I still remember November 2022.
Rather than calculating valuations, I was drawing countless lines to catch the bottom on charts. After months of relentless declines, something in the market was beginning to change.
And before the month was over, another change was about to begin—one that would eventually reach far beyond the stock market.On November 30, OpenAI released ChatGPT.
At the time, few could know how quickly generative AI would move from a technological curiosity to a force reshaping corporate investment, labor, productivity and eventually the debate over the U.S. economy itself or global economies to be extended.
For me, that raised a question I have been watching ever since: if AI was truly a technological revolution, when would it begin to appear in the productivity data?
Nearly four years later, productivity has become one of the most frequently repeated explanations for why the U.S. economy has remained more resilient than many expected.
Executives, economists and investors increasingly speak of an AI-driven productivity boom. But repetition is not evidence.
Before accepting that narrative, I wanted to verify it - How much has U.S. productivity actually improved?
Is the acceleration large enough to be economically meaningful? Is it showing up in output, labor hours and unit labor costs? And perhaps most importantly, is the United States beginning to produce more without requiring the same rate of employment creation?
That is the purpose of this research.
The first question is therefore not whether AI is changing the U.S. economy.
It is whether the productivity data themselves have changed enough since late 2022 to justify asking that question in the first place.
1. Something Has Changed in the Productivity Data
The first finding is relatively clear. U.S. labor productivity has strengthened materially.
According to the Bureau of Labor Statistics, nonfarm business labor productivity has grown at an annualized rate of 2.1% during the current business cycle from Q4 2019 through Q2 2026.
That compares with only 1.5% during the previous business cycle from Q4 2007 through Q4 2019.
The current rate is also equal to the long-run average of 2.1% since 1947.
Source: U.S. Bureau of Labor Statistics, Productivity and Costs, Q2 2026
The change becomes even more interesting when we focus on the period around late 2022.
The BLS productivity index stood at roughly 110.0 in Q4 2022. By Q2 2026 it had risen to about 120.0. That represents an increase of roughly 9% in output per hour over three and a half years. The timing is striking.
November 2022 should therefore be treated as an observation point, not as proof that ChatGPT caused the change.
What the data do show is that productivity growth after late 2022 has been substantially stronger than the low-productivity environment that characterized much of the post-global-financial-crisis period.
2. The Most Important Number May Not Be Productivity Itself
The most interesting part of the data is not simply that productivity rose.
Between Q4 2022 and Q2 2026, nonfarm business output increased by roughly 10%. Hours worked increased by less than 1%.
The gap between those two numbers is essentially the productivity story.
The economy produced considerably more output without requiring a comparable increase in labor input.
That is exactly the type of pattern we would expect to see if an economy were becoming more efficient.
The latest quarterly data continue to show the same basic relationship.
In Q2 2026, nonfarm business output increased at a 1.7% annualized rate. Hours worked increased just 0.3%. Productivity therefore increased 1.4%. Compared with a year earlier, productivity was up 2.2%.
Source: U.S. Bureau of Labor Statistics
This is important because economic growth normally requires some combination of more workers, more hours, more capital or better efficiency.
If output can rise while labor hours barely increase, productivity is doing more of the work.
That does not tell us why productivity improved. But it tells us that the improvement is real.
3. 2024 Was Particularly Strong — and the Decomposition
The annual productivity data provide another clue.
Private nonfarm business labor productivity increased 3.0% in 2024. That was followed by a still-solid 2.2% increase in 2025.
By comparison, productivity grew at an average annual rate of only 1.5% between 2007 and 2019.
Source: U.S. Bureau of Labor Statistics, Total Factor Productivity 2025
The decomposition is even more interesting.
BLS estimates that the 3.0% increase in labor productivity during 2024 reflected approximately:
1.1 percentage points from capital intensity,
0.3 percentage point from labor composition,
and 1.5 percentage points from total factor productivity.
In 2025, labor productivity growth slowed to 2.2%.
