Why AI on one step changes little: the lesson of the dynamo

Electric power reached factories decades before it changed what they produced per hour. The economic historian Paul David told the story in 1990, in a paper called “The Dynamo and the Computer”. It is worth knowing before adding AI to a process.

What happened in the factories

The carbon filament lamp dates from 1879, and the first central generating stations opened in New York and London in 1881. Yet in 1899, David writes, electric motors made up less than 5 percent of the mechanical drive in American factories. It took about two more decades to reach half.

The reason was the factory itself. In a factory driven by steam or water, the engine turned long shafts mounted overhead, and belts carried the power from the shafts down to each machine. To lose less power in the long shafts, such factories were built on several floors.

When electric motors arrived, most owners kept that arrangement. They added motors to turn the shafts they already had, one motor for each section of shafting. David calls this the “group drive”, and says it stayed in fashion from the mid-1890s until the eve of the 1920s. The shafts and belts remained. So did the building.

The change that mattered was a motor on every machine, the “unit drive”. With no shafts to carry, a factory could be a single floor of lighter construction, laid out by the path the material takes and not by the reach of the shaft. Machines could be moved without stopping the plant.

When the gains came

David’s account is that factory electrification had no impact on productivity growth in manufacturing before the early 1920s: four decades after the first power station opened. In the 1920s the growth rate of productivity in American manufacturing rose by about 5 percentage points over the decade before, and he attributes roughly half of that rise, statistically, to the growth of electric motor capacity.

Why so long? Mostly, he writes, because it did not pay to replace plants that still worked and were built for the old kind of power. The industries that moved early were the ones growing fast enough to be building new plants anyway.

Two cautions from the same paper

David himself warns against reading the analogy too literally. His words: “Computers are not dynamos.” And electric motors explain about half of the 1920s rise in his own numbers, not all of it.

So the story is not a forecast. It does not say how long AI will take, or how much it will return. It says something narrower, and more useful: the gain did not come from the new power source. It came from rebuilding the work around it.

What this says about AI in a process

Putting AI on one step of a process, and leaving the rest as it is, is the motor on the old shaft. The step gets faster. The hand-offs, the waiting and the second typing around it stay where they were.

Take an order that arrives by email, is typed into the order system, is passed to the warehouse by message, waits to be seen, and is later typed again from the delivery note into an invoice. AI can read the order from the email. If nothing else changes, the message still has to be sent, the order still waits, and the invoice is still typed by hand.

Rebuilding the same process asks different questions. Which steps exist only because of how information used to travel? Could the warehouse see a new order the moment it is saved? Does the invoice need the delivery note at all, or only a confirmation? Some of the answers are a decision, some are a new setup, and some use AI.

Before you can rebuild, you have to see

A factory owner in 1900 could walk the floor and see the shafts. The shafts of office work are harder to see. They are the hand-offs between people, the inboxes where items wait, and the same detail typed into a second system.

They also do not appear in the procedure, and one person rarely knows all of them. Customer service knows it sends a message. The warehouse knows the message can sit for hours. The bookkeeper knows the invoice is typed from paper. The process is what you get when you put the three accounts together.

That is the order of work: first see how the work is really done, by the people who do it, with times and waits as they report them. Then decide what a decision can fix, what needs rebuilding, and where AI belongs.

Bondi Discovery does the first part. An AI interviewer talks with each person about a real recent case, and the product joins what they said into one map of each process, with a source for every detail.