A jobs shock does not require perfect AI
AI could cause a serious credit shock by reducing the earnings on which households borrowed. A business may produce more with a smaller payroll while its former employees struggle to repay debts taken on at their old salaries. If new work and public support arrive too slowly, lost wages can spread into defaults and tighter lending.
For a household exposed to that transition, access to cash deserves priority over a bet on the recovery. Assets that rebound after a rescue may have to be sold to cover expenses while the household is still waiting for that recovery.
The five-year stress test here assumes that, by 2031, AI becomes cheap and reliable enough to handle substantial design, programming, and administrative workflows with little supervision. Today's unemployment rate tells us little about how households would cope with deployment on that scale.
A design business could retain its creative director and deliver the same work with fewer production designers. A software team could need fewer developers to implement and maintain a comparable product. Professions can survive while headcount and wages fall. Employers still have to pay for checking output, fixing mistakes, integration, and responsibility for failures, but those costs may also fall as systems improve.
An IMF scenario exercise published in April 2026 examines how the pace of automation and institutional responses could shape the transition. Workshop participants raise a difficult point about retraining: moving workers into another occupation offers less protection when that occupation is also becoming automatable.
How much new demand would preserve employment?
Let output mean comparable completed work at the same quality. If AI reduces the human hours required per unit by a fraction a, and hours per worker stay constant, then:
Required workforce after adoption / original workforce = output growth factor × (1 − a).
The reduction must cover the entire workflow after extra review and rework. A faster coding benchmark or a count of exposed tasks cannot be inserted into this equation as if it measured net labor savings.
For a hypothetical business that saves 40% of human hours per unit, a 30% increase in output leaves workforce requirements at 1.30 × 0.60 = 0.78 of the starting level. It can produce more with 22% fewer full-time-equivalent workers. Keeping the original workforce requires output to rise by about 67%.
Some markets could expand enough. Cheaper software can make previously uneconomic products viable; cheaper design can bring professional services to smaller customers. Existing occupations can grow as well as new ones emerge, so counting new job titles misses much of the adjustment.
Demand still has to catch up with technical capacity. A customer may want a better application rather than five applications. A company may buy more design variations without increasing its design budget. If output expands while prices fall faster, industry revenue can shrink.
The Acemoglu and Restrepo task framework separates automation's displacement of labor from productivity gains and the creation of new tasks. Whether new demand and tasks absorb displaced workers fast enough is the question the crisis thesis has to answer.
The workforce calculation applies to one activity. Employers can also shorten hours, stop replacing departing staff, or reassign people. Aggregate unemployment depends on their opportunities elsewhere. The severe case is simultaneous substitution across several of the occupations to which displaced workers would normally turn.
When lost income becomes a financial crisis
A household's mortgage does not fall when its salary does. A worker who finds another job at substantially lower pay may remain employed and still struggle to service the old debt. If enough borrowers fall behind, lenders can tighten credit for households and small businesses beyond the occupations first affected by AI.
The economy's total income may initially hold up. Savings on payroll can become profits, lower prices, or payments for computing and other inputs. How much spending returns depends on who receives the gains and what they do with them.
For a first-round illustration, normalize household disposable income to 100. Suppose five units move from a group that spends 80% of an additional unit of income to a group that spends 40%. Hold prices, investment, taxes, and transfers fixed. Planned consumption falls by 5 × (0.80 − 0.40) = 2, even though total income has not yet changed. Halving the spending-propensity gap reduces the consumption gap to one; equal propensities eliminate this channel of consumption loss.
Profits may be retained by companies, paid as dividends, accrue inside pensions, or be taxed to fund public spending. Those routes return different amounts to household budgets at different times. Higher investment or public transfers could offset the consumption gap. Purchases of existing shares do not themselves finance new production. The timing matters for borrowers whose payments come due before they receive any share of the productivity gains.
Investment can cushion the shock, then amplify it
AI infrastructure spending can replace some lost demand during construction. Trouble develops if the capacity is financed on revenue expectations that customers cannot sustain. The BIS's March 2026 analysis of AI infrastructure financing describes borrowing through separate entities alongside leases, capacity commitments, and guarantees. These arrangements connect technology companies, private lenders, and banks even when the debt sits outside a technology company's own balance sheet.
