Every AI rollout I’ve watched up close follows the same pattern: buy the licences, run the training, keep the process, and wait for the results to show up. We tried that approach once before, with electricity. It took about forty years.

McKinsey’s latest State of AI survey, published in August, puts a number on where we are. Eighty percent of respondents say AI has improved their own productivity. Thirty-seven percent say it has contributed anything at all to their company’s earnings before interest and taxes, a share that is “essentially unchanged” from a year earlier. The tools work for the person using them. So far they mostly don’t show up in the company’s results.

That gap has a precedent, and the precedent tells us what closes it.

For the rest of us: what a line shaft was

Picture a factory in 1890. In the basement or out back, one large steam engine turns a single iron shaft that runs the length of the building, near the ceiling. Leather belts drop from that shaft to every machine on the floor. When the engine runs, every belt turns. When it stops, everything stops.

That arrangement dictated the building. Long shafts lose power to friction, so factories were built tall rather than wide, stacking floors to keep the shaft short. Machines were placed where the belts could reach them, not where the work would flow best. The building’s heavy frame existed partly to carry the shafting.

Now electricity arrives. The obvious move, and the one most owners made, was to take out the steam engine and put an electric motor in its place. Same shaft, same belts, same building. Economists call this group drive: one motor turning a shaft that serves a group of machines.

The less obvious move was unit drive: a small motor on each machine, no shaft, no belts. And once the shaft was gone, the building could change too.

Four decades on the old line shaft

The economist Paul David laid this history out in 1990, in a short paper titled The Dynamo and the Computer. He wrote it because economists were puzzled that computers were everywhere but productivity wasn’t moving, a puzzle Robert Solow had summed up in 1987: “You can see the computer age everywhere but in the productivity statistics.”

David’s answer was that we’d seen this before.

Edison’s central power stations opened in New York and London in 1881. By 1899, electric motors still accounted for less than 5% of the mechanical drive capacity in American factories. Factory electrification had no measurable impact on manufacturing productivity before the early 1920s, when only slightly more than half of factory drive capacity had been electrified. That, David noted, was “four decades after the first central power station opened for business.”

Part of the delay was plain economics. Existing factories still worked, and ripping out serviceable plant didn’t pay. Electricity itself only became cheap enough to change the calculation after 1914–17, when regulated utility rates fell. But the part that should interest anyone running an AI programme is what factories did during those decades. From the mid-1890s to about 1920, group drive dominated. Owners added electric motors to the old shaft-and-belt system, and the old system stayed in place. David points out that this layering actually held productivity growth back: the old equipment still sat on the books, the capital tied up in the plant went up, and the output barely moved.

When the gains finally came, they didn’t come from the motor. They came from what unit drive made possible. Without overhead shafting, factories could be built single-storey and lighter. Machines could be placed to follow the flow of materials instead of the reach of a belt. A line could be rearranged without shutting down the whole plant. Skylights went where the shafting used to hang, and workers were no longer caught in belts. By David’s estimate, about half of the 5-percentage-point acceleration in US manufacturing productivity growth in the 1920s is statistically accounted for by the spread of electric motors in that decade.

The dynamo was the same machine in 1900 and in 1925. What changed was the factory built around it.

Where AI is on that timeline

Read the McKinsey numbers again with that history in mind.

The survey (1,719 respondents in 97 countries, fielded May to June 2026) finds adoption is broad and getting broader. Eighty-nine percent report regular use of AI in at least one business function; 44% say it is scaling across the enterprise, up from 38%. Individual benefit is broad too: the 80% productivity figure holds almost identically from the C-suite down to individual contributors.

But enterprise results are flat. The 37% EBIT figure hasn’t moved. And the group McKinsey calls “AI high performers,” companies attributing at least 5% of EBIT to AI and describing its value as significant, is stuck at about 6% of respondents.

What separates that 6% isn’t the tool they bought. McKinsey’s answer is the unit-drive answer. Nearly three-quarters of high performers report fundamentally redesigning workflows because of AI, up from 55% a year ago. Among everyone else, it’s about one-quarter. In the words of one of the report’s authors, high performers “fundamentally redesign workflows rather than layer AI onto existing ones.”

That’s group drive and unit drive, a century apart. A licence for everyone and a process nobody touched is a new motor on the old line shaft.

One honest caveat. The high-performer group is 92 respondents, and all of this is self-reported survey data. It shows a strong association, not proof that redesign causes the returns. But it lines up with the historical pattern too closely to dismiss.

What “unit drive” means for AI

Nobody in 1900 needed to be persuaded that electric motors worked. What they lacked, David argues, was experience: architects and engineers who had built enough electrified factories to know what the new layout looked like. That knowledge was built one plant at a time.

We’re in the same position. The practical questions for any leader aren’t about models. They’re about the building:

  • Pick one workflow and rebuild it end to end. Not “add AI to step four,” but ask what the process would look like if it had been designed around the capability from the start. Where do approvals, handoffs and rework exist only because the old tools needed them?
  • Measure the process, not the tool. Seat counts and usage dashboards measure how many motors you installed. Cycle time, error rates and cost per case measure whether the factory changed.
  • Expect the old system to fight back. Paper procedures kept alongside computers, David noted in 1990, sometimes hurt “each system’s performance.” AI bolted onto a process that still requires every old sign-off adds work instead of removing it.
  • Build the architects. The scarce resource in the 1910s was people who knew how to lay out an electrified factory. The scarce resource now is people who can redesign a workflow around AI and own the result.

What this means

History rhymes; it doesn’t repeat. AI is diffusing much faster than electricity did, and nobody has to wait for utility rates to fall. David himself warned against two opposite mistakes: undue optimism and unrealistic impatience. Both apply here. The individual gains are real, and the enterprise gains are not automatic.

What the dynamo story settles is where the gains come from. They didn’t come from the dynamo in 1900 and they won’t come from the model in 2026. They come from rebuilding the factory: the process, the roles and the decisions around the new capability. Most organisations are still running AI on the old line shaft. The 6% who aren’t are the ones pulling ahead.

References