The Solow Paradox Returns, and This Time It's Speaking to AI
There is a silent pattern that economic history has repeated at least twice with clarity before the era of artificial intelligence. First with industrial electrification, then with personal computers. In both cases, the technology arrived decades before its impact appeared in productivity statistics. In both cases, the period of "nothing is happening" was precisely the moment when everything was being reconfigured underneath.
Economist Robert Solow captured it with a phrase that was never designed to make anyone laugh: "You can see the computer age everywhere except in the productivity statistics." It was 1987. PCs were proliferating in corporate offices, mainframes were processing transactions at speeds unthinkable a decade earlier, and the embryo of what would become the internet already existed. Yet the aggregate productivity of the American economy did not move. That phenomenon was recorded as the Solow Paradox, and its resolution took nearly ten years to arrive.
What is happening today with artificial intelligence has an almost identical geometry. And the accumulation of recent data — from large-scale surveys to reports from major technology platforms — suggests that the inflection point that took a decade to arrive for computers could be materializing right now for AI.
---
When 90% Say "Nothing Changed" and the Market Says Otherwise
In February of this year, a survey administered to 6,000 business leaders delivered a result that, at first glance, would seem devastating to the arguments of those who have spent years promising the AI revolution: 90% of respondents reported that the adoption of artificial intelligence had had no measurable impact on employment or productivity in their companies. At the same time, 63% declared they had adopted AI in some form.
That is exactly the portrait of 1987. A technology omnipresent in discourse, adopted by the majority, but with no visible footprint in the real economy as measured by conventional instruments.
But there is another number in the same picture that changes the framing. An analysis by the Federal Reserve Bank of Saint Louis found that generative AI produced a 5.4% improvement in the productivity of workers who used it. It is not a figure that justifies the current valuations of AI companies. Nor is it negligible. It is, in historical terms, the kind of weak signal that typically precedes a deeper structural movement.
The distance between the 90% who see no change and the 5.4% that actually measures improvement is not a contradiction. It is the difference between adopting a tool and redesigning the entire process around it. The nineteenth-century factories that installed electric motors on top of the same systems of steam-powered shafts and pulleys did not obtain efficiency gains. Those that demolished the physical architecture of their plants and built from scratch around the individual motor at each workstation did obtain them — but that process took forty years from the moment Edison lit his first generating plant in 1882.
The Deloitte analysis on generative AI adoption adds another fragment to the puzzle: the majority of companies that adopted AI report a positive return, and nearly 25% of adopters report productivity or financial gains exceeding 30%. That quarter of companies is not operating with different tools from the remaining 75%. It is operating with a different organizational logic — which is exactly the type of variable that does not appear in technology adoption surveys but that determines where value will concentrate over the next five years.
---
What the Results of Alphabet and Microsoft Reveal That the Survey Cannot See
While most companies report zero impact, two companies with dominant positions in technological infrastructure are publishing numbers that do not fit that narrative. And the difference is not that they have access to better AI, but that they control the distribution channel through which millions of organizations access it.
Alphabet reported in its most recent quarter a 19% growth in search revenues, attributing part of that increase directly to the integration of AI into its primary search product. Its Google Cloud division grew 63% year over year, and the company noted that large-scale enterprise customers who adopted its AI services generated revenues with growth of 800% compared to the previous year. That last number is not an indicator of absolute volume, but it is a signal of the speed of adoption among the corporate segment that historically takes the longest to move.
Microsoft, for its part, reported that its AI business is currently operating at an annualized revenue run rate of 37 billion dollars. To contextualize that figure: OpenAI, the company that captures the most media coverage in the AI space and that operates with an annualized revenue of around 20 billion dollars, remains smaller in scale than Microsoft's AI segment alone.
The pattern that emerges is not that of a failed technology awaiting validation. It is that of a technology whose economic value capture is concentrating, for now, in the platforms that control the infrastructure and distribution channels to the enterprise customer: Alphabet, Microsoft, and to a lesser extent Salesforce, ServiceNow, and Databricks, which also reported growing monetization of their integrated AI capabilities.
This faithfully replicates what happened in the 1990s with computing. Intel, Microsoft, Cisco, and telecommunications operators captured the majority of the economic value of the digital revolution long before the impact of that revolution was visible in aggregate productivity statistics. Companies that were users of that technology took years longer to translate their investment into real operational gains.
---
The Lag That Nobody Measures Is in the Organizational Architecture
There is a specific friction that explains why the distance between adoption and productivity does not automatically collapse, and that friction rarely appears in market analyses. It is the speed of organizational redesign, which is orders of magnitude slower than the speed of technological adoption.
When a company installs a generative AI tool into the workflow of its content team or customer service operation, the initial gain is marginal. The worker learns to use the tool, but the process within which that tool operates still has the same bottlenecks, the same layers of approval, the same role design that existed before AI. The 5.4% improvement measured by the Federal Reserve Bank is, to a large extent, the impact of the tool upon the existing process.
The leap that converted electrification from a technical data point into a productivity revolution was not the installation of the motor. It was the elimination of the central drive shaft and the decentralized distribution of energy throughout the plant — which involved physically demolishing the previous infrastructure and rebuilding it. The AI equivalent is not "implementing a copilot." It is redesigning which processes exist, which ones disappear, which roles make sense, and which decisions can be made without direct human intervention.
The companies in the top quartile of the Deloitte analysis — the 25% reporting gains exceeding 30% — are doing something different from installing tools. They are redesigning entire workflows around capabilities that previously did not exist. That is an operation that requires tolerance for transitory chaos, willingness to abandon processes that were working, and above all an honest reading of what the end customer actually values and which parts of the internal process generate no value for anyone except the person who designed them.
That redesign is slow, politically costly within organizations, and difficult to measure in the short term. That is precisely why it does not appear in the survey of 6,000 business leaders as visible impact. But it is exactly what, when it reaches critical mass in enough sectors and companies, produces the kind of movement in productivity statistics that economists describe retrospectively as an inflection point.
---
What the Solow Paradox Cannot Resolve on Its Own
The historical analogy has analytical value, but it also has a limit that is worth naming with precision. The latency period between technological adoption and measurable productivity in the 1980s and 1990s occurred in a context of slower technological iteration. The language models that exist today will be primitive versions of those that will exist within three years. The competitive pressure on companies to adopt and redesign processes is more intense now than what organizations faced during the transition to the PC.
This does not mechanically shorten the period of organizational lag, because that lag depends on human and institutional factors that do not accelerate at the same rate as the technology. But it does mean that the distribution of benefits between companies that redesign and those that install without redesigning will become visible on balance sheets with greater speed than it did during the PC revolution.
The 25% of adopters with gains exceeding 30% reported by Deloitte is not a statistical curiosity. It is the first evidence that the separation between both groups is already occurring. If the historical pattern holds, that differential will widen before macroeconomic statistics register it clearly. By the time productivity indices show the leap that Solow had been waiting to see since 1987, the competitive advantage of those who redesigned rather than simply adopted will already be structurally difficult to recover.
The question that the Solow Paradox always leaves unresolved is the same: how much time does an organization have to move from being a user of the tool to being a designer of the processes that the tool makes possible? In the 1990s, that margin was almost a decade. This time, the geometry of the market suggests it will be considerably shorter.
