100,000 Monte Carlo Simulations of AI Layoffs. Not One Beat Keeping the Workforce.
Profit turns negative in year five. Shareholder value by year fifteen. The technology delivers — the restructuring sold on its promise destroys output, earnings, and eventually the value of the firm that ordered it.
The Conclusion, First
A firm that deploys AI and cuts no one earns more in year one than a firm that deploys the same AI and cuts a fifth of its workforce. It earns more in every year after that, too.
Not more eventually. Not more once you count social costs. More immediately, in operating profit, on the firm’s own income statement.
There is no profitable window. Across every configuration tested — every level of AI capability, every severance assumption, every displacement scenario — the number of years in which cutting beats not cutting is zero.
Here is the whole argument in five numbers.
Year one: $47M versus $82M. Universal AI-driven headcount reduction produces $47M of operating profit in its first year for a firm that would have earned $82M by deploying the identical technology and keeping its people. The gap opens on day one and never closes.
Twenty-year value: negative $83M versus $804M. Discounted at 8%, universal cutting destroys shareholder value outright by year fifteen and reaches negative $83M by year twenty. Restraint reaches $804M.
The board’s number is 21%. The evidence supports 6%, or zero. The profit-maximizing cut of roughly a fifth of headcount depends entirely on assuming AI replaces a departing knowledge worker at near-parity. Charge realistic severance and use a replacement rate consistent with the published evidence, and the optimum collapses to 6% of headcount. Use the rate the strongest studies imply and the optimum is to cut nobody at all.
Thirty-eight million people, permanently. Not temporarily displaced. Not retrained. White-collar professionals in their forties, fifties and sixties, for whom no adjacent sector exists.
$1.9 trillion a year, forever. The public cost of supporting them — roughly 6.1% of national income — alongside 38 to 57 million lost life-years.
The strategy fails on its own terms, in its own first year, before a single externality is counted.
Why It Looks Profitable Anyway
If cutting loses from year one, why does every board deck show a gain?
Because the deck compares the wrong two things.
An executive evaluating AI substitution compares the post-restructuring firm to the firm as it was before AI — the $50M baseline. Against that comparison, cutting 21% and deploying AI looks excellent: $47M in year one rising through the transition, on paper a transformation.
The correct comparison is to the firm that deployed the same AI and kept its people — $82M. Against that, cutting is a loss in year one and every year after.
The two comparisons diverge because AI augmentation and AI substitution are separable decisions that get bundled into a single business case. Almost all of the gain comes from deploying the technology. The layoff contributes nothing, and then subtracts.
Two structural features make the error stick.
Cumulative value still rises for four years. Against the pre-AI baseline, the curve climbs until year four, then turns. Nothing in a quarterly report distinguishes “rising against the wrong benchmark” from “working.” By the time the curve turns, the decision is four years old and irreversible — the people are gone, the knowledge with them.
The timing matches an executive tenure. The apparent gains fall inside the term of the person who authorized the cut. The losses fall to a successor. This is not a claim about bad faith. It is a horizon mismatch: the executive and the owner are looking at genuinely different projects, and only one of them is measured on the second.
Sitting underneath both is a coordination trap. A single firm’s laid-off employees are a rounding error in that firm’s own customer base, so the demand feedback from one company’s decision onto its own revenue is approximately zero. Unilateral restraint costs the restrained firm real margin and buys it almost nothing. Every firm’s private calculation stays unchanged no matter how bad the aggregate becomes, because none of them can see it in their own accounts.
Part I: The Number the Board Is Using
Start with the firm’s own problem, modeled generously.
A company is not an aggregate quantity of labor. It is a network of human actors, AI agents, software systems, organizational knowledge, coordination pathways, and exception-handling capacity. Human contribution decomposes into task execution, orchestration, tacit knowledge, structural connectivity, exception handling, trust and innovation. AI substitutes for some of these far better than others.
The Replacement Frontier
Let h be the fraction of headcount removed, g the productivity gain among those remaining, and d the organizational effectiveness lost. Output is preserved when:
(1 − h)(1 + g)(1 − d) ≥ 1
which requires:
g* ≥ 1 / [(1 − h)(1 − d)] − 1
Headcount cut Structural loss Required productivity gain
-----------------------------------------------------------
10% 2% 13.4%
20% 5% 31.6%
30% 7% 53.6%
40% 10% 85.2%
50% 15% 135.3%
60% 20% 212.5%Deep AI substitution is not ordinary automation. As headcount falls, the productivity required from what remains rises almost vertically.
