Companion note. This is the productivity companion to The Slope Is the Cycle, which argues that the AI capex cycle depends on capability slope, compute supply, and demand breadth.
One-sentence read. AI productivity is not missing; it is just not arriving in the clean order the automation story predicted.
Data through the June 14, 2026 build. Nonfarm-business and manufacturing productivity are 2026 Q1, incorporating the BLS revision of June 4, 2026 (nonfarm +0.8% → +0.3%, manufacturing +3.6% → +3.2%). CES employment is May 2026. Annual BLS industry productivity is 2024 for most sectors and 2025 for a few; the hospital proxy is 2022.
The direct-substitution story has an obvious test. Start with the work frontier models overlap with most — accounting, software, banking, professional services, content, and other information-heavy jobs. If AI were already producing a broad labor-productivity boom, those sectors should break out first.
They have not.
That does not make the productivity dividend imaginary. It suggests the first measurable path is less theatrical: AI embedded inside scheduling systems, customer-service workflows, inventory tools, billing software, logistics platforms, and back-office operations.
The substitution test fails
The exposure score that orders the chart is an ordinal task-overlap ranking, built up to the sector level from Eloundou et al. (2023), “GPTs are GPTs”. Their measure is potential economic impact — the share of tasks a model could touch — and it deliberately does not separate augmenting effects from displacing ones. Read it as a way to sort sectors, not as an adoption rate or a forecast of lost jobs.
The data carry a specific limitation. BLS publishes quarterly productivity only at broad aggregates; industry-level productivity is annual and lagged. Most rows below are 2024, a few have just reached 2025, and the hospital proxy is still 2022. So the chart is useful for falsifying the most obvious story, not for proving what AI did in 2025 or 2026.
Sector productivity
The exposure test: high-AI sectors are not leading yet
Each row shows its latest available BLS reading, tagged per row: annual industry series are 2024 or 2025; quarterly aggregates are 2026 Q1. Rows are ordered by approximate AI task exposure. Delta = latest reading minus trailing trend.
High AI exposure (≥60) · n=5
0 of 5 accelerating (≥+1pp vs trend) — Accounting & tax services, Publishing (incl. software), Depository banks, Travel & reservation services, Broadcasting & content
Selected low-exposure industries (<40) · n=3
1 of 3 accelerating (≥+1pp vs trend) — Motor vehicle & parts dealers, Manufacturing (all), Food & beverage stores
Excludes broad nonfarm benchmark.
A = annual industry reading; Q = quarterly annualized aggregate.
AI-exposure scores are rough rollups inspired by Eloundou et al. (2023), 'GPTs are GPTs' — treat as ordinal ranking, not exact tabulation.
The substitution story predicts that high-exposure sectors should show the earliest productivity acceleration. They do not, at least in these lagged BLS readings. Lower-exposure operational sectors looked stronger earlier in the year but have softened on the newest annual data. Treat this as evidence about where to look next, not proof of AI causality.
The high-exposure rows do not look like an automation boom. Accounting is outright weak. Publishing, banks, travel and reservation services, and broadcasting all sit closer to trend than breakout. None clears the +1-point acceleration mark, and the five-sector group runs about −0.6 percentage points below its own trailing trend.
The firmer readings sit lower on the exposure scale, but the picture there has softened since the spring. Manufacturing is still running above trend. Auto dealers have decelerated, and food and beverage stores have turned negative on the latest annual data. The low-exposure group now averages +0.3 percentage points, down from a much wider gap a quarter ago. Wholesale trade is the one strong outlier, and its exposure score lands it in neither bucket.
The result is narrower than it looked last quarter. The data still reject the most obvious version of the substitution story — the highest-exposure sectors are not leading. They no longer make a confident case that low-exposure operations are surging, either. These are lagged, noisy industry prints, and cyclical recovery or ordinary software adoption could account for part of any move.
The dividend may start where software was scarce
The more plausible mechanism is complementarity: AI may matter first not by replacing the most exposed office worker, but by making software work where software used to be brittle.
Many operating businesses still run on exceptions — a late shipment, a missing part, a billing mismatch, a staffing gap, a customer who does not fit the dropdown menu. Traditional software handled the standard cases and pushed the rest back to people. AI helps because more of that middle can now be routed, summarized, classified, or resolved inside the workflow instead of escalated out of it.
Productivity of that kind does not look like a layoff wave. It looks like fewer escalations, steadier scheduling, faster billing, tighter inventory, and less managerial drag — gains that accrue inside operations long before they register as headcount.
This is still a hypothesis, and the softer low-exposure numbers are a reminder of how circumstantial it is. The pattern could reflect cyclical recovery, routine software adoption, labor-market normalization, or measurement noise as readily as AI. The claim is about where to look, not what has been shown.
Weight the economy by workers
Exposure rankings are not economic weights. The economy is weighted by workers, hours, capital, and output, and once the same sectors are sorted by employment the center of gravity shifts.
Where the workforce is
Large employment pools — exposure × productivity proxy
Employment from BLS CES (2026 M05), sorted by jobs; 133.6M total across the listed sectors. Productivity uses the nearest available BLS proxy where a clean sector aggregate is unavailable.
