How to win with AI
What can history tell us about how firms should adopt artificial intelligence?
This is a free post from Our Long Walk, my blog about the economic past, present, and future of South Africa. If you enjoy it and want to support more of my writing, please consider a paid subscription to the twice-weekly posts that include my columns, guest essays, interviews and summaries of the latest relevant research.
When new technologies emerge, even ones with clear benefits, they are not immediately adopted.
Take electricity. While the recent opening ceremony of the Paris Olympic Games may have been dazzling, it pales in comparison to the thrill experienced by attendees of the 1878 Paris Exhibition when electric arc lamps illuminated the Avenue de l’Opéra for the first time. That was the first large-scale public demonstration of electric lighting and it was clear to all Parisians then that electricity would transform society. In America, too, there was much excitement about the economic possibilities of electricity, notably after Charles Brush, two years earlier, demonstrated how electric light could be created by a generator at the Philadelphia Centennial Exposition.
Yet by 1900 – a generation later – electricity was still very much a novelty, with only 10% of American manufacturing powered by it. For many, the optimism of two decades earlier had faded into scepticism, prompting the question: Was electricity really that transformative a technology?
And then, quite suddenly, it happened.
Between 1900 and 1930, the share of electricity in American manufacturing rose from 10% to 80%, with profound economic consequences. We now have the evidence. Economic historian Will Damron uses newly transcribed data from early 1900s North Carolina manufacturers to show that electrification led to a substantial increase in productivity for manufacturers, with electrified factories experiencing a 10% increase in labour productivity and a 9% increase in total factor productivity compared to non-electrified factories. One of the most important benefits was that electricity allowed factories to reorganise workflows. Says Damron:
The ability to power each machine individually allowed manufacturers to reorganise factories. Because energy needs no longer constrained where machines were located, manufacturers could arrange machines so that materials moved linearly through the factory without having to backtrack to previous stages of production. Electricity also made it easier to operate a subset of machines, making it more convenient to operate multiple shifts or shut down part of the factory for repairs or adjustments.
Electrification, Damron finds, also resulted in higher average wages for workers in electrified factories, with these factories paying about 16% more than their non-electrified counterparts. But here’s the key question: Given all the benefits, why did it take so long for manufacturers to switch to electricity as a source of power for their factories?
There are many reasons, of course. Damron argues that the costs of implementation and the need for complementary innovations in transmission and financial markets slowed down adoption. For example, factories located near hydropower plants were more likely to electrify, showing that costs of transmission mattered, while corporations were more likely to adopt electricity due to better access to capital.
But another obvious reason might be that factory owners simply did not understand the potential of this new technology – or how to get the most out of it. If this is true, you would expect to see some experimentation, with some factories more successful than others. Over time, the less successful ones will learn from the more successful ones or be forced to close down.
To find evidence of this, though, is not easy. In fact, we have to go back another century, to the early 1800s, and back to France. In a new paper in the Journal of Political Economy, three economic historians study the slow adoption of mechanised cotton spinning during the first Industrial Revolution in France and the reasons behind it.
Mechanised cotton spinning relied on new inventions forged in Britain during the 1760s and 1770s: the spinning jenny, the water frame, and the mule. These took some time to spread across the Channel to France, because of a British ban on exporting them, but by the early nineteenth century, the French started producing their own spinning machines (based on stolen British blueprints), and a local mechanised cotton spinning industry began to develop. This mechanisation meant that cotton production shifted from the home to the factory, requiring new industry-specific technical knowledge, knowledge that did not yet exist and which necessitated a trial and error process.
It is this process of trial and error that the authors wanted to investigate. Using a hand-collected dataset of these early French factories, they show that initially, the new technology led to large differences in productivity across firms. Factory managers experimented with the best workflows, with some more successful than others. Over time, the authors show, the variation in productivity across factories declined as inefficient plants exited the market and new, more efficient entrants replaced them. What is interesting is that this reduction in the variation of productivity was unique to mechanised cotton spinning; the authors also analyse the metallurgy and paper milling sectors and find that they exhibit far less variation in productivity because they did not require such extensive experimentation of the production process. They also show that factories were learning from each other: those who were located close to a very productive factory were also more likely to see improvements in their productivity.
There are many questions today about the speed of adopting artificial intelligence (AI) technologies in the workplace. The primary obstacle is that AI adoption often necessitates system-wide changes within an organisation, particularly when tasks are interdependent, just like electricity in the early twentieth century or mechanised cotton spinning in the early nineteenth century. This interdependence means that adopting AI for one task can impact the efficiency and outcomes of other tasks, requiring comprehensive organisational redesigns to fully integrate AI. A new paper by economists Ajay Agrawal, Joshua Gans, and Avi Goldfarb points out that AI can lead to more varied decisions, making it harder for an organisation to coordinate its activities. This means AI works best in ‘modular companies’, companies where decisions are mostly independent of each other. In these cases, AI’s ability to predict outcomes can enhance decision-making without affecting other parts of the organisation. In companies where decisions are closely linked, however, adopting AI requires major changes in how teams coordinate. Here, effective communication of AI predictions is essential. The authors suggest that organisations should invest in systems that enable seamless communication of AI-generated insights, allowing decision-makers to learn from each other and coordinate their actions more effectively.
When electricity and mechanised cotton spinning were introduced, factories needed to reorganise and learn new skills to make the most of these technologies. The slow adoption happened because of high initial costs, not fully understanding the technology’s potential and, most importantly, needing to redesign how factories function. For AI, this means companies need to improve how different teams work together to handle the new and varied decisions AI brings. This will take time and resources, but history suggests that those who experiment and learn from the best will ultimately win out.
An edited version of this article was published on News24. Support more such writing by signing up for a paid subscription. The images were created with Midjourney v6.