AI, jobs and income to 2035
Software now does more of the work customers pay for. We call this period the Cybernetic Age. This outlook sets out where it is taking US jobs and household income over the next ten years, the odds we give each path, and the dated calls that will show whether we were right.
Through 2030 the likeliest path is more automation and no new federal response
We put that at 82%. Firms keep automating, labor’s share of output keeps falling, and income support stays temporary. Each column sums to 100, give or take a point of rounding.
Labor market, US
Which scenario the US is on at each date
| Scenario | 2030 | 2035 |
|---|---|---|
| AbsorbedUnemployment stays 3.5–5.5%; labor’s share drifts lower | 47% | 36% |
| Jobless boomOutput grows, pay does not keep up | 26% | 30% |
| Broad substitutionAI becomes a reliable remote worker | 8% | 21% |
| StallProgress or spending breaks first | 19% | 13% |
Primary federal response
What Washington makes its main answer
| Policy response | 2030 | 2035 |
|---|---|---|
| No primary responseChecks and benefit extensions stay temporary | 82% | 62% |
| Ownership stakesA public fund or adult accounts paying each person | 7% | 15% |
| Permanent cash$250 a month or more to most adults | 4% | 12% |
| Enforcement firstEnforcement outgrows income support | 6% | 9% |
| Limits on automationBinding limits on named occupations | 1% | 3% |
Odds are probabilities from a simulation of chained steps, each with a low, central and high value, run over 200,000 paths. The markers that decide each scenario and policy response are in Scenarios and policy responses.
Bottom line
- 01Output is going to profits.Labor’s share of nonfarm business output fell to 54.1% in the first quarter of 2026, the lowest since the series began in 1947, and fell again in the second quarter. Corporate pre-tax profits reached 15.7% of GDP, the highest since 1950.
- 02The damage so far is in hiring, at the first rung.Employment of 22- to 25-year-olds in AI-exposed jobs stands 19% below its expected path, up from 15% a year earlier. Unemployment is 4.2% and weekly jobless claims are 194,000, so employers are taking on fewer young people while keeping the staff they have.
- 03Health care and the build-out are absorbing workers.Health care and social assistance added 1.29 million jobs in two years while the rest of the private sector lost 31,000. Specialty trade contractors employ a record number of people, and data-center construction spending rose 73% in a year.
- 04Most of the effect is still ahead.In computer and math work, AI handles 33% of tasks in observed use against 94% it could handle in principle. A quarter of US firms now use it.
- 05The next recession is when the response gets decided.We put the odds of a recession starting by the end of 2029 at 42%. If job losses reach workers in their prime in exposed occupations, pressure for a federal response rises. Rebate or “dividend” checks are likely in that recession (41% by 2030). Permanent cash for most adults is not (4% by 2030, 13% by 2035).
- 06Ownership is the route to watch.Trump Accounts have enrolled children since 1 October, OpenAI has proposed giving 5% of its equity to a US sovereign wealth fund, and Senator Sanders’ S.4825 would give the public a 50% stake in large AI firms. We give a federal equity fund or adult accounts paying each person 26% by 2035. The amounts have to be real: 5% of OpenAI pays about $8 per adult a year at a 5% payout.
- 07For companies, three things move the price.Some buyers pay more for a company whose work software already does, and others would rather buy it as it runs today and make that change themselves. Companies that build and run the physical base behind AI (power, electrical, cooling, construction) are hiring. And the law still requires a named person to answer for what software decides, so a clear record of who signs for what is part of what lenders and buyers check.
Indicators
The latest numbers, each with its source and a source grade. A record is agency data, a filing or a statute. A party is an interested party’s own data or a reputable secondary source. Crowd is a forecasting aggregate.
