The AI Productivity Curve

In 2024 US output per hour rose 3.0 percent, and in 2025 it rose 2.1 percent, against a late-1990s internet pace of about 2.5 percent a year. Walk up to the companies and almost none can point to where the new machines produced the gain. Productivity is that ratio: more output from the same hours.

The two distances

The popular claim is that the new machines are already raising output faster than the internet did. From far enough away the headline annuals look like they support it. The public chatbot arrived in late November 2022. Official labor-productivity annuals since then have been strong: about 2.0 percent in 2023, 3.0 percent in 2024 as later revised, 2.1 percent in 2025. The 2024 print is the revised official annual, not an early first look.

2025 was softer and bouncy quarter to quarter. One quarter printed 5.2 percent annualized. The neighbors were weaker. Into 2026 the bounce continued: the first quarter revised to 0.3 percent annualized, the second quarter’s first print 1.4 percent, 2.2 percent on a year-ago basis. Those later quarters date the series. They do not rewrite the annuals, and they do not close the bet.

The internet-boom benchmark is about 2.5 percent a year from 1995 through 2000, up from roughly 1.5 percent in the first half of that decade. So 2024 ran above the boom pace. Even the weaker 2025 sat in the same neighborhood, slightly under. On the headline number, matching or beating the internet boom still holds.

Ask the people running the firms and the picture splits. A survey of about six thousand chief executives and finance chiefs across four countries, fielded from late 2025 into early 2026, found 89 percent reporting no effect on labor productivity — sales per employee — over the previous three years. By the end of 2025 about nine in ten companies had the new tools running somewhere, and 94 percent said they were not seeing significant value.

The economy-wide number says something good is happening. The people inside the companies mostly shrug.

Both true at once

Both readings can sit on the table at the same time. That is what a large, slow, genuinely important technology looks like in its early years. The reading is standard for general-purpose tools. It is not proven for this one.

Early on, the expected picture is a strong official series, wide uptake, and most firms still spending before anything comes back. The “good national numbers” half is the half that has not been attributed.

What the lag looks like, and what the numbers are not

The gap has an old name in the form of a 1987 remark: the computer age visible everywhere except in the productivity statistics. Personal computers sat on desks for years before the payoff showed up. It took until the mid-1990s for the official series to move.

A powerful new tool costs before it pays. Early spending can make a firm look less productive for a while. The payoff arrives later, once the new way of working settles in. That J-shape is the mechanism.

One often-repeated figure is about 1.9 percent extra productivity since the public chatbot. That number is a trend-gap, not an estimate of what the tools contributed. From late 2022 through mid-2025, aggregate labor productivity rose 2.16 percent annualized against a 2015–19 trend of 1.43 percent — excess cumulative growth of 1.89 percentage points since the public release. The same shop’s time-savings model puts the tools themselves at about 1.1 to 1.3 percent. Keep both, labeled.

The pessimistic pole is a modest 0.5 percent over the next decade. That is a public rounding of total-factor-productivity gains, not of GDP. GDP in the same paper sits higher.

Nothing in the headline annual proves the new machines caused it. The strong 2024–25 print could be the economy shaking out after the pandemic — shedding inefficient arrangements, reshuffling workers. Official series cannot attribute. An industry-level correlation with time-savings cannot be read as causal.

Optimistic studies tend to look at cases where the tools clearly helped. Broad surveys of everyone are far less flattering. Red Teaming is a worked example of holding two opposing datasets without grabbing the convenient one.

The optimistic story is directionally fair and overconfident. The numbers are young. They wobbled lower in 2025. Most companies cannot feel them. Causation is unproven. That is the ordinary early shape of a general-purpose tool, counting year three from late 2022. Later strong or weak quarters do not turn the shape into a conclusion.

Treat the curve as a live bet. It has not been decided. The thing to watch is whether ordinary companies start reporting real, measurable gains.

The bet flips toward real if national productivity stays strong and the company surveys flip. It flips toward a mirage if the national series drifts back toward the old 1.5 percent pace while firms still cannot find the value.

This resizes the claim’s role in America’s Industrial Revival. The productivity boom is the optimistic cherry on top, not the foundation. The foundation is freight data. Lean lightly.

Related

• America’s Industrial Revival — where this claim first showed up as an amplifier; freight is the foundation, this is the cherry
• The AI Industrial Revolution — under-the-hood reason the tools could eventually move the numbers
• The Age Of Nonlinear Returns — if the late payoff arrives, this is what cashing it in looks like
• Red Teaming — holding two opposing datasets without grabbing the convenient one

Open questions

Is 2024 the new tools, or the post-pandemic shake-out?

How long is the lag — a couple of years, or a decade like the personal computer?

Which kinds of work move first?

Sources

U.S. Bureau of Labor Statistics. Productivity and Costs, annual averages: 2024 3.0%, 2025 2.1% (revised). Quarterly prints through 2026-Q2 used only to date the bounce. https://www.bls.gov/productivity/

Congressional Budget Office. Labor Productivity: Developments Since 1995 (March 2007). Late-1995 to 2001-Q1 averaged 2.5%; first half of the 1990s about 1.5%. Chicago Fed Letter 193 (2003) is the companion cite for the boom window.

Yotzov, I., Barrero, J. M., Bloom, N., et al. (2026). Firm Data on AI. NBER Working Paper 34836. Nearly 6,000 CEOs, CFOs, and senior finance managers in the US, UK, Germany, and Australia; fielded November 2025–January 2026. 89% report no impact on labor productivity over the past three years. https://www.nber.org/papers/w34836

McKinsey & Company. “Where AI will create value (and where it won’t)” (29 April 2026). End-2025: almost nine of ten companies had deployed AI; 94% report not seeing significant value. https://www.mckinsey.com/capabilities/strategy-and-corporate-finance/our-insights/where-ai-will-create-value-and-where-it-wont

Federal Reserve Bank of St. Louis. “The State of Generative AI Adoption in 2025” (13 November 2025). Excess cumulative labor-productivity growth of 1.89 percentage points versus the 2015–19 trend since ChatGPT’s public release — a trend-gap, not a contribution estimate. Time-savings model: about 1.1–1.3%. https://www.stlouisfed.org/on-the-economy/2025/nov/state-generative-ai-adoption-2025

Acemoglu, D. (2024). “The Simple Macroeconomics of AI.” Economic Policy. TFP gains over ten years upper-bounded around 0.55–0.71%; public rounding “0.5 percent in 10 years.” https://economics.mit.edu/sites/default/files/2024-04/The%20Simple%20Macroeconomics%20of%20AI.pdf

Solow, R. (12 July 1987). New York Times Book Review. The computer age visible everywhere except in the productivity statistics. David, P. A. (1990). “The Dynamo and the Computer.” American Economic Review. Brynjolfsson, E., Rock, D., & Syverson, C. (2017, 2021). AI-era restatement and the productivity J-curve.

Maxinomics. Americans Are About to Get a Lot Richer (YouTube, 30 May 2026). Originating popular claim that AI is already outrunning the internet. Treated here as a hypothesis, not a finding.