spb/spboucher.ai Public

1# The Macroeconomics of White-Collar Automation23Artificial intelligence is usually discussed as a labor-market technology.45Will programmers lose their jobs?67Will accountants be automated?89Will lawyers need fewer junior associates?1011Will analysts, consultants, researchers, designers, and managers become less valuable?1213Those are important questions.1415But they may be too narrow.1617If artificial intelligence substantially automates white-collar work, the consequences will not remain confined to individual occupations. White-collar workers sit near the center of modern developed economies. They earn relatively high incomes, pay significant amounts of income tax, hold mortgages, own financial assets, consume services, and support expensive urban housing markets.1819A large shock to cognitive labor therefore becomes a macroeconomic shock.2021The relevant question is not simply:2223> **How many jobs will AI automate?**2425It is:2627> **What happens when the marginal cost of cognitive production falls faster than an economy can redistribute the resulting productivity gains?**2829## 1. AI as a New Factor of Production3031A simple way to think about AI is to introduce machine cognition directly into the production function.3233Instead of an economy producing output using only capital and human labor,3435$$36Y_t = A_t F(K_t, L_t),37$$3839suppose production increasingly depends on three major inputs:4041$$42Y_t = A_t F(K_t, L_t^{H}, L_t^{AI}),43$$4445where:4647* \(Y_t\) is aggregate output,48* \(K_t\) is conventional capital,49* \(L_t^{H}\) is human labor,50* \(L_t^{AI}\) is machine-provided cognitive labor,51* and \(A_t\) captures general productivity.5253This distinction matters because \(L_t^{AI}\) has very different economics from human labor.5455Human labor is scarce.5657It must be educated, recruited, compensated, coordinated, and given time to perform tasks.5859Machine cognition can potentially be replicated at extremely low marginal cost.6061An effective cognitive-labor aggregate might therefore look like:6263$$64L_t^{\text{eff}} = L_t^{H} + \phi_t L_t^{AI},65$$6667where \(\phi_t\) represents the relative effectiveness of AI labor.6869If \(\phi_t\) rises rapidly, the economy can obtain substantially more effective cognitive labor without proportionally increasing human employment.7071That is the central macroeconomic novelty.7273## 2. The First Effect Is a Productivity Boom7475Initially, the effect should look extremely positive.7677A firm that required 100 analysts may eventually produce comparable output with 40 analysts using powerful AI systems.7879A software company may ship more code with fewer engineers.8081A law firm may process thousands of documents in hours rather than weeks.8283An accounting department may automate reconciliation, reporting, and document processing.8485In a simple production setting,8687$$88\frac{\partial Y}{\partial L^{AI}} > 0.89$$9091More machine cognition increases output.9293If AI also improves the productivity of existing workers, there is an additional complementarity:9495$$96\frac{\partial^2 Y}{\partial L^{AI}\,\partial L^{H}} > 0.97$$9899That is the optimistic version of the story.100101AI does not replace workers.102103It amplifies them.104105For some occupations, that may dominate for a long time.106107But complementarity does not guarantee permanent labor demand.108109A technology can initially make workers more productive and later become capable enough to substitute for them.110111## 3. The Important Variable Is the Elasticity of Substitution112113The macroeconomic outcome depends heavily on how easily machine cognition substitutes for human cognition.114115A useful conceptual production function is a CES specification:116117$$118L_t^{\text{cog}} = \left[ \alpha (L_t^{H})^{\frac{\sigma-1}{\sigma}} + (1-\alpha)(\phi_t L_t^{AI})^{\frac{\sigma-1}{\sigma}} \right]^{\frac{\sigma}{\sigma-1}},119$$120121where \(\sigma\) is the elasticity of substitution between human and machine cognitive labor.122123If124125$$126\sigma < 1,127$$128129the two inputs are relatively complementary.130131If132133$$134\sigma > 1,135$$136137substitution becomes economically powerful.138139As AI capabilities improve and the price of machine cognition falls, firms have a stronger incentive to replace human cognitive tasks.140141The relevant price ratio becomes something like:142143$$144\frac{P_{AI}}{w_H},145$$146147where \(P_{AI}\) is the cost of acquiring a unit of AI cognitive output and \(w_H\) is the human wage.148149If150151$$152\frac{P_{AI}}{w_H} \rightarrow 0,153$$154155then even moderate substitutability can produce major organizational changes.156157## 4. Why White-Collar Automation Is Macroeconomically Different158159The automation of some previous forms of labor affected relatively low-paid occupations.160161White-collar automation can hit a very different section of the