Capital intensity contributed 0.9 percentage point.
Labor composition contributed 0.4 percentage point.
TFP contributed 0.8 percentage point.
This matters because the productivity acceleration cannot be explained purely by companies adding more workers. It cannot even be explained entirely by simply adding more capital per worker.
In 2024, measured TFP itself made the largest contribution.
That means the economy appeared to become better at combining capital and labor to produce output. This is the kind of improvement that could eventually be associated with a general-purpose technology.
But again, the data do not identify AI as the cause.
Capital deepening unrelated to generative AI, post-pandemic normalization, organizational changes, older automation technologies, sector composition and business-cycle effects may all be contributing.
The correct conclusion is therefore narrower. The productivity acceleration is real.
4. Unit Labor Costs Are Sending an Equally Important Signal
BLS defines unit labor costs as hourly compensation relative to labor productivity.
If compensation rises 4% but productivity rises 3%, the labor cost associated with producing one unit of output rises much less than the wage itself.
That can allow wages to increase without creating the same inflation pressure that would occur in a low-productivity economy.
This is one of the most important channels through which productivity can change the inflation outlook.
The contrast with 2022 is striking. In 2022, productivity was weak while compensation was rising rapidly, producing substantial unit labor cost pressure.
By Q2 2026, the relationship looked very different.
Hourly compensation increased at a 2.7% annualized rate. Productivity increased 1.4%.
As a result, unit labor costs increased only 1.3%. Over the previous four quarters, unit labor costs increased 1.4%.
Source: U.S. Bureau of Labor Statistics
This does not mean AI has solved inflation. Nor does it mean wages no longer matter. But productivity changes the arithmetic.
A higher-productivity economy can absorb faster nominal wage growth without generating the same increase in production costs.
This is why the productivity question eventually becomes a Federal Reserve question.
5. The Labor Market Is Weakening in an Unusual Way
The employment data make the productivity story more interesting.
Payroll growth has slowed sharply. In July 2026, U.S. nonfarm payroll employment declined by 23,000. More importantly, the average monthly increase over the previous 12 months was only 34,000.
The unemployment rate, however, remained relatively low at 4.1%.
This is not what a conventional recessionary labor market normally looks like.
If the economy were experiencing a severe contraction in labor demand, we would generally expect layoffs and unemployment claims to rise substantially.
That has not happened.
For the week ending August 8, 2026, initial jobless claims were only 209,000.
JOLTS tells a similar story.
In June 2026, job openings stood at 7.4 million. Hires were 5.3 million. The hires rate was 3.4%. Layoffs and discharges remained relatively low at 1.1%.
Source: U.S. Bureau of Labor Statistics, JOLTS, June 2026
The combination is unusual. Hiring has weakened significantly. Payroll creation has slowed dramatically. But layoffs remain contained.
The labor market increasingly looks less like a market in which companies are aggressively firing workers and more like one in which companies are becoming reluctant to add new workers.
That distinction may prove extremely important.
6. What If the First AI Labor-Market Effect Is Not Mass Layoffs?
Much of the public debate about AI and employment has focused on job destruction.
But that may not be the first place an AI productivity effect appears. The firm does not necessarily need to fire employees.
Instead, when demand grows, it may simply hire fewer additional workers.
The sequence could look like this:
AI or automation adoption → existing workers produce more → incremental labor requirement declines → vacancies decline → hiring slows → payroll growth weakens while layoffs remain relatively low.
That would produce exactly the type of labor-market pattern we are currently observing.
But there is a major problem with this interpretation. We do not have strong evidence that AI is causing the hiring slowdown yet.
In fact, recent Federal Reserve research explicitly cautions against making that claim.
A March 2026 Federal Reserve study comparing AI adoption with firm and industry job postings found no evidence that higher AI adoption had reduced overall job postings.
The researchers concluded that the post-pandemic slowdown in national job postings does not appear to have been driven, even modestly, by AI so far.
