceived socially (un)desirability of the answers (Bertrand and Mullainathan, 2001). We take several steps to minimize this respondent bias. First, surveys need to be responded by the appropriate person in the firm that has all the in- formation needed to respond. During the implementation of the screening process we ensure that the interview is arranged with the appropriate person or persons (Bloom et al., 2016). Senior managers (and in larger firms other managers such as plant managers) are asked to respond to the sections that cover the technologies used, and HR managers are asked to respond the questions on employment. Second, when possible use face-to-face interviews, which lead to higher response rates and lower respondent bias and measurement errors than web-based interviews.Only during the pandemic and due to existing mobility restrictions, we implemented surveys on the phone. Third, as discussed above, the use of a closed-ended design in the questionnaire reduces measurement error in the answers as the respondent is questioned about specific technologies (one at a time), and only when the presence of each of the possible technologies is established, the question about the most widely used technology is triggered. While this increases the length of the interview, it also increases the reliability of the data collected. Fourth, also as discussed above, we pre-pilot the questionnaire in each country to ensure that questions are clear in their wording in the specific geographical and cultural contexts, simple, and objective, so that the response does not require any subjective judgement (Bertrand and Mullainathan, 2001). Fifth, and more importantly, to avoid social desirability bias, by which respondents may overstate the use of more sophisticated tech- nologies, the survey avoids the words “technology" and “sophistication" and employs more neutral terms such as “methods" and “processes". In addition, the survey is administered so that the respondent does not know all the possible technologies in a business functions before asserting whether a technology is used in the firm.35 This reduces the risk that managers are framed to bias responses to the more advanced (socially desirable) technology, since they don’t know what they will be asked in advance. Finally, when possible, enumerators are instructed to visually verify the information provided during the interviews. For example, 35 It also allows for “don’t know" options. 104 in the case of use of a sophisticated production technology that can be visually identified in the shop floor. A.6 Ex-post Checks and Validation Exercises In addition to using best practices in survey design and implementation, it is important to perform validation checks once the data is collected. This allows us to measure the effectiveness of all these efforts to minimize bias and measurement error. In what follows, we describe some of the validation tests performed. Minimizing potential non-response bias Our survey implementation was designed to minimize non-response through the use of well-prepared agencies and institutions to adminis- ter the survey and the presentation of adequate supporting letters to encourage participation. Table A.17 shows response rates by country, firm size group and sector. Response rates vary between 15% in Croatia and 86% in Georgia. Table A.17: Response rates (by country) Country Response rate Bangladesh 30% Brazil 39% Burkina Faso 45% Cambodia 16% Chile 40% Croatia 15% Ethiopia 42% Georgia 86% Ghana 49% India 49% Kenya 77% Korea* 24% Poland 47% Senegal 57% Vietnam 80% Average across countries 46% These are unweighted response rates calculated as the ratio between firms that responded the survey and the total number of firms in the sample which we attempted to conduct the interview. The high response rate for Vietnam is associated with the fact that the survey was implemented by the national statistical office. In most cases, these response rates are high relative to typical response rates in firm-level surveys, which for the U.S. are around 5 to 10 percent, and are consistent with response rates observed for WMS and MOPS (Bloom 105 et al., 2016).36 To minimize potential non-response bias, we adjusted the sampling weights for unit non-response. The non-response adjustment was calculated at the strata level, so that the weighted distribution of our respondent sample across strata (sector, size, region) exactly matches the distribution of establishments in the sampling frame. More importantly, to check the reliability of the instrument we implemented a series of ex-post tests in the first phase of the survey, focusing