In 1987 Robert Solow joked that you could see the computer age everywhere except in the productivity statistics. Three years later the economic historian Paul David answered him with a short paper called The Dynamo and the Computer. He went back to electricity.
Edison opened his Pearl Street power station in New York in 1882. American factories took until the 1920s to show the gains. The delay came from the buildings. Steam-era factories drove every machine from one central shaft through belts and pulleys, and stacked their floors close to the engine. Early adopters swapped the steam engine for a big electric motor and left everything else in place. The gains arrived when a new generation of factories gave each machine its own small motor and laid the work out on a single floor, following the flow of materials.
The same lag, with agents
Most AI deployments today look like that first electric motor bolted onto the old shaft. An assistant sits on top of a process designed for people passing documents between desks. The process underneath has not changed.
Redesign takes years because the buildings this time are job descriptions and approval chains. I wrote about the productivity paradox in 2025, and David’s paper is still the clearest explanation I know for the gap.