4.1. Technology-Driven Structural Change in Employment in NAICS Sectors
The BLS data enable us to study the projected structural change of employment across actual sectors and to infer on the different countervailing effects at work in the various types of sectors. Following the method described in Section 3.2
, we obtained Employment Projections for all occupations affected by the focal technologies and broken down by NAICS sector. In Figure A3
in Appendix B
, we report the employment increase for these filtered occupations by major SOC group, for each of the affected sectors. A first observation is that the number of sectors as well as the number of jobs directly affected is limited. In the last two columns on the right, the projections in total number and percentage of increase in number of jobs are reported for all occupations per major group, not just the filtered occupations. A closer look at the projected shifts in employment within the same sector (vertically) and across sectors (horizontally) are in line with displacement and various countervailing effects and fully captured within our multisectoral framework of structural change.
Overall, there is a relatively small job loss (−75.3 k, about 0.0%) in occupations affected by the focal technologies, but a loss is in itself already in stark contrast with the high number of jobs expected to be created in the economy as a whole (11,519 k, about 7.4%). When looking at the projected growth in the number of jobs by major group of occupations, we see that there is almost exclusively a decline in the number jobs in Production (SOC 51) and Office and Administrative Support (SOC 43) occupations. In Production, this is almost all due to a decrease in the Manufacturing sector (NAICS 31–33). Note that the job loss overall (−406.9 k, −4.3%) is much lower than the job loss in occupations affected by technology (−681.9 k people, −13.4%), underlining the effect of technological unemployment. In Office and Administrative Support occupations, the story is quite different: the number of administrative jobs in a wide range of sectors such as Manufacturing (NAICS 31–33), Government (NAICS 90), Education (NAICS 61), Wholesale trade (NAICS 42), Information (NAICS 51), and several others is declining, while the number of administrative jobs in Health Care (NAICS 62), Management support and services (NAICS 56), and Retail trade (NAICS 44–45) is increasing. Detailed analysis reveals that any (partially technology-driven) decrease affects mostly secretaries and office clerks across the aforementioned sectors, while the increase is due to job creation in service representatives, information and stock clerks. That said, the rate of increase in administrative jobs in the economy as a whole (0.6%) is relatively low compared to the total rate of job creation (7.4%), which may well be (partially) caused by a technology-driven productivity increase.
There also are several occupations in which there is a strong increase in the number of jobs, notably Computer and Mathematical (SOC 15), Management (SOC 11), and Architecture and Engineering (SOC 17) occupations. While the biggest increase occurs in the Professional, scientific, and technical services (NAICS 54) overall, there may be a shift of employment across occupations within this sector as well as in the Information (NAICS 51), and Finance and Insurance (NAICS 52) sectors. In these sectors, a technology-driven loss of jobs in Administrative jobs (SOC 43) seem to be offset by a gain of jobs in Computer and Mathematics related jobs (SOC 15), possibly created to reap complementarities (or where people may be hired to automate processes and thus cause technological unemployment). For example, in the Information sector (in which data is created, processed, and transferred), there is a loss of 43.5 k administrative jobs but an increase of Computer and Mathematical jobs.
When looking at the projected growth in the number of jobs by sector, the change is much in line with “classical” effects of automation, demographic developments and progressive outsourcing. There is a particularly strong employment growth in the Professional, scientific and technical services (NAICS 54 with about 331.8 k people), of which the Occupational Utilization reveals that there is an increasing demand for engineers, e.g., due to robotization and automation (SOC 17), but also the use of advanced digital and internet-connected devices, and outsourcing of security (SOC 15). Moreover, we observe a decrease in Office and Administrative Support (SOC 43) due to automation and an increase in Business Operations and Management (SOC 13, 11) due to outsourcing and purchasing training and consulting services. Health care (NAICS 62, 251.5 k) sees a rise on several occupations due to the aging population and several organizational changes such as the introduction of reception services and team-based structures to cope with that. The Administrative and management services (NAICS 56, 105 k people) sector sees an increase in employment, e.g., for customer representatives, as there is more outsourcing in other sectors. Moreover, there is substantial job loss in the Manufacturing sectors (NAICS 31–33, −751.1 k people), notably due to substitution, but also offshoring. Obviously, the job loss due to substitution is not (entirely) offset by intra-sectoral complementary jobs. It may, however, be (partially) compensated by the development of software, robot and AI as well as the production and servicing thereof, which reflects in the statistics for other sectors described before.