on countries we implemented the survey first. First, we study whether, in the sample of contacted firms, there are significant differences between those that responded and those that declined participating or could not be reached. The only information available in all firms we attempted to contact in the three sampling frames is the number of employees. Table A.18 tests whether there are differences in employment between the respondent and non-respondent groups, controlling for characteristics used for stratification. We find no significant differences in firm size between respondents and non-respondents in any of the three countries. Table A.18: Comparison of establishment size between respondents vs non-respondents VARIABLES Brazil Vietnam Senegal Respondents (FAT) 2.52 52.34 -4.92 (22.19) (80.27) (6.63) Observations 1,754 1,500 3,075 R-squared 0.129 0.172 0.237 Controls: Sector FE Y Y Y Size-group FE Y Y Y Region FE Y Y Y Note : *** p<0.01, ** p<0.05, * p<0.1. Data are from the list of establishments contacted by the enumerators. For each country, the level of employment was regressed on a dummy for respondent while controlling for stratification such as sectors, size groups (small, medium, and large), and regions. Estimates for Vietnam are based on the original list of 1500 firms, with 1346 respondents and 154 non-respondents. Robust standard errors in parenthesis. Second, under the premise that any systematic relationship between firm characteris- tics and participation is continuous in their reluctance to participate in the survey, we can learn about sample differences between respondents and non-respondents by comparing firms across different percentiles of the distribution of the number of attempts it took for them 36 The average response rate for the WMS is around 40 percent.The response rate for MOPS, implemented by the United States Census Bureau, was around 80 percent. 106 to respond the survey.37 For Senegal, we explore whether after controlling for observable characteristics, there are significant differences in average technology sophistication in GBFs between firms that required a larger number of attempts to be contacted (top quartile) and those that did not. Table A.19 shows that there are no statistically significant differences in technology sophistication between the two groups. Table A.19: Comparison of technology sophistication between high and low number of at- tempts VARIABLES Senegal Senegal Top quartile of attempts (4 or more) -0.021 -0.027 (0.020) (0.019) Observations 1,753 1,666 R2 0.377 0.437 Controls: Sector FE Y Y Size-group FE Y Y Region FE Y Y Age Y Exporter Y Foreign owned Y Note: *** p<0.01, ** p<0.05, * p<0.1. Data are from the Senegal FAT survey with information on the number of attempts to complete interview at the firm level. Technology sophistication is regressed on a dummy for the top quartile of the number of attempts (4 or more) with controls for the stratification (sectors, size groups, and regions) and/or firm characteristics (age groups, exporter, and foreign owned). Robust standard errors in parenthesis. Third, we compare firms that were in the first sample list provided to enumerators and those in subsequent lists. Table A.20 show that there are no statistically significant differ- ences between the two groups. In each of these exercises, we find no statistical difference in the number of employees, technology sophistication, wages, and share of workers by skill and education between firms in the group that proxies for the response sample and the group of firms that proxies for the non-response sample.38 Minimizing enumerator bias. To minimize the potential for enumerators to introduce biases when administering the survey, we conduct in each country the same standardized training and piloting prior to going to the field. We also conduct ex-post tests to identify 37 Behaghel et al. (2015) infer the reluctance to participate in the survey from the number of attempts that it take for a firm to accept the request. 