We have put an electric motor on the steam shaft

In 1987 the economist Robert Solow, fresh from a Nobel for his work on growth, wrote a line that has outlived most of his equations. You can see the computer age everywhere, he said, but in the productivity statistics.

Firms had spent a decade buying computers. Output per hour had barely moved. The observation was so uncomfortable that it acquired a name, the productivity paradox, and then sat unresolved for years while everyone involved continued buying computers.

I think about that line every time an executive tells me their AI programme is going well.

The explanation arrived in 1990, and it is not about technology

The economic historian Paul David published a paper called The Dynamo and the Computer. His argument was that we had seen all of this before, and recently enough that we should have remembered.

Factories electrified in the 1890s. The productivity gains did not arrive until the 1920s. Roughly forty years separated the technology from the benefit.

The reason is the part worth sitting with. A steam-powered factory was built around one enormous central shaft running the length of the building, with belts dropping off it to drive every machine. Machine placement was dictated by proximity to the shaft, not by the logic of the work. That was the constraint the whole building was designed around.

When electricity arrived, the first thing factory owners did was entirely rational and entirely useless. They ripped out the steam engine and put an electric motor in its place. Same shaft. Same belts. Same layout. Same work sequence. Nothing improved, because nothing had changed except the source of the rotation.

The gains only came when a new generation of managers realised something less obvious: if every machine can have its own motor, then the shaft is no longer necessary, and the factory can be laid out according to the flow of the work rather than the geometry of the power source.

That realisation required new buildings, new job designs, new supervision, new training, and the retirement of the people who had built their careers on the old layout. It took forty years.

What this means for the agent sitting in your service desk

Most enterprises have dropped an AI agent onto a process that was designed around the constraints of human attention. Queues exist because humans cannot hold many things at once. Handoffs exist because specialisation was cheaper than generalism. Approval steps exist because a person can be held accountable and asked to explain.

Then we put an agent on the shaft and are surprised that the numbers have not moved.

Erik Brynjolfsson has a useful refinement of this, which he calls the productivity J-curve. Measured productivity does not merely stay flat during the transition. It dips. The reason is that the intangible investment, the process redesign, the skills, the measurement apparatus, is real expenditure that does not yet show up as output. You are paying for the new factory while still producing at the old rate.

This matters more than it sounds, because it produces the failure mode I see most often. A programme in the dip that cannot prove the dip is temporary gets cancelled, and it gets cancelled by people acting reasonably on the evidence in front of them.

The uncomfortable arithmetic

Gartner expects around 40 per cent of enterprise applications to carry task-specific AI agents by the end of 2026, up from under 5 per cent the year before. Gartner also expects around 40 per cent of agentic AI projects to be cancelled by the end of 2027, citing cost escalation, unclear business value and inadequate risk controls.

Those are usually presented as good news and bad news. They are not. They are one curve.

Adoption is running ahead of measurement. The buying happens in year one and the invoice arrives in year three, by which time nobody in the building can prove the thing worked. Cost escalation, unclear value and weak controls are not three independent risks. They are three symptoms of the same condition, which is that the organisation changed its tooling without changing its process, and then had no instrument capable of detecting that it had made no difference.

Why forty years is the wrong number to take from this

I am not arguing that agentic AI will take forty years to pay off. Software diffuses faster than capital equipment, and a process redesign no longer requires a new building.

The transferable lesson is narrower and more useful. The lag is not caused by the technology maturing. It is caused by the organisation redesigning itself around what the technology makes possible. That work is slow because it is political, not because it is technical. It involves telling people their job shape is changing, and it involves someone senior deciding which function goes first.

If your AI programme consists of buying capability and attaching it to existing process, you have bought an electric motor and bolted it to the steam shaft. The technology works perfectly. The numbers will not move.

What to actually do about it

Three things, none of which require a larger budget.

Find the shaft. For any process you are automating, ask which of its steps exist because of a constraint that no longer applies. Queues, batching, handoffs and staged approvals are usually artefacts of human attention limits. Some of them are still load-bearing. Many are not, and nobody has audited them in a decade.

Measure the unit, not the aggregate. You cannot detect a J-curve you are not instrumented for. What does one ticket, one claim, one invoice cost you to process today? If the answer takes more than a day to produce, you have no way to tell a temporary dip from a permanent failure.

Budget for the intangible. The process redesign, the evaluation apparatus and the training are not overheads on the AI project. They are the project. The technology is the cheap part, and treating it as the whole part is precisely how the 1890s factory owner spent good money on a motor and got nothing back.

Disruption will not come from faster chips. It will come from mathematics, from measurement, and from the unglamorous work of redesigning the factory around the motor.