The AI productivity paradox: why effect lags behind expectation

Few technologies have been announced as unanimously as a productivity leap as artificial intelligence. In the figures, this leap is so far hard to find. The AI productivity paradox names this gap: the tension between the broad expectation of AI's effect and its so far hard-to-measure trace in productivity. The term is a sharpening of an older observation, not a new discovery.

This gap is no reason for disappointment. It is a familiar pattern, and whoever knows it reads the current figures differently.

An observation older than the AI-hype

The economist Robert Solow put the pattern into a formula in 1987. The computer age, he wrote, could be seen everywhere except in the productivity statistics. The machines stood on every desk, and yet measured productivity barely moved. Economists have since called this gap the Solow paradox.

Only years later did the picture reverse. The productivity gains of computerisation appeared in the statistics once companies had learned to rebuild their workflows around the technology. The technology alone had moved little; only its embedding lifted the return, and that took time. The AI productivity paradox is the application of this old observation to the present.

What the current figures show

Two recent studies give the pattern an empirical form. They measure different things and therefore stand side by side, not one inside the other.

The first is a working paper of the National Bureau of Economic Research. It surveyed around six thousand executives in the United States, the United Kingdom, Germany and Australia and found that about eighty-nine per cent reported no productivity effect from AI over three years. This is the gap in figures, broadly gathered and across countries. They are, however, perception data from self-report, not a causal proof. They show that expectation runs far ahead of the measured effect, not why.

The second study is a working paper of the US Census Bureau by Kristina McElheran and Erik Brynjolfsson. It provides a causal finding but applies to a narrower object: industrial AI in manufacturing, not the generative AI of knowledge work. Within this boundary it shows that adoption first costs productivity and only yields gains over the medium term. Productivity falls into a trough and only then rises above the starting level. The researchers call this course a J-curve.

Both papers are preliminary. They are published as working papers and internally reviewed, but not finally peer-reviewed. This does not diminish their value; it only determines how far one may rely on them.

Why expectation runs ahead of effect

Why does the gap open? There are several plausible explanations, and they do not exclude one another.

The first is the J-curve itself. A new technology does not lift the return on its own; only the complementary investments in processes, skills and structure lift it, and these take time. For manufacturing this course is empirically supported. For the broad gap of knowledge work it is a plausible but unproven reading.

The second is a question of measurement. Productivity effects appear late and dispersed in the statistics, long after the technology was introduced. This was Solow's original argument, and it holds today as it did then.

The third is expectation itself. Hype and pilot projects create expectation faster than implementation creates effect. It is possible that the gap reflects less a delayed effect than a persistent optimism running ahead of reality. Which of the two explanations prevails is open; the broad survey leaves precisely this question unanswered.

What this means for judging AI promises

It does not follow from the paradox that AI does not work. It follows that the absence of short-term effect is to be expected and, on its own, no reason to reverse course. Whoever measures an adoption by the productivity figure after one year and declares it failed confuses the trough with the outcome.

The real risk lies elsewhere. It lies in drawing the wrong conclusions in the trough. The causal finding from manufacturing points to this, and here the statement is demonstrable, that it is precisely older industrial plants that lose, because under the pressure of adoption they abandoned proven management practices. At these plants, the abandonment of established practices explains a substantial part of the initial losses. The maturity they had built over years became a burden in the trough, because they threw it overboard under pressure.

That is the real lesson of the paradox. The question is not whether the technology works, but whether an organisation crosses the trough without damaging itself in the process. The questions can be answered with a organizational diagnosis, like our Ambiflow-Diagnosis.

From paradox to adaptability

The AI productivity paradox describes a gap in time. What an organisation does in this gap decides on which side of the curve it emerges. Precisely this capacity, to hold what bears the load under pressure and at the same time to reallocate with intent, is what makes an adaptive organisation. Whoever wishes to understand how to recognise it will find the account on the topic page on the adaptive organisation.

---

Bernhard Nitz is the owner of transformind GmbH and a partner at Königswieser & Network. He works with leadership teams in corporations and SMEs across the DACH region when change threatens to founder on its own complexity.