38 See Table A.18 to A.24 in Appendix A. 107 Table A.20: Comparison of technology sophistication between original and replacement sam- ple VARIABLES Brazil Brazil Vietnam Vietnam Senegal Senegal Original sample -0.014 -0.037 0.030 0.043 0.021 0.028 (0.048) (0.047) (0.050) (0.048) (0.018) (0.018) Observations 638 637 1,484 1,484 1,753 1,666 R-squared 0.299 0.335 0.262 0.320 0.377 0.437 Controls: Sector Y Y Y Y Y Y Size group Y Y Y Y Y Y Region Y Y Y Y Y Y Age Y Y Y Exporter Y Y Y Foreign owned Y Y Y Note: *** p<0.01, ** p<0.05, * p<0.1. Data are from the Brazil, Vietnam, and Senegal FAT surveys. For each country, technology sophistication (M OSTj ) is regressed on a dummy for the original sampling list with controls for the stratification (sectors, size groups, and regions) and/or firm characteristics (age groups, exporter, and foreign owned). Robust standard errors in parenthesis. abnormal interviews or outliers by running regressions of firm-level sophistication on enumer- ator dummies and firm controls as discussed in the text. Table A.21 shows that enumerator dummies are not significant for Brazil, Ghana, India, and Korea. For Bangladesh, Senegal, Vietnam, and Kenya, no more than 20% of enumerator dummies are statistically significant with respect to their distribution towards the type of firms they interview across various strata. Table A.22 compares the average technology sophistication (M OSTj ) for GBF, ex- cluding the firms with abnormal enumerators and in the entire sample. We find no economic or statistical difference between mean sophistication in these countries. Minimizing respondent bias. A critical factor to minimize respondent bias is to identify the right respondent. The protocol for the implementation of the survey required that the survey should be ideally answered by the top manager. About 47% of the survey was answered by the owner or CEOs, while the other responses included factory managers, other managers, administrative staff, and accountants. Almost 80% of the interviews were conducted through one visit in person interview with the main respondent. In circumstances in which the main respondent did not have all the information about a general topic of the questionnaire, especially in modules B and C, they were requested to consult with other colleagues. To assess the relevance of response bias, we conduct a parallel pilot in Kenya where 108 Table A.21: Analysis of enumerator bias distribution VARIABLES Brazil Vietnam Senegal Bangladesh Share of Significantly Different Interviewers 0 0.09 0.08 0.11 Number of Significantly Different Interviewers 0 13 2 4 Number of Interviewers 8 145 25 37 Ghana India Korea Kenya Share of Significantly Different Interviewers 0 0 0 0.2 Number of Significantly Different Interviewers 0 0 0 2 Number of Interviewers 44 18 9 10 Note: Data from the Firm-level Adoption of Technology (FAT) surveys in Brazil, Vietnam, Senegal, Bangladesh, Ghana, India, and Korea. Significantly different interviewers are identified from the regres- sions of employment on interviewer dummies with controlling for stratification information (e.g., sector, size, and region). For each country, the share of significantly different interviewers is computed by di- viding the number of interviews conducted by significantly different interviewers by the total number of interviews. we re-interview 100 randomly selected firms with a short version of the questionnaire. For those firms, we randomly select three business functions and ask about the presence of the relevant technologies.39 Both the original and back-end interviews in the pilot are conducted by phone by different interviewers. Despite using phone interviews which are subject to greater measurement error than face- to-face interviews, comparison of answers from the pilot reveals that 73% of the answers were the same across the two interviews.40 We estimate a probit model to assess the likelihood of consistent answers between the original and the back-check interviews, controlling for firm- level fixed-effect. Reporting the use of a technology in the back-check interview is associated with 80.6% of likelihood of reporting the use of the same technology in the original interview. Conversely, reporting that a technology is not used in the back-check interview, is associated with a 29.3% likelihood of being reported in the original survey. Additional validation exercise with employer-employee census (RAIS) in Brazil Some final ex-post checks were conducted with the Brazil data and takes advantage of the fact that we have access to the RAIS administrative data, which is a matched employer- employee dataset that covers the universe of firms in the sampling frame. This allows us to 39 The pilot coincided with the beginning of the data collection for phase two which includes new countries, and Kenya is one of them. Despite the fact that Kenya is not in the sample, the pilot is informative about the significance of response bias. The re-interviews produce 1,661 answers (106 interviews times 3 business functions times an average of 5.2 technologies per function). 40 The consistency ranges from 65% in business administration to 77% in sales across business functions, and from 85% among the most basic technologies to around 61% in intermediate, and 77% at the most advanced technologies across functions. 