MJ Convento

/

If AI replaces everyone, who is left to buy the product?

2,021 words, about 9 min read

In February we costed out a plan to stop backfilling two support seats on our internal ops app and let a model do the first pass on reconciliation exceptions. The spreadsheet said we would avoid about 1.9 million pesos a year. I approved the pilot. Then I sat looking at the “annual saving” cell and noticed what it was made of: two people’s rent, groceries and fare money, pulled out of the same consumer economy our sellers sell into.

Not a moral observation. An accounting one, and I have spent the months since working out whether it adds up to anything. The thesis: a firm that automates away its customers’ incomes has automated away its own demand. Payroll is somebody else’s revenue.

Payroll is somebody else’s revenue

The identity part is solid. World Bank national accounts data (indicator NE.CON.PRVT.ZS, updated 13 July 2026) puts households and NPISHs at 76.1 percent of Philippine GDP in 2024, 67.9 percent in the United States, 56.6 percent for the world. Sell into the Philippines and three quarters of your economy is somebody’s household budget.

Most household budgets are paid for by work. The ILO’s World Employment and Social Outlook: September 2024 Update puts the global labour income share at 52.3 percent in 2022, flat through 2023 and 2024, which is 1.6 percentage points below 2004. The ILO costs that at 2.4 trillion dollars of labour income foregone in 2024 alone, in constant PPP. US wage and salary disbursements to persons were 42.7 percent of gross domestic income in 2024, against 51.6 percent in 1970 (BEA series W270RE1A156NBEA).

Here is the part people skip. The identity proves nothing on its own, because payroll cut does not vanish. It becomes margin, dividends, capital spending or lower prices, and all of those are demand too. Data centre construction is demand. The thesis only has teeth if a peso of wages and a peso of capital income are not spent at the same rate. Labour income is more evenly distributed than capital income, which points the right way. A claim about magnitudes, not a law.

The Ford story is told backwards

The usual evidence is Henry Ford’s five-dollar day: pay workers enough to buy the cars. It does not survive the record. The Henry Ford’s account of the 5 January 1914 announcement is blunt about context. The moving line had cut Model T build time from twelve and a half hours to 93 minutes and left Ford carrying labour turnover of 370 percent. The five dollars was not even a raise. It was profit sharing: a man on 2.30 dollars a day kept that rate and could earn a 2.70 dollar bonus, conditional on sobriety, a clean house, no boarders and regular saving, checked by inspectors who called at workers’ homes.

Raff and Summers examined this in “Did Henry Ford Pay Efficiency Wages?” (Journal of Labor Economics, vol. 5 no. 4, 1987, pp. S57 to S86) and concluded the decision is most plausibly explained by labour problems of the kind efficiency wage theory describes, with queues of applicants at the gate and measurable gains in productivity and profits afterwards. Ford’s workers were a rounding error in the American car market. The correction matters more than the story: the mechanism cannot live inside one firm’s decision. It can only be a property of the aggregate.

The underconsumption tradition made the aggregate argument long before Ford. Malthus worried about deficient effectual demand in Principles of Political Economy (1820); Hobson and Mummery built a book on oversaving in The Physiology of Industry (London, Murray, 1889). Keynes named our version. In “Economic Possibilities for our Grandchildren” (1930) he described “technological unemployment”, meaning “unemployment due to our discovery of means of economising the use of labour outrunning the pace at which we can find new uses for labour”, then dismissed it in the next sentence as “only a temporary phase of maladjustment”, en route to predicting fifteen-hour weeks. The standard rebuttal is the lump-of-labour objection: the work to be done is not fixed, cheaper output raises real incomes, new tasks appear. Historically it wins, and I would not bet against it lightly.

What the evidence says in September 2026

The cleanest measurement of automation on local labour markets is Acemoglu and Restrepo, “Robots and Jobs: Evidence from US Labor Markets” (Journal of Political Economy, vol. 128 no. 6, 2020, pp. 2188 to 2244): one more robot per thousand workers cuts the employment to population ratio in a commuting zone by 0.18 to 0.34 percentage points and wages by 0.25 to 0.5 percent. Large, local, not aggregate. Acemoglu’s “The Simple Macroeconomics of AI” (NBER 32487, 2024, published in Economic Policy vol. 40 no. 121, 2025) puts total factor productivity gains at no more than 0.66 percent over ten years, under 0.53 percent once you allow that early evidence comes from easy-to-learn tasks. He also predicts AI widens the gap between capital and labour income.

Exposure estimates are wide and not measurements. The IMF’s Gen-AI: Artificial Intelligence and the Future of Work (SDN/2024/001, January 2024) puts about 40 percent of global employment in high-exposure occupations, 60 percent in advanced economies, 26 percent in low-income countries. The WEF Future of Jobs Report 2025, surveying over 1,000 employers, projects 170 million jobs created against 92 million displaced by 2030, net 78 million, with 40 percent expecting to cut headcount where AI automates tasks and 52 percent expecting to spend more of revenue on wages.

