Report from the McKinsey Technical Institute
Automation will be a global force, affecting all countries, all sectors, all jobs, and all work activities. Already today, machines and algorithms are playing a much larger role in the workplace, but how soon will it be before we all feel the impact of automation technologies? Could machines really carry out much or most of the work humans do today—and if so, by when?
In this chapter, we describe factors that can affect the pace and extent to which automation is adopted in the global economy. We also present the results of a model that includes a range of potential scenarios illustrating how the automation of existing work activities could evolve.
FIVE FACTORS INFLUENCING AUTOMATION OF WORK ACTIVITIES
Overall, we have identified five broad factors that can influence the pace and extent of automation of work activities. They are: technical feasibility; the cost of developing and deploying solutions; labor market dynamics; economic benefits; and social and regulatory acceptance.
Technology has been automating human activities for centuries, from the printing press to the steam engine and the internet, and fundamentally reshaping the economy in the process. Over the past two centuries, the share of people working on the land in advanced economies has fallen from a majority to a tiny fraction. More recently, the United States and other advanced economies have seen a decline in the share of the workforce engaged in manufacturing. Technological advances require basic scientific research, but in order for these advances to be adopted, they also require engineering solutions, or “applied research.” Both take time to develop. There is a lag between a technology being demonstrated, and a viable product being developed using that technology. Orville and Wilbur Wright pioneered flying an aircraft in 1903, for example, but it took 11 more years before the first commercial flight, across Tampa Bay, in Florida, took place, and the true birth of commercial aviation in the United States is usually dated back to 1926, when pioneering operators had to begin complying with federal regulations. 62 There was a similar lag in the development of automobiles, between German engineers Nikolaus Otto, Gottlieb Daimler, and Wilhelm Maybach, patenting the compressed charge, four-cycle engine in the 1870s and production of automobiles on a commercial scale some 15-20 years later. Solutions have to be engineered for specific use cases. While from a “scientific” perspective, a light passenger vehicle is conceptually the same as a tractor-trailer, the engineering required to develop these two solutions has very different specifications and each requires time and energy. Similarly, while you could view predictive maintenance of a complex engineered system such as a power plant and preventive health care for a congestive heart failure patient (predictive maintenance of the “human system”) as being conceptually the same problem, actually creating the software and models to prevent a failure in these two systems require quite distinct and considerable engineering efforts.
At times, a paradigm shift is needed in how a new technology should be applied. When steam engines in factories began replacing the water wheel, everything was driven from a central mechanical drive. The first attempts at using electricity tried to duplicate this, but only really made a difference when people realized it would be more efficient to distribute electrical energy through the building, and have separate electric motors in individual machines, rather than one central motor that distributes mechanical energy. 63
Today, technological advances in automation abound in areas from physical hardware and robotics to artificial intelligence and software, as we discussed in Chapter 1. Yet much of the innovation that we are seeing, from self-driving cars to digital personal assistants such as Apple’s Siri, Amazon’s Alexa, and Google Assistant, are still in development—and often imperfect. That means there remains a lot of work to be done by scientists and engineers. 64
Cost of developing and deploying solutions
The cost of automation technologies will affect when and where they are deployed. The costs of developing and deploying solutions have to be recouped. Even when the technology is purchased from a supplier, the supplier will amortize the development costs into the pricing.
Developing and engineering automation technologies takes capital. Some technologies require substantial physical infrastructure such as tooling and laboratories. But even “virtual” solutions that are based on software require real investments in engineers to create solutions. A decade ago, the largest corporate research and development spending was to be found in automotive and pharmaceutical companies; today, technology companies dominate, with companies including Amazon, Alphabet, Intel, and Microsoft spending more than $10 billion apiece on annual R&D. 65 Hiring talent can also be costly. Google, for example, acquired DeepMind Technologies in 2014, at an estimated price of $500 million. With approximately 75 DeepMind employees at the time of the deal, the price tag was nearly $7 million per employee. This is in line with other estimates by experts, who say that “aquihires” of cutting-edge AI startups cost around $5 million to $10 million. 66
Deploying automation technologies also incurs costs. For physical technologies, these are real capital expenditures. Industrial robots cost from tens of thousands to millions of dollars. Replacing an ordinary heavy truck with a self-driving truck requires an expenditure of capital. And the self-driving truck will likely be more expensive than the truck it replaces, at least when the technology is first released.
These deployment costs are lower for software-based solutions, especially when delivered remotely through the cloud, and where the software is sold as a service, thereby turning capital expenditure into operating expenditure. But “the cloud” has a real physical instantiation as data centers and networks, and they in turn represent costs. Moreover, deployment of a software (or hardware) solution can almost never be done without significant implementation services, such as the costs of customizing the software for an individual organization, changing the processes within an organization, and training staff. For enterprise software, the associated implementation services often cost several times the costs of the software itself. All these costs affect the business cases for where and when automation is adopted.
62. Ronald E.G. Davis, “The birth of commercial aviation in the United States,” Revue belge de philologie et d’histoire, volume 78, number 3, 2000.
