Everywhere except in the productivity figures.
In 1900 electricity ran the factories and changed nothing on the balance sheet. The reason it took forty years to pay off is the reason your AI has not yet.
By 1900, electricity had visibly won. Edison's first power station had been running in lower Manhattan since 1882. Electric light was spreading through cities, electric trams moved through the suburbs, electric motors were turning up in more and more factories. An investor reading the trade press would have concluded the age of electricity had arrived.
An economist reading the productivity figures would have concluded that almost nothing had happened. Through the 1890s and 1900s, output per hour in American manufacturing grew at about the same unremarkable rate as before the dynamo. The most important new power source in a century was everywhere except in the numbers.
The economist Paul David laid the puzzle out in a 1990 paper, answering a line Robert Solow had written three years earlier about computers — that the computer age was visible everywhere except in the productivity statistics. David's point was that the delay was not a glitch. It was the rule.
Something sat between the new power and the gain it should have produced, and it took forty years to clear. What it was, and why it took so long, travels straight into the present.
The dynamo paid nothing at first because the factory owners bolted it onto the architecture they already had.
The old architecture was the steam-powered mill. A steam engine delivers its power through one central drive shaft running the length of the building, so factories were built tall and narrow, stacked over several floors to keep every machine close to the shaft. Power ran through a dangerous labyrinth of leather belts and pulleys that turned constantly, whether one machine was working or fifty.
When the owner bought a dynamo, he unbolted the steam engine, bolted the electric motor to the same central shaft, and left the rest of the building as it was. The belts still rattled and the vertical floors still dictated the work. The investment was large and the productivity yield was flat.
The gain arrived only when a later generation saw what the dynamo actually offered. Its point was not that it was a cleaner steam engine. Its point was that power could be decentralised — a small motor placed on each individual machine, the central shaft abandoned entirely. Freed from the shaft, the factory could be rebuilt single-storey, wide and flat, arranged around the natural flow of materials instead of the geometry of a drive train. The owner had to stop automating the old structure and redesign the building around the tool.
The delay between the dynamo and the American productivity boom of the 1920s was roughly forty years — about the time it took for the generation of vertical-factory owners to retire and their capital to be written off.
That forty years is the precedent for the pattern, not for the duration. The delay was generational because its bottleneck was generational: the boom waited on incumbents to age out. The bottleneck now is a different one. The advantage to the firms that redesign early compounds, and it compounds quarterly. The factory owner who waited out the dynamo's forty years still had a factory at the end of them. The firm that waits out the equivalent now will face rivals whose head start was compressed into a few years. The pattern repeats. The clock does not.
The modern enterprise is repeating the unbolted-engine move, and doing it at speed.
It buys an advanced model, provisions access, declares a technological win, and leaves the structure around the tool untouched. The consulting layer encourages this. It has always told boards that adoption is a one-layer race — whoever deploys fastest wins — and it is selling the AI transition on the same terms.
But the dynamo's lesson was that deployment is the cheapest, earliest act of a long change, and the value waited on everything underneath it. So when the model is installed and the core numbers stay flat, the useful question is not which vendor to switch to. It is which layer beneath the tool the flat result is pointing at.
A tool lands in an organisation on top of three layers it does not control, and only the top one is technology.
The buckling is a reading from the structure — proof the old architecture has met its limit.
Beneath the tool sits the data it has to draw on — the institutional memory, scattered in most enterprises across siloed systems and contradictory definitions. Beneath that sits the fluency of the people using it, and whether they meet it as something to explore or something to survive. And beneath all of it sits the workflow architecture: the assumption about how value gets defined, divided and delivered through the organisation. Three of those four layers are human and organisational. The technology is only the first, and it is the one that needed the existing organisation least.
This is why the early months get worse before they get better. The fast new tool pulls on the data layer and hits fragmented records. The workforce, watching a tool deployed to absorb their tasks, uses it to automate their existing output rather than to find new value. And the workflow still locks decision authority above the front line, so a surge of machine-generated volume lands at a middle layer with no power to act on it. Friction rises across all three. The data grows less trustworthy, the workforce more wary, and the executive layer finds itself managing a flood of automated noise while the financial metrics sit still.
The "worse" is not a sign the tool was a mistake. It is the sound of a wooden carriage being driven at Mach 1 by a jet bolted to its axle. The buckling is a reading from the structure — proof the old architecture has met its limit.
For ten thousand years a tool could be adopted by being installed. This is the first that has to be adopted by being built around. A firm cannot build itself around a fluid, adaptive technology while still treating the people who do its work as objects to be synchronised by a downstream change plan. The friction will not yield to a sharper milestone report or a more aggressive adoption dashboard, because the friction is not an execution problem. It is the structure telling the truth about its own shape.
So when the deployment lands and the numbers stay flat, the board has a real choice in how it reads the flatness. It can treat it as an adoption failure and reach for more communication, more training, more dashboards — the instinct that makes the threshold harder to cross, because it spends against the workforce while leaving the architecture in place. Or it can read the flatness the way the dynamo should have been read: the tool is bolted to a structure built for a different power source, and the value is waiting on the redesign, not the deployment.
The factory that paid off was the one rebuilt around the motor. The electricity had been there for forty years. What changed was the building.
The final section — the reframe — is reserved for members. Enter your email to continue reading.