AI and compute economics 2026-04-25 12 minute read

Where AI productivity actually shows up: a sector decomposition

Aggregate US labor productivity grew 2.7 percent in 2024 and 1.9 percent annualized through Q4 2025. The AI signal in those numbers is real but narrow. It lives in five sectors and a dozen occupations, not in the headline.

Two years into broad enterprise AI deployment, the productivity question has moved from speculation to measurement. BLS labor productivity data show nonfarm business output per hour up 2.7 percent in 2024 and roughly 1.9 percent annualized through 2025, above the 2007 to 2019 trend of 1.4 percent. BEA multifactor productivity data lag by a year but show a similar lift in the information sector and professional services. The aggregate signal is real. The distribution is highly uneven. This brief decomposes the productivity gain by sector and occupation, identifies where the Solow paradox is binding now versus resolving, maps the organizational and regulatory bottlenecks that determine speed of diffusion, and lays out a 2026 to 2030 projection grounded in BLS, BEA, OECD, and McKinsey Global Institute data.

The question behind the headline #

US nonfarm business labor productivity grew 2.7 percent in 2024, the strongest annual print since 2010 outside the pandemic distortion years. Through Q4 2025, the four quarter moving average sits at roughly 1.9 percent. The pre pandemic 2007 to 2019 trend was 1.4 percent. The lift is real and statistically meaningful at the aggregate level.

The interpretive trap is reading the aggregate as evidence that AI is broadly transforming the economy. It is not, at least not yet. Decompose the print by sector and the gain concentrates heavily in information, finance and insurance, professional and business services, and a narrow slice of healthcare administrative workflows. Manufacturing productivity remains weak. Construction productivity is negative on a 10 year basis. Retail trade is mixed. The sector dispersion is now wider than at any point since the late 1990s software cycle.

This brief works through how BLS and BEA build the productivity numbers, where the AI signal actually shows up by sector and by occupation, why the Solow paradox is partly binding and partly resolving, what the bottlenecks to diffusion look like, and what 2026 to 2030 realistically delivers under three scenarios.

How the productivity numbers are built #

Two productivity series matter for this question. BLS publishes quarterly labor productivity (output per hour worked) for the nonfarm business sector and for major industries. BEA publishes annual multifactor productivity, which BLS calls total factor productivity (TFP), as the residual after accounting for capital services, labor composition, and intermediate inputs.

Labor productivity moves with capital deepening as well as with technology. If a firm buys a fleet of GPUs and the workers using those GPUs produce more output per hour, labor productivity rises even if the underlying technology is unchanged. TFP strips out the capital deepening effect and isolates the residual gain from better recipes, better organization, and better technology embedded in the inputs. TFP is the cleaner measure of what people mean when they say AI is making the economy more productive.

The catch with TFP is the lag. BLS publishes industry level TFP with roughly a 12 to 18 month delay. The most recent industry TFP release covers 2023, with 2024 expected mid 2026. So the cleanest evidence on AI productivity for 2024 and 2025 has to come from labor productivity series with TFP inferences from the capital and hours data we already have.

A second methodological point matters for AI specifically. BLS measures output for service sectors largely through deflated revenue or, in some cases, through quality adjusted indexes. If AI improves output quality without raising measured revenue (more useful customer service interactions, better written reports, fewer software bugs shipped), the productivity gain is invisible to BLS. This is the modern version of the Hausman critique applied to digital services and is one reason BEA and BLS have ongoing measurement projects on intangible capital and software.

Sector decomposition: where the gain is showing up #

The table below pulls BLS labor productivity by major sector for 2024 and the four quarter average through Q4 2025, against the 2007 to 2019 trend. The pattern is the headline of this brief.

Information (NAICS 51), which includes software publishing, data processing, and telecommunications, is running far above its long term trend. Finance and insurance is running well above. Professional and business services is running modestly above. Healthcare is mixed with hospital productivity weak and ambulatory care stronger. Retail trade is roughly on trend. Manufacturing and construction are below trend.

The information sector lift is the cleanest read of AI productivity in the official data. Software publishing labor productivity grew above 7 percent in 2024 and tracked above 6 percent annualized through 2025. This sector employs the most concentrated population of AI tool users (developers using Copilot, Cursor, and now agentic coding tools) and the smallest organizational gap between tool capability and tool deployment. The signal here is what the rest of the economy can get to once diffusion catches up.

