What's changing

From the dynamo to AI: why new technologies change tools first — and economies later

Electricity took four decades. The lag was never the technology — it was the architecture of the firm.

Who this is for Owners and executives weighing what AI has actually changed in their company, and what has only changed in their tools.

7 October 2026

In 1990, economist Paul A. David offered a historical explanation for what had become known as the productivity paradox. Computers were spreading rapidly through businesses and investment in information technology was rising, yet productivity statistics showed little of the expected improvement.

David argued that this was not particularly unusual. He compared the computer revolution with the electrification of industry at the end of the nineteenth and the beginning of the twentieth centuries. His central argument was that a general-purpose technology does not generate its full economic value when the technology itself arrives. The real gains emerge later, once businesses redesign their processes, organisational structures and operating models around it (David, 1990).

Today, the same framework offers a useful way of thinking about artificial intelligence. The analogy is not perfect, and David himself warned against taking it too literally. But the emerging evidence points towards a similar sequence:

rapid technology diffusion → local productivity gains → organisational redesign → economy-wide productivity impact

We may currently be somewhere between the second and third stages.

Electricity: the motor was new, but the factory was still old

The electric motor existed long before electricity transformed industrial productivity. In the United States in 1899, electric motors accounted for less than 5% of factory mechanical drive capacity. It took roughly another two decades for electrification to reach the 50% level. Factory electrification had no visible impact on manufacturing productivity growth before the early 1920s, roughly four decades after the first central power station opened (David, 1990, pp. 356–357).

The reason for the delay is particularly relevant to AI. Early adopters introduced electricity into factories that had been designed around steam and water power. Under the ‘group drive’ system, common from the mid-1890s to the eve of the 1920s, electric motors were added to turn sections of the existing shafting, each driving a group of machines. The original shafts and belts stayed in place.

The technology had changed. The operating model had not. The result was a new source of power embedded inside an old production architecture (David, 1990, p. 357).

The major productivity gains appeared only later, when manufacturers stopped treating electricity as a better substitute for steam. Under the ‘unit drive’ system, individual electric motors were installed directly on machines. That changed almost everything.

Factories no longer needed to be built around a mechanical transmission system. They could become lighter, single-storey buildings. Machinery could be arranged around the optimal flow of materials rather than the location of shafts and belts. Production lines could be reconfigured more easily. Individual sections could be maintained without shutting down the entire plant (David, 1990, p. 358).

Electricity did not simply improve the engine. It changed the architecture of production. This is where the parallel with AI becomes particularly important.

Computers went through much the same transition

Computers entered commercial use in the 1950s and personal computers spread rapidly during the 1980s. Yet the major acceleration in US productivity did not become clearly visible until the mid-1990s. Federal Reserve economists Stephen Oliner and Daniel Sichel later estimated that the use of information technology and the production of computers accounted for about two-thirds of the one-percentage-point step-up in productivity growth between the first and second halves of the 1990s (Oliner and Sichel, 2000).

Before that happened, many companies had effectively done with computers what early factories had done with electric motors: they automated elements of an existing system without fundamentally redesigning the system itself. The paper document became a digital document. The typewriter became a word processor. Paper accounting became accounting software. Existing approval chains were transferred into ERP systems.

David recorded the same overlay in his own time: old paper-based procedures were being retained alongside the new electronic methods, sometimes degrading the performance of both (David, 1990, p. 357).

We are now seeing the same pattern with AI.

Most businesses are still at the ‘group drive’ stage

AI is spreading much faster than previous general-purpose technologies. According to the Stanford AI Index 2026, generative AI reached roughly 53% population-level adoption within three years, faster than the personal computer or the internet. The share of surveyed organisations reporting AI use reached 88% in 2025 (Stanford AI Index 2026).

Those numbers require some care. When actual business use is measured across the wider economy rather than among surveyed organisations, penetration remains considerably lower.

A US Census Bureau working paper based on the Business Trends and Outlook Survey found that between November 2025 and January 2026, 18% of US firms used AI in at least one business function. Those firms represented around 32% of employment, reflecting much higher adoption among larger firms. Among very large firms in information, professional services and finance, use rates reached 50–60% (US Census Bureau, 2026).

Europe shows a similar pattern. According to Eurostat, 20% of EU enterprises with ten or more employees used AI technologies in 2025, up from 13.5% in 2024 and 8.1% in 2023 (Eurostat, 2025).

The technology is therefore spreading extraordinarily quickly. But the redesign of businesses around the technology is only beginning. There is an important distinction between using AI and building an organisation around AI.

Suppose an employee previously spent two hours preparing a report and can now create the first draft with ChatGPT in ten minutes. That is a genuine productivity improvement. But the organisational structure may remain completely unchanged: the same employee produces the same report, sends it to the same manager, through the same approval process, using the same systems.

That is the equivalent of fitting an electric motor to the old line shaft while leaving the shafts and belts in place.

The local productivity gains are already measurable

Unlike the early decades of computerisation, there is already reasonably strong evidence of AI productivity gains at the level of individual tasks and workers.

A study of 5,172 customer-support agents found that access to a generative AI assistant increased the number of issues resolved per hour by 15% on average. For less experienced and lower-skilled workers the improvement was substantially larger, around 30%, while the most experienced workers gained little. The authors conclude that the tool appears to capture and spread the practices of the most productive agents (Brynjolfsson, Li and Raymond, 2025).

A separate field experiment by researchers from Harvard Business School and Boston Consulting Group involved 758 consultants. For tasks within the capabilities of GPT-4, consultants using AI completed 12.2% more tasks, worked 25.1% faster and produced results rated more than 40% higher in quality than the control group.

The same experiment produced another result that matters just as much. For a task selected to lie outside the model’s capabilities, consultants using AI were 19 percentage points less likely to produce the correct answer than those working without it (Dell’Acqua et al., 2023).