income distribution.162163Consider aggregate labor income:164165$$166Y_L = \sum_i w_i L_i.167$$168169A relatively small number of high-income workers can contribute disproportionately to \(Y_L\).170171If AI primarily affects occupations with wages above the economy-wide average, then employment losses need not be enormous to create a significant decline in aggregate labor income.172173For example, suppose ten percent of workers generate twenty-five percent of total labor income.174175Automating a substantial fraction of their tasks creates a much larger demand shock than the headline employment number alone would suggest.176177This is one reason measuring only unemployment could be misleading.178179The economy might experience limited unemployment while still seeing:180181* slower wage growth,182* fewer entry-level positions,183* declining hours,184* weaker bargaining power,185* lower professional-service prices,186* and a shift from labor income toward capital income.187188The transition can occur through wages before it occurs through unemployment.189190## 5. Labor Share Becomes a Critical Variable191192Define labor's share of aggregate income as:193194$$195s_L = \frac{wL}{Y}.196$$197198If AI increases production faster than labor compensation, then:199200$$201\frac{\dot{Y}}{Y} > \frac{\dot{(wL)}}{wL},202$$203204and consequently:205206$$207\dot{s}_L < 0.208$$209210The economy can therefore become substantially more productive while the fraction of output paid to workers falls.211212This is not logically contradictory.213214Suppose output rises from 100 to 150 while total compensation rises only from 60 to 65.215216Workers are receiving slightly more aggregate income.217218But labor's share falls from219220$$221\frac{60}{100} = 60\%222$$223224to approximately225226$$227\frac{65}{150} \approx 43.3\%.228$$229230The economy is richer.231232Labor receives more income in absolute terms.233234And yet capital captures most of the incremental output.235236That distinction is central to understanding AI.237238## 6. Productivity Growth Does Not Automatically Become Demand Growth239240GDP expenditure can be written as:241242$$243Y = C + I + G + NX.244$$245246AI primarily attacks the supply side first.247248It increases what firms are capable of producing.249250But additional productive capacity does not automatically create proportional demand.251252Suppose household consumption depends on labor and capital income:253254$$255C = c_L Y_L + c_K Y_K,256$$257258where \(Y_L\) is labor income and \(Y_K\) is capital income.259260If workers have a higher marginal propensity to consume than capital owners,261262$$263c_L > c_K,264$$265266then shifting one dollar of national income from labor toward capital can reduce immediate aggregate consumption.267268Now suppose:269270$$271Y = Y_L + Y_K.272$$273274An AI-driven redistribution can occur even while total income increases:275276$$277\Delta Y > 0,278$$279280but if281282$$283\Delta Y_L < 0284$$285286and287288$$289\Delta Y_K \gg 0,290$$291292consumption may grow much more slowly than productive capacity.293294This creates one of the central macroeconomic tensions of advanced automation:295296$$297\boxed{ \text{AI productivity growth} \not\Rightarrow \text{proportional aggregate demand growth} }298$$299300The economy can become capable of producing more than households are willing or able to purchase at previous prices.301302## 7. AI Could Be Structurally Disinflationary303304This leads naturally to prices.305306Suppose potential output rises rapidly because of AI:307308$$309Y_t^* \uparrow.310$$311312If aggregate demand grows more slowly,313314$$315AD_t < Y_t^*,316$$317318the output gap becomes negative relative to productive capacity.319320In a simplified Phillips-curve framework,321322$$323\pi_t = \pi_t^e + \kappa (Y_t - Y_t^*) + u_t.324$$325326If AI causes \(Y_t^*\) to rise faster than actual spending \(Y_t\), then:327328$$329Y_t - Y_t^* < 0,330$$331332placing downward pressure on inflation.333334This means sufficiently powerful AI may be structurally disinflationary in many cognitive services.335336Software development becomes cheaper.337338Legal analysis becomes cheaper.339340Translation becomes cheaper.341342Marketing production becomes cheaper.343344Financial analysis becomes cheaper.345346Certain forms of education become cheaper.347348Content production becomes cheaper.349350The economic problem could eventually shift from producing enough to maintaining sufficient nominal demand.351352## 8. But Asset Prices Could Move in the Opposite Direction353354At the same time, AI does not necessarily make every scarce asset cheaper.355356In fact, the opposite may occur.357358If AI increases the returns to capital ownership, income may flow toward owners of:359360* AI companies,361* semiconductor firms,362* energy infrastructure,363* data centers,364* proprietary datasets,365* intellectual