Sector (NAICS)Labor productivity 2024Annualized through Q4 20252007 to 2019 trend
Information (51)+5.8%+6.2%+2.6%
Finance and insurance (52)+4.1%+3.7%+1.8%
Professional and business services (54, 55, 56)+2.9%+2.4%+0.9%
Wholesale trade (42)+2.6%+2.1%+1.6%
Retail trade (44, 45)+1.7%+1.5%+1.4%
Healthcare and social assistance (62)+0.6%+1.1%+0.3%
Manufacturing (31, 32, 33)+0.4%+0.8%+0.7%
Transportation and warehousing (48, 49)+1.2%+1.4%+1.1%
Construction (23)minus 0.9%minus 0.4%minus 0.5%
Accommodation and food services (72)+0.3%+0.7%+0.5%
BLS labor productivity by major sector, full year 2024 and four quarter moving average through Q4 2025, against the 2007 to 2019 trend. Source: BLS Major Sector Productivity and Costs and Industry Productivity programs, with author rounding.

Occupation level: the gain shows up here first #

Sector aggregates wash out the occupation level signal. AI productivity gains are sharply skewed toward occupations whose work product can be encoded as text, code, or structured data. The 2024 BLS Occupational Employment and Wage Statistics universe of roughly 830 detailed occupations is the right denominator.

Three classes of occupations show measurable gain in 2024 and 2025. Software developers and related computing occupations are the cleanest case. Customer service representatives, claims adjusters, and similar text intensive support roles are the second. Writers, editors, paralegals, and analyst occupations whose deliverable is a document are the third.

MIT and Microsoft Research field studies on Copilot users, the OECD AI and Future of Work program, and McKinsey Global Institute's June 2023 generative AI report each converge on a similar pattern: gains of 20 to 50 percent on specific narrow tasks, gains of 10 to 30 percent at the occupation level on tasks that bundle the assistable work with the human only work, and gains in the low single digits at the firm level once organizational frictions are added back in.

Occupation clusterMeasured task time savingOccupation level liftFirm level lift (current)
Software development (general purpose code)26 to 55 percent10 to 25 percent3 to 8 percent
Customer service and support14 to 35 percent8 to 18 percent4 to 10 percent
Document review and paralegal work20 to 45 percent12 to 22 percent5 to 12 percent
Financial analysis and reporting18 to 40 percent9 to 18 percent3 to 7 percent
Marketing copy and content production30 to 60 percent15 to 30 percent4 to 9 percent
Translation and localization35 to 70 percent20 to 40 percent8 to 18 percent
Radiology and pathology image triage10 to 25 percent5 to 12 percent1 to 4 percent
Sales prospecting and outreach20 to 45 percent8 to 16 percent2 to 6 percent
Indicative AI productivity effects by occupation cluster, derived from MIT and Microsoft Research Copilot field studies, the BCG GPT-4 consultant experiment, McKinsey Global Institute estimates, and OECD AI and Future of Work surveys 2023 to 2025. Firm level effects are smaller because of organizational and process frictions.

The Solow paradox in this cycle #

Robert Solow's 1987 quip that we see computers everywhere except in the productivity statistics framed the late 1980s and early 1990s puzzle. The paradox resolved roughly a decade later when the late 1990s productivity boom showed up clearly in BLS data, with information technology producing sectors leading and IT using sectors following with a lag.

The current AI cycle has a similar shape, compressed in time. The information sector productivity lift is already in the official data. Finance, professional services, and select healthcare workflows are visible at the leading edge. Manufacturing, construction, hospitality, and most of healthcare delivery show essentially no AI signal yet. The diffusion curve is steeper than the 1990s curve in the leading sectors and similar or shallower in the lagging ones.

Three reasons the lag persists in lagging sectors. First, the work in those sectors is largely physical, not text or code. Foundation models do not pour concrete or restock shelves. Second, the organizational change required to capture the gain in laggard sectors is larger because the workflow is less text mediated. Third, the regulatory and liability constraints in healthcare delivery, construction inspection, and certain financial activities raise the cost of deploying AI at scale and slow the curve.

The right framing is not whether the paradox is resolved or unresolved. It is partly resolved in five sectors and not yet resolved in eight others. The aggregate productivity number averages over both groups.

Bottlenecks to diffusion #

Three bottlenecks dominate the gap between occupation level lift and firm level lift, and between firm level lift and sector level lift.

Organizational capacity. Most enterprises lack the change management, process redesign, and middle management capacity to absorb the workflow changes that capture AI value. Erik Brynjolfsson's productivity J curve framing is the cleanest description: investment in intangibles (training, process redesign, organizational restructuring) precedes the measurable output gain by 18 to 60 months. Census Bureau Annual Business Survey data on AI adoption, released through 2025, shows that the share of firms using AI for core business processes (not pilots) crossed 12 percent in late 2025, against 5 percent at the start of 2024. The diffusion is real but the absolute level remains low.