63. Nicholas Carr, The big switch: Rewiring the world from Edison to Google, W. W. Norton, 2008.
64. For recent research on the modeling of enterprise 2.0 adoption, see Jacques Bughin, “Taking the measure of the networked enterprise,” McKinsey Quarterly, October 2015, and Martin Harrysson, Detlef Schoder, and Asin Tavakoli, “The evolution of social technologies,” McKinsey Quarterly, June 2016.
65. Justin Fox, “The big spenders on R&D,” Bloomberg News, April 29, 2016.
66. The age of analytics: Competing in a data-driven world, McKinsey Global Institute, December 2016.
Labor market dynamics
The labor costs associated with work activities that could potentially be automated are another factor that will affect the pace and extent of automation. These costs are affected by the complex dynamics of labor markets.
Even for the same activity, for example, entering data into a financial system, there is a wide range of wages paid, across different positions—from accounting clerks to chief financial officers—and in different companies, for example a small, family-owned business compared with a Global 50 corporation. Wage rates also vary by geography.
Labor supply is determined by demographics, with the share of the working-age population potentially declining in many countries in coming decades. But it also varies in terms of skills, which are affected both by intrinsic talents of individual human beings, as well as by education and training that people receive. People can learn new skills but it takes time and money (see Box 3, “Technological change and skills”).
Labor markets are dynamic systems. Supply, demand, and wages all vary over time. If automation frees up human capital, then supply will increase, which could be redeployed into other positions if the demand exists. But there could be a skills mismatch, which will require time and training, delaying redeployment. There could also be information asymmetries, which digital labor markets could help address. 67
Box 3. Technological change and skills
Economists who study the impact of technological innovation on the workforce have noted varying effects on workers at different skill levels, depending on the time period. 1 In the 19th century, for example, technological changes raised the productivity of lower-skill workers and created new opportunities for them, at times replacing the craftsmanship of higher-skill artisans. This so-called unskill-biased technical change reduced the value of some high-skill workers even as it boosted lowerskill ones.
With the advent of information technology and the internet, the reverse has happened: the productivity of higher-skill workers, especially those engaged in abstract thinking, or with creative and problem solving skills, has increased, while the relative demand for lowerskill workers has not. This phenomenon of skill-biased technical change is manifested in a number of ways. In advanced economies, median income households have been receiving a lower share of the total wage share of GDP, in part because demand for less-skilled workers has dropped, even as demand for high-skill labor has risen. 2 In 1981, college-educated workers in the United States earned a 48 percent wage premium over high school graduates. By 2005, that premium had risen to 97 percent—in other words, an American college graduate earns almost twice as much as a high school graduate. 3
What will the spread of automation mean for workers at different skill levels? Erik Brynjolfsson and Andrew McAfee discuss a new shift to “talent-biased technical change” which has created very high demand for superstars such as experts in artificial intelligence or data scientists. 4 Our analysis suggests that all workers at all skill levels have the potential to be affected at least partially by automation based on currently demonstrated technologies.
1. See David H. Autor, Frank Levy, and Richard J. Murnane, “The skill content of recent technological change: An empirical explanation,” Quarterly Journal of Economics, November 2003, and Daron Acemoglu and David H. Autor, “Skills, tasks, and technologies: Implications for employment and earnings,” in Handbook of Labor Economics, volume 4B, David Card and Orley Ashenfelter, eds., Elsevier, 2011.
2. See, for example, Global wage report 2012/13, ILO, December 2012; OECD economic outlook 2012 volume 1, OECD, June 2012; Andreas Hornstein, Per Krusell, and Giovanni L. Violante, The effects of technical change on labor market inequalities, Center for Economic Policy Studies working paper number 113, July 2005.
3. David Autor, “Skills, education, and the rise of earnings inequality among the ‘other 99 percent,’” Science, volume 344, issue 6186, May 2014.
4. Erik Brynjolfsson and Andrew McAfee, The second machine age: Work, progress, and prosperity in a time of brilliant technologies, W. W. Norton & Company, 2014.
Demand is also not static. New types of work activities and jobs are created all the time, even as new technologies come on the market. Bank tellers and ATMs are one example. ATMs were first installed in the United States and other developed economies in the 1970s, and from the mid-1990s, banks rapidly increased their use of ATMs. Contrary to some expectations, their advent actually increased the demand for human tellers, for two reasons. First, ATMs reduced the cost of operating a bank branch, and banks responded by opening more branches. Fewer tellers were needed in each branch, but more branches meant that teller jobs did not disappear. Second, the tasks that ATMs could not do—in particular, developing relationships with clients such as small businesses—became more valuable. For tellers, the nature of their activities changed, with cash handling becoming less important and human interaction more important. 68
The relative costs of labor vs. automation will affect the pace and extent of adoption. If workers are in abundant supply and significantly less expensive than automation, this could be a decisive argument against it. For example, food service is one activity with a high automation potential based on adapting currently available technologies. However, current wage rates for this activity are among the lowest in the United States, reflecting both the skills required and the size of the available labor supply. Since restaurant employees who cook earn an average of about $10 per hour, a business case based solely on reducing labor costs may be unconvincing.
