The Productivity Paradox Returns

In 1987, the economist Robert Solow wrote a sentence that became one of the most cited observations in the history of technology and economics. It appeared in a New York Times Book Review article, almost as an aside: “You can see the computer age everywhere but in the productivity statistics.”

The personal computer was transforming offices across the developed world. Spreadsheets replaced ledgers. Word processors replaced typewriters. Databases replaced filing cabinets. The investment was enormous, the enthusiasm boundless, and the productivity statistics stubbornly flat. The computers were everywhere. The gains were nowhere.

In February 2026, a team of researchers published a paper through the National Bureau of Economic Research that made the same observation about a different technology. Ivan Yotzov, Jose Maria Barrero, and their colleagues surveyed nearly 6,000 CEOs and CFOs across the United States, the United Kingdom, Germany, and Australia. The finding was stark: more than 80 percent of firms reported no impact from AI on either employment or productivity over the previous three years. Nine in ten executives said AI had not moved the needle.

The parallel is not coincidental. It is structural. And understanding why it is structural — rather than dismissing it as a timing problem or a measurement error — is the difference between organisations that will eventually capture AI’s value and those that will spend another decade wondering where the returns went.

The Paradox Has a Pattern

Erik Brynjolfsson has been thinking about this problem longer than most. In 1993, he coined the term “productivity paradox” in a paper that took Solow’s quip and turned it into a research programme. Three decades later, working with Daniel Rock and Chad Syverson, he offered a framework that explains why general-purpose technologies — technologies that reshape entire economies — consistently fail to show up in productivity statistics during their early adoption phase.

They called it the Productivity J-Curve.

The argument is precise. When a general-purpose technology arrives, organisations begin making significant complementary investments — new processes, new business models, new training, new organisational structures. These investments are largely intangible. They do not appear in national accounts as capital formation. They appear as costs. The organisation is spending money, time, and cognitive effort on restructuring, and the restructuring produces no immediate output. Measured productivity dips. The J-curve descends.

Later — sometimes much later — the intangible investments mature. The new processes become routine. The new structures become second nature. The complementary innovations begin producing output. Measured productivity rises sharply. The J-curve ascends.

Brynjolfsson, Rock, and Syverson published this model in the American Economic Journal: Macroeconomics in 2021. They showed that adjusting for intangible investments related to computer hardware and software yielded a total factor productivity level 15.9 percent higher than official measures by the end of 2017. The productivity gains from the computer age were real — they were just invisible for years, hidden inside investments that the measurement system could not see.

The question for AI is whether the same pattern is repeating. The answer, based on the data available in early 2026, is yes — but with a complication that the J-curve framework alone does not capture.

The Data Says Not Yet

The NBER study is not the only evidence. It is part of a convergence.

PwC’s 29th Global CEO Survey, published in January 2026 and covering 4,454 CEOs across 95 countries, found that 56 percent reported no significant financial benefit from their AI investments. Not “modest gains.” Not “early returns.” No benefit. Only 12 percent — one in eight — reported that AI had both increased revenue and reduced costs. The rest were either still waiting or had already concluded that the investment was not paying off.

The San Francisco Federal Reserve, in Economic Letter 2026-06, noted that AI-related investment had surpassed the contribution of IT components to real GDP growth made during the dot-com boom — both in absolute levels and as a share of GDP. The money is flowing. The capital expenditure is real. Knowledge-intensive industries with surging AI-related job postings accounted for 50 percent of output growth in the third quarter of 2025, while representing just over a quarter of total output. But the Fed’s own assessment was cautious: most macro-level studies of productivity growth find limited evidence of a significant AI effect.

The investment is visible. The productivity is not. Solow’s sentence, rewritten for 2026, would read: you can see the AI age everywhere but in the productivity statistics.

The Four-Percent Question

The European Investment Bank added a crucial piece to this puzzle. EIB Working Paper 2026/02 analysed matched data from over 12,000 non-financial firms in the European Union and the United States. The finding: AI adoption increases labour productivity by an average of 4 percent.

Four percent sounds meaningful until you examine the distribution. The gains concentrate in medium and large firms — the ones that already have training budgets, data infrastructure, and the organisational capacity to absorb a new technology. Small firms — the firms that make up the vast majority of the European economy — see less. The 4 percent is an average that describes almost nobody accurately.

More importantly, the EIB found that the productivity gain materialises only in firms that made what the researchers called “complementary investments” — in software, data infrastructure, and workforce training. Without those investments, the gain approaches zero. The AI itself does not produce the 4 percent. The AI plus the organisational restructuring produces the 4 percent. The technology is a necessary condition. It is not a sufficient one.

This finding aligns precisely with the Productivity J-Curve framework. The intangible investments — the training, the workflow redesign, the process re-engineering — are the mechanism through which the technology’s potential becomes actual output. Skip the intangibles and the technology sits idle. Or worse: it sits active but unproductive, generating outputs that nobody uses, automating processes that nobody restructured, and creating the illusion of transformation without the substance.