109 Table A.22: Difference in technology sophistication in general business functions with and without outlying enumerators All Sample Sample Without Difference Different Enumerators Vietnam Mean 1.934 1.947 -0.013 SE (0.012) (0.012) (0.017) Observations 1,499 1,341 Senegal Mean 1.406 1.404 0.002 SE (0.011) (0.011) (0.016) Observations 1,786 1,784 Bangladesh Mean 1.482 1.458 0.024 SE (0.015) (0.015) (0.021) Observations 903 798 Kenya Mean 1.938 1.936 .002 SE (0.020) (0.020) (0.029) Observations 1305 1296 Note: *** p<0.01, ** p<0.05, * p<0.1. Data from the Firm-level Adoption of Technology (FAT) surveys in Vietnam, Senegal, Bangladesh and Kenya. Brazil, Ghana, India and Korea are excluded because they do not include significantly different interviewers. The average of technology sophistication in general business functions (M OSTj ) is compared between all sample and sample excluding significantly different enumerators. Standard errors in parenthesis. compare variables in RAIS with variables we collected in FAT for the same firms. First, we analyze the correlation between sales per worker and our technology measures (GBF) and (SBF) from FAT and average wages from RAIS. Table A.23 reports the point estimates of regressing firm-level FAT variables on the log or average wages per worker from RAIS and a set of firm-level controls. The FAT variables are log of sales per worker (column 1), and average technology sophistication (GBF, column 2, and SSBF, column 3). In all three cases we find strong positive associations between the FAT and the RAIS variables. Second, we compare the differences between labor-related indicators from a matched employer-employee administrative data for firms in FAT versus the universe of firms. To perform this comparisons we obtained the weighted average for firms in FAT, using the weights we constructed as described in section A3 and compare it with the average for all 110 Table A.23: Relationship between FAT survey variables and log of wages from administrative data for Brazil (1) (2) (3) Variable log(sales per worker) GBF SSBF ln(Wage) RAIS 0.882*** 0.400*** 0.299*** (0.157) (0.111) (0.101) Observations 592 675 674 R-squared 0.346 0.364 0.800 Controls: Sector FE Y Y Y Region FE Y Y Y Size-group FE Y Y Y Age Y Y Y Exporter Y Y Y Foreign owned Y Y Y Note: *** p<0.01, ** p<0.05, * p<0.1. Average wage information for each establishment is obtained from the 2017 Relação Anual de Informações Sociais (RAIS) merged with the Firm-level Adoption of Technology (FAT) data used in this exercise, including sales per worker, the technology adoption index (M OSTj ) for GBF and SSBF, and firm characteristics used as controls. Regressions estimated using establishment-level sampling weights. Robust standard errors in parenthesis. firms in RAIS that are part of our universe for the State of Ceará, in Brazil 41 . We then perform a t-test to compare the differences. Table A.24 shows that the differences are not statistically significant. Overall, these ex post checks appear to validate the quality of the data collected. 41 The variables are number of workers, average wages, share of workers with college degree, share of low skilled, and share of high-skilled workers, where high- and low-skilled workers are defined as in Autor and Dorn (2013). 111 Table A.24: Comparison between FAT sample and RAIS data (universe) Number of Average Share Share Share high employees wage college low-skill high-skill FAT Average (weighted) 28.55 1,311.89 0.05 0.16 0.42 RAIS Average (universe) 23.85 1,349.29 0.05 0.17 0.39 Estimate (RAIS - FAT) -4.70 37.40 0.00 0.00 -0.03 Standard Error (3.08) (29.77) (0.01) (0.01) (0.02) T-Statistic -1.52 1.26 0.55 0.20 -1.64 Note: *** p<0.01, ** p<0.05, * p<0.1. Data from the 2017 Relação Anual de Informações Sociais (RAIS) and the Firm-level Adoption Technology (FAT) survey in Brazil. The estimates from RAIS data are unweighted, and those from FAT surveys are weighted by the sampling weights. Robust standard errors in parenthesis. 112 B Construction of Measures This section provides details about the construction of the various technology. B.1 Technology Measures The technology measures at the business function (BF) level that are discussed in the text are as follows - Nf , AN U Mf,j , N U Mf,j , M AXf,j , and M OSTf,j . The means of all of these BF-level measures are reported in ??. To better understand the construction of these variables, it is helpful to understand the structure of the corresponding questions asked. The figure below provides an example question asked for a particular business-function (BF) in the "Livestock" Sector. The presence of each of the technologies is calculated as a binary variable taking values 0 and 1.42 Figure B.1: Example question for the presence of technologies and most-used technology If a respondent answers that they use more than one technology in the BF, then they are asked about the most-used technology in that BF.43 42 If a respondent answers "Don’t know", that is coded as missing. We also do not take into account technologies outside the grid (Other). 