The potential economic benefits from automation are not limited to labor cost reductions. As we noted in the case studies in Chapter 3, performance gains include increased profit, increased throughput and productivity, improved safety, and higher quality, which are harder to quantify than labor costs but no less tangible. In our hypothetical look at automation of an oil and gas control room, for example, performance gains accounted for more than three-quarters of the total benefits. There are also indirect benefits, such as wage growth, and automation’s potential to create business and economic incentives that drive corporate decision making, and unleash entrepreneurial energy. Automation could also spur policy makers. However, these indirect benefits will have to be netted out against indirect costs caused by automation, such as those associated with labor displacement.
Regulatory and social acceptance
Automation faces some significant regulatory and social barriers to implementation. They include safety and liability issues. One accident could trigger stringent regulations. Artificial intelligence used in military robots or autonomous vehicles may have to make judgments that harm people, creating moral controversy. Technology makers could also be exposed to legal product liability if robots malfunction.
Privacy is also a potential barrier, especially in areas where personal data is highly sensitive, such as health care; already, health care IT can struggle to link together different data sets due to mandatory anonymization of data.
From a social perspective, too, automation will need to overcome some barriers. If many workers lose jobs and are unable to find new ones, the social and political pressures against automation could become significant. Humans may not want to adopt new technologies or work with automated products due to fears about job security.
Finally, personal preferences and discomfort with technologies could prevent automation in settings where human relationships are important, such as for caregivers.
68. James Bessen, “Toil and technology,” Finance & Development, volume 52, number 1, March 2015.
MODELING SCENARIOS FOR POTENTIAL PACE AND EXTENT OF ADOPTION
To analyze a range of potential scenarios for the pace at which automation could affect activities across the global economy, we constructed a model that synthesizes the effects of these five factors into four timing stages. We estimate when automation technologies will reach each level of performance across 18 capabilities, the time required to integrate these capabilities into solutions tailored for specific activities, when economic feasibility makes automation attractive, and the time required for adoption and deployment (Exhibit 14). We modeled scenarios incorporating these stages for each individual activity in every occupation for all sectors across 46 countries that account for about 80 percent of the world’s workforce.
Five factors affect the pace and extent of automation; we model using four stages.
NOTE: Economic benefits affect both when adoption will begin and its pace. For determining economic feasibility, we assume that decision-makers discount the uncertain benefits of initial labor cost savings by roughly the same amount as they believe the also uncertain non-labor cost-related benefits will be captured.
SOURCE: McKinsey Global Institute analysis
The scenarios we have modeled create a time range for the potential pace of automating current work activities. We have created two theoretical bookends, an “earliest” scenario, in which all of the modeling parameters are flexed to the extremes of the set of plausible assumptions that would result in faster automation development and adoption, and a “latest” scenario, in which we flex all of the parameters in the opposite direction. The reality will likely fall somewhere between the two. Modeling all of these factors, the date at which 50 percent of the world’s current work activities are automated could be around 2055, but we posit possible scenarios where that level of adoption occurs up to almost 20 years earlier or later. (Exhibit 15).
Automation will be a global force, but adoption will take decades and there is significant uncertainty on timing.
1. Forty-six countries used in this calculation, representing about 80% of global labor force.
SOURCE: McKinsey Global Institute analysis
We stress that we are not making specific point predictions; we use scenarios to describe the envelope of potential outcomes, and we fully acknowledge the simplification that comes with any modeling exercise. For example, we do not account for all of the dynamics of the labor market described above, including whether wages for specific occupations would decline because displaced workers would increase labor supply.
The model is likely to be precisely wrong but we hope it is directionally right with regard to overall findings. Occupations within sectors with high automation potential today, jobs that involve types of activities that we categorize as easiest to automate, such as physical activities in a predictable environment or data collection and processing, will most likely be among the first to feel the impact of automation. From a geographical perspective, advanced economies are likely to deploy automation ahead of many emerging economies, because of higher wage levels, which make a stronger business case for deployment, as well as the nature of the solutions that will be needed to integrate the technologies into the workplace.
Furthermore, while the macro effects on a nation’s entire economy may be quite slow (we model the adoption of automation across several decades), the micro impact on particular work activities, in certain sectors, and in individual countries, could be quite fast, as competitive pressures come together with advances in technological development. That is, for an individual worker who is displaced by automation, or a company whose industry is disrupted by competitors using automation to shift the basis of competition, these effects could occur quickly indeed.
DEVELOPING AUTOMATION CAPABILITIES IS A KEY INITIAL DETERMINANT OF ADOPTION TIMELINES
The deployment of automation in the workplace can begin only when machines have the capabilities required to carry out particular work activities. Technological innovation must first deliver the technical capability before a workplace solution can be developed and deployed.
In Chapter 2, we outlined 18 performance capabilities required to carry out the range of work activities and the current state of the technology for those capabilities as measured against human performance. Along with our assessment of the current state, we developed progression scenarios for each of these capabilities. We did so through surveys of academic and industry experts and through an extrapolation of metrics including recent commercial successes and the historical trajectory of the capabilities, along with a range of other predictors. Details of our methodology can be found in the technical appendix.
