In the past few years, US labour productivity growth has been strong relative to its pace in recent decades (Abdelrahman and Foerster 2026). Figure 1 shows the growth rate of labour productivity, measured as business-sector output per hour worked, for selected periods through the first quarter of 2026. Growth rates are expressed at an annual rate. 

From 2005 through 2019, productivity growth averaged about 1.5% per year. The pace remained similarly slow during and immediately after the pandemic (2020-2022), despite wild swings as the economy shut down and reopened (Fernald and Li 2022).  

The post-2022 surge stands out against these earlier baselines. The dashed line shows the average productivity growth rate of 2.5% from the beginning of 2023 through the first quarter of 2026. Looking year-by-year, growth reached about 3¼% in 2023. It then eased to about 2½% in 2024 and 2% in 2025–26 – still faster than the pre-2023 period. 

To trace the sources of the surge, we use growth accounting to split productivity growth into the contributions of different economic drivers, shown by the coloured bars in the figure. Labour composition (orange) captures changes in worker skills, such as education and experience. Capital deepening (green) is growth in the equipment, structures, and intellectual capital available for each (composition-adjusted) hour of work. The final factor (blue) is total factor productivity (TFP), defined as the residual that absorbs whatever capital deepening and labor composition cannot explain. 

Most of the acceleration in labour productivity growth comes from faster growth of TFP. Comparing the entire 2023-26 period to the pre-pandemic period, TFP picked up by about 0.8 percentage points; the contribution of capital deepening rose by 0.3 percentage points; and labour composition was little changed. Investment patterns help explain the modest role of capital deepening. Optimism about AI lifted investment in computing equipment (Kalyani and Li 2026). Yet investment elsewhere, including office buildings, retreated, damping the growth of total capital. 

Figure 1 Contributions to growth in US output per hour

Business sector, percent change, annual rate

Source: Fernald (2014), updated. 
Note: Each bar is labelled by the calendar year(s) whose Q4-over-Q4 growth it contains — e.g., 2005–19 averages quarterly (log) growth from the beginning of 2005 to the end of 2019. 2025–26* is the average annual growth rate over the five quarters ending 2026Q1 (the latest data). Dashed line: average labour-productivity growth, 2023–26. Capital deepening is the contribution of capital relative to quality-adjusted hours; total factor productivity is measured as a residual. 

AI and TFP: Technical efficiency or more intensive input usage? 

In the long run, the most important reason why TFP rises is technological efficiency, or innovation, broadly construed. New products, improved methods, or reallocation toward more productive firms let the economy produce more from the same inputs.

AI is a plausible new source of such gains. Generative tools can draft text, write code, and summarize documents, letting workers produce more from the same hours. The timing seems to fit. The acceleration began around ChatGPT’s release at the end of 2022. A natural conjecture, then, is that AI raised the speed of technological progress.

Yet surveys of firms often find only small productivity gains so far from adopting AI (e.g. Yotzov et al. 2026). Though not focused on AI per se, the regime-switching model in Kahn and Rich (2026) also finds limited likelihood that the productivity trend has so far picked up.  Turning new tools into output gains takes changes in workflows, organisation, and skills. Those changes take time. 

At the same time, higher measured TFP need not signal greater efficiency, especially in the short run. TFP is a residual – the labour productivity growth that measured labour composition and capital deepening cannot explain. As such, it varies with the intensity of input use. 

Higher input utilisation lets firms meet a demand surge even without efficiency gains. If firms can’t (or choose not to) immediately hire or add capital, they can push existing employees to work harder or run plant and equipment longer. Neither margin shows up in measured hours or capital, so the extra output lands in the TFP residual even though no true innovation has occurred.

Two features of AI may raise utilisation, pushing up measured TFP and labour productivity.

The first is demand. AI has spurred investment in data centres, chips, and power. In addition, an AI-driven surge in stock prices has likely stimulated consumption. Since AI’s efficiency gains have probably not yet arrived, meeting that demand requires more inputs.

The second is uncertainty, which shapes how firms choose to obtain those inputs. AI holds promise, but it clouds how production should be organised. Outside the AI buildout itself, firms hesitate to make hard-to-reverse investments and hold off on hiring or shedding workers because the right workforce size and mix remain unclear. 

On its own, an increase in uncertainty tends to reduce spending and hence utilisation (Leduc and Liu 2016, Amberg et al. 2026). But paired with strong demand, it squeezes firms. They face more orders than their frozen input plans can meet at normal intensity. So they work existing inputs harder. The extra intensity appears in the data as faster measured productivity growth without faster technological progress. We find this channel strong enough to account for the entire acceleration.

Measuring utilisation

Input utilisation is hard to measure. It is not observed directly for the entire economy, so it must be inferred. We use the Fernald (2014) utilisation measure which, in turn, follows Basu et al. (2006). 

To see the intuition, consider a firm that wants to meet strong demand but, for now, has a given capital stock, workforce, and technology. It can ask its existing workers to work longer (which we observe) and harder. It can also run its capital longer. Each margin is costly, for example through overtime pay, so firms typically use all of them at once. The observed margin serves as a proxy for the unobserved ones. 