Eurostat’s December 2025 data makes the gap concrete. Twenty percent of EU enterprises with 10 or more employees use AI technologies. But only 17 percent of small enterprises do, compared with 55 percent of large ones. The adoption itself is stratified. And within the 20 percent that adopt, the EIB data tells us that only those making complementary investments are capturing returns.

The technology is not evenly distributed. The returns are even less evenly distributed. And the organisations most likely to adopt AI without the complementary investments — the small and medium-sized firms that represent the backbone of the EU economy — are the ones least likely to see the paradox resolve in their favour.

The Decision Architecture Problem

Ajay Agrawal, Joshua Gans, and Avi Goldfarb anticipated this in their 2022 book Power and Prediction: The Disruptive Economics of Artificial Intelligence. Their framework is built on a simple decomposition. Every decision has two components: prediction — estimating what will happen — and judgment — deciding what to do about it. AI dramatically improves prediction. It drops the cost of forecasting, pattern recognition, and probabilistic estimation toward zero.

But prediction without reorganised judgment is just cheaper data. The value of AI comes not from better predictions alone, but from reorganising the decisions that those predictions feed into. A logistics company that uses AI to forecast demand but leaves its routing, staffing, and inventory decisions unchanged has improved one input to a process that it did not restructure. The prediction is better. The decision architecture is the same. The output barely moves.

Agrawal, Gans, and Goldfarb use a term that captures the problem: “system-level redesign.” The gains from a general-purpose technology do not come from inserting it into existing systems. They come from redesigning the systems around the technology’s capabilities. The steam engine did not transform manufacturing because it was a better power source. It transformed manufacturing because factories were redesigned around centralised power — and then redesigned again around distributed electric motors, which required an entirely different factory layout, a different workflow, different skills, and different management structures.

The economic historians have documented this. Paul David, in his 1990 paper “The Dynamo and the Computer,” showed that it took approximately 40 years from the introduction of the electric dynamo to the realisation of its full productivity potential — because the complementary organisational innovations took that long to develop and diffuse. The factories had to be rebuilt. The workers had to be retrained. The management systems had to be reinvented.

The parallel to AI is direct. Most organisations have inserted AI into existing workflows. They have not redesigned the workflows around AI’s capabilities. The prediction engine is running. The decision architecture is untouched. The productivity statistics reflect this accurately: the technology is doing what it does, but the organisation has not changed what it does.

The Measurement Trap

There is a subtler problem that compounds the structural one: the way we measure productivity may be systematically blind to the value AI creates.

Productivity, in the national accounts, is output per unit of input. But what counts as output? If a marketing team uses AI to produce five drafts instead of one, and only one of those drafts is used, the productivity statistic sees the same output — one published document — produced with the same input. The AI’s contribution is invisible. The four unused drafts are not waste in the traditional sense; they are options. The team chose better because it could evaluate five alternatives instead of one. The quality of the decision improved. The quantity of the measured output did not.

This is the mismeasurement hypothesis that Brynjolfsson has been exploring since the 1990s. When technology improves quality, variety, or decision-making rather than quantity, the productivity statistics miss it. GDP measures transactions. It does not measure the quality of those transactions, or the decisions that preceded them, or the options that were evaluated and rejected.

The NBER study’s own measure — sales per employee — is revealing. Sales per employee captures volume. It does not capture whether the sales were better targeted, more profitable per unit, or required less post-sale support. A firm that uses AI to improve customer segmentation might sell the same volume to better-matched customers, reducing churn and increasing lifetime value — but sales per employee remains flat. The productivity statistic says: no impact. The income statement, eventually, says something different.

This does not mean the paradox is merely a measurement artefact. The NBER finding that 80 percent of firms see no impact is too broad to explain away with mismeasurement alone. Many firms genuinely are not capturing value from AI. But the measurement system’s blindness to quality improvements means that even the firms doing it well may not show up in the statistics — yet.

The Executive Expectation Gap

The NBER study surfaced another finding that deserves attention. Despite reporting no impact over the past three years, the same executives predict substantial AI effects over the next three years: a 1.4 percent boost to productivity, a 0.8 percent increase in output, and a 0.7 percent reduction in employment.

This is the expectation gap that Solow’s paradox feeds on. The technology is perpetually about to deliver. The returns are always three years away. The investment continues because the promise continues, and the promise continues because the technology demonstrably works at the task level — it writes the email, it drafts the report, it analyses the data faster. The micro-evidence of capability sustains the macro-expectation of transformation, even as the macro-evidence of transformation fails to materialise.

The PwC data sharpens this. The 12 percent of CEOs who report both cost and revenue gains from AI — the “vanguard” — are not using different technology. They are using the same models, the same tools, the same platforms. What distinguishes them, according to PwC, is that they have embedded AI extensively across products, services, demand generation, and strategic decision-making. They have established responsible AI frameworks. They have built technology environments that enable enterprise-wide integration.