While machines can already match median human performance or even exceed the top levels of human performance in some of the 18 capabilities, such as, for example, information retrieval, gross motor skills, and optimization and planning, many other capabilities require more technological development, for example to raise natural language understanding and logical reasoning to a median human level.
Scenarios for achieving higher levels of performance capabilities
Exhibit 16 shows a potential range of time frames for technology to attain the next level of performance for each capability. From a technical standpoint, some of these performance capabilities are quite far advanced, and developing top-quartile human performance may be relatively fast, although there are some significant variations among the different capabilities.
- Sensory perception is already at median human performance. A robot “tongue” can already detect the color index and alcohol content of beer with more than 80 percent accuracy, for example. 69 For tactile perception, which can already exceed top-quartile human performance for several dimensions, key challenges include miniaturizing the size of hardware and adapting the sensors to function in different environments. 70
- Cognitive capabilities are considerably more varied in their performance compared to humans. Information retrieval, optimization and planning, and recognizing known patterns and categories have already reached the top quartile of human performance, whereas generating novel patterns, creativity, coordination with multiple agents, and logical reasoning and problem solving are still at a relatively early stage. In terms of coordinating with multiple agents, robots already have demonstrated the ability to coordinate with similar types of robots. 71 Their ability to collaborate with humans is still at an early stage. Automated creativity is perhaps the furthest away; computer creativity for now requires human involvement to judge the quality of work or to provide direction.
- Social and emotional capabilities for now are below median human performance levels. Despite advances in artificial intelligence, machines still have difficulty identifying social and emotional states (sensing), drawing accurate conclusions about them (reasoning), and responding with emotionally appropriate words or movements to them (output).
- Physical capabilities are already at top-quartile human performance for gross motor skills and navigation, which has enabled wide deployment of robots in industrial automation, military, and defense—and given consumers turn-by-turn navigation apps for their smartphones. For fine motor skills, we estimate that top-quartile performance could be achieved when robotic hands have the same degrees of freedom as human hands. Mobility remains a challenge, especially vertical mobility such as climbing stairs and ladders.
69. Carlos A. Blanco et al., “Beer discrimination using a portable electronic tongue based on screen printed electrodes,” Journal of Food Engineering, volume 157, July 2015.
70. Giorgio Cannata and Marco Maggiali, “Design of a tactile sensor for robot hands,” in Sensors: Focus on tactile, force and stress sensors, Jose Gerardo Rocha and Senentxu Lanceros-Mendez, eds., InTech, 2008.
71. Julia Sklar, “Making robots talk to each other,” MIT Technology Review, August 2015.
Ranges of estimated time frames to reach the next level of performance for 18 human-related performance capabilities.
SOURCE: McKinsey Global Institute analysis
Natural language capabilities are critical for automation across many activities
Natural language understanding is of critical importance for a wide range of applications in the workplace. Despite some technical advances in this capability, including the accuracy of machine translation, machines still have a long way to go to achieve median human performance. The speed of technological advances in natural language capabilities in the future will substantially affect the timelines for automation overall.
The development of higher levels of natural language understanding is the single most important factor constraining the technical automation potential of activities for which the current performance level of technologies is insufficient. In the earliest scenario for automation, the current level of performance in natural language constrains the proportion of all activities that can be automated by 18 percent. In the latest scenario, the current level of natural language understanding would hinder automation of 40 percent of activities, since other capabilities will have reached their required performance levels by then.
Natural language processing capabilities have advanced in recent years and will likely continue to develop and improve, led by their use in auto (in-vehicle speech recognition), health care (clinical documentation improvements), and general personal use (virtual assistants in both mobile and fixed devices). Technology companies and venture capitalists continue to invest heavily in technologies related to natural language processing. Algorithms can write passages and new articles that are largely indistinguishable from those written by humans, and many experts expect continued rapid technical advances; by 2018, according to one prediction, machines could write as much as 20 percent of all business content. 72
HISTORICAL TIME FRAMES FOR DEVELOPING SOLUTIONS THAT INTEGRATE MULTIPLE CAPABILITIES
Many of the 18 performance capabilities in our framework operate together. To arrange tables or dining areas as well as a human waiter, for example, machines will need some capacity for natural language understanding and generation, an ability to recognize known patterns, retrieve information, coordinate with multiple agents, and navigate. Furthermore, machines will need optimization and planning capabilities, fine motor skills, gross motor skills, and mobility. Developing solutions for this type of multi-capability activity requires time and technology.
To develop scenarios for the times required for solution development, we looked at historical precedents, examining almost 100 existing automation solutions that use hardware or software or both. We collected details of the development time and the technical capabilities that were integrated, recording the number of years from the initial research to the product launch, and identifying as many as three of the most relevant capabilities out of our 18 that were used. For example, Viv, an artificial intelligence technology that aspires to be the “intelligent interface for everything,” combines logical problem solving with natural language understanding and generation, and took four years to integrate. 73 PillPick, a pharmacy automation system that helps hospitals eliminate the risk for medication errors during packaging and dispensing, combines fine motor skills with sensory perception, and took two years to develop. 74 For each activity for which a solution needs to be developed, we used the 25th and 75th percentile time frames for the capabilities whose historical development times at those points were longest. For more details of our methodology, see the technical appendix.