property,366* land,367* and productive businesses.368369Capitalized asset values depend on expected future cash flows:370371$$372P_t = \sum_{j=1}^{\infty} \frac{E_t[CF_{t+j}]}{(1+r)^j}.373$$374375If AI dramatically raises expected profits,376377$$378E_t[CF_{t+j}] \uparrow,379$$380381then asset prices can rise even while wage growth weakens.382383We could therefore see a strange combination:384385$$386\text{consumer-price disinflation} + \text{asset-price inflation}.387$$388389Digital goods become cheaper.390391Labor becomes less scarce.392393But ownership of productive capital becomes more valuable.394395That environment would feel very different depending on whether a household primarily earns wages or owns assets.396397## 9. White-Collar Automation Could Hit Housing Through Income, Not Supply398399Housing introduces another transmission mechanism.400401In many major cities, expensive housing is supported by large populations of high-income professional workers.402403Housing demand can be represented loosely as:404405$$406D_H = f(Y_L, r, E[P_{H,t+1}], N, \ldots),407$$408409where labor income \(Y_L\) is an important driver.410411If AI weakens professional income or reduces the number of workers required in expensive commercial centers, demand for certain housing markets may change.412413This would not necessarily reduce housing prices everywhere.414415Housing remains constrained by land, regulation, geography, and construction.416417But AI could produce substantial relative-price effects.418419Cities whose valuations depend heavily on concentrations of white-collar employment may respond differently from locations where housing value derives primarily from physical scarcity, amenities, or population growth.420421Remote AI-enabled work could amplify this effect.422423If one highly productive worker can coordinate many digital agents from almost anywhere, the economic premium attached to physically locating large teams in expensive cities may decline.424425## 10. Credit Markets Create an Amplification Channel426427White-collar workers do not merely consume.428429They borrow against expectations of future income.430431A standard household borrowing constraint might look like:432433$$434B_t \leq \theta E_t \left[ \sum_{j=1}^{T} \frac{Y_{L,t+j}}{(1+r)^j} \right].435$$436437Credit capacity depends partly on expected future labor income.438439If households begin to believe that certain professional incomes are less secure, lenders may eventually incorporate that risk into underwriting.440441That could lower borrowing capacity even before actual job losses occur.442443The mechanism becomes:444445$$446\text{AI exposure} \rightarrow \text{lower expected lifetime income} \rightarrow \text{lower credit capacity} \rightarrow \text{lower consumption and housing demand}.447$$448449Because developed economies are highly leveraged, this expectation channel could matter.450451Automation does not need to eliminate millions of jobs overnight.452453It only needs to alter expectations about the persistence of future earnings.454455## 11. The Entry-Level Problem May Arrive First456457One especially important transition mechanism concerns junior workers.458459Organizations often use junior employees for tasks that are simultaneously productive and educational.460461Drafting.462463Research.464465Data preparation.466467Basic programming.468469Financial modeling.470471Document review.472473Presentation preparation.474475Those tasks also happen to be among the easiest to delegate to AI.476477Suppose firms require experienced workers \(L_S\) and junior workers \(L_J\).478479Traditionally, today's junior employees become tomorrow's senior employees:480481$$482L_{S,t+1} = (1-\delta)L_{S,t} + \gamma L_{J,t}.483$$484485If AI substantially reduces demand for juniors,486487$$488L_{J,t} \downarrow,489$$490491then eventually:492493$$494L_{S,t+1} \downarrow.495$$496497This creates an institutional problem.498499Automation can remove the training layer from a profession before it removes the profession itself.500501In the short run, senior professionals become more productive.502503In the long run, fewer humans accumulate the experience required to become senior professionals.504505That dynamic could reshape law, finance, consulting, software engineering, academia, medicine, and many other knowledge industries.506507## 12. The Corporate Sector Could Become Extremely Profitable508509From the firm's perspective, AI can reduce unit labor costs.510511Define unit labor cost as:512513$$514ULC = \frac{wL}{Y}.515$$516517If output rises while payroll falls or grows slowly:518519$$520ULC \downarrow.521$$522523Holding prices approximately constant, operating margins increase.524525A simplified profit function is:526527$$528\Pi = PY - wL - rK - C_{AI}.529$$530531If AI substitutes for expensive labor such that:532533$$534\Delta C_{AI} < -\Delta(wL),535$$536537then profits increase.538539At