The Power Loom Principle: How Work Redesign Drives Productivity Surges
Productivity Has Stalled. We don’t know what to do about it.
Firms are desperate for productivity gains. After multiple cost-cutting cycles and disappointing returns from AI,1 we’re running out of ideas. The problem isn’t a lack of technology—it’s a misunderstanding of where productivity actually comes from. Not in a macroeconomic sense. It’s clear that technological innovation is the most significant driver of growth, and a great deal of effort is put into policy and planning to ensure that capital (both human and physical) is directed to inventing tools, techniques, infrastructure that will (hopefully) drive the next wave of productivity. Our lack of understanding is in a grounded, work-centric sense. Or, more pointedly, we don’t understand where the innovation that drives productivity comes from.
Our predilection for lone geniuses—the brilliant scientist, insightful inventor, or brave entrepreneur—is no more than an appeal to the “great men” of history, and ignores how innovation actually unfolds. From the steam engine’s iterative refinements,2 innovation emerges from repurposing existing tools.3 Breakthroughs emerge from systems that pool and refine knowledge, as Silicon Valley has done historically.4 Most major discoveries (calculus, telephones) occur independently by multiple people simultaneously, suggesting context matters more than genius.5
Robert Gordon argued (in The Rise and Fall of U.S. Growth)6 that the great productivity boom of 1870–1970 was a one-time event, driven by general purpose technologies like electricity and sanitation. Many have disputed his pessimism, pointing to the potential of future breakthroughs—humans are very creative after all. But both Gordon and his critics miss the deeper point: the true source of past progress wasn’t the technologies—it was how work was reorganised to make those technologies productive. The real revolution happened in the workplace, not the lab.
Productivity is seen as the universal solvent—it can dissolve many of our current problems, problems associated with low growth and tight resource constraints. It can increase profits, reduce costs, and improve competitiveness of businesses. Economics sees increased productivity as a key driver of economic growth, higher living standards, and improved national wealth. It can help address issues such as inflation, unemployment, and budget deficits by creating a larger economic pie. As Paul Krugman opined “Productivity isn’t everything, but in the long run, it’s almost everything.”7 Even individuals seek productivity (or feel crushed by it) in the quest for meaningful and manageable work—the quest for work-life balance.
This shared conviction hides a deeper uncertainty: where does productivity actually come from? Not at the macro level, where it’s measured, but at the micro level, where it is made. While everyone is right to care about productivity, we don’t agree on how it’s produced. Especially not in today’s post-industrial economy. If we want more of it, then what do we invest in?
Productivity is assumed to come from innovation, which is itself assumed to come from technology, which is assumed to come from research. Hence calls for government to invest in research, and government calls for business to invest in technology (capital deepening). We want a technology hero to save us from the low-productivity villain who seems to be stalking us. The current rush to invest in LLMs is a obvious symptom—being startling and unusual, LLMs are assumed to be a disruptive, a ‘hero’ technology. Much as steam was the hero of the First Industrial Revolution and electricity of the Second, the telegraph, printing press, internal combustion engine, and personal computer. More recently we’ve looked to blockchain and then the metaverse. LLMs are just the more recent hero that we’ve fixated on.
We have this all backwards though: we think that innovation is the cart and research is the horse, when it’s the other way around. Steam power became before thermo dynamics, praxis before research, and this is a common pattern we see across history.8 Learning-by-doing, bottom up trial and error is our major source knowledge, and so innovation. We’ve forgotten this. Or, more accurately, we’ve buried the evidence.
Science—or, more accurately scientists and the institutions they create—has obfuscated the contributions of craft and labour to technological progress,9 with scientific knowledge institutionally elevated and insulated from other knowledge systems. These dominant institutions have marginalised local, experiential, or embodied knowledge—learning by doing.10 While empirical studies show that major industrial innovations (e.g., in textiles, metallurgy, machine tools) were frequently developed by practitioners experimenting and tinkering, typically well before scientific theories could explain them.11 Complex systems evolved through interactive learning, feedback, and situated knowledge rather than theoretical breakthroughs.12
Science has become more important as our technology has become more complex,13 but we still have the cart (learning by doing) preceding the horse (research). By discounting the role of learning by doing we limiting ourselves to seeing only a minor part—the science part—of the productivity problem. No where is this more obvious than in Robert Gordon’s one-time technologies, which has a key lesson we have failed to learn.