The firm’s profit function
Raw capacity is retained human capacity plus machine replacement. But raw capacity is not organizational performance:
P(h) = C_raw × Q × K × F × X × I × C Q = quality retention K = tacit-knowledge retention F = network integrity X = exception-handling capacity I = innovation retention C = coordination efficiency
Simulating 100,000 versions of a 1,000-person, $500M-revenue knowledge firm gives a profit-maximizing reduction of 21% of headcount — against an at-risk pool of 60% of staff, about 35% of the exposed roles.
That is the number in the deck. It is also almost entirely an artifact of three assumptions, and it is worth being precise about which.
Part II: The Assumption Holding It Up
The model assumes each displaced worker is replaced by 0.94 worker-equivalents of AI — 86% task coverage at 110% efficiency. Read that again: the machine is assumed to be ten percent better than the person it replaced, across nearly nine-tenths of their work.
No study supports this. The published evidence measures something else entirely — humans made faster by AI, not humans replaced by it — and where replacement or near-replacement has been examined, the results run the other way.
Brynjolfsson, Li and Raymond studied 5,172 customer-support agents and found a 15% average gain in issues resolved per hour, concentrated among the least experienced. The most skilled workers saw small speed gains and small declines in quality [1].
Noy and Zhang found ChatGPT cut completion time about 40% on bounded writing tasks while raising quality about 18% [2] — an augmentation result on short, self-contained work.
Dell’Acqua and colleagues, with 758 BCG consultants, found 12.2% more tasks completed and 25.1% faster work inside the technology’s capability frontier. On a task deliberately placed outside it, AI assistance made consultants substantially more likely to be wrong, driven by overreliance — by people failing to supervise output that required supervision [3].
METR found experienced developers working on repositories they already understood took 19% longer with early-2025 AI tools, while believing they had been sped up [4]. Their 2026 follow-up could not produce a reliable estimate, because too many developers now refuse to work without AI to sustain a control group [5].
More than 80% of nearly 6,000 executives across four countries reported no measurable effect of AI on either employment or productivity over the prior three years [10].
And Klarna ran the actual experiment. It announced in February 2024 that an AI assistant was doing the work of 700 customer-service agents. By May 2025 chief executive Sebastian Siemiatkowski told Bloomberg the company had gone too far and was rehiring humans. His diagnosis is the degradation term in plain language: cost had become “a too predominant evaluation factor,” and “what you end up having is lower quality.”
Two further assumptions compound it. Severance is charged at zero — Meta booked $148,000 per head in Q2 2026, Oracle roughly $86,000, against a loaded cost of $150,000 per employee in this model. And the degradation terms are quadratic, so cutting one worker in five is assumed to cost 2.6% of institutional knowledge, which is the least defensible number in the firm layer.
What happens when you fix them
Effective AI replacement rate Severance charged Profit-maximizing cut -------------------------------------------------------------------------- 0.94 (as originally modelled) no 20% 0.94 yes 14% 0.71 yes 10% 0.52 (evidence-consistent) yes 6% 0.19 (METR / Klarna implied) yes 0%
Charge realistic severance and the optimum falls from 20% to 14%. Use a replacement rate the evidence actually supports and it reaches 6%. Use the rate implied by the studies of experienced workers on familiar work and the profit-maximizing action is to cut nobody.
The 21% figure is not a finding. It is a vendor claim propagating through a spreadsheet.
Part III: What Happens When Everyone Does It
The calculation above holds baseline revenue constant. For one firm, correct. For all firms, a fallacy of composition.
Consumption is roughly 68% of US GDP; labor’s share of national income about 60%. When employers simultaneously convert wage payments into profit, they change the income distribution that generates the demand they all sell into.
Scarring. Davis and von Wachter, using Social Security records, find present-value earnings losses of 1.4 years of pre-displacement pay when unemployment is below 6%, and 2.8 years when it exceeds 8% [22]. That doubling is measured, and it means the cost of displacement rises with how many other firms are displacing simultaneously. The feedback is self-amplifying.
Ganong and Noel, using bank data from nearly 200,000 unemployment-insurance recipients, find nondurable spending falls 6% at job loss and a further 12–13% at benefit exhaustion — over 30% for low-liquidity households [23].
Functional redistribution. The stronger channel, and it requires no one to be unemployed. Wage income has a marginal propensity to consume near 0.8. Profit income accrues overwhelmingly to the top wealth decile, closer to 0.1. Moving a dollar from wages to profits destroys roughly seventy cents of demand, even when the accounting shows no loss.