| Indicator | Latest | Year ago | Before the pandemic | Source |
|---|---|---|---|---|
| Unemployment rate | 4.2% Sep | 4.4% | 3.6% | BLS · record |
| Initial jobless claims, weekly | 194k | 233k | 207k | Labor Department · record |
| Hires, monthly (JOLTS) | 5.19M Aug | 5.24M | 5.95M | BLS · record |
| Ages 22–25 in AI-exposed jobs, against expected path | −19% | −15% | — | Stanford Digital Economy Lab, ADP payrolls · record |
| Unemployed 27 weeks or more, share of unemployed | 27.1% | 23.6% | 20.2% | BLS · record |
| Private jobs outside health care, change over 24 months | −31k | — | — | BLS · record |
| Health care and social assistance, change over 24 months | +1.29M | — | — | BLS · record |
| Labor share, nonfarm business | 54.1% Q1 | — | — | BLS; index 93.4 in Q2 · record |
| Corporate pre-tax profits, share of GDP | 15.7% Q2 | 13.1% | 11.1% | BEA · record |
| Productivity growth, year on year | 2.2% Q2 | — | 1.5%/yr 2007–19 | BLS; total factor productivity slowed to 0.8% in 2025 · record |
| Task coverage, computer and math | 33% observed | — | 94% possible | Anthropic · party |
| Firms using AI in the last two weeks | 24.6% | 17.3%* | — | Census Business Trends and Outlook Survey · record |
| Length of task AI finishes at 50% reliability | 17.4 h | — | — | METR; doubling about every 129 days · party |
| Corporate income tax receipts, fiscal year | $381.4B | $452.1B | — | Treasury Monthly Treasury Statement · record |
| Big-tech capital spending 2026, against OpenAI and Anthropic revenue run-rate | ≈$725B | — | — | Company filings and reports, against ≈$115B · record / party |
| Americans who expect fewer jobs from AI | 71% | — | — | Pew; AI a top issue for 35% against 83% for the cost of living (AP-NORC) · record |
| Forecasters’ 2035 view: employment against 2025 | −3.3% | — | — | Metaculus Labor Hub; real median wage +3.3%, labor share 54.9% · crowd |
*November 2025, when Census changed the question wording. BLS and BEA series were read through FRED on 24 October 2026.
Labor’s share keeps falling
Publicly observableLabor share of nonfarm business output, index 2017 = 100, quarterly
Profits take a record share
Publicly observableCorporate profits before tax, share of GDP, quarterly
Care and the trades grew
Publicly observableEmployment, December 2019 = 100, monthly to September 2026
Information and temp work shrank
Publicly observableEmployment, December 2019 = 100, monthly to September 2026
What is still open. Whether AI’s observed share of tasks climbs toward what it could do, or stays held back by checking, software and liability. Whether the hiring gap for young graduates turns into job losses for workers in their prime. Whether the next recession arrives before that coverage closes. Other causes are live in every indicator above: limits on immigration, an energy shock from the war in Iran, a Federal Reserve rate rise and federal job cuts. Yale’s Budget Lab still finds no clear AI footprint in occupational data, and the Dallas Fed puts AI at no more than 0.1 point of unemployment.
Driving forces
The pressures moving the system now, and which way each one is pushing.
Output moving to capitalaccelerating
Labor’s share is at a record low and profits at a 75-year high, with productivity near 2% and total factor productivity slower. That pattern reads as more capital per worker rather than a jump in efficiency.
A broken first rungwidening
The gap for 22- to 25-year-olds went from 15% to 19% in a year. Firms that adopt AI show 9% fewer junior staff after six quarters while senior employment holds, so the effect runs through hiring.
Broad, shallow adoptionrising
A quarter of firms and 45% of workers use AI, hours per user are flat, and observed coverage is a third of what is possible. Most of the savings buyers are paying for has yet to arrive.
Absorption through care and the build-outrising
Care adds jobs because the population is aging. Data-center construction spending rose 73% in a year, and Meta’s 5-gigawatt Louisiana campus needs about 7,500 construction workers at peak and about 1,000 permanent staff. The build-out hires, and each site’s hiring is temporary.
Accountability written into lawhardening
California, Utah and Colorado now bar “the AI did it” as a defense. Connecticut and California require layoff notices to say whether AI played a part. Each law keeps a named person answerable for what software decides.
Ownership vehicles appearforming
Children’s investment accounts are law, an AI lab has proposed giving equity to a public fund, and a Senate bill would take half of large AI firms. None of it is large yet.