Illustratively, of the thirty fastest growing occupations reported in the BLS Employment Projections (see Figure A4
in Appendix C
), a total of seventeen occupations are directly or indirectly related to healthcare, either as healthcare practitioner (SOC 29) or support (SOC 31), personal care (SOC 39) or by providing training for healthcare (SOC 25). This, again is to be attributed to demographics and aging, but arguably also local demand spillover, which is related to disposable income. Five occupations pertain to computers and mathematics (SOC 15). Of the thirty fastest declining occupations reported, sixteen are Production-related occupations (SOC 51) such as operators, assemblers, setters, etc., and six Office and administrative occupations (SOC 43) such as computer and telephone operators, typists, data keyers, etc. In addition, in our subset of occupations (that are affected by the focal technologies), we see a confirmation of the trend that there is a decline in low- and medium-skilled jobs (such as assemblers) and growth in high-skill, high-productivity jobs (such as software engineers) (see e.g., [6
]). Surprisingly, there is an increase in medium- to low-skilled jobs for occupations such as stock clerks and order fillers, receptionists and information clerks, which appear to be automatable. Supposedly, this is, on the one hand, due to a demand-side increase and, on the other hand: (i) a low acceptance for digital services in occupations in which human contact is appreciated; (ii) yet relatively poor performance of automated services (so limited substitutability); and (iii) lowering of unit costs actually compensating automation with higher demand.
Conclusively, the occupational outlook data features structural change in line with displacement and countervailing effects discussed above, but also reveals that several countervailing effects and other factors (demographics and offshoring) may be counteracting.
4.2. Automatability of Occupations
For a socially and economically sustainable growth path, the labor displacement in the sectors of application must be counterbalanced by job creation within the same and other sectors. Following our approach described in Section 3.3
, we construct a comprehensive data set and compute automatability scores for major occupation groups as in formula in Equation (3
). As the actual values of the automatability rely on the activity automatability score values, the absolute values have no particular meaning and are not reported. However, the relative
values of course indicate relative automatability. Arguably, not only the automatability, but also the absolute number of jobs determine the actual probability that particular occupations face automation. As such, we plot the automatability against the number of jobs in Figure 3
plots the number of jobs
for which the normalized automatability
exceeds the threshold x
on the X
-axis. Only 25.5% of jobs have a normalized automatability score of
(interpreted as that half of the activities of an occupation can be fully automated within 10 years). Moreover, only a relatively low percentage of 4.65% of jobs has a normalized automatability score of
or higher. Thus, in contrast to the figures reported in [8
], we find that very few jobs are highly automatable. This is largely in line with findings in [40
]. Moreover, Manyika et al. [52
] (see p. 5), found that, for about 60% of occupations at least one third of the activities can be automated. We find something related: just below 60% of jobs (so not occupations) has a normalized automatability of one-third or more. Thus, for roughly 60% of the jobs, roughly one third or more of all activities can be automated.
Using a straightforward translation of the automatability of work activities into automatability of the occupation, we find that the high level of automatability reported in [8
] may well be overestimated, and actually have a much lower level such as those reported in [40
]. In addition, neglecting the variability of tasks between workplaces also causes overestimation of automatability [51
]. Thus, we have added another estimation method to the toolkit of researchers using the O*Net database (see e.g., [8
]), although further refinement may well be needed.
4.3. Classification of Occupations on Type of Sector and Type of Technology-Driven Change
In Section 2
, Figure 1
contains the changes in occupations subject to automation (and notably conceptions of the use of robotics and advanced software such as AI) on the one dimension and various types of sectors affected on the other dimension. As described in Section 3
, we have filtered the occupations in the BLS Occupational Outlook Handbook on keywords related to our focal technology. Using the job activity and job outlook descriptions (and notably the expected changes in these occupations), the occupations are manually classified in Figure 1
. In the classification of occupations, we have used the automatability scores of occupations as a crosscheck, because highly automatable occupations are: (i) found in sectors in which the technology is applied; and (ii) feature a substantial number of activities which can be substituted so are “old” occupations with routinized, readily standardized tasks. Note that the OOH contains existing occupations. For occupations which are supposedly emerging, we filtered occupations from the Atlas database (following the method described in Section 3
). These occupations are subsequently also manually classified, see Figure A6
). Figure 5
contains descriptive names for aggregated groups of occupations (which are our own labels, not the SOC major or minor groups) found in Figure A6
. The color shading indicates whether employment in the occupations in particular sectors is assessed to be decreasing
(applying sectors with existing occupations, these are subject to substitution), balancing
different counteracting developments (e.g., complementary and spillover sectors), or increasing
(e.g., producing and developing sectors). We now discuss the employment for each type of sector and type of technology-driven change in occupations in detail.