The escape valve is competition: if savings pass into prices, real income is restored broadly. This model assumes zero pass-through, on the grounds that a company oriented to shareholder value does not voluntarily surrender margin. That assumption is contestable and testable — watch whether unit costs and prices diverge in AI-intensive sectors.
The offsets are weak. Acemoglu’s task-based estimate caps aggregate AI productivity at 0.71% of TFP over ten years, and argues even that is optimistic [24]. AI capital expenditure is currently a large demand impulse — Oracle’s fiscal-2026 capex reached $55.7 billion, pushing free cash flow to negative $23.7 billion [7]; Meta guided 2026 capex to $130–145 billion [9] — but that is a build-phase effect. Once built, the same spending becomes amortized cost rather than incremental demand.
Anyone observing healthy aggregate demand today and concluding the substitution is costless is reading the build phase and calling it the steady state.
Part IV: There Is No Reabsorption
Everything turns on what happens to the displaced.
The standard displacement literature measures idiosyncratic job loss — one plant closes while the surrounding labor market stays intact. That is the wrong instrument here.
For category-wide destruction the evidence is unambiguous. Autor, Dorn and Hanson, studying the China trade shock, found local employment-to-population rates fell at least one-for-one with the sector’s decline, and that the declines were “not, over the course of a decade, largely offset by sectoral reallocation or labor mobility.” Little migration. Sharply higher uptake of unemployment, disability, retirement and healthcare transfers. Effects still measurable through 2019 — nearly two decades — alongside rising poverty, single parenthood, and drug and alcohol mortality [29].
Reallocation did not happen for manufacturing workers. Expect it to happen less here, not more.
These displaced are white-collar professionals, disproportionately in their forties, fifties and sixties. Farber finds reemployment rates substantially lower for workers 55 and over, with correspondingly larger long-term earnings losses [30]. Johnson and Mommaerts find men under 50 return to nearly their prior wage, while those over 62 take an average 35.6% wage cut [31]. Older workers dominate the ranks of the long-term unemployed and of discouraged workers who leave the labor force entirely — constrained by health, caregiving, community ties, and employers unwilling to hire “overqualified” candidates with nontransferable skills [32].
There is no adjacent sector absorbing a fifty-four-year-old financial analyst.
This model therefore assumes permanent displacement: zero reabsorption, an 80% permanent loss of prior earnings, public assistance replacing 28%.
The Claim That Has No Evidence Behind It
Every defence of large-scale AI substitution eventually arrives at the same sentence: the technology will create new roles to replace the ones it destroys.
There is currently no evidence of AI-driven job creation at a scale that offsets AI-driven displacement. Not weak evidence. Not contested evidence. None.
Where new roles should appear first, employment is falling fastest. If AI were generating novel work, entrants would capture it — cheapest, most AI-fluent, least encumbered by legacy practice. The opposite is happening. Employment for workers aged 22–25 in the most AI-exposed occupations has fallen 13–16% since late 2022, with young software developers roughly 20% below their 2022 peak, while workers 30 and over in the same occupations grew 6–12% [28].
Hiring has frozen rather than rotated. The US hiring rate fell to 3.1%, lowest since April 2020. Openings dropped from 10.3 million in January 2023 to 6.88 million. Quits sit at 1.9%. Work reappearing elsewhere would show up as churn. The market has stopped moving.
The productivity gains that would fund new roles are not there. Task creation at the required scale would appear in aggregate productivity. Against Acemoglu’s 0.71% ceiling and 80% of firms reporting no effect, it has not.
The burden of proof belongs to whoever asserts the offset. Until it is met, a reinstatement effect is a forecast, not a finding, and forecasts do not belong in models.
“They’ll find something else” is not a rebuttal either
What displaced professionals can get is service work, and here the standard statistics actively mislead. A displaced analyst taking retail or food-service work is recorded as employed, which makes the labor market look like it cleared. Their income has collapsed, pension contributions stopped, health coverage changed, and consumption fallen to a level indistinguishable from unemployment. Workers who give up searching leave the labor force and stop being counted as unemployed at all.
Underemployment at this scale is unemployment with better statistics. The model treats it accordingly.
Part V: The Long-Run Result
Running the dynamic model — firms re-optimizing annually, no rehiring, unemployment insurance exhausting after year one, capital-expenditure support ending after year three, and the expertise pipeline degrading as junior roles disappear:
5-yr 10-yr 15-yr 20-yr ---------------------------------------------------------------------------- Board deck (others hold steady) $471M $792M $1,010M $1,159M Deploy AI, cut nobody $327M $550M $701M $804M Universal cutting, no fiscal charge $103M $64M −$6M −$83M with 30% of public cost recovered $70M $6M −$81M −$171M with 60% of public cost recovered $37M −$52M −$157M −$259M
Annual profit is below the no-cut alternative in year one and turns negative outright in year five. Cumulative value turns negative in year fifteen — without charging firms a cent of the public cost.