High-exposure workforce
Jobs in sectors with AI exposure ≥ 60 — where rising productivity means fewer labor hours per unit of output, though not necessarily fewer workers overall.
Low-exposure workforce
Jobs in sectors with AI exposure < 40 — where productivity gains can come alongside steady or rising headcount rather than instead of it.
High exposure marks sectors where direct task substitution should show up earliest if the automation-first story is right; it does not prove displacement. A possible tool dividend — low exposure (<40) with the proxy accelerating ≥+1pp vs trend — is flagged only as a candidate, not a confirmed AI effect. Productivity proxies use the nearest available BLS series, labeled per row, since BLS Industry Productivity rarely publishes 2-digit NAICS aggregates.
Employment levels from BLS CES, May 2026. Each productivity figure uses the nearest available BLS proxy, labeled inline, because BLS Industry Productivity does not publish 2-digit NAICS aggregates for most of these sectors. AI-exposure scores use the same Eloundou-style rollup as the chart above.
Health care and government together hold roughly 47 million jobs in this cut. Retail, accommodation and food services, construction, transportation and warehousing, and manufacturing add another large block of low- and moderate-exposure work. These are not the sectors where the chatbot demo looks most impressive, but they are where small workflow gains compound at macro scale.
The high-exposure rows are real and smaller. Professional and technical services, financial activities, and administrative and support services come to about 29 million jobs here. That is where “fewer labor hours per unit of output” is the more direct mechanism — and it is also where the productivity proxies are decelerating or flat rather than accelerating.
The workforce-weighted read is not that AI does nothing. It is that the largest labor pools may feel AI first as workflow leverage, while the sectors where AI competes most directly with tasks show friction before they show acceleration.
Healthcare is the missing giant
Healthcare is too large to treat as a footnote and too poorly measured to read at face value. It is labor-intensive, heavy with administrative cost, rich in AI use cases — documentation, billing, coding, scheduling, triage, prior authorization, call-center load — and badly served by standard productivity statistics. BLS builds hospital output from the quantity of patients served and treatment intensity, counted through patient stays and visits, and states plainly that a lack of data prevents broad measurement of outcomes or quality of care.
Healthcare scenarios
Four possible paths for healthcare productivity
Employment is still growing; measured hospital productivity remains weak and lagged. The question is whether AI shows up as administrative compression, clinical leverage, prevention, or not at all.
← Output / productivity declining
Output / productivity growing →
Employment
growing ↑
Status quo / measurement gap
Available data hereMeasured data points here for now
Employment keeps growing because demographics demand it. Measured productivity stays negative because BLS builds the hospital series from quantity of patients served and treatment intensity — patient stays and visits — and says lack of data prevents broad measurement of outcomes or quality of care. AI is being piloted in clinical and back-office workflows but hasn't reached scale. The available reading — -2.0% labor productivity at hospitals (2022 proxy), 23.9M health-care jobs and still rising — sits here. The open question: is this a pre-diffusion lag that transitions to #1 over 5–10 years, or a lasting measurement gap where outcomes improve but the productivity statistic never reflects it?
Productivity inflects
The optimistic macro path
AI materially reduces administrative work — documentation, coding, billing, prior authorization, scheduling, and call-center load. Clinical labor is redeployed toward care, and output grows faster than hours. Demand keeps rising with the aging population, so headcount still grows, just slower than output. Productivity prints finally bend positive after a decade of negative readings, and quality holds or improves in parallel. Sector output expands because volume grows, not because per-unit prices rise. This is the trajectory the optimistic case needs from healthcare specifically.
Employment
shrinking ↓
Healthier population
20+ year horizon — a welfare story, not a productivity-stat story
AI-assisted prevention, drug discovery, and chronic-disease management mean people need less medical intervention over time. Demand falls on the supply-of-disease side, so both measured output (interventions performed) and employment shrink. The wrinkle is that this is a welfare gain the statistics invert: measured as healthy life-years delivered rather than procedures performed, output would still be rising even as the visible numbers decline. A labor-productivity series built on procedures cannot capture this case at all.
Admin automation wins
Rationalization scenario
AI displaces back-office and administrative roles — billers, schedulers, coders, prior-auth specialists. Clinical headcount holds because of credentialing floors and demand. Total sector employment dips even as services delivered rise, and productivity prints jump. The caveat worth naming: because healthcare prices are mediated by insurers, government programs, negotiated rates, and local market power, cost savings may not flow quickly or directly to patients — they can land as margin for systems and insurers instead. Good for the productivity statistic, mixed for consumer surplus.
Transition to watch: #4 → #1 (top-left to top-right) is the constructive path — the same growing headcount, but output finally outruns it. #4 → #2 (top-left to bottom-right) is the rationalization case, where AI mostly lands in admin roles. Staying in #4 indefinitely while capex keeps climbing is the path that worries investors.
A scaffold for thinking about health care's trajectory, not a forecast. The available measured data put us in quadrant 4: more workers, weak measured productivity. Which way it moves depends on whether AI lands in administrative workflows, clinical work, prevention, or stays stuck in pilots.
