A similar discussion has been taking place in government offices, corporate boardrooms, and economics departments for a number of years: AI is obviously doing something remarkable in the lab, in the demo, and in the meticulously staged product announcement, but where exactly is it in the numbers? The expenditures are huge. The zeal is unrelenting. For the majority of that time, the productivity increases were either negligible or undetectable. Then, in early 2026, some information surfaced that altered the discussion, or at the very least, provided one side with a lot more information.
In a Financial Times opinion piece, Erik Brynjolfsson, who oversees Stanford’s Digital Economy Lab and has spent years creating what he refers to as the J-curve theory of technology adoption, made the case that the waiting is essentially over. According to his revised analysis, US productivity increased by about 2.7% in 2025, which is almost twice as much as the 1.4% annual average of the preceding ten years. It’s not a rounding error. If the underlying data holds up, which there are legitimate doubts about, that’s a significant change. However, the movement’s direction and timing align with the pattern Brynjolfsson has been forecasting for years: general-purpose technologies like artificial intelligence (AI) typically result in lengthy delays between initial investment and quantifiable output because companies need time to reorganize workflows, retrain employees, and rebuild processes around the new capability before the gains actually appear on a government spreadsheet.
AI Productivity Dividend — Data, Analysts & Economic Context (2025–2026)
| Key analyst | Erik Brynjolfsson — Director, Stanford Digital Economy Lab; co-founder of Workhelix; author of the “J-curve” productivity theory for general-purpose technologies |
| US productivity growth (est. 2025) | ~2.7% — nearly double the 1.4% annual average of the prior decade, per Brynjolfsson’s updated analysis |
| BLS payroll revision | Bureau of Labor Statistics benchmark revisions reduced previously reported payroll growth by approximately 403,000 jobs — even as GDP remained strong |
| GDP growth (Q4 2025) | 3.7% real GDP growth — strong output alongside near-flat job creation of ~15,000 jobs/month in 2025 |
| The J-curve pattern | Named for historical lag seen with past general-purpose technologies (steam engine, computer) — gains trail investment by years as businesses reorganize and retrain workers before measurable output improvements emerge |
| Harvest phase | Brynjolfsson’s term for the shift from AI investment/experimentation into measurable economic return — he argues 2025 data suggests the US has entered this phase |
| Labor market signal | Entry-level hiring in AI-exposed sectors fell ~16%; workers using AI tools to augment skills saw stronger demand — early evidence of structural labor market adjustment |
| Skeptical view | The Economist (Feb 22, 2026): AI improving fast but effect on output “not so much” — productivity data volatile and boom not yet confirmed |
| Corporate survey finding | Fortune / CEPR research (Feb–Mar 2026): nearly 90% of firms said AI had no impact on employment or productivity over the past three years — though most expected future gains |
| Policy stakes | Kevin Warsh, Trump’s Fed nominee, is counting on AI-driven productivity to help contain inflation; Treasury Secretary Scott Bessent predicted AI would soon start “biting” into output figures |
| SAP / agentic AI outlook | SAP forecasts Agentic AI will drive further productivity and GDP gains in developed economies through 2026 and beyond — scale varying by region and sector |
The same cycle was experienced by the steam engine. The computer did the same. It’s easy to forget now, but serious economists in the 1980s and 1990s were genuinely perplexed that decades of computer adoption hadn’t resulted in noticeable productivity gains. Robert Solow’s dry observation that computers appeared everywhere but in productivity statistics gave rise to the term “Solow Paradox.” After years of quiet organizational change, a productivity boom that no one could have predicted in terms of timing or scope finally occurred. According to Brynjolfsson, AI is following a similar trajectory, and the harvest phase—his term for the point at which earlier investment starts producing quantifiable returns—started in earnest in 2025.
Even though real GDP continued to grow, including 3.7% in the fourth quarter of 2025, the Bureau of Labor Statistics significantly strengthened that argument by revising its payroll data downward by about 403,000 jobs. By definition, higher productivity results from a combination of reduced labor input and steady or increasing output. Rather than being a clear indication of AI efficiency, the revision might be partially a measurement artifact. Revisions to data occur for a variety of reasons. However, the evidence’s trend—lower headcount growth, higher economic output—is in line with what would be expected if AI technologies were actually enabling businesses to produce more with fewer workers.
There’s something subtly startling about the details when you look at the labor market data alongside this. While demand for workers who actively use AI tools to expand their capabilities increased, hiring for entry-level positions in AI-exposed industries fell by about 16%. That is not a minor or transient variation. It appears to be the beginning of a structural change, the transition from “AI as a tool some people use” to “AI as a baseline expectation woven into how entire categories of work get organized.” Businesses aren’t just experimenting at the margins anymore. At the very least, some of them are rebuilding around the technology in ways that manifest in their hiring practices and personnel requirements.
Nevertheless, there is an important counter-narrative that merits careful consideration. In a February 2026 article, The Economist contended that although AI was developing quickly, its quantifiable impact on economic output was still small and that the productivity boom was definitely not here yet. Nearly 90% of businesses said AI had no appreciable effect on employment or productivity over the preceding three years, according to a comprehensive survey cited by Fortune. These are neither technophobes nor fringe voices. These are businesses that are in operation and share their personal experiences. It’s quite possible that productivity gains are concentrated in a relatively small subset of industries or firm types, resulting in aggregate statistics that appear promising but don’t accurately represent what the majority of businesses are actually experiencing on the ground.
The truth is that both can be true simultaneously, but no one finds this to be very satisfying. While most businesses may feel that AI hasn’t changed much yet, the early signal in the data may be genuine. This scale of transition does not occur consistently. It frequently takes years for the lagging sectors to catch up enough to appear in the headline figures. They first arrive in pockets and then in waves. The results of 2025 may not provide a complete picture of how AI is affecting the economy, but they do provide a clear first look. The question is whether the trend will continue in 2026 or if the revision is altered and the discussion is restarted.
