A new motor on an old chassis: why AI still doesn't show up in productivity

In 1899, twenty years after the first central power stations went live, electric motors accounted for less than 5% of factory mechanical drive in the United States. The 50% mark was only reached in the early 1920s. And the acceleration in manufacturing productivity showed up only in the 1919–1929 decade.

Forty years between the technology existing and the gain appearing. Paul David told this story in 1990, and the interesting part is not the delay — it is the reason for it.

The factory that swapped the motor and changed nothing

A steam-powered factory had a layout dictated by physics. One huge engine at the centre, a line shaft crossing the ceiling, leather belts dropping to each machine. Since power was lost over distance, everything had to sit near the shaft. Multi-storey buildings, machines stacked by proximity to power rather than by sequence of work.

When electricity arrived, the first generation of factories did the obvious thing: they replaced the steam engine with one big electric motor and kept everything else. Same shaft, same belts, same building. It was called group drive. It worked, it was cleaner, and the productivity gain was close to nothing.

The leap came with unit drive: a small motor on each machine. Then the shaft went away, the belts went away, and the building could be light and single-storey. Only then could a machine be placed where the flow of work wanted it, instead of where the power reached.

The motor was not the innovation. The motor was permission to redesign the factory. Whoever only swapped the motor paid for the investment and kept the old layout.

The same layout, under a new name

Today AI gets installed on the old chassis with remarkable ease.

The assistant writes the email nobody needed to write. The copilot summarizes the meeting that did not need to happen. The model drafts the document that exists because a 2011 process demands that document. Approval still passes through the same four levels, rework still happens at the same point, the bottleneck is still where it was — now with a layer of AI on top, faster and more expensive.

That is group drive. New motor, old shaft.

An MIT Media Lab report from August 2025 estimated that around 95% of enterprise generative AI pilots produced no measurable impact on results. The number made headlines and the methodology was publicly challenged — treat it as an indication, not a law. But the direction matches what you see in the field, and what happened with the electric motor, the personal computer and ERP.

The bill nobody puts in the proposal

Here come the numbers that change the conversation, and they are not new.

Brynjolfsson and Hitt, in a 2000 paper, estimated that the ratio of intangible assets to IT assets may be 10 to 1. For every dollar in technology, something like ten dollars in redesigned process, training, firm-specific human capital and organizational change.

And there is a figure in that same paper that anyone who has lived through an ERP rollout recognizes instantly: across a sample of projects, hardware was less than 4% of typical start-up cost, licenses and development were 16%, and everything else — roughly 80% — was consulting, process redesign and training. The authors add that employees' own time was not included. So the real ratio is worse than the table says.

Sit with that. You did not buy a system. You bought a process change, and a system came with it.

Everyone who approves an ERP budget knows this in practice, and almost nobody plans it that way. The project is measured by system milestones, invoiced by system deliverables, and the part that consumed 80% of the money — the new process working in the hands of real people — appears as a line called "training", near the end.

What the J-curve adds

Brynjolfsson, Rock and Syverson described in 2021 what they call the productivity J-curve: when a firm invests heavily in intangibles — redesigning process, training, reorganizing — measured productivity falls before it rises. The effort is real, the spending is real, and the result has not arrived yet.

For AI specifically, their estimate is that complementary intangible investment needs to be 2.7 to 4.1 times the observable investment in technology. Less than the 10-to-1 of the computer era, and still most of the money.

The practical reading is harsh: an organization investing correctly in AI will look worse before it looks better. And most companies kill the initiative at exactly that stretch of the curve, because they demand returns in the quarter where the cost lands and the gain has not.

If the spending is process, the deliverable is process

If 80% of the money goes into changing how people work, then the deliverable is not the system. It is the new procedure, executed by real people, with the exceptions life brings.

And here the conversation stops being economic and turns operational, because process change has a known behaviour: it reverts. Three months after go-live, half the team is back to the old way — with a side spreadsheet, an exception that became the rule, a workaround someone invented in week two and taught to everyone else.