In The Rise and Fall of American Growth,14 Robert Gordon argued that the 1870–1970 boom in productivity growth, with productivity growth jumping from roughly 1.5% p.a. to 3% before falling back again—was a one-time event driven by a select group of general purpose technologies: electricity, sanitation, combustion, and communication. Gordon considered the slowdown to be likely permanent—the great age of progress is behind us. Critics will often acknowledge Gordon’s historical analysis but conclude by stressing the inherent difficulty in forecasting innovation and placing faith in ongoing human creativity and the potential for unforeseen breakthroughs to drive future growth. This is missing the point, as we need to ask why there was an exceptional 100 years. Pointing to the unpredictable nature of human creativity is not a theory. It’s a vibe.
The real lesson of Gordon’s work is not that the well of progress has run dry, but that we have misunderstood where past leaps in productivity truly came from. The technologies he identified—electricity, sanitation, combustion—were not standalone breakthroughs. They became transformative only when combined with other technologies, and when work systems were reorganised around them, redistributing expertise and enabling new affordances. In that sense, it’s not that we haven’t learned Gordon’s lesson—it’s that Gordon and many others learned the wrong lesson. The power loom was not revolutionary on its own; it was the shift from home-based putting-out systems to factory-based wage labour that unlocked its true productivity potential. The same dynamic applies today.
Let’s turn back to productivity. Take a procedure, a process, a sequence of tasks. If we approach the work via science, then there’s two obvious things we can do to improve productivity. First is to optimise the process, removing variation to increase efficiency as variation is waste. Variation represents inconsistency in outputs, increased cycle times, excess inventory (to buffer against unpredictable demand or supply), extra labor (for inspections, troubleshooting, or handling exceptions) and wasted materials (due to defects or overproduction), and so poor resource utilisation as well as potentially hiding other problems. Eliminating variation removes waste and so improves productivity. The second is to add new, and more productive technology. Technology can improve efficiency and accuracy (reducing defects and waste) while freeing up workers for higher-value activities. Both of these can provide a modest boost. We might get a slightly larger (but still modest) boost when adding new technology by redesigning the process to better integrate technology. However, both of these approaches are limited due to how they approach work, as a procedure—a process broken into a sequence of tasks.
A better approach is to look back into history and see how some of the technology-infused booms in Gordon’s 100 exceptional years developed. To do this we need to consider the work system rather than a process.
In this framing improvement then comes from within the system, rather than without. Workers learning by doing is how the insights and innovation are generated. New technology is developed when knowledge—the worker’s know how—is abstracted, reified in tools and techniques. In weaving, for example, the flying shuttle15 (not much more than a stick, and a couple of boxes) abstracted the problem of moving the shuttle through the shed, a temporary opening created between raised and lowered warp threads when the harnesses (or heddles) are lifted or lowered.
The creation of knowledge—and so productivity—is an iterative process. Worker knowledge is abstracted into tools, and then these tools then drive a redistribution of expertise in the work system.16 A classic example of this redistribution is how a cockpit remembers it speeds,17 where the introduction of speedbugs—small movable markers on an aircraft’s airspeed indicator—externalised memory, turning a mental calculation (“What’s V₂ again?”) into a perceptual task (“Look where the bug is”). The dial itself “remembers” speeds, reducing individual cognitive load which frees them to focus on more important aspects of the work, while experienced pilots no longer need to coach juniors on speed recall as the interface scaffolds performance. Most interestingly expertise becomes a property of the system (pilot & speedbug & checklist), not just the person. Similarly checklists in healthcare reduce reliance on heroic recall by surgeons. Or most recently GitHub’s “Copilot” suggests code, redistributing programming knowledge. The cockpit with speedbugs then becomes the new baseline work system, and the process starts again.
The internal operation of a work system is strongly influenced by its boundary—how it interacts with the other systems around it, exchanging information and resources. Indeed, the structure of the work is the product of an emergent process, a response to the constraints placed on the system, a co-creative interchange between our work system’s agency, the drive and motivation within the system, and the constraints of the system, the boundaries and limitations of the container itself.18 We can see this in the development of the power loom, where the incremental mechanisation of the manual loom (via learning by doing) reached the point where all the common actions of the weaver were supported by tools, mechanisation.