The best available outcome is $804M — deploy the technology, keep the people. Every cutting strategy underperforms it, in every year, on every horizon. And no individual firm can choose it alone.
Part VI: The Bill
Thirty-eight million people, permanently out of the workforce or severely underemployed, do not disappear.
At a 28% replacement rate across disability, food assistance, Medicaid and ACA subsidies, plus the income and payroll taxes they no longer pay:
$1.9 trillion per year. Roughly 6.1% of national income. About $18 trillion over a decade.
That annual figure exceeds the entire US defense budget several times over and would be the largest transfer program in the country’s history by a wide margin. Somebody funds it, and it is difficult to construct a politics in which the beneficiaries of the displacement are not asked to contribute. At 60% recovery, the twenty-year NPV is negative $259M.
Part VII: What Doesn’t Recover
The expertise pipeline. Experts are manufactured through work. Senior engineers once debugged simple failures; senior analysts once built ordinary models. Let organizational knowledge evolve as:
KH(t+1) = (1 − δ)·KH(t) + α·L(t) − D(t)
AI-driven restructuring simultaneously reduces L(t), the learning generated through human work, and raises D(t), expertise lost through departure. Run the model to twenty years and firm profit falls below what it earned before any of this began, with demand long since stabilized. The demand channel is eventually self-limiting. Expertise destruction is not.
A firm eliminating its junior tier is not only externalizing a social cost. It is consuming its own future senior tier.
Human life. Sullivan and von Wachter, matching Pennsylvania earnings records to Social Security death records, find mortality 50–100% higher in the year after displacement, still 10–15% elevated twenty years later, implying 1.0–1.5 years of life expectancy lost for a worker displaced at forty [33]. Replicated in Denmark, Sweden, the Netherlands and Finland, with elevated deaths from alcohol, substance use, cardiovascular disease and suicide persisting two decades out.
At 38.1 million displaced: 38 to 57 million life-years.
Part VIII: What Would Have to Be True
AI would have to replace knowledge workers at something near parity. This is the load-bearing assumption and it has never been demonstrated. Every measurement of experienced workers on familiar work — METR’s 19% slowdown, the quality declines among Brynjolfsson’s most skilled agents, Dell’Acqua’s errors outside the frontier, Klarna’s reversal — points the other way. Anyone defending large-scale substitution is defending this number, whether or not they know it.
Competition would have to pass savings into prices. This model sets pass-through to zero. In genuinely competitive markets savings do reach consumers. This is the most contestable assumption here, and the most easily tested.
Reabsorption would have to work. The China shock says it does not when a whole category goes, and the older-worker literature says it works least well for exactly this cohort. Note what does not depend on this: cutting lost to restraint at 5%, 15%, 30% and 50% annual reabsorption. Zero reabsorption changes severity, not direction.
There are also limits this model cannot represent. A $1.9 trillion annual transfer bill and 19% sustained unemployment have no US precedent outside the Great Depression. Long before year twenty there would be defaults, asset repricing, political rupture and emergency policy — none of it in these equations, and none of it obviously helping the case for cutting.
The Structural Conclusion
There is an AI Replacement Frontier, and it is convex. There is an output-maximizing workforce and a profit-maximizing workforce, and they are not the same number.
But the finding that matters is simpler than any of that. The profit-maximizing action, for a firm that has deployed AI, is to keep its people. The layoff was never the source of the gain. The technology was. Bundling them into one business case obscured a comparison that, once made correctly, is not close — in year one or in year twenty.
This is not an argument against AI. Deploying the technology and keeping the workforce is the highest-value strategy in the entire model, on every horizon tested.
It is an argument that the decision cannot be left where it currently sits. The firm’s own accounts will not surface the error: they compare the restructured company to its own past rather than to the company it could have been, and by the time the cumulative curve turns, the knowledge is four years gone. Closing the gap requires something outside the firm’s objective function — competition policy that raises pass-through, tax treatment that stops favoring capital substitution over labor, severance obligations that internalize scarring, and an unemployment system built for the fact that displacement costs double when synchronized.
The central executive question is no longer how many jobs can AI automate.
It is: am I comparing this to the right alternative?
On every calibration this model admits, the answer has been no.