A thinner tax base under the build-outworsening
Corporate income tax receipts fell 14.7% and then 15.6% while big-tech capital spending reached about $725 billion, most likely because full expensing lets builders deduct the spend. Paying for any new transfer gets harder just as the case for one grows.
The loop
How one stage feeds the next, and what feeds back.
- 1
Capability
The length of task AI can finish doubles about every 129 days; reliability lags.
17.4 h at 50% - 2
Adoption
Firms sell the work as a service, buy services firms and automate them, or build the physical base.
24.6% of firms · 33% of 94% - 3
Distribution
Savings the firm keeps show up as labor’s share moving to profit, and the first rung thins.
54.1% labor · 15.7% profits - 4
Pressure
Weak while the harm is a missing first job. Rises if a recession brings losses for workers in their prime.
4.2% unemployment - 5
Policy response
None, ownership, cash, enforcement first, or limits on automation.
none binds · forming
↻ The response chosen changes who owns the capital and how fast automation runs, which feeds back into what gets built and adopted next.
MIT economists Daron Acemoglu, Furkan Gitmez and Mehdi Shadmehr explain why the loop can run faster than anyone would choose (NBER working paper 35336, June 2026). Each firm automates past the point its owners as a group would pick, because it ignores the political risk its own automation adds. New labor-intensive work lowers that risk, and firms invest too little in it because they do not count the stability it buys. Their model gives the state three responses. We add a fourth it leaves out: giving people a direct stake in the capital.
Scenarios and policy responses
Scenarios are the four ways the US labor market can go. Policy responses are what Washington can make its main answer. Each one has markers that let us say which we are on.
Absorbed
Productivity grows about 2–2.5% a year and unemployment stays between 3.5% and 5.5%. Entry-level hiring in exposed fields stays weak, care and the trades grow, and labor’s share drifts lower.
Markers: unemployment averaging 5.5% or less in 2033–35, labor share index at 90 or above, no permanent universal cash. 2035 range 17–59%.
Jobless boom
Output grows and pay does not keep up. Productivity runs at 2.5–4%, profits take a rising share, and unemployment reaches 5.5–8% after a recession, concentrated among new graduates and routine service work. In Britain from 1780 to 1840, output per worker rose 46% and real wages 12% (Allen; Crafts argues the gap was smaller).
Markers: labor share index below 90, or a prime-age employment rate below 78%, by 2035. 26% in both extreme runs.
Broad substitution
AI becomes a reliable remote worker and spreads. Most computer-based tasks cost less done by software than by a person, and robots follow in logistics and driving early in the 2030s. Without policy, employment falls sharply.
Markers: a prime-age employment rate below 75%, or unemployment above 9% before any policy offset. 2035 range 8–42%.
Stall
Progress on long tasks stalls, or the spending cycle breaks first. A correction cuts investment and the construction jobs it created; AI’s effect on jobs stays small and any unemployment is cyclical.
Markers: AI’s labor effects stay small through 2035 and hyperscaler capital spending falls 30% or more from its peak. 2035 range 7–14%.
| Policy response | Markers | 2030 | 2035 |
|---|---|---|---|
| No primary response | Checks and benefit extensions stay temporary; no permanent transfer, public stake or limit on automation | 82% | 62% |
| Ownership stakes | A public AI or compute equity fund, or government-seeded accounts extended to adults, paying or crediting each person | 7% | 15% |
| Permanent cash | A permanent transfer funded from capital, such as $250 a month or more to most adults | 4% | 12% |
| Enforcement first | Spending on enforcement outgrows income support for three fiscal years running while labor’s share falls, and federal forces are used against labor or economic protest | 6% | 9% |
| Limits on automation | Binding federal limits on automating named occupations | 1% | 3% |
Timeline
Four stages to 2035, each named for what changes in it.
The first rung keeps thinning
Unemployment stays between 4% and 6% unless an ordinary recession arrives. AI use reaches about a third of firms on trend. The 3 November midterms decide which income route has the votes: dividend checks and accounts, or longer unemployment benefits paid for with taxes on AI and data centers. Builders of power and data-center capacity keep hiring.
