Who Buys After the Layoff

On 27 February 2024, Klarna said an OpenAI-powered assistant had handled 2.3 million conversations in its first month, two-thirds of customer-service chats, work equal to 700 full-time agents, and an estimated $40 million of 2024 profit improvement. The company also said customers could still choose a live agent. American Banker later reported Klarna's own clarification: the 700 figure was an equivalent, not a same-day headcount cut, and outsourced agents could be moved to other clients. In May 2025, Fortune and Fast Company reported CEO Sebastian Siemiatkowski telling Bloomberg that cost had been too dominant a metric, quality had fallen, and Klarna would hire humans again while still using AI. That sequence is not a morality play. It is a ledger problem. The cost line moved first. The customer line moved later, and not always in the same direction.

Tap to enlarge

The public argument about AI and work is usually binary: jobs die, or jobs multiply. The record, as of August 2026, supports neither slogan. Some tasks are automated. Some jobs change composition. Some new roles appear, especially around energy, chips, and data centers. Some entry-level hiring stalls. The neglected identity is simpler. A wage is also a shopping budget. If the wage leaves a household faster than another budget replaces it, someone who used to buy groceries, software seats, and sneakers buys less. If a new paycheck, a transfer, or an authorized agent replaces that budget, receipts can hold even as the original job does not. The rest of this essay is an audit of which of those three replacements the data currently support, and which they do not.

What the numbers say, and what they do not

Start with exposure, which is not the same as unemployment. Eloundou, Manning, Mishkin, and Rock estimated that around 80% of the US workforce is in occupations where at least 10% of tasks could be affected by large language models, and about 19% in occupations where at least 50% of tasks could be. The Science version of the same project is more conservative on "over half the tasks" once you restrict to simple LLM interfaces, then jumps when you add likely software wrappers. The authors do not date adoption. They do not count layoffs. They score tasks.

The IMF Staff Discussion Note SDN/2024/001 by Cazzaniga, Jaumotte, Li, Melina, Panton, Pizzinelli, Rockall, and Tavares puts global employment exposure near 40%, about 60% in advanced economies, 40% in emerging markets, and 26% in low-income countries. In advanced economies they split high-exposure work roughly into 27% high complementarity and 33% low complementarity. Kristalina Georgieva's IMF blog repeats the 40% global figure and stresses that AI hits cognitive jobs, so rich countries are both more exposed and more able to use the tools. India's exposure in that note is lower than Brazil's on the country sample. That is occupational mix, not immunity.

Forecasts of net job change still disagree, which is a fact about models, not a license to average them into comfort. Goldman Sachs Research has used a global exposure figure of about 300 million jobs, a US estimate that AI could automate tasks equal to 25% of work hours, and a base case of 6-7% of workers displaced over roughly a decade, with unemployment up about 0.6 percentage points if the transition is spread out. Faster adoption, they say, would hit harder. The World Economic Forum Future of Jobs Report 2025 surveys employers and publishes a 2030 churn of 170 million jobs created and 92 million displaced, a net plus 78 million, with clerical work on the decline list and AI, data, and energy roles on the growth list. Employer surveys are not labor-force surveys. Treat the WEF net as a stated hiring intention, not a census.

Announcement data is messier still. Challenger, Gray & Christmas counted 33,429 US job-cut announcements in July 2026, the lowest monthly total in two years, and 477,033 year to date, down 41% from 806,383 in the first seven months of 2025. AI was the leading cited reason in July (10,970 cuts, 33% of the month) for a fifth straight month, and 112,713 year to date, about 24% of 2026 announcements. Since the firm began tracking AI as a distinct reason, the cumulative AI-cited total is 184,538. Technology still led industries (149,023 cuts in 2026 through July, 31% of the total). Andy Challenger also said hiring plans were up 25% year to date and that AI is "not dismantling" the labor market. The same report created a bucket, "Technological Update (possibly AI)," because firms allude without pinning the model. HR Executive quoted Avature CEO Dimitri Boylan warning that "AI" is becoming shorthand investors like. Visa's 7% reduction, in that write-up, was an efficiency story in which AI would change work. Citation is not causation. Non-citation is not innocence.