43 Again here, if the answer is "Don’t know", or "Other", we assign it a missing value. 113 After constructing these technology-level binary variables and the variable for most widely-used technology (which are at establishment-BF level), we move on to constructing the Nf , AN U Mf,j , N U Mf,j , M AXf,j , and M OSTf,j , variables. The details and formulae used to construct each of these variables are listed below - Nf : This simply denotes the number of technologies in the grid that exist for business function f . This does not depend on the answers provided by the survey respondents, and takes the same value for a particular BF, across all establishments44 . AN U Mf,j : AN U Mf,j denotes the absolute number of technologies used by an estab- lishment in a specific business function. This variable is calculated by counting the number of technologies in a BF that the establishment confirmed using. Note that AN U Mf,j will always be less or equal than Nf . N U Mf,j : This denotes the relative number of technologies and is calculated using the AN U Mf,j −1 following formula: N U Mf,j = Nf −1 ∗ 4 + 1. This variable is an afine transformation of AN U Mf,j . We subtract 1 from both the numerator and denominator of the ratio in the formula so that this ratio ranges from 0 to 1. By multiplying this ratio by 4 and adding 1, the resulting measure ranges from 1 to 5 in all business functions of all establishments, hence allowing for comparability. M AXf,j : M AXf,j denotes the sophistication level of the most-sophisticated technol- ogy used by the establishment in the BF. It is calculated using the formula: M AXf,j = M AX −1 rf,j Rf −1 ∗ 4 + 1, where rf,j M AX is the absolute rank of the most sophisticated technology used in a BF, and Rf is the rank of the most sophisticated technology possible in the BF f . For most BFs, Rf equals Nf .45 However, there are a few exceptions where the number of technologies in the grid for ta BF does not equal the rank of possibly most sophistication technology. These exceptions are discussed in Appendix B.1.1. We apply the same afine transformation to the relative ranks so that the range of M AXf,j is [1, 5]. M OSTf,j : M OSTf,j denotes the sophistication level of the most widely-used technology in a particular BF by the establishment. It is calculated using the formula: M OSTf,j = 44 For instance in the example in Figure B.1, the value that Nf takes here is 4 for all establishments. Again, this does not include the "Other" technology, as it is not specified in the grid. 45 For example, if a respondent answered "Yes" to "Breed substitution", "Inbreeding", and "Artificial M AX Insemination", and "No" for all other technologies, rf,j would be 3 for that establishment for this BF. Here Rf = Nf = 4, so M AXf,j would be 3.67. 114 M OST −1 rf,j Rf −1 ∗ 4 + 1, where rf,j M OST is the absolute rank of the most widely-used technology in a BF, and Rf is the rank of the most sophisticated technology possible in BF f .46 The relative sophistication ranking is subject to an afine transformation so that the range of M OSTf,j is [1, 5]. After constructing the measures at the establishment-BF level, we construct establishment- level measures by averaging them across all the BFs of each establishment. In particular, N U M j , M AX j , and M OST j are calculated as follows - Nj Sf,j Sj = (B.5) f =1 Nj where S ∈ {N U M, M AX, M OST }, and Nj is the number of BFs conducted in establishment j. B.1.1 Exceptions As mentioned above, there are some exceptions to the calculation of M AXf,j and M OSTf,j statistics for particular business functions. These exceptions are limited to two sectors - Automotives (Motor Vehicles) and Health. Automotives - As specified in Figure A.5, in BFs "Body Pressing", "Painting", and "Plastic Injection Molding", although the number of technologies is more than 2, there are sub-BFs, where the ranking of technologies is only 2 (basic and advanced). For instance, in "Body Pressing", there are three distinct sub-BFs - "pressing of skin panels", "structural components", and "welding of main body" (refer to Figure A.5). As such, if the ranking