72. Bernard Marr, “Why management dashboards and analytics will never be the same again,” Forbes, January 21, 2016; Gartner predicts our digital future, Gartner, October 6, 2015.
73. Lucas Matney, “Siri-creator shows off first public demo of Viv, ‘the intelligent interface for everything,’” TechCrunch, May 9, 2016.
74. Anthony Vecchione, “New automated packaging system ensures patient safety,” Drug Topics, November 22, 2004.
Development timelines for solutions associated with each capability.
SOURCE: McKinsey Global Institute analysis
MODELING THE ECONOMIC FEASIBILITY OF AUTOMATION
Our model assumes that the adoption of automation will begin after it becomes economically feasible that is, when the economic benefits exceed the costs. These benefits include not only labor cost reductions, but also other performance improvements such as higher throughput or improved quality, as discussed earlier in this chapter. However, particularly early in the adoption cycle, the certainty that these benefits can be captured will be lower until they are demonstrated at scale. Thus, for the purposes of modeling, we assume that adoption begins when the cost of automating a particular activity (for a specific occupation in a country) is at parity for the cost of human labor for the same activity. To be fair, the certainty that labor cost savings can be completely captured will also be somewhat lower early in the adoption cycle. On a net basis, our model assumes that decision-makers discount the benefits of initial labor cost savings by roughly the same amount as they believe the also uncertain non-labor cost-related benefits will be captured. In our experience, business leaders tend to view labor cost savings as being more predictable than other benefits.
Moreover, as we will discuss in the section on the international impact of automation, the overall lower level of wages in some emerging economies could mean that automation proceeds more slowly there than in advanced economies, where wages are higher than the global average. The cost of labor is not necessarily fixed, however. The related labor supply and demand dynamics are also a factor in the timeline for automation, as discussed above, but we do not model these dynamics.
ADOPTION OF AUTOMATION TECHNOLOGIES AFTER COMMERCIAL AVAILABILITY CAN TAKE YEARS
Even when a solution to automate an activity is commercially available, and implementing the solution makes rational economic sense, adoption still takes time—and can be very costly. Full adoption across an entire sector of any technology, particularly those that are integrated into the workplace, takes years. Individual decision makers must become aware of the potential, and there is always a spectrum of willingness to adopt, a phenomenon well documented in the research studying the diffusion of innovations. 75 Capital has to be deployed, technology acquired and installed, and processes transformed, within or across enterprises. Often, regulations need to change, and individual workers and employees have to become accustomed to the new technologies and processes.
From a review of the historical rate of adoption of previous technologies, the time from commercial availability to 90 percent adoption ranges from approximately eight to 28 years (for 50 percent adoption, the range is between about five and 16 years). This lag applies not only to hardware-based technologies that are capital-intensive and require physical installation; even technologies that are made available purely online take years to be adopted. A fast-growing consumer service such as Facebook began in 2004, and as of this writing, has not yet reached full adoption, even of non-China internet users (see Box 4, “Adoption of hardware vs. software/cloud-based technologies”). Furthermore, that decade of adoption does not even take into account previous instances of social networking services. Automation technologies that would be incorporated into the workplace require even more user and process changes than a consumer service that individuals can adopt independently.
To create scenarios for the rate of automation adoption, we analyzed adoption rates of 25 previous technologies to establish a range of timelines. We incorporated these historical examples into classic S shaped adoption curves. (Exhibit 18).76 We analyzed both hardware and software/online technologies, and we divided them into groups representing technologies with the fastest and the slowest adoption. Technologies with the fastest adoption rates include stents, airbags, MRIs, TVs, and online air booking. The slower adoption category include cellphones, personal computers, dishwashers, and pacemakers.
75. Everett M. Rogers, Diffusion of innovations, The Free Press, 1962.
76. S-shaped adoption curves were studied for consumer technologies by Frank Bass, “Comments on ‘A new product growth model for consumer durables,’” Management Science, volume 50, number 12, December 2004.
Box 4. Adoption of hardware vs. software/cloud-based technologies
Hardware-based automation technologies, such as robots and self-driving vehicles, have requirements that can lengthen the time of adoption, that is, substantial capital requirements and the need to physically produce and deploy these technologies.
These requirements are much lower for software-based technologies, particularly those that are deployed through cloud technologies, that is, where the bulk of the computing occurs at centralized data centers that are accessed through networks. For these technologies, customers often pay for these services on an on-demand basis, reducing the need for the customer to deploy capital expenditure or manage for peaks in capacity (the cloud provider takes on these tasks). In general, the marginal cost of producing one more instance of a piece of software tends toward zero. And there are several cases of specific pieces of consumer applications, particularly those that have “gone viral,” whose adoption seems to have been extraordinarily fast, such as Pokemon Go.