Payroll microdata is the closest thing to a current labor fact. Brynjolfsson, Chandar, and Chen at the Stanford Digital Economy Lab, August 2026 revision, use ADP payroll covering millions of US workers through June 2026. They report six facts. There is no evidence of economy-wide job destruction. Employment of workers aged 22-25 in AI-exposed occupations stands 19% below the path of less-exposed peers; experienced workers show no comparable gap. The gap has widened since their 2025 drafts (earlier public versions said 13%, then 16%). The channel is reduced hiring more than increased separations. Declines concentrate where AI usage substitutes for tasks; where it complements, employment is flat or rising, especially for experienced workers. Adjustment is in employment, not base pay. They flag that the pattern attenuates with education controls, that some divergence predates ChatGPT, and that ADP looks sharper than some national survey benchmarks. They call the facts canaries, not causal estimates.

That sits beside an older productivity paper, not a contradiction of it. Brynjolfsson, Li, and Raymond in the Quarterly Journal of Economics (2025; NBER w31161) studied 5,172 customer-support agents given a generative assistant. Issues resolved per hour rose about 15% on average, with the largest gains among less experienced agents (on the order of one-third in the working-paper versions). Quality and customer sentiment moved with speed for novices. That is augmentation inside a job. The Stanford ADP paper is about whether the junior seat exists to be augmented. Both can be true in different firms in the same quarter.

The ILO's Generative AI and Jobs analysis (Gmyrek, Berg, Bescond) likewise stresses occupation-level exposure and clerical concentration, not a single global unemployment rate. The OECD Employment Outlook line has been similar: adoption is real, aggregate employment effects through the mid-2020s are still modest, and the distribution inside occupations is the live issue. Autor's 2015 JEP essay, "Why Are There Still So Many Jobs?", remains the right prior: machines take tasks, remaining tasks often become more valuable, and new work appears. Whether that prior still holds is empirical, not theological.

Three things can happen to a job. Only one of them is disappearance.

David Autor, Caroline Chin, Anna Salomons, and Bryan Seegmiller's New Frontiers (QJE 2024) tracks new work in US Census and SOC titles from 1940 to 2018. New occupational titles keep arriving. Demand shocks and technological complementarity create them. Automation also destroys old titles. The net has not been a one-way slide into idleness. Acemoglu and Restrepo (2018, 2019) split the same history into displacement (automation of existing tasks) and reinstatement (new tasks for labor). They have also argued, in later work, that displacement intensified in recent decades while reinstatement did not always keep up. That is a warning about speed, not a proof that 2026 is 1933.

James Bessen's ATM history is the standard counterexample to "the machine ate the occupation." As Bessen and Autor both recount, US ATM counts rose from roughly 100,000 to 400,000 between the mid-1990s and 2010. Tellers per urban branch fell (on the order of 20 to 13). Branch counts rose enough that teller employment did not collapse; Autor cites teller employment moving from about 500,000 to 550,000 between 1980 and 2010, even as the occupation's share of all jobs slipped. The remaining job mixed less cash and more customer work. That is job change. It is not a law that every AI deployment will clone the ATM. Demand for the output has to be elastic, and the leftover human tasks have to be worth staffing.

Displacement, when it does happen, is expensive for the person even after reemployment. Jacobson, LaLonde, and Sullivan (AER 1993), using Pennsylvania administrative data, found high-tenure workers separating from distressed firms suffered long-term earnings losses averaging about 25% per year. Losses began before separation. They were not confined to a few industries. They remained large even for workers who found jobs in similar firms. Couch and Placzek (2010) revisited the result with different data; the qualitative finding, large and persistent earnings losses, survived. Unemployment insurance replaces a fraction of lost wages during joblessness. It does not restore the firm-specific component after the new job starts. For consumer demand, that gap matters. A rehired worker is not a fully restored shopper.

Historical panic is also data. David Ricardo added the machinery chapter to the Principles in 1821 because he changed his mind: machines could hurt labor in some cases. Keynes, in Economic Possibilities for our Grandchildren (1930), coined "technological unemployment" as a temporary mismatch of finding new uses for labor. Both texts are still quoted by people who have not read the surrounding paragraphs. Ricardo was not predicting a jobless Britain. Keynes was writing a century-scale essay about abundance. Use them as history of thought, not as 2026 forecasts.

The worker was also the customer. That was not a slogan. It was a P&L.