scale, this could produce unusually high corporate profitability.540541The macroeconomic distribution becomes:542543$$544Y = W + \Pi + R + T + \cdots545$$546547with the profit share \(\Pi/Y\) potentially rising relative to the wage share \(W/Y\).548549Equity markets might therefore respond positively to developments that create significant anxiety in labor markets.550551That is not paradoxical.552553It reflects the difference between owning firms and working for them.554555## 13. GDP May Become a Worse Measure of Economic Security556557Imagine an extreme scenario.558559AI allows the economy to produce twice as many valuable services with half as much human labor.560561GDP increases enormously.562563Productivity increases enormously.564565Corporate profits increase.566567Consumer prices decline.568569Measured economic efficiency looks excellent.570571Yet many households experience lower labor income.572573We could then observe:574575$$576\frac{Y}{N} \uparrow577$$578579while, for a substantial fraction of households,580581$$582Y_{L,i} \downarrow.583$$584585GDP per capita rises while perceived economic security falls.586587That is possible because GDP measures production, not the distribution of claims on production.588589An AI-intensive economy could therefore force economists and policymakers to pay much more attention to:590591* median disposable income,592* household wealth,593* labor share,594* capital ownership,595* consumption distribution,596* and access to essential goods.597598Aggregate abundance does not automatically imply household security.599600## 14. Monetary Policy Would Face a Strange Problem601602Central banks are accustomed to recessions caused by weak productivity, financial crises, demand collapses, or supply shocks.603604AI could create something unusual:605606**a positive supply shock combined with a negative labor-income shock.**607608Potential output rises:609610$$611Y^* \uparrow.612$$613614Inflation falls:615616$$617\pi \downarrow.618$$619620But labor-market weakness increases:621622$$623u \uparrow624$$625626or wage growth slows:627628$$629\dot{w} \downarrow.630$$631632A standard policy rule such as:633634$$635i_t = r^* + \pi_t + \phi_{\pi}(\pi_t - \pi^*) + \phi_y(Y_t - Y_t^*)636$$637638would likely imply easier monetary policy.639640But lower interest rates do not solve the underlying distribution problem.641642Cheap credit can support aggregate demand.643644It cannot necessarily restore the bargaining power of labor if cognitive work has become structurally abundant.645646And lower rates may further increase asset prices, benefiting those who already own capital.647648Monetary policy could therefore become less capable of addressing the core economic consequences of automation.649650## 15. Fiscal Policy Becomes More Important651652Governments are heavily dependent on labor income.653654A simplified tax system can be written as:655656$$657T = \tau_L Y_L + \tau_K Y_K + \tau_C C.658$$659660If AI shifts income from labor toward capital,661662$$663Y_L \downarrow, \qquad Y_K \uparrow,664$$665666the tax base changes.667668This matters because many government systems were implicitly built around abundant employment.669670Payroll taxes finance social programs.671672Income taxes finance public services.673674Workers accumulate pension rights through employment.675676Health and other benefits can be linked directly or indirectly to work.677678If labor's share falls significantly, the fiscal architecture may eventually need to follow the income.679680The problem becomes less:681682> How should we tax robots?683684and more:685686> Where is national income actually accruing?687688A tax system designed for an AI-intensive economy may need to rely more heavily on profits, capital income, consumption, land, or broad ownership structures.689690## 16. The Real Problem Is the Transition691692The long-run equilibrium could be extraordinarily prosperous.693694Imagine a world where machine intelligence makes:695696* legal services,697* software,698* financial analysis,699* education,700* design,701* administration,702* and research703704dramatically cheaper.705706Real living standards could increase enormously.707708The difficult part is getting from the current economy to that economy.709710The transition can be represented as a race between productivity and redistribution:711712$$713g_{AI} \quad \text{versus} \quad g_R,714$$715716where \(g_{AI}\) is the rate at which AI changes productive capacity and \(g_R\) is the rate at which institutions, ownership, wages, prices, and new industries redistribute those gains.717718If719720$$721g_{AI} \gg g_R,722$$723724the transition becomes economically unstable.725726Not because AI fails.727728But because it succeeds faster than social and economic institutions adapt.729730## 17. New Jobs Are Not a Complete Answer731732A common response is that technology always creates new