Does this mean that we should expect the adoption of software/cloud-basedautomation technologies to be much faster than for other technologies? In examining the historical record, we find that the adoption of software based technologies falls within the envelope of the adoption rates of other technologies, that is, ranging from eight to 28 years to reach close to full adoption. In particular, for technologies that will be implemented in the workplace, the technical deployment of the technology represents only a fraction of the time and cost necessary to embed the technology into the processes and practices of an enterprise. We examined the adoption of cloud-based versions of enterprise resource planning (ERP), supply chain management (SCM), and customer relationship management (CRM). While these technologies have not yet reached a plateau in adoption, their adoption curves all fall within the range of adoption of other technologies. For instance, cloud CRM was first offered by Salesforce in 1999, and adoption of that technology still continues to grow.
Even for consumer online technologies, for which the bureaucracy of an enterprise is not a barrier to adoption, the behavior changes necessary for adoption take time to proliferate, particularly when these technologies are appropriately viewed by category, rather than a specific service (for example, Pokemon Go was not the first mobile game, and has not yet come close to full adoption). Take consumer social networking as another example. Facebook, only one example of a social networking platform, and not the first, was launched in 2004, and over a decade later, adoption continues to grow, both of that platform and other social networking platforms. Peer-to-peer (P2P) mobile payments is another consumer technology whose adoption rate falls within the envelope of other technologies.
Historic adoption curves for technological innovations.
1. Technologies considered include airbags, antilock braking systems, cellphones, cloud CRM, cloud ERP, cloud SCM, color TVs, copper production through leaching, dishwashers, electronic stability control, embolic coils, Facebook, instrument landing systems, laparoscopic surgery, Lithium-ion cell batteries, microwaves, MRI, online air booking, P2P remote mobile payment, pacemakers, PCs, smartphones, stents, TVs, and VCRs.
SOURCE: McKinsey Global Institute analysis
The adoption of technologies within enterprises include factors beyond those that underpin consumer adoption. For example, the technology adoption literature discusses rank effects (that is, how the different individual characteristics of firms, such as their size, can affect the rate and extent to which they adopt new technologies) and the effects of competitive dynamics (that is, how the adoption of new technology by one company in an industry could influence the adoption of technology by other companies in that industry). 77 We do not model firm-level adoption dynamics; our model takes advantage of the fact that at a high level for economies and industries, the net result of the enterprise adoption factors are S-shaped curves that resemble those for consumer adoption.
This combination of technical feasibility, solution development, economic feasibility, and adoption enables us to model a set of scenarios that encompass various time frames for the pace of automation. To illustrate, we detail a specific example, driving heavy trucks (see Box 5, “Driving heavy trucks: Modeling scenarios for adoption of automation”).
Adoption will ultimately depend on several factors include overcoming possible policy barriers and resistance to automation, which are meant to be captured in scenarios in the adoption stage of the model. For instance, given the importance of technological advances, intellectual property or regulatory issues could delay certain companies from being able to deploy specific technologies.
77. See Massoud Karshenas and Paul Stoneman, “Rank, stock, order, and epidemic effects in the diffusion of new process technologies: An empirical model,” RAND Journal of Economics, volume 24, issue 4, 1993, and Marc Anthony Fusaro, “The rank, stock, order and epidemic effects of technology adoption: An empirical study of bounce protection programs,” The Journal of Technology Transfer, volume 34, issue 1, February 2009.
Box 5. Driving heavy trucks: Modeling scenarios for adoption of automation
To illustrate how we model scenarios for the pace and extent of automation, let us consider the single but widely performed work activity of driving, for one common occupation: heavy and tractor-trailer truck drivers. 1 This occupation is a significant source of employment, with more than 20 million workers worldwide, of whom 3.2 million are in India, 2.9 million are in China, 1.6 million are in the United States, and one million are in Japan. 2
From a purely technical standpoint, the technologies to automate the required capabilities for this activity already exist. Some of the 18 individual performance capabilities we use in our framework are not required for this activity, including the three social and emotional capabilities, but for other capabilities, existing technology has demonstrated the necessary constituent performance levels to potentially enable driving. However, in order to automate the activity of driving (“Level 4” autonomy as classified by the Society of Automotive Engineers), these individual capabilities must be integrated into a solution, in other words the hardware and software that would constitute an autonomous driving truck must still be engineered. 3 Based on the historical time frames required to create solutions that involve the capabilities necessary to automate this activity, we estimate the time required to engineer such a solution could take more than seven years, which we use as basis for our latest adoption scenario. Conversely, because existing automation technology has met the levels of performance required across all of the necessary capabilities, and we do not know how far along pre-existing engineering efforts might be, our model also accepts the possibility that an organization could announce a Level 4 autonomous truck immediately, which we use as the basis for our earliest adoption scenario.