The flying shuttle is interesting in not for its technical sophistication, but for being the last piece of mechanisation before human power could be swapped for mechanical power. This enabled the weaver to oversee 2.5 looms (on average) rather than manually operate one—an instant 250% increase in productivity, though one requiring the loom to be moved from the weavers home to a suitable source of mechanical power (initially waterwheels, which were soon swapped for more convenient steam power)—the boundary for the work system moved. This shift in where the weaving was done required a radical transformation in the work system: the destruction of the traditional putting-out based industry where work was taken to the weaver’s home (the work came to the weaver), and the creation of the factory system and wage labour where the weaver went to the factory (where the work was). (It’s worth noting that subsequent learning by doing with the power loom delivered a further factor of 20 in productivity improvement.)
We can see a similar dynamic with autonomous taxis, which have been heralded as a transformative ‘hero’ technology. This is not the case though as autonomous taxis are similar to the power loom. The don’t enable us to do something new, but there is the potential make something more productive. With the power loom it was replacing human with mechanical power, enabling a worker to oversee multiple looms. Similarly with autonomous taxis, we’re replacing the human driver in the vehicle with an autonomous driver overseen by a remote human who can step in when the autonomous driver is overwhelmed. The viability of autonomous taxis depends the balance between the increased cost of the autonomous vehicle, vs the savings of having a remote driver oversee a number of vehicles. Like the power loom, autonomous taxis are not a revolutionary technology but a productivity multiplier—their viability hinges on optimising the ratio of vehicles to remote overseers, transforming labor efficiency rather than eliminating it, with current estimates that remote operators managing 10–50 vehicles will replace in-car drivers.19 This cost productivity could also be leveraged to dramatically improve services such as public transport and patient transport.
Gordon’s identification of a set of one-time technologies is an astute one. Transforming the work system—as the power loom did and it’s likely autonomous cars will (allowing for the fact that they are not fully autonomous)—marks the end of one long period of incremental improvement, gives us an immediate and massive productivity boost (a factor of 2.5 for the power loom, a factor of 10-50 for autonomous cars), and puts us on a new long path of incremental learning-by-doing. These transitions are one time, as we’re not adding something to the system, we’re using technology to move the boundary that constraints the system. The motor car transformed the urban landscape, opening up vast amounts of cheap in-fill land between public transport line and facilitated the creation the modern suburb.20 The networked home (where power, water, and waste are delivered via utility networks) transformed domestic life. And so on. They are one-time technologies because they are associated with one-time events. We can’t move the boundary in the same way again—any subsequent boost relies on us finding a new way to move the boundary.
This isn’t just a historical curiosity—it’s a call to action. If we continue to treat productivity as something delivered by new tools alone, we’ll keep mistaking flash for substance. The real challenge is not technological adoption but organisational transformation. To unlock the potential of today’s technologies—LLMs, robotics, autonomy—we must reimagine the systems of work around them. That’s where productivity lives and dies.
Productivity improvement—the growth that isn’t everything, but in the long run, it’s almost everything—is a layered challenge. At the top is the transformative work-system change we saw in the power loom and can anticipate with autonomous vehicles, delivering step-changes in productivity. These, however, are one-time changes. Once we’ve made taxis autonomous we’ll have a long period of incremental improvement in front of us—what we can’t do is play the autonomous taxi card again. At the other end we have process optimisation, removing variation to increase efficiency as variation is waste. We can pull the process optimisation level every so often to realise modest gains. The bulk of the opportunity, though, is in learning by doing. Again, the power loom is clear example. The development of the power loom delivered an immediate factor of 2.5 improvement, but this needs to be compared to the factor of four learning by doing delivered to the manual loom prior to the power loom, and the subsequent factor of 20 after it.
In our current quest to improve productivity, we’re putting the cart before the horse. We’re forcing ‘hero’ technologies into every nook and cranny in an attempt to drive productivity up. Investments in these technologies, and the companies developing them, is reaching eye watering levels. Evidence is growing that this investment is not paying off.21 We’re clearly in a bubble.
What we’re failing to grasp is that productivity breakthroughs demand more than new technologies—they require new ways of working. Just as the power loom’s transformative potential only emerged when we reorganised labor around factories rather than cottages, today’s innovations—from autonomous vehicles to AI—will stagnate until we redesign the work systems that give them meaning. Gordon’s lesson about one-time technologies isn’t that progress is exhausted, but that we’ve exhausted the models that drove past progress. The next productivity revolution won’t come from hero technologies, but from reinventing the cognitive and social architectures that make them productive. The critical shift isn’t asking “How does this replace workers?” but “What new system of work does this enable?” Only then can we harness these tools to begin another century of learning-by-doing.