As expected, we see that occupations that have been classified as “substituted for” in the “applying sectors” typically are in Office and Administrative Support (SOC 43), Transportation and Material Moving (SOC 53), and Production (SOC 51) major groups which are classified as highly automatable (for a quick reference, see Figure 3
, for a detailed analysis, see Appendix D
). That said, particular occupations that are economically attractive to be automated (because of both a high automatability and high number of jobs involved) according to Figure 3
do not yet show up in the OOH assessments and hence are less affected by our focal technology (i.e., automated) than expected so far. This definitely holds for Construction and extraction (SOC 47); Building and ground cleaning and maintenance (SOC 37); Installation, Maintenance and Repair (SOC 49); and Farming, fishery, and forestry (SOC 45), as well as for Food preparation and serving (SOC 35) and Sales (SOC 41). However, note that, with regard to the occupations that are at risk of being automated, our findings are largely in line with those in [8
]. These authors also expected Production, Construction and Extraction, Transportation and Material Moving, Office and Administrative Support, but also Sales and Maintenance and Service to be subject to computerization and robotization (see, e.g., [8
], p. 37). Manyika et al. [52
] found high automatability for the industrial sectors such as Transporting and Warehousing, Manufacturing, but also for those sectors with yet limited automation, such as Accommodation and Food Services, Agriculture, and Retail and Wholesale Trade. Note that our automatability scores and those in the aforementioned studies are based on expert evaluation of whether robots and AI can or cannot perform particular activities within ten years
. If the automation does not materialize in the forthcoming decade, we may have to attribute this discrepancy to a technologically optimistic expert classification of what robots and software can actually do (cf. [40
For occupations in the “making” sectors (producing and developing as well as supplying and supporting), it is not possible to determine how much growth in employment is to be attributed to automation, because of compounding of factors that affect employment (e.g., demographics, outsourcing and offshoring, technological change, and demand dynamics). As such, any figures from the BLS Employment Projections are to be interpreted prudently. As described in the Methodology Section, we made a selection of BLS SOC occupations by filtering on technology-related keywords in the factor descriptions reported in the Occupational Utilization dataset in the Employment Projections. In the descriptions of factors determining utilization, there is an increase reported for occupations headed under Computer and Mathematical Occupations (SOC 15) due to the introduction of software into devices and the use of and need to analyze more data, but also a loss due to cloud computing and offshoring as well as under Architecture and Engineering Occupations (SOC 17) due to the automation of production to cut labor costs. Looking at the OOH figures for the change in employment 2016–2026 for these filtered occupations, employment in the “making” sectors (SOC 15, 17) increases with 371.1 k jobs, while employment in the applying sectors (SOC 43, 45, 51, and 53) decreases with 464.9 k jobs. If we include the manually selected occupations related to computers (SOC 15–11xx), engineering (SOC 17–302x), and management thereof (SOC 11–30xx) in the figure for the “making” sectors, the earlier figure changes to an increase with 659.2 k jobs. Given that the effect of the introduction of robotics and AI (taking only these “applying” and “making” types of sectors) changes from a net loss to a net gain, a detailed analysis is required for further conclusions.
Thus, already in the descriptions of factors causing an increase in projected employment for existing
occupations, there is mention of data collection and analysis as well as development and implementation of advanced information as well as production and logistics systems. However, for emerging
occupations, projections for employment are fraught with uncertainty. The jobs that are typically reported to be created in the nearby future are in the IT sector (e.g., in Big Data, AI, IoT, sensor technology, and general software engineering) and industrial sectors (e.g., in the automotive and robotics sector, advanced transportation and warehousing, and 3D printing) (cf. [53
]). The Atlas database refers to more specific occupations, typically concerned with developing, producing, or applying robots, information systems, software, etc. to specific sectors (e.g., agriculture, health care, construction). In addition, consulting companies gradually start to distinguish robots (e.g., industrial versus service robots) and AI software (e.g., vision or language processing, planning, expert systems and knowledge management) on the basis of (field of) application. This underlines the progressive diversification and institutionalization central to the structural change literature.