Regulation and liability. Healthcare delivery, financial services prudential regulation, legal practice rules, education accreditation, and aviation safety all impose either direct constraints on AI use or significant due diligence requirements that slow deployment. The EU AI Act, in force since August 2024 with phased implementation through 2027, raises the compliance overhead for high risk applications. The US executive order landscape changed in early 2025 and is still in flux. Sector specific regulators (FDA, OCC, FINRA, CMS, state medical boards) move on their own timelines.

Data quality and integration. The single most consistent finding across enterprise AI deployment audits is that the constraint is data plumbing, not model capability. Customer data sits in siloed CRMs. Operational data sits in legacy ERPs. Document corpora sit in SharePoint with no consistent metadata. The integration work to make AI useful at the firm level often exceeds the cost of the AI itself, by 3 to 7 times in McKinsey and Bain's deployment studies. Firms that built a coherent data platform 2018 to 2023 are 12 to 24 months ahead of those that did not.

Adoption by sector: pilot, deployed, scaled #

The Census Bureau's Annual Business Survey added an AI use module starting in 2018 and significantly expanded it for 2023 and 2024 reference years. Combined with the New York Fed's regional surveys and McKinsey's annual State of AI report, a useful three stage adoption picture emerges. Most public commentary collapses these stages, which produces the misleading impression that AI is universally deployed.

Pilot means at least one production use case at small scale. Deployed means AI integrated into a defined business process with measurable output. Scaled means AI use is the default workflow for the relevant function across the firm. The deployment to scaled gap is large in every sector. The scaled column is what shows up in BLS productivity data.

SectorShare piloting AIShare with at least one deployed use caseShare with scaled use
Information and software82%61%34%
Finance and insurance76%47%19%
Professional and business services64%38%14%
Healthcare delivery48%22%6%
Retail and consumer goods55%28%9%
Manufacturing42%19%5%
Construction21%8%2%
Transportation and warehousing38%17%6%
Accommodation and food services29%11%3%
Public administration (federal and state)33%12%3%
Share of firms by AI adoption stage, late 2025 estimates. Sources: Census Bureau Annual Business Survey 2023 and 2024 reference years, McKinsey State of AI 2025, New York Fed regional business surveys, author synthesis. Pilot is at least one production use case at small scale. Deployed is at least one defined business process integration. Scaled is default workflow for the function.

The first place gains show up: occupation, firm, sector, economy #

The propagation order matters for forecasting. AI productivity gains move through four nested layers, and each layer attenuates the signal.

At the occupation level, the gain is measured in field experiments and is large for assistable tasks. The BCG GPT-4 consultant experiment, the Microsoft Copilot studies, the Stanford and MIT customer service experiment with Klarna and Octopus, and the GitHub Copilot developer studies all sit in the 20 to 50 percent range on specific tasks.

At the firm level, the gain attenuates because each occupation only spends a fraction of its time on AI assistable tasks, because the assistable tasks are bundled with non assistable ones, and because the gain has to clear coordination costs across teams. Firm level gains in early adopter case studies sit in the 5 to 15 percent range on relevant cost or revenue metrics.

At the sector level, the gain attenuates further because only a fraction of firms have deployed AI at scale, the scaled deployments concentrate in larger firms, and competitive pressure forces price reductions that transfer some of the gain from firms to customers. Sector level labor productivity lifts in 2024 and 2025 are running 1 to 4 percentage points above prior trend in the leading sectors and roughly 0 in the lagging ones.

At the economy level, the aggregate lift is the share weighted average. Information, finance, and professional services together account for roughly 28 percent of nonfarm business GDP. A 3 percentage point lift across that bloc, weighted accordingly, contributes roughly 0.8 to 0.9 percentage points to aggregate productivity. The remainder of the 2024 aggregate lift comes from cyclical factors, the post pandemic productivity reset, and a smaller diffuse gain across the rest of the economy.

Realistic projections 2026 to 2030 #

Three scenarios bracket the 2026 to 2030 productivity path. The central scenario is the most likely. The upside requires both organizational capacity and regulatory clarity to move faster than the historical baseline. The downside requires either a binding compute or energy constraint, a financial cycle disruption, or a regulatory reversal in major economies.

The central scenario delivers nonfarm business labor productivity growth of 1.9 to 2.2 percent annualized through 2030, against the 1.4 percent 2007 to 2019 trend. AI contributes 0.6 to 0.9 percentage points of that, with the rest from continued capital deepening and the lagged effect of the post pandemic labor market reset. TFP growth, the cleaner measure, runs 0.9 to 1.2 percent annualized against the 0.5 percent prior trend, reflecting genuine intangible capital accumulation and process redesign. By 2030, the share of nonfarm employment in occupations where AI is the default workflow rises from roughly 6 percent in late 2025 to 18 to 25 percent.

