The AI productivity paradox: why effect lags behind expectation

Few technologies were announced as unanimously as a productivity leap as artificial intelligence. In the numbers, that leap is hard to find. The AI productivity paradox describes this gap: the tension between the broad expectation of AI's impact and its still barely measurable trace in productivity. The term sharpens an older observation. It does not announce a new discovery. The pattern behind it is long established. Anyone who knows it reads today's numbers differently.

An observation older than the AI-hype

The economist Robert Solow put the pattern into words in 1987. The computer age, he wrote, was visible everywhere except in the productivity statistics. Computers sat on every desk. Measured productivity barely moved. Economists have called this gap the Solow paradox ever since.

Years later the picture reversed. The productivity gains of computerisation showed up in the statistics once companies had learned to rebuild their processes around the technology. The technology alone had moved little. Its embedding into broader business processes lifted the return, and that took time: the workflows had to adapt, and adapting added complexity to the firms. The AI productivity paradox applies this nearly forty-year-old observation to the present.

What the current figures show

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

The first is a working paper from the National Bureau of Economic Research. It surveyed around six thousand executives in the United States, the United Kingdom, Germany and Australia. About eighty-nine percent reported no productivity effect from AI over three years. That is the gap in numbers, broadly collected and across countries. These are perception data from self-report, not causal evidence. They show that expectation runs well ahead of measured impact. They do not show why.

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

Both papers were still preliminary at the time this post was published. They have been released as working papers and reviewed internally, but not yet through final peer review. That does not lessen their value. It only sets how far one may lean on them.

Why expectation runs ahead of effect

Several plausible explanations exist for the gap between expectation and measurable impact. They tend to complement each other rather than compete.

The first is the J-curve itself. A new technology does not lift the return on its own. The complementary investments in processes, qualification, acceptance and structure lift it, and those take time. For manufacturing, this trajectory is empirically supported. For the broad range of knowledge work it remains a plausible interpretation, not an established reading.

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

The third is expectation itself. Hype, pressure for quick wins and a wave of pilot projects generate expectation faster than implementation generates impact. The gap may reflect less a delayed effect than an optimism that outpaces reality. Which of the two explanations dominates remains open. The broad survey leaves precisely this question unanswered.

What this means for judging AI promises

The paradox does not let us conclude that AI fails to raise productivity. It follows that few short-term effects can be expected, and that on its own is no reason to reverse course. Anyone who measures an adoption by its productivity figure after one year and declares it failed mistakes the valley for the result.

Experienced asset managers know the J-curve effect. Certain investments produce a negative return in their first years before they generate positive yields. In capital markets this logic is planned for and priced in. In AI investment, that awareness is still largely missing.

The real risk lies in drawing the wrong conclusions in the valley, because the J-curve was never expected. The causal finding from manufacturing points to it, and here the claim is supported: older industrial firms in particular lose ground because, under the pressure of adoption, they abandoned proven management practices. For these firms, dropping established practices explains a substantial part of the initial losses. The management and leadership maturity they had built over years became a burden in the valley, and under the pressure of temporarily reduced productivity they threw it overboard.

The paradox does not give every answer. It sharpens the essential question: not whether technology and process improvement can create value, but whether an organisation crosses the valley of the J-curve without damaging itself or giving up too soon. Answers to these questions can be found, for instance, in a focused Ambiflow-Diagnosis.

From paradox to adaptability

The AI productivity paradox describes a gap in time. What an organisation does inside that gap decides whether it emerges from the curve rising, level or falling.

This capability — holding what carries under pressure and transformation, while deliberately changing what must change — is what defines 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.