of the technologies were to take the value of the technology as per the questionnaire, that would be erroneous. Another example is provided in Figure B.2 for the question related to Body Pressing. Consider for example, an establishment answering "Yes" to 1st, 2nd, 3rd, and 6th technologies, and for intensive margin answering "3". If we were to calculate scaled-up M AXf,j and M OSTf,j values for this establishment in body pressing, as we do for all other BFs, it would be equal to "5" and "2.6" respectively (if we take Rf = Nf = 6). However, this is incorrect, as the 1st, 2nd, 3rd, and correspondingly, 4th, 5th, and 6th technologies are related to the parallel sub-BFs respectively. Hence, we assign 46 For example, if an establishment that uses "Breed subsbtitution", "Inbreeding" and "Artifical insem.", M OST but uses "Inbreeding" most often, rf,j would be 2. Consequently, M OST f,j would be 2.33, for this establishment in this BF. 115 Figure B.2: Question for Extensive and Intensive Margins of Tech. in Body Pressing rank "1" to the first three technologies, and rank "2" to the last three technologies, and take an arithmetic mean to calculate rj M AX . So in the example, the most sophisticated technology for the establishment in "pressing of skin" would be rank 1, and both, in "pressing of structural panels" and "welding", would be 2. Effectively rjM AX would then be an arithmetic mean of {1, 2, 2} which is "1.6". Accordingly, value of Rf would be 2, and hence the corrected scaled-up value of MAX and MOST for this estab. in this BF, would be "3.4" and "3" respectively. Table B.2 shows the adjustment made to the ranks in these BFs before calculation of M AXf,j and M OSTf,j . Health - The exceptions in this sector come from two BFs - "Health Equipment" and "Procedures". Firstly, in "Health Equipment" the number of technologies is 11 (see ??), but there exists no clear ranking. For that reason, to calculate M AX f,j and M OST f,j for this 116 Table B.2: Adjustment of Ranks in Automotive BFs Business Function Sub-Business Functions Technologies Rank in Adjusted Rf Nf Data Rank Body Pressing and Welding Pressing of Skin Panels using Operators 1 1 2 6 using Robots 4 2 2 6 Pressing of Structural Components using Operators 2 1 2 6 using Robots 5 2 2 6 Welding of Body Parts using Operators 3 1 2 6 using Robots 6 2 2 6 Painting Water-Based using Operators 1 1 2 4 using Robots 3 2 2 4 Solvent-based using Operators 2 1 2 4 using Robots 4 2 2 4 Plastic Injection Modelling Molding of non-visible interior plastic using Operators 1 1 2 4 using Robots 3 2 2 4 Molding of plastic exterior body parts using Operators 2 1 2 4 using Robots 4 2 2 4 Notes : The table shows the ranking of technologies in these three BFs of Automotives sector. BF is difficult. As a result, the value of M AX f,j and M OST f,j is taken to be the same as the value of N U M f,j (which is the relative number of technologies used by an establishment in the BF), for this particular BF. Coming to "Procedures", both in the data and questionnaire, the questions regarding different procedures are asked individually (see Figure A.10). There are 4 types of procedures - Sepsis Treatment, Childbirth, Trauma, and treatment of Myocardial Infarctions/Strokes. For each of these procedures, there are two corresponding technologies. When the question is asked to the respondent, there are 5 possible options that they could reply, namely - "It is always available", "It is NOT always available", "It is NEVER available", "Don’t know", and "Not applicable". For the purpose of classification, firstly the responses for each technology are collapsed into binary variables, taking values 1 for the "always available" option, and 0 for "NOT always" and "NEVER available" options. Just like in other BFs, value of "Don’t Know" and "Not Applicable" is considered as missing. After the re-coding, these 4 questions are collapsed into one "Procedures" question that would have 8 technologies. Similar to "Health Equipment", there is no clear ranking to technologies there, hence the value of M AX f,j and M OST f,j is taken to be the same as the value of N U M f,j for this particular BF. 117 C Additional Figures and Tables C.1 Technological Sophistication Figure C.3: Distribution of M AXj and M OSTj across establishments Notes : This figure represents the distribution of sophistication measures at the establishment level - M AXj and M OSTj . The histogram and kernel density curves are calculated using establishment-level sampling weights. 