Skip to content
Fritz Gerald ZephirinMenu

Social

The Dynamo Problem

4 min read

The most important paper for understanding enterprise AI in 2026 was written in 1990, about the 1890s.

The economist Paul David was trying to explain the productivity paradox of his own day — computers everywhere, productivity gains nowhere — and he reached for the strangest episode in industrial history. Electric motors began replacing steam engines in American factories around 1890. The technology was obviously superior: cheaper power, no boilers, no fire risk. Adoption proceeded for forty years with almost no measurable productivity gain. Then, in the 1920s, manufacturing productivity exploded.

What happened in between is the whole lesson. A steam-powered factory was built around a line shaft — one great rotating axle running the length of the building, with every machine belted to it, floors stacked vertically to stay close to the shaft, workflow dictated by belt geometry rather than by the logic of the work. When electricity arrived, factory owners did the obvious thing: they unbolted the steam engine and bolted a giant electric motor in its place. Same shaft, same belts, same layout. The “electrified” factory was a steam factory with a different engine — and it produced steam-factory output, minus the boiler costs. Rounding error.

The gains came only when a generation of engineers asked a different question: if every machine can have its own motor, why does the shaft exist? Unit drive let machines sit in the sequence the work wanted. Factories went single-story. Materials flowed in lines instead of around belts. The assembly line itself — the productivity miracle of the 1920s — was downstream of removing a shaft that had been invisible because everyone had built around it for a century. The technology took five years to install. The unlearning took forty.

Now look at the numbers this site keeps citing. MIT: 95 percent of enterprise AI pilots, no P&L impact, on $30–40 billion of spend. NBER: roughly 90 percent of six thousand executives report no productivity effect — while 69 percent of their firms actively use AI. Adopted everywhere, measured nowhere. We have seen this exact silhouette before. These are line-shaft numbers. The tools are bolted where the old engine was — AI drafting emails inside an unchanged approval chain, a copilot attached to a workflow designed around the scarcity it just eliminated — and the layout has not moved. Economists even have the formal version: the productivity J-curve, in which general-purpose technologies depress measured productivity while firms make invisible, intangible investments — process redesign, complementary skills — before the gains appear on the books.

I run operations, and I can report the line shaft from inside. The pattern I keep encountering is the AI assistant that lands in a support queue and gains nothing — the bot answers, then the conversation flows into the same review step, the same routing, the same human re-reading everything, because that’s where the queue has always sent things. The gain appears only when the queue itself is redrawn: what the AI may finish alone, what goes to a human, and how the handoff carries context. The tool was the motor. The queue was the shaft. This is why our procurement now runs on a rule blunt enough to enforce: redesign the workflow or skip the tool.

The analogy has limits, and honesty requires them. Electricity was capital-embodied — rebuilding meant new buildings, so forty years was partly construction time. AI is software-speed, adopted faster than any technology in history (100 million users in two months), and some gains — coding, support — are already measured and real. The lag this time is not concrete and steel; it is purely organizational: org charts, approval chains, job definitions, the entry-level structure, pricing models. That should compress forty years considerably. But organizational unlearning has a stubborn clock of its own, because the people who must redraw the workflow are often the people the workflow exists to employ. A fair reading of the 2026 layoff wave is that the reorganization has finally, brutally, begun — though “we cut staff and kept the same processes” is just the line shaft with fewer people belted to it.

The dynamo problem, stated as strategy: the prize of a general-purpose technology goes not to the earliest adopters but to the first reorganizers. Whoever asks “why does this shaft exist?” — this approval chain, this queue, this org shape designed around expensive cognition — while competitors are still bolting motors to it, collects the 1920s. Buying the technology is the entry fee. The redesign is the business.

Sources & further reading