occupations.733734Historically, that is largely true.735736But the important question is not whether new jobs exist.737738It is whether they appear:739740* quickly enough,741* at sufficient scale,742* with similar compensation,743* and with skills that displaced workers can realistically acquire.744745Suppose AI destroys cognitive tasks at rate \(\delta_A\) and new labor demand emerges at rate \(\lambda_N\).746747A smooth transition requires something like:748749$$750\lambda_N \geq \delta_A.751$$752753But if:754755$$756\delta_A \gg \lambda_N,757$$758759even temporary adjustment can become a major macroeconomic event.760761A ten-year transition may eventually look harmless in a century-long economic chart.762763It can still define an entire generation's economic experience.764765## 18. The Most Important Variable May Be Ownership766767All of these mechanisms ultimately converge on one issue.768769Ownership.770771Suppose productive output increasingly depends on AI capital:772773$$774Y = F(K_{AI}, K_O, L_H).775$$776777As778779$$780\frac{\partial Y}{\partial K_{AI}}781$$782783grows, ownership of \(K_{AI}\) becomes increasingly important.784785If ownership is highly concentrated, AI may produce high growth alongside high inequality.786787If ownership is broad, households can receive AI-generated income through dividends, retirement funds, public investment vehicles, employee ownership, or direct entrepreneurship.788789The distributional question therefore changes.790791Instead of asking only whether people can continue selling labor, we may need to ask whether people participate in ownership of productive capital.792793A highly automated economy with broad capital ownership looks radically different from a highly automated economy with concentrated capital ownership.794795The technology can be identical.796797The macroeconomics are not.798799## 19. The Abundance Paradox800801This leads to a strange possible future.802803The economy becomes capable of producing unprecedented amounts of cognitive output.804805The supply of software, analysis, designs, research, administration, and digital services explodes.806807The marginal cost of many services approaches zero.808809Yet households can still feel financially insecure because access to that abundance remains mediated by income and ownership.810811The paradox can be expressed simply:812813$$814\text{productive abundance} \neq \text{income abundance}.815$$816817An economy can solve the technical problem of scarcity before it solves the institutional problem of distribution.818819That may be the central macroeconomic challenge of advanced AI.820821## 20. The Macroeconomic Question Is Larger Than Employment822823The debate over AI and jobs often assumes a binary outcome.824825Either AI replaces workers or it does not.826827The actual transition will probably be much more complicated.828829AI can simultaneously:830831* increase worker productivity,832* reduce the number of workers required per unit of output,833* lower consumer prices,834* raise corporate profits,835* reduce labor's income share,836* increase the value of scarce assets,837* weaken some urban housing markets,838* increase others,839* reduce inflation,840* alter tax revenues,841* increase wealth inequality,842* and raise real GDP.843844None of these outcomes are mutually exclusive.845846The most important macroeconomic identity remains simple:847848$$849Y = C + I + G + NX.850$$851852But AI potentially changes the mechanisms behind every component.853854Consumption depends on who receives the income.855856Investment depends on the expected returns to AI capital.857858Government spending depends on a tax system increasingly disconnected from employment.859860Net exports depend on which economies control the productive technologies.861862And aggregate supply could expand at a speed rarely seen outside major industrial transformations.863864The defining economic question of artificial intelligence may therefore not be whether machines become capable of doing white-collar work.865866That seems increasingly like a technical question.867868The harder question is what happens afterward.869870If cognitive labor becomes abundant, then the institutions of an economy built around scarce human cognition will begin to look increasingly strange.871872Wages.873874Careers.875876Mortgages.877878Education.879880Taxation.881882Retirement.883884Corporate organization.885886Even the relationship between employment and economic survival.887888All of these institutions were designed around a simple historical assumption:889890> **Useful intelligence is scarce because useful intelligence requires human time.**891892Artificial intelligence attacks that assumption directly.893894And if that assumption breaks, white-collar automation will not simply change the labor market.895896It will change the macroeconomic architecture built on top of it.897