When someone asks why the gain never appeared, nobody can answer. Not because people hide it: because there is no record of the new process being executed. There is the design of how it should be, there is the system, and there is what actually happens — and that third thing was never written down.

Process change is proven, or it did not happen

It is the same reasoning as the 1920 factory. Swapping the motor is verifiable: the invoice proves it. Redesigning the layout too: the floor plan proves it. But changing how people work is only verifiable if someone records the work being done the new way.

Without that record, three things happen and all of them are expensive.

First: you cannot tell whether the change stuck. Adoption becomes the manager's opinion, and a manager's opinion is the most biased data point in any project.

Second: you cannot train the next person. Whoever found the right path leaves the team, and the knowledge leaves with them. The document that remains describes the happy path, written from memory by someone who already knew how.

Third: you cannot prove anything later. When a result is questioned — by an auditor, a client, a new director — the answer depends on the memory of whoever was there.

That is the link between the economics and the operation: 80% of the budget buys a process change, and a process change that is not recorded while it happens is not manageable, not teachable and not auditable. It is the definition of money spent without an asset formed.

Where this reasoning does not apply

• Not every gain requires redesign. Replacing a discrete task — transcribing, translating, classifying — pays off immediately without touching the process. The motor analogy does not hold there, and forcing it is the mirror-image error of this article.
• The analogy has a time limit. Software diffuses far faster than electric motors. Forty years was that transition's timeline, not a law of nature. Anyone quoting "forty years" as a forecast for AI is using history as an oracle.
• Not every company is on the low part of the curve. Some simply bought badly, and no J-curve saves that.
• And a record is not a result. Documenting execution does not make the new process good. It makes it observable, which is a condition for improving it — not a substitute.

What I would do

• Treat the project as what it is. If 80% of the cost is process, 80% of the governance should be about process, not system milestones.
• Write the new process from execution, not from intention. The design says how it should be; execution shows how it is. The gap between them is your work plan.
• Record execution while it happens, in the pilot and in the first weeks after go-live. It is the only moment when the workaround shows up raw, before it becomes a silent habit.
• Measure adoption with evidence, not with an opinion survey. "How many executions followed the new procedure" is answerable; "did the team adopt it?" is not.
• Manage expectations on the low stretch of the curve. Whoever approved the budget needs to know, in advance, that the indicator will get worse before it gets better. Explaining that afterwards sounds like an excuse.

Where Walkstamp fits

Walkstamp exists for the step almost nobody performs: recording execution while it happens.

The person records the screen and works normally, narrating. The tool isolates the moments when the screen changes — including the error that appeared and the workaround that was used — and builds a document with clock time per step, the environment, and the spoken audio paired with each segment. Output is PDF, Word, HTML or Markdown, and it serves both as test evidence and as a work instruction born from what was actually done.

It does not design processes, does not measure adoption on its own, and does not grant compliance with anything. It does one thing: it turns execution into a record, without asking anyone to stop and document afterwards.

And the video never leaves your machine: the whole process runs in the browser, with no server in the path. It is all in Security. It is free and needs no account — you can open it now.

The 1900 factory had the right motor and the wrong chassis. It took forty years for someone to decide to change the chassis. The question I would leave is whether your AI project is buying a motor or paying to redesign the factory — and whether anyone is recording the difference.

---

Sources: Paul A. David, *The Dynamo and the Computer* (American Economic Review, 1990), for the electrification figures and the group-drive/unit-drive distinction. Erik Brynjolfsson and Lorin Hitt, *Beyond Computation* (Journal of Economic Perspectives, 2000), for the 10-to-1 ratio and the implementation cost breakdown. Erik Brynjolfsson, Daniel Rock and Chad Syverson, *The Productivity J-Curve* (American Economic Journal: Macroeconomics, 2021), for the J-curve and the 2.7–4.1x estimate for AI. The 95% figure comes from an MIT Media Lab report released in August 2025, whose methodology has been publicly challenged.