118 Table C.1: Summary Statistics All Small (5-10) Medium (11-99) Large (100+) N Mean p50 N Mean N Mean N Mean Total # of employees 18570 33.95 9.00 9909 8.45 5477 38.03 3184 492.04 % of workers with college 17686 30.15 20.00 9520 30.79 5144 27.71 3021 28.12 Management practice (z-score) 20739 0.00 0.27 11029 –0.09 6098 0.25 3610 0.53 Sales per worker 15355 11.45 11.43 8121 11.47 4545 11.39 2689 11.25 Share Share Share Share Multi-establishment 18.4% 12.1% 20.5% 33.9% Multinational 17.6% 16.4% 16.4% 23.7% Exporter 16.8% 8.1% 18.1% 41.0% Age: 1–5 Years 17.9% 22.2% 15.2% 9.3% 6–10 Years 21.3% 24.1% 19.6% 15.4% 11–15 Years 17.9% 18.4% 17.2% 17.4% 16+ Years 42.9% 35.3% 47.9% 57.9% Has electricity, computer and internet 77.1% 67.0% 85.7% 95.0% Sector: Agriculture 5.1% 5.1% 5.2% 5.1% Livestock 1.9% 2.2% 1.7% 1.4% Food Processing 10.2% 8.9% 11.8% 11.3% Apparel 7.4% 6.9% 6.4% 10.5% Motor vehicles 2.9% 2.0% 3.2% 5.0% Pharmaceuticals 2.7% 1.5% 3.2% 5.3% Wholesale or retail 14.7% 18.0% 12.4% 8.6% Financial services 3.4% 3.2% 4.0% 2.6% Land Transport 5.0% 5.0% 5.2% 4.3% Health Services 4.4% 2.1% 5.5% 9.5% Leather 2.2% 2.2% 1.8% 2.7% Other Manufacturing 15.8% 16.1% 15.1% 16.1% Other Services 24.4% 26.7% 24.3% 17.6% Notes : The table reports summary statistics of establishment-level measures and distribution of establishments, in the overall sample, and by size-groups. The top panel consists of summary statistics calculated using establishment-level sampling weights. The bottom panel reports the unweighted shares of establishments belonging to the various groups. 119 Table C.2: Average level of technology measures Business Function N U Mf,j M AXf,j M OSTf,j Nf AN U Mf,j Business Function N U Mf,j M AXf,j M OSTf,j Nf AN U Mf,j General Business Functions Automotive Business Administration 2.1 2.1 3.0 2.5 5 Vehicle Assembly 2.5 2.2 2.9 1.2 6 Production Planning 1.9 1.9 2.7 2.2 5 Body pressing and welding * 2.3 2.0 1.2 1.3 6 Sourcing 1.9 1.9 2.3 1.7 5 Painting * 1.6 1.8 1.1 1.1 4 Marketing 2.0 2.0 2.3 1.7 5 Plastic injection molding * 1.7 1.9 1.2 1.1 4 Sales 2.2 1.9 2.4 1.6 6 Productive assets management 1.8 2.6 3.2 2.1 3 Payment 2.9 2.2 3.5 2.7 7 Fabrication 2.1 1.9 2.2 1.6 6 Quality Control 1.6 1.8 2.1 1.6 4 Agriculture Pharmaceuticals Land Preparation 1.9 2.2 3.6 3.0 4 Facilities 1.6 1.8 3.0 2.6 4 Irrigation 1.9 1.7 2.7 2.4 6 Weighing scale 1.9 2.2 4.1 3.7 4 Weeding 1.9 1.9 2.4 2.0 5 Mixing/Compounding 1.8 1.8 2.9 2.4 5 Harvesting 1.9 1.9 2.7 2.1 5 Encapsulation 1.8 2.1 3.8 3.1 4 Storage 1.8 1.8 2.7 2.3 5 Quality control 1.6 2.3 3.3 2.7 3 Packaging 1.6 1.8 2.1 1.7 4 Packaging 1.5 2.0 3.3 2.8 3 Fabrication 2.0 1.8 2.3 1.9 6 Livestock Wholesale and Retail Breeding 2.0 2.4 3.0 2.5 4 Customer service 2.3 2.3 2.5 1.5 5 Feeding 3.0 2.3 3.5 2.7 7 Pricing 1.9 1.9 2.3 1.7 5 Animal healthcare 3.2 3.2 4.3 3.0 5 Merchandising 1.6 1.8 2.1 1.6 4 Herd management 2.0 1.7 2.0 1.3 7 Inventory 1.7 1.7 2.4 2.0 5 Transport of livestock 2.1 2.4 3.5 2.8 4 Advertisement 2.5 2.2 3.0 2.3 6 Food Processing Financial Services Input testing 1.7 2.0 2.4 1.6 4 Customer service 3.3 4.1 4.7 2.1 4 Mixing/cooking 2.2 2.6 3.2 2.2 4 Avoid fraud 3.4 2.9 3.5 1.7 6 Anti-bacterial 1.9 2.2 2.8 2.1 4 Loan applications 2.7 3.3 3.8 1.7 4 Packaging 1.8 2.1 2.5 1.9 4 Credit applications 2.3 2.7 3.1 1.8 4 Food storage 1.9 2.2 3.1 2.6 4 Operational support area 2.2 3.4 4.3 3.0 3 Fabrication 1.8 1.6 1.9 1.5 6 Wearing Apparel Transportation Design 1.5 2.0 2.6 1.9 3 Planning 1.6 1.8 2.1 1.6 4 Cutting 2.0 2.0 2.4 1.9 5 Plan execution 1.9 1.9 2.2 1.5 5 Sewing 2.1 2.1 2.8 2.4 5 Monitoring 1.9 1.9 2.4 1.6 5 Finishing 1.7 1.7 2.1 1.7 5 Performance measurement 1.8 1.8 2.2 1.7 5 Fabrication 1.9 1.7 1.9 1.6 6 Maintenance 1.6 1.8 2.3 1.7 4 Leather and Footwear Healthcare Design 1.5 2.0 2.4 2.0 3 Infrastructure and Machines * 5.5 2.8 2.8 2.8 11 Cutting 2.2 2.2 2.6 2.1 5 Appointment and Scheduling 2.0 2.3 2.6 1.7 4 Sewing 2.2 2.2 2.5 2.1 5 Patient records management 1.9 2.2 3.0 2.4 4 Finishing 1.6 1.6 1.8 1.5 5 Healthcare management ** 1.2 1.9 2.7 0.0 2 Fabrication 1.7 1.5 2.1 1.9 6 Procedures * 4.4 3.0 3.0 3.0 8 Other Manufacturing Fabrication 1.9 1.7 2.1 1.7 6 Notes: The table reports the weighted average of the variables listed in the top row for each business function. The weights used are the sampling weights. 