On 5 January 1914, Ford Motor Company announced a $5 day, up from about $2.34, with an eight-hour shift. The Henry Ford documents the context: assembly-line time for a Model T had fallen from 12.5 hours toward 93 minutes, turnover was extreme (the museum's artifact notes cite figures around 370-380% annually), and the raise was profit-sharing with a Sociological Department attached. Ford later called it a cost-cutting move because it bought retention. The museum also records the second-order effect Ford himself noted: better-paid workers were a pool of Model T buyers. Kellogg's teaching case states the same mechanism without romance. Mass production needs mass buying. If you pay the people who make the good, some of the demand is internal.

That identity is weaker in 2026 than in 1914, and pretending otherwise is propaganda. Ford workers were a visible slice of the car market. A SaaS company's support agents are a rounding error in global seat demand. A retailer that cuts store labor still sells mostly to non-employees. The identity is strongest for mass-market categories whose customers are wage earners as a class, not as badge-holders of one firm: groceries, mass apparel, fast food, consumer credit, mid-market autos, budget telecom. It is weakest for luxury and for producer goods sold to other firms.

The spending arithmetic is in household surveys, not in LinkedIn posts. The BLS Consumer Expenditure Survey for 2024 shows average annual outlays of $35,046 in the lowest income quintile and $150,342 in the highest, against an all-consumer-unit average of $78,535. High-income households spend more in dollars and a smaller share of income. Fisher, Johnson, Smeeding, and Thompson at the Boston Fed, using PSID data, find that the marginal propensity to consume is lower in higher wealth quintiles; in their summary, the MPC for low-wealth households is an order of magnitude larger than for wealthy households, and the overall two-year MPC is about 10%. Fisher et al. (2016) report average propensities above 0.8 for the bottom 10% of income and below 0.6 for the top 10%. Dynan, Skinner, and Zeldes (2004) documented that saving rates rise with income. Kaldor's 1956 stylized fact, workers consume more of a marginal dollar than owners, is still the right first approximation, even if the micro estimates vary with horizon and instrument.

Move income from a junior support wage to retained earnings and three things can happen to sectoral demand. The household cuts outlays. The owners consume, but a smaller share, and often different baskets (Matsuyama's non-homothetic demand; Comin, Lashkari, and Mestieri on income-driven demand). Or the firm invests the saving in equipment, which shows up in GDP as investment, not as the original consumer category. None of those is "demand vanishes from the universe." All of them can empty a particular merchant's register.

Macro totals in 2026 do not show a consumption crash. They also do not show a fat household buffer. The BEA put the personal saving rate at 2.7% in June 2026, with personal saving of $646.1 billion. The BEA saving-rate page lists March 2026 at 3.5%, April 3.0%, May 2.8%. FRED PSAVERT matches those prints. The BEA GDP advance for Q2 2026 has real GDP up 1.5% annualized after 2.1% in Q1. Consumer spending and investment both contributed; government spending subtracted; imports rose. Real final sales to private domestic purchasers rose 3.9% in Q2 after 1.7% in Q1. Secondary write-ups of the same release (for example this recap) put Q2 consumer spending growth at 3.2% after 0.5% in Q1, and equipment spending at a 15.2% pace, with AI infrastructure named as a driver. Use the BEA tables as the authority; use the recap only for the equipment color. A low saving rate plus strong equipment investment is consistent with households stretching and firms buying GPUs. It is not a verdict on next year's retail same-store sales.

Falk and Tsoukalas (2026) formalize a related idea: if firms capture the full wage saving from automation and only a fraction of the lost product-market demand, competition can push automation past the level that maximizes even owners' joint profits. That is a model, not a measurement. It is useful as a hypothesis about incentives. It is not evidence that US consumption has already fallen by the size of AI-cited Challenger cuts. Those cuts are still small relative to a labor force above 160 million.

Replacement one: another human job

This is the reinstatement path. The WEF's +170 million / -92 million split is the corporate version. Challenger's 2026 hiring rebound, concentrated in what Andy Challenger called floor work (aerospace, energy, manufacturing), is a weaker, nearer-term version. Data-center construction, grid interconnection, and chip packaging are real hiring categories. They do not automatically rehire the 24-year-old who is not getting the customer-success seat. Occupation mismatch is how you get both "labor shortage" headlines and "entry-level ice" headlines in the same month.