This is critical evidence against an overly aggressive AI narrative. The labor-market pattern is consistent with what AI-driven productivity could eventually produce.
For now, we should say employment creation is weakening much faster than layoffs are rising. We should not yet say AI caused employment creation to weaken.
7. At the Worker Level, However, AI Productivity Effects Are Already Real
The aggregate evidence remains inconclusive.
The microeconomic evidence is stronger.
One of the best-known studies comes from Erik Brynjolfsson, Danielle Li and Lindsey Raymond.
Using data from more than 5,000 customer-support agents, the researchers found that access to a generative-AI assistant increased productivity by approximately 14% on average.
The gains were much larger among novice and lower-performing workers.
Other experiments involving writing, coding and professional tasks have also found meaningful productivity improvements.
These studies establish something important.
Generative AI can increase worker productivity in specific tasks and environments.
What they do not establish is the size of AI’s current contribution to aggregate U.S. productivity.
That aggregation problem is central.
A technology can deliver a 10%, 20% or even larger productivity improvement for workers performing specific tasks while having a much smaller effect on national productivity if adoption remains limited or if those tasks represent only a small share of economic activity.
8. AI Adoption Is Growing Rapidly — But It Is Not Yet Universal
Census Bureau data provide useful context.
Between December 2025 and May 2026, roughly 17% to 20% of U.S. businesses reported using AI. Around 20% to 23% expected to use it within the following six months.
But adoption differs dramatically by company size and industry.
A 2026 Census working paper found that around 18% of firms used AI in at least one business function during the survey period. When weighted by employment, adoption rose to 32%.
This provides a plausible explanation for why aggregate productivity effects may still be difficult to identify. AI adoption is substantial and growing rapidly.
History suggests that general-purpose technologies often require complementary investment, organizational redesign and time before their full productivity effects become visible in aggregate statistics.
The Federal Reserve made exactly this point in a July 2026 research note examining the AI buildout.
9. Productivity Is a Slow Indicator of a Fast-Moving Technology
This is one reason I have continued watching productivity data since late 2022. Productivity is economically fundamental.
But it is not a real-time indicator. It is reported quarterly. It depends on estimates of output and labor hours. It is revised.
By the time a structural productivity acceleration becomes statistically obvious, companies and labor markets may already have been adapting for years.
How do we recognize a productivity transition before the official productivity statistics fully confirm it?
There may not be one perfect leading indicator. Initial claims alone are insufficient. AI productivity does not require mass layoffs.
A better real-time dashboard may combine: Initial claims, continuing claims, payroll growth, JOLTS hires, job openings, aggregate labor hours, and eventually productivity itself.
The signal to watch may not be an explosion in unemployment.
It may instead be the widening gap between output growth and labor-input growth.
10. There Is an Important Alternative Explanation
Before attributing this change to AI, several competing explanations have to be taken seriously. The first is post-pandemic normalization.
COVID created enormous distortions in labor composition, hours worked, supply chains and business operations.
Some of the post-2022 productivity improvement may simply represent normalization from unusually weak 2022 conditions.
If lower-productivity workers leave employment or hiring becomes concentrated in higher-productivity sectors, measured aggregate productivity can rise without individual workers becoming more productive.
And Cloud computing, automation, enterprise software, robotics and digitalization were already transforming businesses.
Companies often reduce hiring faster than output during periods of economic uncertainty, temporarily raising measured productivity.
And finally, a shift toward high-value-added technology and services can raise aggregate productivity independently of generative AI. These explanations are not mutually exclusive.
Then was productivity caused by AI?
How much of the productivity acceleration is being amplified by AI on top of a broader capital-deepening cycle?
At present, the data do not allow us to answer that precisely.
11. AI May Be Deflationary and Inflationary at the Same Time
There is another complication. AI is not purely a deflationary technology. On the operating side, the mechanism is straightforward.
But building the AI infrastructure itself requires enormous amounts of physical capital. The investment boom therefore creates a second mechanism which can be inflationary, particularly during the buildout phase.