120 Table C.3: Percentage of establishments with technology gaps Business Functions GAP Business Functions GAP General Business Functions Automotive Business Administration 34% Vehicle assembly 91% Production Planning 27% Body pressing and welding 0% Sourcing 15% Painting 0% Marketing 15% Plastic injection molding 0% Sales 28% Productive assets management 3% Payment 48% Fabrication 21% Quality Control 5% Agriculture - Crops Pharmaceutical Land Preparation 37% Facilities 7% Irrigation 48% Weighing 27% Pest Control 23% Compounding 24% Harvesting 58% Encapsulation 35% Storage 11% Quality Control 12% Packing 9% Packaging 15% Fabrication 11% Livestock Wholesale and Retails Breeding 13% Customer Service 13% Nutrition 60% Pricing 15% Animal healthcare 58% Merchandising 7% Herd management 67% Inventory 7% Transport of Livestock 30% Advertisement 39% Food Processing Finance Input Test 17% Customer Service 27% Mixing Blending Cooking 18% ID Verification 33% Anti-bacterial 13% Loan Application 23% Packaging 5% Loan Approval 8% Food Storage 8% Operational Support Area 0% Fabrication 8% Wearing Apparel Transportation Design 8% Planning 8% Cutting 6% Execution 28% Sewing 10% Monitoring 39% Finishing 9% Performance Measurement 26% Fabrication 7% Maintenance 13% Leather and Footwear Healthcare Design 2% Infrastructure and Machines 81% Cutting 13% Scheduling Appointments 5% Sewing 2% Management of Patient Records 5% Finishing 4% Healthcare Management 0% Fabrication 3% Procedures 51% Other Manufacturing Fabrication 22% Notes: The table reports the percentage of establishments that experience sophistication gaps in each business function. The percentages are calculated using sampling weights. 121 D Detailed Acknowledgments The preparation of the Firm-level Adoption of Technology survey questionnaire involved the contribution of several sector experts within and outside the World Bank. First, we would like to thank the following World Bank Group colleagues: Erick C.M. Fer- nandes (Lead Agriculture Specialist), Holger A. Kray (Manager, Agriculture), Michael Mor- ris (Lead Agriculture Economist), Madhur Gautam (Lead Agriculture Economist), Robert Townsend (Lead Agriculture Economist), Ashesh Prasann (Agriculture Economist), Apara- jita Goyal (Senior Economist, specialized in Agriculture), Harish Natarajan (Lead Financial Sector Specialist), Erik Feyen (Lead Financial Sector Economist), Laurent Gonnet (Lead Financial Sector Specialist), Arturo Ardila Gomez (Lead Transport Economist), Victor A Aragones (Senior Transport Economist), Edson Correia Araujo (Senior Health Specialist), Irina A. Nikolic (Senior Health Specialist), Brendan Michael Dack (Chief Industry Specialist at IFC), Emiliano Duch (Lead Private Sector Specialist), Blair Edward Lapres (Economist), Kazimir Luka Bacic (Operations officer), Justin Hill (Senior Private Sector Specialist), Eti- enne Raffi Kechichian (Senior Private Sector Specialist), Justin Yap (Private Sector Special- ist), Austin Kilroy (Senior Economist). Similarly, we would like to thank several external experts. From Embrapa-Brazil, we would like to thank Alexandre Costa Varella (Head of Embrapa research unit in the South region, expert on livestock), Flávio Dessaune Tardin (Researcher, Maize and Sorghum), Edi- son Ulisses Ramos Junior (Researcher, Soybean), Isabela Volpi Furtini (Researcher, Rice and Beans), Alberto Duarte Vilarinhos (Researcher, Cassava and Fruits), Carlos Estevão Leite Cardoso (Researcher, Cassava and Fruits, Technology transfer), and other participants of the internal seminars to validate the sector-specific questionnaire for agriculture and live- stock. We also would like to thank a group of senior consultants from Patina Solutions, who contributed with the preparation and validation exercise for several sectors, including Daren Samuels (Consultant, Manufacturing), Christ Baughman (Consultant, Food Process- ing), Sandra Aris (Consultant, Wearing Apparel and Lather), Steve Zebovitz (Consultant, Pharmaceutical sector), Shelly Wolfram (Consultant, Retail and Wholesale), James M. Ked- ing (Consultant, Transport and Logistics), as well as Justin Barners (Consultant, Automo- tive). Finally, we would like to thank Sudha Jayaraman (Associate Professor, Department of Surgery, Virginia Commonwealth University), Christina Kozycki (Infectious Disease Fel- low, NIH), Jonathan Skinner (Health Economist, Dartmouth College), and Elizabeth Krebs (Assistant Professor of Emergency Medicine at the Jefferson Center for Injury Research and Prevention), for their comments and feedback on health services. We also would like to thank the National Statistics Agencies and local partner institutions 122 that provided access to the sample frame and/or support the implementation of the survey. They include: Statistics Korea, Statistics Poland, National Statistics and Demography of Senegal (ANSD), the Federation of Industries of the State of Ceará (FIEC), and the General Statistics Office of Vietnam (GSO), the Kenya National Bureau of Statistics (KNBS), the Bangladesh Bureau of Statistics, and the Central Statistics Office of India. 123