Unemployment under 5% through 2026: about 99% (Polymarket) · Census AI use at 35% or more by October 2027: 35%
Agents reach the back office, and the first AI-era recession
The agent capability of 2025–26 shows up in hiring and pay, first in contact centers and back offices. The next recession is when routine white-collar jobs leave and do not come back, as they did after earlier downturns (Jaimovich and Siu), and when checks go out. EU rules for AI in employment decisions start on 2 December 2027.
Recession starting by end-2029: 42% (30–55) · checks by end-2030: 41% · unemployment averaging 6% or more in a year 2027–30: 31%
A response gets chosen
If labor’s share keeps falling while pressure stays weak, the cheapest course is to leave things as they are, and ownership vehicles are the cheapest addition to it. If a recession has raised the pressure, the choice narrows to paying people more, giving them a stake, or enforcing order, and current budgets make some of those routes cheaper than others.
AI as a drop-in remote worker by end-2028: 22%, by end-2030: 42% · a primary response in place by 2030: 18%
The scenarios separate
Absorbed or jobless boom if AI stays short of reliable drop-in work; broad substitution if it does not, with physical work automating at scale around the middle of the decade. A universal high income needs broad substitution and a large transfer together; savings task by task do not produce that much growth.
Permanent near-universal cash by 2035: 13% · ownership route enacted by 2035: 26% · universal high income: under 5%
Who is affected
The parties whose position changes on the central scenario, and what each needs.
| Who | What changes | What they need |
|---|---|---|
| Owners of services companies | Work that follows a set procedure gets cheaper to deliver, and the customer book stays the asset. Buyers discount a claimed software margin unless headcount has stopped growing with customers. | To know which buyers and lenders already pay for a company like theirs, and what each one checks. |
| Advisers and bankers | This year’s deals are sales, recaps and loans of companies whose work software already does, of builders, and of services firms that can show the change. | A current map of who buys and lends against each kind of company. |
| Buyers rolling up services firms | The savings are real only where headcount stopped growing with customers. | Evidence from the company’s own numbers that the work changed. |
| Buyers who price on cash earnings | Still the main buyers of lower-middle-market services firms. | The same company priced on cash earnings, with AI as a question of margin. |
| Lenders | Equipment finance and data-infrastructure credit are growing, and expensing rules change corporate tax shields. | A named borrower, and backlog in a form they already underwrite. |
| Builders of power and data centers | Power is the scarce input. Construction hiring peaks at each site and moves on. | Time to power, the right to bid, and credit against backlog. |
| People who sign: officers and licensed professionals | The law keeps requiring a named person. Judgment moves into software and the signature stays with a person. | Separate company finances, insurance that names the entity, and a written record of who signs what. |
| Workers aged 22–25 in routine jobs | Fewer first jobs, with few layoffs. They carry little political weight today. | A way in: training, apprenticeships and first roles that AI does not yet fill. |
| Care and trades workers | Demand rises. Aide-level pay stays low and largely publicly funded. | Pay that tracks demand, and a path to owning a trade business. |
| Federal and state government | Transfers are flat, the corporate tax base is thinner, and ownership vehicles exist. | A choice of response before the next recession forces one. |
How money reaches people
When labor earns a smaller share of output, households can still get income four ways: wages, transfers, ownership, or lower prices. These are the routes now in play, ranked by our odds.
Rebate or “dividend” checks
41%by end-2030The promised $5,000 “dividend” (about $1.3 trillion) depends on the midterms, and 71% of voters doubt it. Every downturn since 2000 produced checks, and every round stopped.
Longer unemployment benefits
35%extension by 2030Senator Wyden (23 October) proposed up to 52 extra weeks plus a data-center excise tax. Senator Kelly proposed benefits at 75% of wages, backed by Microsoft and the Teamsters. Extensions mostly come in recessions.
Public ownership stakes
26%enacted by 2035Trump Accounts are in force for children, OpenAI has proposed a 5% stake for a public fund, and S.4825 would take 50%. To resist capture, units need to be non-transferable, held on a public ledger and paid out per person; Russia’s 1990s vouchers show what happens otherwise.
Disclosure of AI layoffs
25%federal, by 2028Law in Connecticut and California. The federal bill S.3108 would require quarterly reports. Disclosure informs; it pays nobody.