Then comes the question of economic feasibility. In our model, adoption begins when the cost of automation reaches parity with the cost of human labor, for an individual occupation within a specific market (which we model at the level of countries). Of course, there are other economic benefits which could contribute to the business case for automation. In the case of driving heavy trucks, for example, these could include higher fuel efficiency, improved safety, increased asset utilization, and so on. However, particularly early in the adoption cycle, the certainty that these benefits can be captured will be lower until they are demonstrated at scale. To be fair, the certainty that labor cost savings can be completely captured will also be somewhat lower early in the adoption cycle. On a net basis, our model essentially assumes that decision makers discount the benefits of initial labor cost savings, because of initial uncertainty, by roughly the same amount as they believe the also uncertain non-labor cost-related benefits will be captured.
Using these assumptions, the time at which we model adoption beginning is sensitive to the cost of labor in different markets. For our example, we compare the United States with China (Exhibit 19). Total wages paid to heavy and tractor-trailer truck drivers exceed $328 billion worldwide, of which almost $60 billion is paid in the United States, and more than $33 billion is paid in China. Because of the significantly higher level of wages in the United States, automation is modeled to become economically feasible faster there. Under our scenarios, the cost of the automation could be below US wage levels within three to ten years after Level 4 autonomy is available, whereas in China it could take 10 to 16 years.
Finally, once automation becomes economically feasible, adoption can still be a long process, even when the business case is compelling. In this case, there are approximately two million tractor-trailers in the United States, each of which is typically on the road for about 20 years, and cost about $160,000 apiece. Replacing the current fleet would cost $320 billion, not including the additional technology for autonomous driving (whose cost we do model as declining over time), and would be certain to take many years.
1. The detailed work activity title is “operating vehicles or material-moving equipment” in the US Bureau of Labor Statistics taxonomy.
2. US Bureau of Labor Statistics; Oxford Economic Forecasts.
3. Taxonomy and definitions for terms related to driving automation systems for on-road motor vehicles, Society of Automotive Engineers, September 30, 2016.
Modeled adoption of operating vehicles or material-moving equipment, United States and China 1
1. The detailed work activity (DWA) is 4.A.3.a.4.I01.D06. Occupations using this activity include: Heavy and Tractor-Trailer Truck Drivers, Industrial Truck and Tractor Operators, etc.
SOURCE: US Bureau of Labor Statistics; McKinsey Global Institute analysis
MODELING THE PACE AND EXTENT OF AUTOMATION ACROSS ACTIVITIES AND SECTORS
Each of the four stages affects the overall pace of automation reflected in the model. Technical feasibility accounts for much of the variance in our modeled scenarios, but economic feasibility is also a significant factor, especially in the earliest scenario, where our modeling suggests it could hold up adoption and deployment for eight to nine years for some activities (Exhibit 20).
Several of the 18 performance capabilities are potential bottlenecks in the model. The types of activities that today have among the lowest potential for automation, such as managing and developing people, applying expertise and experience, and interfacing with stakeholders, will all require substantial further advances in social and emotional capabilities, as well as in natural language understanding and creativity. These activities typically require at least a median-level human performance in social and emotional capabilities, for example. As we have seen, these capabilities are particularly difficult to develop, and most related technologies are in nascent stages and may take several decades to mature. This potentially could delay the development, integration, and deployment of automation as a whole.
Along with timelines for capability development, whether hardware or just software is required is also a key determinant of scenarios in the model, because this will affect the economic viability of automation. Hardware requires significant capital spending, and thus we model with relatively higher initial costs. For example, sensory perception capabilities need cameras and sensors. Mobility requires wheels or other hardware that enable machines to move. Software solutions, by comparison, are relatively less expensive to deploy and thus we model with relatively lower costs compared to hardware solutions. Some 98 percent of the solutions for predictable physical activities are hardware ones, whereas hardware represents only about 30 percent of processing data solutions.
This hardware-software distinction helps explain differences in the pace of automation adoption across different types of activities. For example, under our earliest scenario, the model suggests that unpredictable physical labor activities—which require substantial hardware for successful automation will have much slower adoption rates than for activities that consist of processing data, which primarily require software solutions.
As we will detail in our discussion of country differences below, the modeled time frame for physical activities is longer in emerging economies largely because of lower wages compared to the cost of hardware-based automation solutions.
Differences in the adoption timing are primarily driven by technical feasibility, and in the earliest scenario, also by economic feasibility.
1. Percentage of time automatable by adapting currently demonstrated technology. NOTE: Numbers may not sum due to rounding.
SOURCE: US Bureau of Labor Statistics; McKinsey Global Institute analysis
GLOBAL TIMELINES: MODELING ADOPTION OF AUTOMATION BY COUNTRY
Automation will be a global force that affects all countries, whether they are emerging economies or advanced ones. In the earliest scenario we modeled, automation could account for more than 50 percent of working hours in two-thirds of countries within just 20 years, by 2036. In the latest scenario we modeled, more than half of all countries will have 50 percent automation or more within 50 years, by 2066 (Exhibit 21).
In the early years of automation adoption, our model shows the impact being felt most strongly in a few advanced economies, especially Germany, Japan, and the United States. These countries have both high wages and major industries that already have a high potential for automation based on existing technologies. As automation is adopted in more countries around the world, the impact will become especially pronounced in China and India, because of their large workforce. Initially, our model shows automation affecting workers there in manufacturing and retailing because of their high automation potential, but in the longer term its largest impact will be in agriculture, which is where hundreds of millions of Chinese and Indians earn their livelihood.