Disappointed AI revenues could then hit heavily financed infrastructure projects just as lost wages weaken household balance sheets. These simultaneous losses could become systemic if they force lenders to withdraw credit from otherwise viable borrowers, making it harder to finance investment or refinance existing debts.
Countries would experience different shocks
A country that owns major AI businesses could gain capital income while its workers lose bargaining power. A service-exporting economy could lose foreign orders without owning much of the technology that replaced them. For borrowers earning local currency, depreciation can add to the cost of servicing foreign-currency debt. The importer could capture the cost savings while the exporter absorbs the income loss, leaving a stable global aggregate to conceal severe regional distress.
Income support, shorter hours with broadly maintained household income, and investment in new activities could sustain demand. Governments with limited fiscal capacity have less room to absorb a prolonged transition. Even with positive GDP growth, declining career access, lower wages after retraining, and concentrated ownership could provoke political conflict.
The response would also affect shareholders. Transfers funded by taxes on concentrated profits could support customers' spending while reducing investors' after-tax gains. Restrictions on adoption could slow displacement and delay productivity gains. The return to owning an AI business will partly depend on how its society distributes the income it produces.
Cash and sovereign bonds defend against different risks
Someone who may lose their salary has to consider the risk of selling assets to pay bills. Cash provides room to wait. The comparisons below use a dollar spending baseline and separate the initial liquidation from the following one to three years. Investors with expenses in another currency would also face exchange-rate risk.
Insured dollar deposits within applicable coverage limits and short-term U.S. Treasury bills provide liquidity with limited price sensitivity. Uninsured bank balances and funds left on trading platforms do not belong in the same category. In deflation, cash can buy more without a nominal gain. Inflation erodes that purchasing power, and falling rates reduce the return available when bills mature or deposits reprice.
Longer sovereign bonds carry a much larger interest-rate exposure. Suppose a fixed-rate bond has modified duration of eight. A one-percentage-point fall in its yield implies an approximately 8% price increase; the same rise implies an approximately 8% loss. This first-order calculation excludes convexity, interest income, and currency effects. Duration of eight does not mean eight years to maturity.
A deep demand recession with anchored inflation could reward duration. But a fiscal rescue may raise inflation expectations or the premium investors demand for holding long debt. Long bonds can then lose value even as the central bank cuts short-term rates. Buying them requires a view on the rates at their own maturities.
Inflation-linked sovereign bonds index cash flows to an inflation measure. Their market prices still respond to real yields, and the index may differ from an investor's actual expenses. Sovereign credit and currency risk remain.
Gold has a stronger case when real yields fall
For a prolonged loss of confidence in the policy response or the currency, gold has a stronger capital-preservation case than Bitcoin, especially if real yields fall. Gold's value does not depend on a particular company's profits, which makes it useful to consider when both employment income and corporate credit are under pressure.
A Chicago Fed study examines real interest rates, inflation expectations, and pessimism about the economy. Their relationships with gold vary across historical periods. The practical implication is to watch the response in real yields and monetary confidence rather than assume that rising unemployment will lift gold.
The liquidation phase can be painful. In his review of the March 2020 liquidity shock, Jon Cunliffe, then a Bank of England deputy governor, recorded an approximately 12% fall in the gold price during the dash for cash. Investors who need money urgently can sell gold alongside riskier holdings.
Rising real yields, an expensive entry price, or a successful productivity expansion could weaken the preservation case. Physical bullion, bullion funds, and gold-mining shares also involve different custody, fee, and operating risks. Miners remain equities, with costs and financing needs of their own.
Bitcoin needs buyers with spare liquidity
Households losing income may sell assets; leveraged investors may face collateral demands. Bitcoin's supply rules do not create the dollars those sellers need. A larger AI economy does not mechanically supply new Bitcoin buyers either, since software agents can use many payment systems.
An IMF working paper on the crypto cycle finds that U.S. monetary tightening weakened a common crypto-price factor through investors' willingness to bear risk. That historical result gives little support to treating Bitcoin as insurance for spending needs during a contraction.