In other words, they have done the system-level redesign. They have made the complementary investments. They have reorganised the decision architecture. They are not on the downward slope of the J-curve. They have invested through it and are beginning to climb.

The other 88 percent have bought the prediction engine and left the factory floor unchanged. They are experiencing the paradox not because the technology does not work, but because they have not done the organisational work that allows the technology to work at the system level.

The Organisational Bottleneck

Here is the sentence I want to hold, because it reframes the entire conversation: the machine is not the bottleneck. The organisation is.

The AI models are capable. The computational infrastructure is available. The tools are increasingly accessible. The technical barriers to AI adoption have fallen dramatically. A 200-person company can access the same language models, the same vision systems, the same analytical tools that a Fortune 500 company uses. The technology has been democratised.

What has not been democratised is the organisational capacity to absorb the technology. The capacity to redesign workflows. The capacity to retrain workers — not in how to use the tool, but in how to make decisions differently now that the tool provides better predictions. The capacity to restructure incentives so that the time AI saves is not simply filled with more of the same work. The capacity to build the intangible capital that the J-curve requires.

This is where the productivity paradox becomes an organisational psychology problem — my territory. The complementary investments that Brynjolfsson describes are not software purchases or hardware upgrades. They are changes in how people work, how decisions are made, how roles are defined, and how performance is measured. Every one of these is a human change. Every one encounters human resistance. Every one requires what Amy Edmondson calls psychological safety — the belief that you can try something new, fail, and not be punished for it.

An organisation that deploys AI without creating the conditions for experimentation — without making it safe to restructure, to change roles, to redefine what “productive” means — is an organisation that will remain on the downward slope of the J-curve. The technology will sit idle, like those covered machines on a factory floor. The old workbenches — the familiar processes, the known workflows, the comfortable routines — will continue to bear the marks of heavy use.

The Time Horizon Problem

The historical parallels offer both comfort and warning.

The comfort: the productivity paradox resolved itself before. The computer age did eventually show up in the statistics. The mid-1990s to early 2000s saw a surge in measured productivity growth that economists attributed, in part, to the maturation of IT investments made in the 1980s. The complementary innovations caught up. The J-curve climbed. Solow’s paradox dissolved — not because the observation was wrong, but because the time lag was longer than impatient executives and impatient economists expected.

The warning: the resolution was not automatic. It did not happen simply because time passed. It happened because organisations eventually restructured. They redesigned workflows. They retrained workers. They changed management practices. And the firms that did this first captured disproportionate returns, while the firms that waited — or never restructured at all — were left behind.

Paul David’s 40-year estimate for the electric dynamo is sobering. But the timeline for the computer age was shorter — roughly 15 to 20 years from widespread adoption to measured productivity gains. The question is whether AI’s timeline will be shorter still, or whether the complexity of the organisational changes required will extend it.

My reading of the evidence is cautious. AI requires deeper organisational restructuring than the personal computer did, because AI affects decisions, not just tasks. The PC automated typing. AI automates prediction — and prediction feeds into every decision the organisation makes. Restructuring around better typing was relatively simple: the same documents, produced faster. Restructuring around better prediction requires rethinking which decisions are made, by whom, and how. That is a more fundamental change, and fundamental changes take longer.

The Integration

Here is the tension I want to hold without collapsing, because collapsing it would be premature.

The productivity paradox is real. The data is unambiguous. Eighty percent of firms see no productivity impact from AI. Fifty-six percent of CEOs report no financial benefit. The macro-statistics show investment booming and productivity flat. The paradox is not a narrative. It is a measurement.

The productivity paradox is also, potentially, temporary. The J-curve framework is well-supported by historical evidence. The firms that have made complementary investments are capturing returns. The 12 percent vanguard in PwC’s data are not lucky — they are structurally different. They did the organisational work.

Both things are true simultaneously. AI is not yet delivering on its promise at scale. AI has the structural characteristics of a technology that eventually will. The question is not whether the paradox resolves. The question is who does the organisational work to resolve it — and who waits for a resolution that will not arrive on its own.

The machine works. The organisation has not changed. The productivity statistics are not wrong. They are reflecting, with uncomfortable accuracy, the gap between technological capability and organisational readiness.

Robert Solow could see the computer age everywhere. He could not see it in the productivity statistics. The statistics were not lying. They were telling a truth that executives did not want to hear: the technology is not the transformation. The transformation is the transformation. The technology is just the catalyst that makes the transformation possible — and whether the catalyst produces a reaction depends entirely on the conditions in the vessel.

The vessel is the organisation. The conditions are the culture, the incentives, the decision architecture, the willingness to restructure. Most vessels have added the catalyst and changed nothing else. The paradox is the predictable result.

The question, for any organisation reading this, is not “when will AI deliver?” It is “what have we changed to make delivery possible?” If the answer is “we bought the tools” — that is not a change. That is a purchase. And purchases, without the organisational redesign to absorb them, are how paradoxes are made.