4.4. Spillover and Disposable Income
In our multisectoral perspective, an important countervailing effect is local demand spillover [13
], which is closely related to disposable income. Arguably, people spend their disposable income (to a certain extent) in what we consider the quaternary sectors, i.e., leisure and traveling; arts, theater and culture; sports and lifestyle; books and magazines; entertainment sectors making music, series, film and computer games; gambling; etc. As such, the employment in these “spillover” sectors evolves with the total disposable income made in the sectors in which our focal technologies are applied, developed and produced, supplied and supported as well as in the complementary sectors (e.g., education, legal support, and business consulting). Taking the BLS occupation group SOC 27 as an indicator of the employment in the quaternary spillover sectors, we see that the BLS OOH reports an estimated growth of 2158.5 k to 2276.2 k jobs over the years 2016–2026, which, with a growth rate of 5.5% over 10 years, actually has a lower than the estimated average growth rate of the employment in the economy as a whole, which is 7.3%. Arguably, we are considerably underestimating the spillover effect, because people also spend their disposable income on personal and health care service, both more luxurious consumable and durable goods, etc., all accounted for in figures of other sectors. If we consider health and personal care also a spillover category and account for change in the occupation groups SOC 29, 31, and 39, the rate of growth over the years 2016–2026 is a much higher 17.3%, outweighing the average employment growth rate.
reports the development of the (projected) disposable income over the years 1996–2026. Note that the disposable income increased even during the financial crisis. Moreover, we observe that the increase in employment in the spillover sectors is lower than the increase in disposable income. As argued, part of the income may be spent on more luxurious consumables and durable goods (or saved, possibly effectively deferring consumption to the future). That said, with a continued growth in disposable income, we also expected an increase in employment in the quaternary sector. The development of disposable income for occupations differs for the various types of sectors. In a nutshell, we expect that, in the sector of application, there is pressure on the wages (and thereby disposable income) in occupations facing substitution, but not if tasks are added that require upskilling. Similarly, disposable income in facilitating and inhibiting sectors is expected to decrease (increase) for occupations related to old (new) technology. Finally, the wages and thereby disposable incomes in the “making” sectors are expected to increase.
4.5. Macro-Level Development of Employment: A Structural Rebound?
In Section 2.3
, we report three basic scenarios on the macro-economic development of employment. Empirical evidence in [45
] indicates that a short-run dip in employment is followed by a bounce back to “regular” levels of (frictional) unemployment. This is corroborated by the long run employment levels, see for instance the levels reported for the UK by the Bank of England in Appendix A
. Empirical evidence shows that the unemployment rate, although fluctuating, remains fairly stationary for a century and a half, effectively absorbing major sectoral shifts, a rapid increase of women’s and migrant workers” participation on the labor market (see Figure A1
). That said, at the same time, the hours worked annually decreased substantially in OECD countries (see Figure A2
), effectively countering an increasing in unemployment otherwise observed (see Appendix A
). A critical remark is, hence, that the atomistic view on single jobs as employment measure is somewhat misleading; there may be covert technological unemployment. Note that the drop in the number of working hours reduces disposable income, yet frees up time to spend that disposable income, e.g., on products and services provided by quaternary sectors.
Our preceding analysis reveals that the span of occupations that are partly or entirely automatable as well as the set of sectors thus affected by automation are limited. Consequently, the “end of work” scenario manifests itself in a limited part of the economy at best. We also argued that the loss of jobs due to automation is compensated in various ways. We found that an increase in employment in the “making” as well as “applying” sectors is projected, which includes new occupations related to data, advanced information and production systems, advanced applications of software, etc. The rapid rates of growth of jobs in these occupations may well be able to compensate for the destroyed jobs in “old” occupations (particularly those facing substitution). On top of that, we have observed a strong increase in disposable income which is expected to sustain the high growth rates in employment in quaternary sectors and personal care and healthcare sectors.
Thus, on top of the fact that the number of sectors affected by automation is limited, we have (mostly qualitative) observations that job loss in the applying sectors is limited. This may be well counterbalanced by job creation both in directly related (new) sectors as well as in the spillover sectors. As such, we expect that, if we see a dip in employment in the forthcoming years, this will merely be temporary and in a limited set of sectors, and that, ultimately, there is a rebound to frictional unemployment levels. In the next section, we provide recommendations on policy measures that enhance inter-occupational and geographical mobility and thus assist in such a rebound.