Its opportunity in this scenario comes with renewed liquidity, wider adoption, or stronger demand for assets outside sovereign balance sheets. Those forces could produce a large rebound. They depend on buyers arriving with money to invest after the initial selling pressure, potentially before employment itself recovers.
For a household funding living costs, that timing is decisive. An eventual recovery is of little use if unemployment has already forced a sale. Bitcoin fits the speculative recovery part of this analysis; relying on it for near-term expenses would combine income risk with the risk of a deep drawdown.
Technology stocks divide by cash flow and entry price
A software company paid per employee can lose billable seats when its customers automate, even if it uses AI to cut its own costs. An infrastructure supplier depends on customers financing the next capacity expansion. A business adopting AI outside the technology sector may capture savings without funding frontier models. The label "technology" groups together very different exposures to the same employment shock.
What reaches shareholders depends on competition and the spending required to sustain the business. Firms may pass savings to customers through lower prices. An infrastructure owner can report revenue growth while replacement hardware and new construction absorb its cash. Easily replicated digital work may face falling prices and margins even as output expands.
Valuation can overwhelm an operational success. Consider a hypothetical stock earning 10 per share at a price-to-earnings multiple of 30, giving a price of 300. If AI raises earnings to 12 but the multiple falls to 20, the price becomes 240: earnings rise 20% while the share price falls 20%. If earnings instead fall to eight at the same multiple of 20, the price becomes 160, a 46.7% decline. If earnings reach 12 and the multiple remains at 30, the price reaches 360, a 20% increase. These price changes exclude dividends and trading costs.
Within equities, the more defensible candidates have limited refinancing needs, customers who can keep paying through weaker demand, and cash left after necessary reinvestment. Large appreciation also requires an entry price that leaves room for the business to outperform expectations. A profitable automation business can still be a poor investment if the purchase price assumes growth that leaves little room for disappointment.
What would make the crisis case more convincing?
The warning would be a sequence across several datasets: sustained reductions in human hours per completed workflow, weaker hiring across exposed occupations, and displaced workers taking longer to find jobs at comparable wages. Deteriorating consumption and debt performance among those households would show the shock moving beyond the workplace.
Attributing that sequence to AI requires comparison with less-exposed workers and firms, allowing for interest rates, outsourcing, and industry cycles. A broad fall in hiring cannot identify the cause on its own. The crisis case strengthens when new spending and income support fail to cover lost earnings and lenders begin withdrawing credit.
The asset implications depend on what happens next:
| Conditional path | Assets with a clearer role | What could defeat that role |
|---|---|---|
| Demand contraction with falling inflation | Cash and short bills; longer high-credit-quality bonds if long yields fall | Inflation or fiscal risk keeps long yields high |
| Policy restores demand and market liquidity | Viable businesses at reasonable prices; Bitcoin as a speculative recovery exposure | Earnings disappoint or funding stress persists |
| Rescue undermines currency confidence | Gold; inflation-linked debt only for the domestic inflation component | Real yields rise; market losses before maturity |
| Productivity expands income and spending | Companies retaining durable cash-flow gains | Competition gives the gains to customers; entry prices already assume success |
Markets can price recovery before unemployment peaks. A rescue may also restore spending while leaving the employment structure permanently changed. Asset prices and job statistics need not turn together.
The strongest challenge to the crisis thesis would be displaced workers regaining comparable earnings, or productivity gains reaching household spending through lower prices and transfers before defaults spread. The evidence to watch is the income workers recover after displacement and whether it arrives before arrears force a further contraction in credit.
Sources
The numerical examples use the assumptions stated beside them. They illustrate mechanisms rather than estimate global unemployment or GDP. Historical asset studies inform the analysis of those mechanisms. Sources were reviewed on September 14, 2026.
- IMF, Global Economic and Financial Implications of Artificial Intelligence: Lessons from a Scenario Planning Exercise, April 2026
- Acemoglu and Restrepo, Automation and New Tasks, Journal of Economic Perspectives, 2019
- BIS, Financing the AI infrastructure boom: on- and off-balance sheet borrowing, March 2026
- Federal Reserve Bank of Chicago, What Drives Gold Prices?, 2021
- Jon Cunliffe, Learning from the dash for cash, February 2022
- IMF Working Paper 23/163, The Crypto Cycle and US Monetary Policy, 2023