Humlum, Anders, and Emilie Vestergaard. “Large Language Models, Small Labor Market Effects.” SSRN Scholarly Paper. Rochester, NY: Social Science Research Network, April 15, 2025. https://doi.org/10.2139/ssrn.5219933.
Joel Mokyr, The Lever of Riches: Technological Creativity and Economic Progress (1990).
Michel Callon (ed.), The Laws of the Markets (1998).
Annalee Saxenian, Regional Advantage: Culture and Competition in Silicon Valley and Route 128 (2000).
Merton, Robert K. “Singletons and Multiples in Scientific Discovery: A Chapter in the Sociology of Science.” Proceedings of the American Philosophical Society 105, no. 5 (1961): 470–86.
Robert J. Gordon, The Rise and Fall of American Growth: The U.S. Standard of Living since the Civil War (2016).
pp. 11 in Paul R. Krugman, The Age of Diminished Expectations: US Economic Policy in the 1990s. 3. ed., 3. print. Cambridge, Mass.: MIT Press, 1998.
We’re spoilt with examples of the technological horse preceding the research cart. Ancient civilisations discovered how to smelt and work metals like copper, tin (to make bronze), and later iron, long before they had any sophisticated understanding of the chemical and physical processes involved. The development of farming techniques, including crop rotation, irrigation, and selective breeding, predates modern botany and genetics by millennia. Gunpowder was invented in China centuries before a scientific understanding of its chemical reactions existed. Johannes Gutenberg’s innovation was a mechanical marvel that revolutionised information dissemination, and was developed through practical ingenuity, combining existing technologies like the screw press and movable type, without a prior scientific theory of information or communication.
See Bruno Latour, We Have Never Been Modern (1994), where he discusses the historical construction of the divide between nature and society and its implications for scientific knowledge.
See, e.g., Sandra Harding, Whose Science? Whose Knowledge? (1991) and Donna Haraway, Situated Knowledges (1988)
Nathan Rosenberg, Perspectives on Technology (1976).
See Thomas Hughes, Networks of Power (1993) and W. Brian Arthur The Nature of Technology (2009).
Joel Mokyr, The Gifts of Athena (2002).
Robert J. Gordon, The Rise and Fall of American Growth: The U.S. Standard of Living since the Civil War (2016).
John Kay, 1733.
This is the reflective structure of abstraction and cognitive structure developed by Peter Damerow. See Peter Damerow, Abstraction and Representation (2013).
Hutchins, Edwin. “How a Cockpit Remembers Its Speeds.” Cognitive Science 19, no. 3 (July 1995): 265–88. https://doi.org/10.1207/s15516709cog1903_1.
pp. 19 in Benyamin B. Lichtenstein Generative Emergence: A New Discipline of Organizational, Entrepreneurial and Social Innovation (2014).
The economic case for autonomous taxis mirrors historical productivity breakthroughs (e.g., power looms) where tools redistributed rather than replaced human labor. Key parallels:
Power Loom (18th c.): Mechanised weaving, enabling one worker to oversee 2.5 looms (a 250% productivity gain).
Autonomous Taxis (21st c.): Replaces in-car drivers with remote operators managing 10–50 vehicles (potential 90–98% labor reduction).
Both innovations required:
Capital intensification (looms → factories; taxis → AVs & fleet hubs).
Labor reorganization (home weavers → wage labor; drivers → remote ops).
The true challenge lies not in the technology itself, but in reconfiguring work systems to exploit its leverage.
History shows that transformative productivity gains occur when three conditions align:
Tools abstract expertise (flying shuttle → power loom; AVs → remote ops)
Labor reorganizes around new constraints (cottage industry → factories; drivers → fleet managers)
Institutions adapt to scale learning (apprenticeships → technical schools; driver training → simulation centers)
The 1870–1970 boom succeeded because it reinvented all three simultaneously. Today’s stagnation reflects, in part, our myopic focus on tools alone—a failure to recognise that productivity is a property of systems, not technologies. Autonomous taxis, like power looms before them, will only deliver their promised gains when we stop asking “How does this replace workers?” and start asking “How does this reconfigure work?”
Peter Evans-Greenwood, “The Geography of Desire.” Substack newsletter The Puzzle and Its Pieces (blog) (2025) https://thepuzzleanditspieces.substack.com/p/the-geography-of-desire.
Humlum, Anders, and Emilie Vestergaard. “Large Language Models, Small Labor Market Effects.” SSRN Scholarly Paper. Rochester, NY: Social Science Research Network, April 15, 2025. https://doi.org/10.2139/ssrn.5219933.