Modeling Note
A structural model, not a forecast. It simulates 100,000 parameter draws for a hypothetical 1,000-person digitized knowledge firm, then embeds that firm in a general-equilibrium demand layer calibrated to published estimates.
Measured parameters: displacement scarring and its state dependence (Davis & von Wachter); consumption response to income loss (Ganong & Noel); mortality (Sullivan & von Wachter); category-wide reallocation failure (Autor, Dorn & Hanson); older-worker reemployment penalties (Farber; Johnson & Mommaerts); the aggregate TFP ceiling (Acemoglu); labor and consumption shares of GDP (BEA); severance per head (Meta and Oracle disclosures).
Assumed parameters, which drive the results and should be attacked first: zero competitive pass-through; zero reabsorption with an 80% permanent earnings loss; a 28% public-assistance replacement rate; the marginal-propensity-to-consume differential between wage and profit income; and the organizational degradation exponentials, which generate the nonlinearity in the firm layer and have no firm-specific calibration.
The AI replacement-efficiency parameter is treated as a swept variable rather than a fixed input, because no published study measures it directly. Results are reported across its range.
Figures beyond the point where profit turns negative are extrapolations well outside any observed range. Treat crossing points as meaningful and magnitudes past them as directional. A firm-specific implementation should replace every distribution here with measured organizational data before informing any actual workforce decision.
References
[1] Brynjolfsson, E., Li, D., and Raymond, L. “Generative AI at Work.” The Quarterly Journal of Economics, 140(2), 889–942, 2025.
[2] Noy, S., and Zhang, W. “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.” Science, 381, 187–192, 2023.
[3] Dell’Acqua, F., McFowland, E. III, Mollick, E., et al. “Navigating the Jagged Technological Frontier.” Organization Science, 37(2), 403–423, 2026.
[4] Becker, J., Rush, N., Barnes, B., and Rein, D. “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.” METR, July 2025.
[5] METR. “We Are Changing Our Developer Productivity Experiment Design.” February 2026.
[6] Klarna Group plc. Q4 and FY2025 results, February 2026; Bloomberg interview with S. Siemiatkowski, May 2025.
[7] Oracle Corporation. Fiscal 2026 Form 10-K and FY2026 results, June 2026.
[9] Meta Platforms, Inc. Second Quarter 2026 Results, July 2026.
[10] Yotzov, I., Barrero, J. M., Bloom, N., et al. “Firm Data on AI.” NBER Working Paper 34836, 2026.
[22] Davis, S. J., and von Wachter, T. “Recessions and the Costs of Job Loss.” Brookings Papers on Economic Activity, Fall 2011.
[23] Ganong, P., and Noel, P. “Consumer Spending during Unemployment.” American Economic Review, 109(7), 2383–2424, 2019.
[24] Acemoglu, D. “The Simple Macroeconomics of AI.” Economic Policy, 40(121), 2025.
[28] Brynjolfsson, E., Chandar, B., and Chen, R. “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.” Stanford Digital Economy Lab, August 2025.
[29] Autor, D. H., Dorn, D., and Hanson, G. H. “The China Shock: Learning from Labor Market Adjustment to Large Changes in Trade.” Annual Review of Economics, 8, 205–240, 2016.
[30] Farber, H. S. “Job Loss in the United States, 1981–2001.” Research in Labor Economics, 2004.
[31] Johnson, R. W., and Mommaerts, C. “Age Differences in Job Loss, Job Search, and Reemployment.” Urban Institute, January 2011.
[32] Economic Policy Institute, “Older Workers,” 2022; National Academies, Understanding the Aging Workforce, 2022.
[33] Sullivan, D., and von Wachter, T. “Job Displacement and Mortality: An Analysis Using Administrative Data.” The Quarterly Journal of Economics, 124(3), 1265–1306, 2009.
About the Author
Brian Curry is a Kansas City–based Machine Learning and AI researcher, data scientist, and the founder of Vector1 Research, where he works on the intersection of cognitive AI architecture, causal inference, and economic systems. His research introduced Memory-Node Encapsulation (MNE), an original data structure for artificial episodic memory, and the NeoCortex-M architecture for memory-driven AI. He is the creator of Papilon, an open-source framework for complex system optimization with agentic orchestration, PyCausalSim, a Python framework for causal discovery through simulation, and Daedalus, a platform for causal and economic analysis.
His practitioner work focuses on building production intelligence systems — causal marketing analytics, agentic systems, and the serving infrastructure that makes them deployable inside enterprise environments. He writes about that work at vector1.ai and on Medium.