Bessen's ATM case is the optimistic template: cheaper branches, more branches, different teller. The pessimistic template is typesetting, which Bessen also uses: desktop publishing reduced compositors and increased designers. Work moved across occupations. The people were not always the same people. Jacobson losses apply here. Even successful movers often earn less for years. For a grocer, that is a quieter customer. For a luxury house, it may be noise.

Goldman Sachs's 0.6 percentage-point unemployment bump in a slow-adoption base case is a reminder of scale. US unemployment moving from the low 4s toward the mid-4s is not mass joblessness. It is also not "no one is affected." The Stanford 19% relative gap for young workers in exposed occupations is compatible with a calm national rate. Averages hide the first rung of the ladder.

Replacement two: a transfer

If the wage does not come back as a wage, it can come back as unemployment insurance, SNAP, wage insurance, a negative income tax, or a dividend on a public or private claim. Each has a different MPC and a different lag. Jacobson and LaLonde were blunt in The Costs of Worker Dislocation: existing programs do not, and probably cannot, cover most of the loss, because much of it arrives after reemployment. A UBI, in the mechanical sense of an unconditional add-on to disposable income, raises the consumption floor. It does not, by itself, change a firm's incentive to automate the next task. That is accounting, not a political endorsement or rejection.

Worker equity is the other transfer. If displaced staff still hold stock, some profit returns to a high-MPC household. Employee stock ownership is uneven, concentrated in listed tech, and usually too small to replace a salary. Co-determination and profit-sharing have a long literature (Weitzman 1985 is the classic). Voluntary adoption that surrenders a large profit share is rare. Mandates are politics. For this essay the only claim required is narrower: equity can recycle some demand; it has not been shown, in US data, to recycle enough to offset a 25% earnings scar.

Replacement three: an agent spending someone else's mandate

This is the path that did not exist as infrastructure in 2014. Agentic commerce is the practice of software discovering, comparing, and paying within rules a person or firm set. McKinsey has published a 2030 range of about $900 billion to $1 trillion of US B2C retail "orchestrated" by agents, and $3 trillion to $5 trillion globally. Those are research estimates under stated assumptions about adoption and merchant readiness, not observed GMV. McKinsey's later automation-curve note keeps the same global band.

The rails are no longer hypothetical. Visa Intelligent Commerce is a product family for agent-initiated payments with controls and tokens. Visa's corporate perspective cites 14,500 financial institutions and 175 million merchant locations as the network the agents would ride. Mastercard has public Agent Pay work. OpenAI and Stripe shipped an Agentic Commerce Protocol for in-chat checkout. Google has pushed AP2-style credentialed mandates and, with retailers, a Universal Commerce Protocol. Anthropic's MCP and Google's A2A are not shopping protocols; they are how tools and agents talk. Together they make it possible for purchasing power to move without a browser session.

Possible is not equivalent. An agent does not create a budget. It spends a budget that already exists: a wage, a credit line, a corporate treasury, a parent's card, an owner's draw. If the junior wage is gone and no transfer replaces it, the agent has nothing to spend unless capital income is assigned to it. That assignment is a distributional choice: family sharing, employer stipends, public transfers, or owners instructing agents to buy mass-market goods they themselves would not browse. Non-homothetic demand says the last of those is the least likely for groceries.

Trust is the measured brake. Earlier consumer surveys (including figures used in the agentic commerce essay: high use of AI tools, much lower trust to complete a purchase) should be updated as products ship, but the gap is the right object. Fraud, chargebacks, and "the agent bought the wrong SKU" are not academic. They are why Visa talks about tokens and mandates instead of scraping card numbers into a prompt.

There is a further mismatch of baskets. A wage-funded human buys on habit, brand, and proximity. An agent buys on constraints: price, delivery slot, spec, return rate. Merchants that spent a decade optimizing landing pages lose that surface. Merchants that publish machine-readable inventory, policy, and price may keep the ticket even if the human job that used to click "buy" is gone, provided the budget remains. Protocol fit does not replace income.

What the counter-arguments get right

First, general equilibrium. Income that leaves a call-center payroll does not leave Earth. It may appear as lower prices, higher profits, cheaper goods, and more real consumption for whoever still has a claim. If AI lowers the price of software support, users buy more of something else. That is the standard productivity defense, and it is often correct with a lag. The lag is the investigative object. Ramey and Shapiro's work on costly capital reallocation is the industrial version: factories do not instantly become other factories. Occupations do not instantly become other occupations.

