OBEN
Past arrivals of general-purpose technologies (GPTs) offer insights into potential implications of the diffusion process of generative artificial intelligence (AI). Among these insights are three lessons for transformation: the benefits of labor augmentation, the velocity of the feedback cycle between ongoing technical improvement and innovation in application sectors, and the possibility of a J‑curve effect in productivity.
What do these three empirical observations imply for the insurer’s technology choices in the generative AI adoption process? How can business leaders steer the diffusion of this technology in the organization?
The arrival of generative AI offers opportunities for automation. Automation can augment labor by enhancing the skills of humans, and automation can substitute labor in performing individual tasks.1
Erik Brynjolfsson, a Stanford University scholar of technology diffusion processes, invites us to imagine that Ancient Greece had been able to automate all tasks in its economy by substituting labor for technology, thus forgoing any future labor augmentation. Although its citizens would keep benefiting from the additional leisure time created by the absence of labor, there would have been no further increase in the standard of living—to this day, this society would be without the motor car, the computer, and modern medicine.2
The history of automation demonstrates that although substitution and augmentation both occur, the latter offers far greater economic benefit. We can note, for example, how GPT adoptions from the dawn of the Industrial Revolution (e.g., the steam engine) to the modern era (e.g., the electric motor, the semiconductor) have significantly augmented human labor. One consequence is, as Brynjolfsson reminds us, that the value of an hour of human labor in units of consumer purchasing power is more than ten times higher now than it was in 1820.
The thought experiment of a static society points to labor substitution as a concept that delivers one-off and immediate gains. Labor augmentation, on the other hand, is a dynamic concept that keeps delivering productivity gains over time. Missing the greater dynamic benefits of labor augmentation for the lesser static benefits of labor substitution is known as a Turing trap.3
GPTs are defined as technologies that are widely used, are capable of ongoing technical improvement, and enable innovation in application sectors.4 As generative AI improves, the insurer benefits from additional opportunities for co‑invention in applications and workflow design. The velocity of this feedback cycle between technical improvement and downstream innovation is expected to be higher for generative AI than it was for older GPTs such as the steam engine, electricity, or electronic computing.5
In 1996, Timothy Bresnahan (Stanford University) and Shane Greenstein (University of Illinois, now Harvard University), also scholars of technology diffusion processes, published a seminal study on the transition of U.S. corporations from mainframe computing to C/S (client/server) computing.6 The researchers observed that the main bottleneck in adopting and benefiting from C/S computing was the cost of co‑invention. As a result, organizations slowest to transition were those with the highest costs of adoption rather than the lowest benefits of adoption.
The study on the diffusion of C/S computing holds an important lesson for the insurer’s technology choices. Real option theory has inspired principles of decision-making that emphasize the importance of learning and irreversibility. Here, learning refers to the rate of technological advance, and irreversibility relates to the cost of co‑invention. Real-option principles of decision-making favor technology choices that promote a high degree of reversibility and expand the option set available to the decision-maker in the future. Both aspects work in favor of low costs of co‑invention and, as a result, early adoption of technological advances.
Financial Options – Option theory was developed in finance for the purpose of valuing derivatives — a derivative is a financial instrument the price of which derives from the value of an underlying asset, such as a stock. A financial option accords its holder the right (but not the obligation) to buy (call option) or sell (put option) the underlying asset at a pre-specified point in time (European-style exercise) or within a pre-specified time window (American-style exercise).
Real Options – The concept of real options applies option theory to investment decisions on real (as opposed to financial) assets. Real options originate in the simultaneous presence of learning and irreversibility of decisions. Learning is associated with future changes in the environment, which may arise from new technology arriving or business conditions changing. There is complete irreversibility of a project that is not pursued but there is typically some irreversibility of a project that is executed. The extent to which an executed project is irreversible is related to sunk costs, which are costs that cannot be recovered in the event of a project being terminated or its course being changed.
Value of Waiting – Learning and irreversibility create value to holding off with the decision to execute a project. This is because executing a project today deprives the decision-maker of the option of executing this (or a competing) project at a future date after additional information has arrived. This value of waiting subtracts from the (traditionally calculated) net present value of the project.
Cost of Waiting – Executing a project today determines the set of real assets in place in the future, and the decision not to execute today is always irreversible. Business conditions may change and not having executed the project may be a regret. Also, executing a project today determines the set of projects available for execution in the future. This is particularly important where technology is lexicographic in nature (“without A no B, and without B no C”). A prominent example is a cloud infrastructure (A) supporting a state-of-the art data architecture (B) that is foundational to cloud-based AI systems (C). The (absolute value of) the cost of waiting adds to the (traditionally calculated) net present value of the project.
Real Options Approach in Practice – Real options are challenging to calculate in complex business environments. The practical value of real option theory derives from establishing principles of decision-making that emphasize the importance of learning and reversibility. The value of waiting favors technology choices that come with a high degree of reversibility. The cost of waiting prefers technologies that expand the option set available to the decision-maker in the future.
The arrival of a GPT requires an organization to make complementary investments. In the case of generative AI, these complementary investments may be related to cloud infrastructure, a cloud-based data architecture, and workflow orchestration engines where these foundations are not yet in place. Importantly, the process of co‑invention requires investments at the application layer in software engineering, workflow orchestration, and organizational design, all of which contribute to the formation of intangible assets.
At the early stage of the GPT adoption process, these intangible assets are output that the organization produces alongside its goods and services. Once built, the assets serve as input to the organization’s production process. This intangible capital tends to escape traditional measures of productivity, both as output and as input. As a result, when these assets are formed, an organization’s productivity is underestimated, and measured productivity may even decline. Later, when this intangible capital serves as production input over the longer run, the organization’s productivity is overestimated.7
For an insurer, a crude measure of productivity is the expense ratio.8 While intangible assets are formed, they consume resources and contribute to the numerator of the expense ratio only. Following their formation, when they serve as a resource in production, they contribute to the denominator only. Awareness of a potential J‑curve effect in productivity is important for quantifying the contribution of generative AI to the insurer’s productivity in the short and the long runs.
The arrival of a GPT is rare event, even in modern times. This makes studies on the diffusion process of older GPTs a valuable source of information. Lessons of past arrivals of GPTs demonstrate the importance of an organization’s technology decisions in steering the adoption process. Technology choices that promote a high degree of reversibility and expand the option set available in the future will allow the insurer to benefit from the expected high velocity of technical improvement and downstream innovation.