Permanent near-universal cash
13%by 2035 · 4% by 2030$1,000 a month per adult would cost about 58% of FY2026 federal receipts. Support is 23% and opposition 46%. The OpenResearch study found $1,000 a month cut other income by $1,900 a year and work participation by 4.2 points.
Taxes on AI itself
10%federal, by 2030The one statewide candidate who ran on a token tax finished third. Senator Rounds: “If you start to tax it, you slow it down.”
High income arrives in prices first. AI makes advice, software and paperwork cheaper, while care, housing and power stay expensive, and those three dominate the cost of living. A dividend buys more only if those costs fall too, which Sam Altman’s own 2021 proposal assumed.
New work
Most jobs in 2018 had titles that did not exist in 1940 (Autor and co-authors). From 1987 to 2017, though, automation displaced work about twice as fast as new tasks brought it back (Acemoglu and Restrepo). So far the AI era looks like the second pattern.
| Kind of work | Evidence now | Scale | How long it lasts |
|---|---|---|---|
| Care and health | +1.29M jobs in 24 months; BLS projects the sector will add 37% of new jobs to 2035 | About 24M jobs | Durable; driven by aging and largely publicly financed |
| Building and running the physical base | Data-center construction spending +73%; specialty trades at a record; one campus needs 7,500 at peak and 1,000 permanently | Hundreds of thousands | 2–4 years per site; grid trades 5–10 years |
| Signing for what software decides | California, Utah and Colorado bar “the AI did it”; insurers and lenders price the entity and the people who sign | Every automated unit needs one | Durable while the law requires a human principal |
| Teaching and checking AI | Mercor pays about $4M a day at roughly $100 an hour, about 7,000 full-time equivalents | Tens of thousands | 2–4 years; the generalist tier is shrinking |
| New firms and tiny teams | A record 5.67M business applications in 2025; Lovable reached about $400M in revenue with 146 staff | Large revenue, few jobs | Unproven until applications become payroll |
Key assumptions
The assumptions the odds depend on, and the evidence that would break each one.
| Assumption | Breaks if |
|---|---|
| AI’s observed share of tasks climbs slowly. | Observed coverage in computer and math moves from 33% toward 60% and employment of prime-age workers in those occupations falls. |
| The harm stays at the first rung. | Separations rise for prime-age workers in exposed occupations and unemployment holds above 6% while openings in unexposed work also fall. |
| No primary response takes hold before 2030. | A capital-funded transfer, a public AI stake paying each person, or another response meeting its markers becomes law. |
| The build-out keeps hiring. | Data-center construction employment rolls over and other work does not pick up the hours. |
| Software cannot be the legal principal. | A US bank, card network or the IRS accepts an AI agent as principal with no human responsible party. |
Signposts
A signpost is a number that moves the odds if it crosses the threshold shown.
| Indicator | Now | Threshold | Moves odds toward | Checked |
|---|---|---|---|---|
| Hiring gap, ages 22–25 in exposed jobs | −19% | Beyond −25%, or spreads to ages 26 and up | Jobless boom | Yearly, August |
| Prime-age separations in exposed occupations | flat | Rising two quarters running | Higher pressure; a response gets chosen | Monthly |
| Labor share index | 93.4 | Below 90, or back to 96 and holding | Jobless boom, or absorbed | Quarterly |
| Observed task coverage, computer and math | 33% | 60% | Broad substitution | Each Anthropic release |
| Firms using AI (Census) | 24.6% | 35% by October 2027 | Broad substitution | Every two weeks |
| Ownership vehicle scaled up | children only | Accounts extended to adults, or a public AI fund paying each person | Ownership stakes | Congress.gov |
| Guaranteed-income bill leaves committee | in committee | A floor vote, or a statewide program | Permanent cash | Congress.gov, state legislatures |
| Control of Congress | vote 3 Nov | Which party holds each chamber | Checks and accounts, or benefits and AI taxes | Once |
| Capital spending against AI lab revenue | ≈$725B vs ≈$115B | Spending cut by 30% or more | Stall | Quarterly earnings |
Calls
Each call is dated, carries odds and names the source that settles it. We score each one when it resolves, with the Brier score: the squared gap between our odds and what happened, where 0 is perfect.