That combination makes the hospital productivity line dangerous to read literally. A weak print could mean AI has not landed. It could mean AI is improving quality, complexity management, or clinician burnout in ways the statistic was never built to capture. Or it could mean deployments are still stuck in pilots and point solutions.
Healthcare could matter enormously for AI diffusion and still fail to show up early in the available measured data. If AI materially compresses administrative work and that capacity is redeployed to care, the macro effect can be large without resembling a layoff cycle. If prevention reduces unnecessary intervention, welfare improves while the output statistics get harder to read, not easier.
The capex implication
This is where the productivity question reconnects with the capex question. If AI value is created inside workflow software, internal tools, and operational systems, then direct model-lab revenue — the AI subscription and compute-provider line items investors can actually see — will understate the value being created.
The same fact cuts the other way on timing. Value that is diffuse, indirect, and captured by customers or software intermediaries does not quickly become revenue for the firms financing the most expensive compute. A factory scheduling tool, a retail inventory assistant, a hospital documentation workflow, or a logistics exception handler can all consume AI through a software vendor, a cloud platform, or an internal system, and little of that has to surface as model-lab revenue on a hyperscaler’s investment horizon.
That is the uncomfortable middle: AI diffusion can be real, productivity-enhancing, and still too indirect to close the capex loop on schedule.
What would change my mind
The substitution story gets stronger if the next industry-productivity releases show high-exposure sectors moving from mixed to sustained acceleration. The complementarity story gets stronger if lower-exposure gains survive revision and can be tied to specific AI-enabled software and operational systems. The evidence I am watching:
- high-exposure sectors begin a sustained acceleration, not a one-print jump;
- lower-exposure acceleration persists after BLS revisions rather than fading the way it did this quarter;
- enterprise AI deployments move from pilots to budgeted, line-item systems;
- healthcare administrative productivity improves in ways the measured data can register;
- productivity gains map to revenue capture somewhere in the AI stack, not just to customer surplus.
A real productivity dividend can still arrive through the side door — embedded tools, operational software, internal workflows. If that is the path, direct model-lab revenue will understate the value created, and may also overstate how quickly that value can rescue the current capex wave.
Source and data appendix
BLS Productivity & Costs / Industry Productivity
- Source type: Official API data.
- Endpoint:
https://api.bls.gov/publicAPI/v2/timeseries/data/. - Fetched at: 2026-06-14.
- Trend definition: the trailing average behind each delta is a trailing 8 quarters (2 years) for quarterly aggregates and the trailing 5 annual readings for industry series. The same window drives the delta chips in both charts.
- Notes: Nonfarm-business and manufacturing series use PRS prefixes and reflect the BLS revision published June 4, 2026; sector productivity uses IPS prefixes. Industry data are annual and lagged. CES employment series are monthly levels in thousands of jobs, not rates.
| Series | Latest period | Latest value | Trailing trend |
|---|---|---|---|
| Nonfarm business productivity | 2026 Q01 | +0.3% annualized q/q | +2.40% (2-yr, 8Q avg) |
| Manufacturing productivity | 2026 Q01 | +3.2% annualized q/q | +1.40% (2-yr, 8Q avg) |
| Publishing, including software | 2024 | +5.1% | +4.22% (5-yr avg) |
| Accounting and tax services | 2024 | -2.4% | +0.38% (5-yr avg) |
| Depository banks | 2024 | +1.6% | +1.46% (5-yr avg) |
| Travel and reservation services | 2024 | +10.5% | +10.76% (5-yr avg) |
| Broadcasting and content | 2024 | +7.0% | +7.72% (5-yr avg) |
| Wholesale trade | 2025 | +4.4% | +0.92% (5-yr avg) |
| Motor vehicle and parts dealers | 2025 | +1.7% | +0.78% (5-yr avg) |
| Food and beverage stores | 2025 | -1.7% | +0.02% (5-yr avg) |
| General merchandise stores (retail) | 2025 | -2.0% | +1.08% (5-yr avg) |
| Food services (accommodation proxy) | 2024 | +1.5% | +3.04% (5-yr avg) |
| General freight trucking | 2024 | +6.3% | +1.84% (5-yr avg) |
| Hospitals proxy | 2022 | -2.0% | -0.98% (5-yr avg) |
CES employment
- Source type: Official BLS Current Employment Statistics via the same API endpoint.
- Latest period: 2026 M05.
- Notes: Employment levels are millions of jobs after converting from thousands.
| Sector | Jobs |
|---|---|
| Health care and social assistance | 23.86M |
| Government, all levels | 23.39M |
| Retail trade | 15.46M |
| Accommodation and food services | 14.38M |
| Manufacturing | 12.61M |
| Professional, scientific, and technical | 10.81M |
| Financial activities | 9.10M |
| Administrative and support services | 9.04M |
| Construction | 8.34M |
| Transportation and warehousing | 6.60M |
AI exposure scores
- Source type: Author rollup inspired by Eloundou et al. (2023), “GPTs are GPTs”.
- Notes: Scores are approximate 0-100 task-overlap rankings. Treat them as ordinal sorting tools, not exact measurements or causal estimates, and note that the source measure does not distinguish augmentation from displacement.