The productivity paradox and the J-curve

While we invest record amounts in AI, productivity statistics remain flat. Where did the return get buried?

Since the generative AI boom began, global infrastructure investment has surpassed one trillion dollars ¹. Goldman Sachs projects this technology will inject 7 trillion into global GDP ². And yet productivity remains flat: G7 productivity growth averages an anemic 1.4% annually, identical to the previous decade ³.

This paradox is technically baffling but historically predictable. Because it has happened before. In 1900, factories began to electrify. Large electric motors were installed to replace central steam engines. They kept the same design: a drive shaft connecting all the machines via belts. The result was disappointing. Costs dropped slightly, but productivity didn’t budge.

It took 40 years to understand what electricity actually made possible: eliminating the central shaft and putting a small motor in each individual machine. This was not an incremental improvement — it was a radical redesign that allowed the factory floor to be reorganized. And it was that redesign, not electricity itself, that sent productivity surging in the 1920s.

Today we are repeating the mistake. We use AI to write emails faster, not to eliminate the need for email. To code more quickly, not to redefine what we build. To answer support chats in half the time, not to make a product that doesn’t need support. We are in 1900, not 1920. And we wonder why the numbers won’t move.

The “Productivity J-Curve” theory ⁴ explains the phenomenon perfectly: general purpose technologies require massive intangible investment — retraining, process redesign, organizational restructuring — that accounting records as current expenditure, not as an asset. During this phase, productivity stalls or falls. Only when these invisible investments mature does the curve rise sharply. For AI, McKinsey estimates that inflection point will begin to materialize around 2028 ⁵.

That time horizon will only deliver value to organizations capable of distinguishing between two types of investment. One means buying technology to do the same things, only faster. The other requires admitting that current processes were designed to solve problems from another era, and that optimizing them is the most expensive way to perpetuate mediocrity. The companies that capture exponential gains will not be those investing the most billions in GPUs. They will be those with the courage to ask what to eliminate before automating. Five years from now, when productivity finally takes off, the gap between the two groups will not be incremental. It will be unbridgeable. And by then, there will be no budget left to close it.