| Call | Resolves | Odds | Settled by |
|---|---|---|---|
| The labor share index for the third quarter of 2026 (first release) is below 93.4 | 30 Nov 2026 | 55% | BLS |
| The hiring gap for ages 22–25 at the August 2027 Stanford revision is 19% or wider | Aug 2027 | 60% | Stanford Digital Economy Lab |
| Firm AI use is 35% or more in a Census survey period ending in October 2027 | 31 Oct 2027 | 35% | Census BTOS |
| A recession dated by NBER begins between November 2026 and December 2029 | 31 Dec 2029 | 42% (30–55) | NBER |
| Federal rebate, stimulus or “dividend” checks to most households become law | 31 Dec 2030 | 41% (22–61) | Congress.gov |
| Unemployment averages 6.0% or more in any year from 2027 to 2030 | Jan 2031 | 31% (16–49) | BLS |
| No primary federal response is in place by the end of 2030, on the markers above | 31 Dec 2030 | 82% | Markers in Scenarios and policy responses |
| Permanent federal cash of $250 a month or more reaches most adults | 2030 / 2035 | 4% / 13% | Congress.gov |
| A federal AI equity fund, or adult accounts paying or crediting each person, becomes law | 31 Dec 2035 | 26% (11–46) | Congress.gov |
| The labor share index for the second quarter of 2030 is below 93.4 | Sep 2030 | 55% | BLS |
| Labor scenario at the end of 2030 and 2035: absorbed / jobless boom / broad substitution / stall | 2031 / 2036 | 47/26/8/19 · 36/30/21/13 | Markers in Scenarios and policy responses |
Questions to test this
Questions a sharp reader would use to break the argument.
- 01A shortage of workers alone would raise real wages and labor’s share; both are falling. What besides AI produces falling real pay, a record-low labor share and a 75-year-high profit share at once, and what would separate those causes over the next four quarters?
- 02A one-time ownership stake avoids a tax bill that grows with automation. Does it escape that cost, or move it to the moment of transfer, when capture is likeliest?
- 03The people hurt most today are young and educated. Is their political weight small because there are few of them, or because what they lost is a first job rather than income, and which of those changes in a recession?
- 04If the law keeps requiring a named person for every decision, does that slow automation, or create a thin layer of signers on top of it?
Sources and method
We saved the raw text of every source and checked each figure against it; where sources disagreed, we went back to the original. Labor, profit and productivity series are BLS and BEA data read through FRED on 24 October 2026. Federal receipts are from the Monthly Treasury Statement, Table 9. Odds are chains of steps, each with a low, central and high value, simulated together over 200,000 paths; the response odds add a judgment for each labor scenario. Ranges in brackets come from runs with every step at its low or high value.
- Acemoglu, Gitmez and Shadmehr, June 2026: NBER working paper 35336
- BLS labor share, BEA corporate profits and hires, via FRED: PRS85006173; A053RC1Q027SBEA; JTSHIL
- Treasury: Monthly Treasury Statement
- Brynjolfsson, Chandar and Chen: “Canaries in the Coal Mine,” August 2026
- Anthropic, 5 March 2026: labor-market impacts
- Indeed Hiring Lab: on the 54.1% labor share
- Bill text: S.4825
- TechCrunch: OpenAI’s 5% equity proposal
- Washington Examiner: the $5,000 dividend promise
- OpenResearch cash study: NBER working paper 32719
- Metaculus: Labor Hub, October 2026; AGI date; US UBI by 2035
- Forecasting Research Institute: economic effects of AI
- Jaimovich and Siu: jobless recoveries
- Allen and Crafts: “Engels’ Pause”; working paper 1268
- Autor and co-authors; BLS: new work; employment projections
Frontier Desk maps who’s buying, lending and paying for companies, data and compute. We put the right parties together.
Frontier Desk outlook, 26 October 2026. Figures are as of 24 October 2026. Forecasts are probabilities, and none of this is investment, legal or tax advice. Frontier Desk is not a broker-dealer, lender or investment adviser.