The upside scenario requires three conditions. First, agentic AI moves from demo to production at scale, capturing multi step workflows rather than single task assistance. Second, the data plumbing constraint resolves through better tooling, standardization, and a generation of enterprise architecture catching up. Third, regulatory clarity in healthcare, financial services, and the EU lets deployment accelerate rather than slow. Under these conditions, labor productivity growth runs 2.5 to 3.2 percent annualized and the AI contribution rises to 1.2 to 1.6 percentage points.

The downside scenario has labor productivity growth reverting toward the 2007 to 2019 trend by 2028 if the leading sectors saturate before diffusion to the lagging ones accelerates. A binding power constraint that limits frontier compute, a credit cycle that compresses enterprise capex, or a major liability event in a regulated AI deployment would each pull in this direction.

Metric2007 to 2019 baseline2024 actualCentral 2026 to 2030Upside 2026 to 2030Downside 2026 to 2030
Labor productivity growth (nonfarm business, annualized)1.4%2.7%1.9 to 2.2%2.5 to 3.2%1.2 to 1.5%
TFP growth (nonfarm business, annualized)0.5%estimate 1.1%0.9 to 1.2%1.4 to 1.9%0.4 to 0.7%
AI contribution to labor productivitynegligibleestimate 0.6 ppt0.6 to 0.9 ppt1.2 to 1.6 ppt0.3 to 0.5 ppt
Share of employment in scaled AI workflows0%6%18 to 25%30 to 38%10 to 14%
Information sector labor productivity+2.6%+5.8%+4.0 to 5.0%+5.5 to 7.0%+2.5 to 3.5%
Productivity scenario table 2026 to 2030. Sources: BLS Major Sector Productivity, BEA TFP series, OECD productivity panel, McKinsey Global Institute generative AI estimates, author synthesis. The 2024 TFP figure is an estimate pending official BLS release in 2026.

What enterprises and investors should plan for #

Five working assumptions hold for the next 36 to 60 months of AI productivity planning.

One. Aggregate productivity statistics will continue to understate sector dispersion. Information, finance, professional services, and select healthcare and retail subsegments will run 2 to 4 percentage points above their long term trends. Most other sectors will run on or near trend. Investment theses that assume uniform AI productivity gains across the economy are wrong.

Two. The gap between occupation level lift and firm level lift is the strategic battleground. Firms that close that gap by investing in process redesign, change management, and data plumbing capture the full value. Firms that simply distribute AI tools and expect the productivity to show up on its own do not. The Brynjolfsson J curve is the right framing.

Three. The bottleneck composition shifts over time. In 2024 and 2025, model capability was a binding constraint for many use cases. By 2026 and beyond, model capability is rarely the binding constraint. Data quality, organizational capacity, and regulation are. Capital allocation should move accordingly.

Four. Occupation level disruption runs ahead of sector level disruption. Workforce planning, labor market regulation, and education and training systems should anticipate occupational reshuffling within sectors before they see large sector level employment changes. The sector that shows the smallest aggregate productivity lift can still see significant occupation level reshuffling under the surface.

Five. The 2026 to 2030 productivity path is bracketed but not predetermined. Policy choices on permitting, immigration, regulation, and education materially move the realized path within the bracket. The largest single lever is whether enterprises build the organizational capacity to absorb AI value, which is a management problem rather than a technology problem.

How we work this problem #

Two of our anchor platforms are built directly against this question. Athena (see /platforms/athena) models AI economics end to end, from chip and energy cost curves through deployment paths to occupation and firm level productivity outcomes by sector. Aegis (see /platforms/aegis) models the policy, regulatory, and labor market response, including BLS and BEA data integration, sector specific regulatory tracking, and workforce transition modeling. The two platforms are designed to plug into each other so that an AI deployment decision can be tested against both the productivity side and the policy and labor side of the equation in a single workflow.

Engagements typically combine a sector and occupation level productivity assessment, a bottleneck audit covering organizational, regulatory, and data dimensions, and a deployment roadmap with dated milestones tied to measurable productivity outcomes. Reach out at /engage if a decision is coming up.

Sources #

Cite this brief

@misc{hossen2026aiproductivitydecomposition,
  author = {Hossen, Md Deluair},
  title  = {Where AI productivity actually shows up: a sector decomposition},
  year   = {2026},
  url    = {https://deluair.com/consultancy/insights/ai-productivity-decomposition},
  note   = {Deluair Consultancy briefs}
}