Automation impact will be global under multiple modeled scenarios.
SOURCE: US Bureau of Labor Statistics, 2014 O*Net database; McKinsey Global Institute analysis
Adoption of automation could be faster initially in advanced economies than emerging ones because of wage levels and integration solution costs. Higher wages in advanced economies and hardware costs will likely make automation economically viable faster there than in emerging economies. In Japan, the United States, and Europe’s five largest economies—France, Germany, Italy, Spain, and the United Kingdom—this is likely to mean earlier adoption in a range of sectors, especially manufacturing and services. In the United States, for example, our model shows manufacturing, retail, health care, transportation, accommodation and food services, and administrative services to be among the first sectors affected. In terms of the impact on workers, sectors that are large employers, such as health care, will be affected even though their expected automation adoption rate (48 percent) is lower than some other high-adoption sectors such as accommodation and food services (83 percent) in 2036 under earliest scenario.
Initially, technologies with physical capabilities are the most likely to become available, and will be increasingly adopted in sectors such as manufacturing and retail. As adoption in these sectors approaches 100 percent, their contribution to overall adoption rates will hit a plateau. Once the technology to replicate cognitive and natural language understanding capabilities has been mastered, services sectors globally will be affected, and the pace and extent of overall automation will pick up, especially in advanced economies with large service sectors.
For all advanced economies, the business case for automating will become stronger as solutions for technology integration in the workplace become cheaper. Exhibit 22 shows the estimated wages corresponding to automation adoption by country.
In China, India, and other emerging economies, cost and relative lower wage levels will likely delay adoption. Manufacturing relies heavily on predictable physical activities, and automating them will require hardware solutions that require considerable up-front capital investment. This may not be economical in emerging economies, given the lower cost of labor there, until the cost of the solutions drops sharply.
Emerging economies could achieve a pace of automation similar to that of advanced economies if solutions become cheaper, possibly through localized innovation. Adoption could also be accelerated as a result of policy measures, increased competition, a lack of legacy systems that could be a brake on automation implementation, and a high degree of technological literacy.
In sectors where software solutions to integrate automation technologies will be required, the pace of automation across emerging and advanced economies could be similar. For example, we model similar rates of automation adoption in the finance and insurance sectors—which are characterized by a high proportion of data processing and collection—in both the United States and China. Software, which has a relatively minimal marginal cost, accounts for just over half the technology integration solutions needed in this sector, and the global disparity in wages is not as pronounced in some areas. For example, architects or investment bankers in emerging economies are relatively well paid. Moreover, wage distribution globally is fairly similar for real estate, rental and leasing activities, health care, and social assistance.
Wages associated with automated activities are modeled to be highly concentrated in countries with higher wages and populations.
1. Forty-six countries used in this calculation, representing 78% of global labor force; calculations based on current FTEs in each sector, not considering FTE growth and sector migration. NOTE: Not to scale. Numbers may not sum due to rounding.
SOURCE: US Bureau of Labor Statistics; McKinsey Global Institute analysis
The largest impact on workers could be felt in labor-intensive sectors in China and India, whereas in other countries it will be felt across multiple sectors
In analyzing the potential impact of automation on workers, we make a number of assumptions. We calculate the impact based on the current number of full-time equivalents in each sector and do not take into account any possible growth in their numbers per sector or the movement of labor from one sector to another. We also assume that the technologies we identify will emerge throughout the world and that they will be adopted only when they become economically feasible.
China and India both have very large farming sectors, with about 230 million people out of a working population of 450 million in India alone working in agriculture. In China, more than 200 million work on the land out of a total workforce exceeding 770 million. Given the employment size of this sector, even a relatively low rate of automation adoption of about 10 percent could have significant employment consequences in both countries.
In both China and India, the impact of automation on employment could also be felt in the retail and manufacturing sectors, as both have a relatively high potential for automation and a sizable labor force. In the five major European countries, Japan, and the United States, the employment impact will likely be spread across multiple sectors, especially in the event that large-scale automation begins relatively soon. A detailed view of our model of automation impact on individual countries is available via an interactive graphic online. 78
Automation technologies will be increasingly adopted in every industry, every sector, and every country in the world, but there is considerable uncertainty about the speed and intensity with which they will arrive. These will depend on several key variables, both technical and economic. Machines will need to be able to simulate the full range of human performance capabilities, and solutions to integrate the technology into the workplace will need to be adopted. Only when costs have fallen below wage levels will automation become economically viable, and ultimately adopted. All of this could happen within two decades for a wide range of activities and sectors, but it could also take much longer. Whatever the time frame, the consequences will be significant not just for individual workers and sectors but also for the global economy as a whole. In the next chapter we look at how automation could upend some cherished notions about productivity, growth, and development.
78. The data visualization can be viewed online at http://public.tableau.com/profile/mckinsey.analytics#!/