Simply giving people AI, therefore, is not enough. Productivity depends on which work is delegated to AI, which work remains with humans, how outputs are checked, how exceptions are handled and how the overall process is designed.

The Stanford AI Index 2026 reaches a broadly similar conclusion from the accumulated research. The largest gains come in structured work whose output is easy to measure and monitor: 14–15% in customer support, 26% in software development. In tasks that need deeper reasoning or judgement the gains are smaller and, in some studies, negative (Stanford AI Index 2026, pp. 174, 219).

This is close to the transition from the early ‘group drive’ electrical factory to the individually powered factory.

The next stage is not better AI tools. It is the AI-native organisation.

The most important innovation produced by electrification was ultimately not the electric motor. It was the new type of factory that became possible because of it.

Similarly, the computer revolution ultimately mattered because organisations became digital. Databases, ERP systems, electronic commerce, automated financial systems, digitally coordinated global supply chains and entirely new business models followed.

The same logic suggests that the largest economic impact of AI will not come from copilots, chatbots or faster content generation. It will emerge when businesses begin changing the unit around which work itself is organised.

A conventional business process might look something like this:

information → person → software system → another person → approval → another system → decision → execution

AI allows a fundamentally different architecture:

event → AI interprets context → retrieves relevant information → applies rules → performs the work → escalates only exceptions or decisions requiring human accountability

The distinction is fundamental. In the first model, AI helps a person perform the existing job faster. In the second, AI becomes part of the operating process itself. That is the point at which automation becomes organisational transformation.

The largest obstacle is also familiar

David made another observation that is particularly relevant today. Physical factories depreciate. Eventually, machinery and buildings have to be replaced, and that creates a natural opportunity to redesign the factory around a new technology.

Information structures do not depreciate in the same way. An Excel workbook can survive for twenty years. An approval process can survive for thirty. An ERP workflow can remain operational long after the assumptions on which it was designed have disappeared.

David therefore expected a strong inertial component in the evolution of information-intensive organisations: the mere passage of time does not create occasions to redesign them (David, 1990, p. 360).

AI exposes that problem very quickly. A business can purchase ChatGPT Enterprise, Microsoft Copilot or an AI agent within weeks. But its data may still sit across multiple disconnected systems. Business rules may exist primarily in employees’ heads. Exceptions may still be handled through email and WhatsApp. Responsibility may be distributed according to job structures designed decades ago. And nobody may have a complete picture of where one process ends and another begins.

AI therefore reveals not only an organisation’s technological opportunities. It reveals its organisational debt, to borrow Steve Blank’s term.

We may therefore be entering an AI productivity paradox

In 1987 Robert Solow observed: “You can see the computer age everywhere but in the productivity statistics.” David noted that an observer in 1900 could have said the same of the dynamo (David, 1990, p. 356). Over the next few years we may hear the same remark about AI.

The idea is not new. Brynjolfsson, Rock and Syverson described an AI productivity paradox in 2017 and attributed it mainly to implementation lags. They later modelled the pattern as a ‘productivity J-curve’: measured productivity growth is understated while firms make the complementary investments a general-purpose technology requires, and overstated later, when those investments pay off (Brynjolfsson, Rock and Syverson, 2017; 2021).

At the level of individual tasks, the gains are already visible. At the level of individual business processes, they are beginning to emerge. At the level of the overall economy, there is still limited evidence that AI has produced a major step-change in aggregate productivity.

The Stanford AI Index 2026 draws the same line: task-level gains are measurable, while the economy-wide evidence is still early and inconclusive (Stanford AI Index 2026, pp. 171, 220).

That does not necessarily mean AI is overhyped. Electricity and computers both passed through periods in which investment and adoption became visible well before the corresponding productivity gains.

But history also warns against the opposite assumption. A powerful technology does not automatically create productivity growth. Complementary changes are required: new processes, new skills, new organisational structures, better data, new governance mechanisms, new standards and, eventually, new forms of enterprise.

AI is different in one important respect

There is nevertheless a strong reason to believe that the current transition could happen faster. A dynamo had to be manufactured, transported, installed and physically connected. Computers required hardware, local infrastructure, software and significant capital investment. AI reaches users over networks and devices they already have. Millions of workers can gain access to a new model almost simultaneously.

But this accelerates the distribution of the technology. It does not necessarily accelerate the redesign of the organisation at the same rate.

That may define the AI productivity paradox. Model capabilities may improve extremely quickly. Adoption may spread equally quickly. Yet the architecture of businesses — their processes, responsibilities, systems and decision structures — may change far more slowly.

Conclusion: the real question is not how capable AI becomes

The history of electricity and computing suggests that the most useful question is not how powerful is AI? A better question is: how far have we moved from using a new technology inside an old organisation towards designing the organisation around the new technology?

Electricity became transformative not when an electric motor was attached to the old line shaft, but when the shaft around which the factory had been designed disappeared. Computers became transformative not when the typewriter became a word processor, but when digital technology changed how organisations stored information, coordinated work and conducted business. AI is likely to follow the same path.

We already have credible evidence that AI can increase individual worker productivity by double-digit percentages in suitable tasks. We also have evidence of exceptionally rapid adoption. But much of the corporate world is still using AI as a more powerful tool inside processes designed for human workers and conventional software.

If David’s historical analogy continues to hold, the largest economic impact will come later. It will come when businesses stop adding AI to existing work and start redesigning the work itself around AI.

At that point, the most important outcome of the AI revolution may not be artificial intelligence as a technology. It may be a new architecture of the firm.

References

Start with a conversation.

Book a 45-minute call to discuss what keeps you up at night as a business owner.

Book a call hello@afanasyevs.com