That leads to one of the more interesting macroeconomic questions of this cycle: is AI simultaneously a deflationary productivity shock and an inflationary investment boom?
The answer may be yes.
And the relative strength of those two channels could change over time. The investment shock may arrive first and the productivity dividend may arrive later.
12. This Is Where the Federal Reserve Problem Begins
The Federal Reserve ultimately cares about the sustainable speed at which the U.S. economy can grow without generating inflation.
That sustainable speed depends partly on productivity.
Suppose potential output growth is lower than policymakers assume.Then strong GDP growth would rapidly tighten labor markets and increase inflation pressure.
But suppose productivity has structurally accelerated. Then the economy could grow faster while requiring fewer additional labor hours and producing less unit-labor-cost pressure.
In that world a strong GDP does not automatically mean overheating. This distinction may become increasingly important.
Real GDP increased at an annualized rate of 1.5% in Q2 2026 after 2.1% in Q1. At the same time, payroll growth has weakened sharply.
Productivity remains above the post-2007 trend. Unit labor cost growth has moderated substantially. That combination complicates conventional interpretations of the economy.
If they focus primarily on slowing payroll growth, they may underestimate the economy’s capacity to continue producing output.
The missing variable may increasingly be productivity.
13. Could the Fed Be Underestimating U.S. Potential Output?
This remains a hypothesis, not a conclusion.
But it deserves serious consideration.
If the U.S. economy can sustainably produce more output with slower growth in labor hours, then potential output may be rising faster than models based heavily on historical productivity trends would suggest.
That would have significant consequences.
It could mean stronger real GDP growth is compatible with lower inflation. It could mean labor-market cooling does not necessarily imply recession meaning the equilibrium relationship between GDP, employment and inflation is changing.
And it could mean the data dependent Fed faces a difficult measurement problem. Potential output cannot be observed directly.
If a general-purpose technology is changing productivity faster than historical models can capture, policy could temporarily be calibrated to an economy that no longer exists in quite the same form.
This is why the AI productivity question is not merely a technology story.
It is a monetary-policy story.
Conclusion: The Productivity Shift Is Real. The AI Attribution Is Not Yet Proven.
Nearly four years after November 2022, the data finally allow us to say something more confidently.
U.S. productivity has improved. The improvement is economically meaningful.
The post-2007 productivity regime has, at least so far, given way to stronger output-per-hour growth. Perhaps most importantly, output has increased significantly while labor hours have risen only modestly.
Unit labor cost pressure has fallen sharply from 2022 levels. Meanwhile, employment creation has weakened far more dramatically than layoffs have increased.
Those are real observations.
What we cannot yet prove is that generative AI caused them. Firm-level studies show substantial productivity gains from AI. Business adoption is spreading rapidly. Capital investment in AI infrastructure is enormous.
The timing is consistent with the beginning of a broader productivity transition.
But Federal Reserve research has not found evidence that AI adoption explains the aggregate slowdown in job postings, and the macroeconomic productivity data remain influenced by many forces besides AI.
The appropriate conclusion is therefore neither skepticism nor certainty. It is observation.
The productivity acceleration is real enough that the AI hypothesis can no longer be dismissed. But the causal evidence is not strong enough to declare victory.
That makes the next few years unusually important.
If output continues rising while labor-hour growth remains weak, productivity remains above the pre-pandemic trend, unit labor costs stay contained and AI adoption continues spreading, the case for a structural change in the U.S. economy will become increasingly difficult to ignore.
And if that happens, one of the most important assumptions in macroeconomics may need to change.
The U.S. economy may be capable of growing faster without generating the same amount of employment growth or inflation pressure that similar GDP growth would have produced in the past.
The question I began watching in late 2022 was simple:
If AI was truly a technological revolution, when would it begin to appear in the productivity data?
The answer today is more interesting than I expected. We may already be seeing the first signs.
What we do not yet know is how large they will become.