The Myth Of Full Automation

Just as there is no perfect person, there is no perfect algorithm. JL

Claas Ludwig reports in Bank Underground:

The lack of perfectly automated systems reflects it being difficult to capture all possible eventualities by algorithmic automation. The need for human labor (also) reflects the inability to create perfectly automated systems in real-world environments, the need to keep innovating, directed against automation and economic constraints. Human labor is often cheaper, especially for tasks which are complex and/or small scale, a decision between a fixed cost investment versus the human worker, who constitutes variable cost. The number of staff required actually grew as the degree of automation increased.

Recent developments in digital technology fuel the notion that we are at an inflection point in human history, where fully automated robots are on their way to permanently replacing humans at work. To better understand the dynamics between automation and the demand for human labour, I undertook a case study on financial advice robots – colloquially known as roboadvisors. For the roboadvice firms examined, I found that human involvement is still crucial. Full automation is thus a myth, at least for now, in this industry. But roboadvisors do demonstrate that some cognitive ‘non-routine’ tasks can be automated. Previously, ‘non-routine’ tasks had been widely considered as non-automatable. Roboadvisors demonstrate how the frontier of potential automation is not limited to menial, routine tasks.

The debate on human replacement by machines

Technology can replace human work – and the prospect of being replaced causes fear in humans. This fear fuels an ongoing debate in society, which asks whether machines will outperform humans and make us permanently redundant. In this vein, one previous Bank Underground blog concluded that economists should seriously consider the possibility that millions of people may be at risk of unemployment. In another blog however, it was argued that in the long run robotisation doesn’t create unemployment.

Human replacement by technology is real and long established, having already revolutionised occupations from agriculture, manufacturing and clerking to service and management occupations. However, up to now there is no evidence of permanent mass unemployment from automation. Rather, the type of jobs people undertake have changed over time, across many economies and industries. But this does not mean that the permanent replacement of human workers by machines could not happen in the future. There is no natural or economic law that would rule out such a scenario.

A renewed wave of angst about human replacement came with the appearance of new technologies such as computers, the Internet and developments in artificial intelligence (AI). Some see this as posing a threat to social stability, citing drastic scenarios such as mass unemployment.

Understandably, the issue has attracted the attention of policymakers and scholars. Among the scholars, different schools of thought have emerged. On one hand, some believe that we are currently witnessing a singular event in human history, in which robots will permanently replace humans, creating no (or few) complementary jobs (Brynjolfsson and McAfee 2014; Ford 2015; Frey and Osborne 2017). I call this school the machinists. On the other hand, there are scholars who believe in the continuation of past trends and the current balance between technological progress and the creation of complementary jobs (Autor et al. 2003; Goldberg 2015; Mindell 2015; Shestakofsky 2017). I call these the humanists.

The roboadvice case

To investigate evidence for the machinists versus humanists debate, I conducted a case study on a recent innovation in retail finance, known colloquially as ‘roboadvisors’. Roboadvisors seek to transform the investment management industry, away from the traditional human-to-human advisory process into a digital, human-to-computer process. Similar to going online to get advice on which music you may like based on simple preferences, clients can go online and invest money through roboadvisors.

As such, roboadvisors are an example of automation that could replace jobs, and in this case, finance professionals.

How automated is roboadvice in reality?

First, I scrutinized whether the roboadvisor is really fully automated, without needing any human assistance or supervision. I examined this from two perspectives: for clients and from inside the firm.

To understand the clients’ perspective, I conducted self-tests. Thereby, I went through the onboarding process of different roboadvisors, simulating the willingness to invest money through their platforms. During the onboarding process, I had no human contact with the staff of the firms, no online chat nor phone contact. All that was needed was internet access, ability to navigate the simple user interface, and willingness to provide personal information to the system. From the client’s perspective, the process thus appears fully automated, except that it still needs input from the client, indicated as orange squares in Figure 1. Policies that support data sharing, like Open Banking, might mean that even less user input is necessary in the future.                             