Second, measurement. Challenger counts press releases. ADP is a selected payroll processor. BLS household and establishment surveys disagree with private payroll in ordinary years. Eloundou scores tasks with humans and with GPT-4; the Science and working-paper summaries do not match digit-for-digit depending on software assumptions. IMF complementarity is a constructed index. WEF is a survey of HR intentions. Goldman is a staff scenario. Anyone stacking these into one "AI jobs number" is inventing a statistic.

Third, labeling. Firms have an incentive to say "AI" when the real story is interest rates, overhiring in 2021-2022, or a failed product. They also have an incentive not to say "AI" when they fear brand damage, which is why Challenger added the "possibly AI" bucket. Duolingo's AI-first contractor policy drew public pushback; that is reputation, which can feed back into consumer demand without a single layoff showing in BLS.

Fourth, emerging markets. IMF exposure is lower where work is still manual. A layoff in Bangalore IT services is a demand event for urban consumption in India; a harvest-season labor pattern is not. Do not export the Stanford 22-25 finding onto the Indian Periodic Labour Force Survey without doing the work.

Fifth, quality reversal. Klarna's 2025 rehiring is evidence that some "700 FTE equivalent" claims were cost theater that failed a CSAT test. IBM's own survey, as cited by Fortune, found that a minority of AI projects delivered the promised return. Automation that does not stick does not remove the wage, and therefore does not remove the shopper.

A checklist that does not require a worldview

If you run a consumer business, the question is not "will AI take jobs." It is three measurable ratios.

1. What share of your category's buyers are wage-dependent in the occupations your own sector is automating? A luxury watch brand and a prepaid wireless carrier do not share an answer.
2. How fast is income replaced for those occupations in your catchment: new jobs (watch vacancies and wage offers in Lightcast or JOLTS), transfers (UI claims), or other household income (second earners, credit)?
3. If an agent becomes the buyer, do you have a product record an agent can purchase without a human landing page, and is there a funded mandate behind it?

If you run a labor ministry or a central bank, the analogous trio is: occupation-age cells (Stanford's canary), consumption by income quintile (BLS CE), and the saving rate plus equipment investment split (BEA). A calm unemployment rate with a collapsing 22-25 employment ratio in exposed jobs and a 2.7% saving rate is a different economy from a calm unemployment rate with stable junior hiring.

If you write model policy, tax the right margin or admit you are doing something else. A profit tax and a UBI can be justified on equity grounds. They do not, as a matter of first-order conditions, change the private return to automating the next task. A tax or a standard on the automation itself does. Whether that is wise depends on complementarity, measurement, and evasion. The IMF note is more useful here than a slogan: prepare infrastructure and training where complementarity is high; do not pretend exposure is destiny.

The receipt is the test

Ford paid $5 in part because turnover was eating the line, and in part because mass production without mass buying is a pile of unsold cars. Klarna published a 700-agent equivalent and then discovered that some of the demand it needed was the feeling of a competent human on the other end of a dispute. Stanford's 19% gap says the first rung is already different for a subset of occupations. BEA says households are still spending, and saving little. Challenger says firms keep typing "AI" on layoff notices even as total notices fall from 2025. McKinsey says agents might sit in front of trillions of dollars of checkout by 2030 if merchants and shoppers allow it.

None of that licenses a single headline. AI will remove some jobs, change some jobs, and create some jobs. The proportions are not known, and anyone who states them as a point forecast is selling certainty. The part that can be investigated without waiting for 2030 is narrower. When a wage stops, does a receipt continue, and through which of the three pipes: a new wage, a transfer, or a funded agent. If the pipe is empty, the firm that cut the cost also cut a customer, usually someone else's, which is why the incentive to keep cutting does not police itself.

The next useful dataset is not another exposure index. It is a merge of occupation-level hiring, household consumption by those occupations, and the share of checkout that arrives with a machine mandate rather than a human session. Until that merge exists, the honest position is conditional. If reinstatement is fast, the 1914 logic still holds. If transfers arrive, demand becomes a political calendar. If agents spend capital income on mass goods, the shopper is no longer the worker, and storefronts built for human habit will fail even when GDP looks fine. If none of the three pipes fill, the layoff is also a demand forecast, and it will show up first in the categories wage earners actually buy.