The measurements are mixed. Brynjolfsson, Chandar and Chen’s “Canaries in the Coal Mine?”, revised 12 August 2026 on ADP payroll data through June 2026, finds no economy-wide displacement, but employment of 22 to 25 year olds in highly AI-exposed occupations now sits about 19 percent below where it would be had it tracked their less-exposed peers, up from 15 percent at the July 2025 vintage. It runs through reduced hiring rather than separations, concentrates where observed AI use automates rather than complements, and shows up in employment, not base pay. The authors call these descriptive patterns rather than causal estimates, and the gaps shrink once you control for education. Lee Tucker’s “You’re (not) Hired” (US Census, CES-26-27, April 2026) finds the same shape in administrative data: early-career hires down 9 percent, employment down 12 percent in the most exposed industry-state cells over the ten quarters after ChatGPT.

The firm side says the cutting is not happening at scale. The 2026 AI supplement to the Census Bureau’s Business Trends and Outlook Survey (CES-26-25, reference period November 2025 to January 2026) finds 18 percent of firms using AI in a business function, 32 percent employment-weighted, 66 percent of users relying on AI solely to augment tasks, and AI-related employment decreases in only 2 percent of firms. Anthropic’s Economic Index report of 26 June 2026, on about 9,700 surveyed users, found 10 percent rating their own job loss next year as likely, 38 percent of those blaming AI, and over a third putting a junior colleague’s risk above 60 percent. The much-quoted MIT NANDA claim that 95 percent of enterprise GenAI pilots produced no measurable profit and loss impact is only distributed by request form, so I am not leaning on it.

The objections that keep me honest

New task creation. Acemoglu’s own framework has a reinstatement effect, and over long horizons that has been the larger term. Nothing in the data rules out AI creating categories of work nobody has named.

Cost disease. Baumol’s “Macroeconomics of Unbalanced Growth: The Anatomy of Urban Crisis” (American Economic Review, vol. 57 no. 3, June 1967, pp. 415 to 426) says employment migrates toward sectors where productivity growth is slow, because that is where labour is still needed. The WEF’s absolute-growth list is farmworkers, delivery drivers, construction workers, nursing professionals, personal care aides. Cost disease as a jobs forecast, and wage income can outlive the automation of the cognitive middle.

Capital income is demand too. If returns shift to owners, owners spend and reinvest, and current AI infrastructure spending is a large wage bill for people pouring concrete and running substations. The shortfall needs the propensity gap to be big, and I do not have that number.

Complementarity and policy. The Census 66 percent augment-only figure and the Canaries finding that employment rises where AI complements both point at assisted work, not replaced work. Transfers, taxation and shorter hours are all available. Keynes’s fifteen-hour week was a proposal, not a prophecy.

A coordination failure, not a collapse

So I land narrower than where I started. The demand collapse story is not evidenced. Nobody has shown aggregate consumption falling because of AI, and the careful researchers say plainly that there is no economy-wide displacement in the data as of mid-2026.

The firm-level incentive problem is evidenced, and it is a different claim. A firm cutting payroll captures the whole saving and bears almost none of the demand loss, because its former employees were a negligible share of its own market. True of Ford in 1914, true of us. Nobody internalises the cost, so the aggregate outcome is not what any firm chose. A coordination failure, and those are not fixed by one participant behaving well.

What would falsify the worry: the labour income share rising while adoption deepens, the entry-level gap closing, real median wages growing with productivity rather than lagging it. Four numbers, in order of how much they would move me:

  • The ILO labour income share, stuck at 52.3 percent since 2022.
  • The Stanford Canaries dashboard, monthly, for whether 19 percent widens past 20.
  • Real median earnings against productivity, US and Philippine series both.
  • Household consumption as a share of GDP, which for the Philippines climbed from 72.5 percent in 2015 to 76.1 percent in 2024.

What we decided

The pilot ran nine weeks, then broke in a way that settled it. A model was triaging reconciliation exceptions, and an AI-assisted refactor of the payout fee calculation shipped in the same release train. The rewrite replaced integer centavo arithmetic with float pesos:

// before: exact, centavos as int throughout
$fee = intdiv($order->amount_centavos * $rateBp, 10_000);

// after the assisted rewrite: pesos as float, rounded at the end
$fee = round($order->amount_centavos / 100 * ($rateBp / 10_000), 2);

Both lines pass the unit tests we had. On the 18 May batch, 4,193 rows, the nightly job logged PayoutMismatch: expected 1482331.00, computed 1482330.97 and finance spent two days reconciling by hand. Three centavos. The model had flagged the batch as a rounding anomaly and closed it low priority, which is what it was told to do and the wrong call. The person who used to catch that was a junior analyst reading the same queue and asking why the number looked ugly.

So the rule now, across the engineering team and everyone inside the ops app: AI is pointed at throughput, not at headcount. Assistants write our tests, draft migrations and triage log noise, which has genuinely raised what the team ships. We kept both support seats and we still hire juniors. Not on principle. The Canaries revision found declines concentrated in work built on codified knowledge while employment rose for experienced workers in tacit-knowledge roles, and juniors are the only mechanism I know for turning the first kind into the second. Cutting the bottom of the ladder to save 1.9 million pesos buys a shortage of people who can smell a wrong number in six years.

The ILO publishes its next labour income share estimate later this year. If it has moved off 52.3 percent in either direction, I will write the follow-up.

Sources