Figure 1: Degree of automation of roboadvisors

To examine how automated the process appears from inside the firm, I conducted interviews with CEOs of roboadvice firms and also their technical and support staff. In total more than thirty experts were interviewed. Thereby, I found out that in many instances human involvement was still crucially needed, and that the number of staff required actually grew as the degree of automation increased. The need for human labour reflected several factors: the inability to create perfectly automated systems in real-world environments, the need to keep innovating, innovations which are directed against automation and economic constraints.

The lack of perfectly automated systems partly reflects it being difficult to capture all possible eventualities by algorithmic automation. For example, three CEOs of very advanced roboadvice firms told me that, even for the largely automated onboarding process of new clients, between 10% and 20% of these have to be processed manually. Also, between 5% and 10% of existing customers still get in touch with human support staff, even though most information is provided on-line. However, these numbers can vary widely, depending on the maturity of the firms. Plus, the real-world environment is not static, so automated systems had to be constantly adapted by human engineers through adjustment loops. Thereby, I found out that innovations do not have to be ‘big’ or ‘revolutionary’ to create labour demand. Even small changes and adjustments in the process created new needs for human staff.

Besides, innovations that happen in tech firms can be directed towards generating more human involvement, not only towards more automation. For example, some of the roboadvice firms develop their product in the direction of a hybrid solution, by reintroducing human advisors. Finally, economic constraints are a big factor too. It turns out that human labour is often just cheaper, especially for tasks which are complex and/or which have to be processed on a small scale. Often the management has to make a decision between a fixed cost investment (an engineer programming a program, testing etc.) versus leaving the task to the human worker, who constitutes variable cost. The limitation for automation is usually more pronounced in start-ups than at mature and well-funded companies.

To sum up, these findings support the humanist school of thought, which believes in the enduring relevance of human labour in highly automated systems, both from the client’s perspective and especially activities inside roboadvice firms.

Non-routine processes can be automated

The roboadvisor case study does, however, show that cognitive ‘non-routine’ tasks can be automated. This finding contradicts some longstanding views in the debate, which see cognitive ‘non-routine’ tasks as the last bastion of human superiority over machines (Autor et al. 2003). The analogue individualised financial advice from one human to another clearly is a cognitive ‘non-routine’ process. It involves several ‘non-routine’ characteristics explicitly named by Autor and colleagues (2003: 1322), such as flexibility (‘I read an article in the news, now you should change your investment strategy’) and complex communication (making jokes to keep the client happy and investing more). Besides it also incorporates human assigned actions like intuitive/emotional decisions (‘You should buy that stock because I believe in it’). Roboadvisors do not reproduce all such ‘non-routine’ abilities exactly in the way humans do. Instead, they have created a model that does not need many of these ‘non-routine’ human abilities to deliver financial advice.

The roboadvisors can do so because the digital world has created new opportunities for workarounds. The potential for workarounds is made possible not only by digitalization but also because of the intensifying role that metrics play in financial markets and the world around us (Beer 2016). This has been accompanied by a cultural change in which there is more trust placed in quantitative approaches and less in human intuition and emotion. In combination with technological progress and the Internet as a direct touchpoint to retail clients, this has paved the way for roboadvisors to introduce new automated advice models successfully. These models are now completely routine processes. So, the roboadvisory case has shown how formerly cognitive ‘non-routine’ tasks can, through technological and cultural change, become automated ‘routine’ tasks.

To sum up, these findings show the potential for the automation of a large set of tasks, which have been previously considered as non-automatable. Thus, these findings also offer support to the machinists.

Conclusion

My findings on the roboadvice business model provide support for both schools of thought, machinists and humanists. On the one hand, roboadvisors have succeeded in automating what were considered ‘non-routine’ cognitive tasks, which gives support to the machinists. On the other hand, my findings from the self-test and the expert interviews, combined with the sharp rise in the number of staff in roboadvice firms over the past few years, point in the other direction, suggesting that complementary job creation is taking place. This trend lends support to the humanists.

These opposing results give no indication on the overall net effect on human labour demand in the light of progressing automation. However, they make clear, that the ability to automate ‘non-routine’ cognitive tasks will, for a wide set of tasks, lead to significant further changes in the types of jobs humans will perform in the future.