The method infers changes in industry utilisation from observed movements in hours per worker, scaled by an estimated pass-through of those movements into labour productivity. Industry estimates are aggregated using Domar weights, which reflect each industry’s importance.

Utilisation explains most of the productivity surge

The stacked bars in Figure 2 show measured TFP growth split into the Fernald (2014) measure of utilization growth in light grey and utilisation-adjusted TFP growth in dark grey. In 2023, utilisation growth was a drag on measured TFP growth and utilisation-adjusted TFP growth soared. Since the beginning of 2024, however, utilisation accounts for essentially all TFP growth, leaving little growth in utilisation-adjusted TFP. 

Figure 2 Utilisation and utilisation-adjusted TFP

Source: Fernald (2014), updated. 
Note: See notes to Figure 1. For each bar, the contribution of utilization and utilization-adjusted TFP sums to the residual TFP growth in Figure 1. 

Superficially, aggregate weekly hours appear to contradict the utilisation estimate. Average weekly hours in private industries declined in 2024 and began rising only in mid-2025. Two features of the utilisation measure resolve the tension. First, industry hours are filtered to isolate cyclical movements rather than long-run trends. Second, aggregation relies on pass-through and output weights rather than employment shares. Hence aggregate utilisation can climb even while economy-wide average hours stay flat.

The utilisation measure is course, and productivity statistics are volatile and subject to revision as new data become available. Hence, any conclusions about the recent past are necessarily tentative.

Growth and inflation implications of higher utilisation 

Utilisation gains are inherently temporary, since intensity of input use can rise only so far. Hence, utilization growth alone cannot sustain elevated measured productivity growth. Past episodes of elevated utilisation, such as recoveries from recessions, unwound as hiring and investment caught up with demand. 

Rising utilisation is also costly. More intense input use raises overtime pay, for example, which businesses may pass through to prices. Thus, whether the recent productivity gains reflect genuine technological improvements or rising utilisation matters for monetary policy. Technological improvements allow faster non-inflationary output growth. Intensity-driven gains do not. Hence the recent figures may overstate the economy’s non-inflationary speed limit.

Conclusion

Intensity rather than efficiency may be driving the recent US productivity acceleration. Higher utilisation can account for essentially all of the exceptional rise in productivity growth over the past two years without any acceleration in underlying technological progress. AI could have mattered indirectly, by raising demand and uncertainty and pushing firms to work existing inputs harder rather than adjust them. That channel raises costs as it raises output. For now, the measured productivity gains appear to derive from “working harder” rather than “working smarter”.

Still, much is unknown. Explosive growth in the future is possible even if near-term productivity gains from AI are small (Jones and Tonetti 2026). Firms themselves expect large gains over the next three years (Yotzov et al. 2026). At the same time, if working existing inputs harder also teaches workers and firms how to deploy new AI tools, then today’s intensity could boost tomorrow’s efficiency. Utilisation-adjusted TFP growth over coming releases, together with new rounds of firm surveys, can help track the gains.

Authors’ note: The views in this column are those of the authors and do not necessarily reflect those of the Federal Reserve Bank of San Francisco the Federal Reserve System.

References

Abdelrahman, H and A Foerster (2026), “Have we entered an era of high productivity growth?”, FRBSF Economic Letter 2026-14.

Amberg, N, R Friberg, X Liu and C Syverson (2026), “Have we got news for you: Theory and evidence on firms’ optimal expected capacity utilization”, Working Paper.

Basu, S, J Fernald and M S Kimball (2006), “Are technology improvements contractionary?”, American Economic Review 96(5): 1418–1448.

Fernald, J G (2014), “A quarterly, utilization-adjusted series on total factor productivity”, Federal Reserve Bank of San Francisco Working Paper 2012-19.

Fernald, J and H Li (2022), “The impact of COVID on productivity and potential output”, in Reassessing Constraints on the Economy and Policy, Jackson Hole Economic Symposium Proceedings.

Jones, C I and C Tonetti (2026), “Past automation and future A.I.: How weak links tame the growth explosion”, Working Paper, Stanford Graduate School of Business.

Kahn, J A and R W Rich (2026), “Tracking the new economy: Using growth theory to detect changes in trend productivity”, Journal of Monetary Economics 54(6): 1670–1701 (updated 10 June 2026).

Kalyani, A and H Li (2026), “Is optimism for artificial intelligence boosting investment?”, FRBSF Economic Letters 2026-13.

Leduc, S and Z Liu (2016), “Uncertainty shocks are aggregate demand shocks”, Journal of Monetary Economics 82: 20–35.

Yotzov, I, J M Barrero, N Bloom, P Bunn, S J Davis, K M Foster, A Jalca, B H Meyer, P Mizen, M A Navarrete, P Smietanka, G Thwaites and B Z Wang (2026), “Firm data on AI”, NBER Working Paper 34836.