What Will Be Scarce When AI Gets Cheap?

The easy story about AI and work is a sorting exercise. Put each task in a column: automatable, not automatable; exposed, safe; cognitive, manual. It feels useful until the thing being sorted is a cafe, a classroom, a clinic, or a consulting relationship.

A model can produce the lesson plan, the summary, the first legal draft, the translation, the diagnosis checklist, the data memo, and the customer-support answer. In many cases that is exactly the point: the output gets cheaper. But some services are not bought only for the output. People also buy the person or institution that stands behind it.

Here is the governing distinction: a human in the loop is not the same as a human in the product. A human in the loop checks, corrects, or approves machine output. A human in the product is part of what is valued: presence, provenance, trust, care, taste, judgement, accountability, relationship.

That distinction is already under pressure. AI companions, tutors, coaches, and support agents can imitate attention cheaply. So the scarce object cannot simply be warmth, conversation, or a friendly interface. It has to be narrower: provenance, accountability, embodied trust, liability, and stakes.

The first article in this sequence argued that AI adoption is not one curve. People use chatbots before firms reorganize workflows. Platform traces move before productivity statistics. Exposure measures tell us what AI could touch, not what it has already changed. Social pressure can push adoption before anyone has clean evidence that the tool is making them better off.

Once that map exists, the next question is not only whether AI is being used. It is where value moves if use deepens.

This is where Alex Imas’s essay “What will be scarce?” becomes useful. Imas is not making the familiar claim that some jobs are safe because AI cannot do them. His argument is sharper than that. He asks what happens if AI makes commodity-like output cheap and real income rises. In that world, people may spend more on goods and services where the human element is not a production flaw to be automated away, but part of what is being bought. ¹

The payoff is a cleaner way to read AI labor claims. Ask less often which jobs are “AI-proof.” Ask where human involvement is an attribute of the product, where synthetic substitutes are good enough, and where the people doing relational work can capture the value they create.

The claim is compelling. It is also easy to overstate. “Relational work becomes more important” is not the same as “relational workers become better paid.” A great deal of care work is socially vital and badly compensated. Teaching, elder care, hospitality, social work, and service work already show that importance does not automatically become income.

That gap — between importance and income — is what this essay takes up. The first article mapped what we can observe about AI diffusion. This one asks what becomes scarce if diffusion keeps going.

The Human In The Product

Imas starts from a strange fact about abundance. Some services are easy to mechanize in the narrow sense, but not easy to improve by removing people.

A coffee shop can standardize drinks, automate ordering, optimize throughput, and treat each customer as a unit in a queue. But part of what people buy in a cafe is not only caffeine. It is hospitality, atmosphere, recognition, a small ritual, a place to sit, the sense that someone is attending to them. Once the functional product is cheap and good enough, the nonfunctional layer can become the point.

That intuition travels beyond coffee. Education is not only information transfer. Therapy is not only advice. Care is not only task completion. Consulting is not only analysis. Leadership is not only instruction. In each case, some of the value comes from a person being responsible to another person in a particular context.

The boundary matters because many future-of-work claims smuggle one case into the other. Human verification can be useful without being what the buyer values. Human provenance can matter even when the functional output is easy to distribute or imitate.

The map avoids the lazy question: which jobs can AI do? The better question is: in which goods and services is human involvement itself a valued attribute?

For some tasks, human involvement is a cost. If a model can summarize a meeting, convert a spreadsheet, draft a first version, or answer a routine support question cheaply and reliably, the human labor saved is part of the gain. For other services, removing the human may remove part of the value. A patient may want a diagnosis that is technically correct, but also a clinician who listens, explains, notices fear, carries responsibility, and can be held accountable. A student may need information, but also attention, discipline, encouragement, and an adult who knows when they are bluffing. A client may need a model output, but also someone whose judgement they trust enough to act.

That does not mean AI is irrelevant to these settings. It means the relevant scarcity may not be “can the machine produce the answer?” It may be “who is trusted to stand behind this answer in a relationship?” Imas’s model is one way to make that intuition less sentimental and more economic.

The Model, And The Catch

The argument would be too soft if it stopped at examples. Imas’s technical note formalizes it as a structural-change story. ²

There are two sectors. One is an automated sector, where AI lowers prices. The other is a relational sector, where the human aspect is part of the value. If the relational sector has higher income elasticity than the automated sector, then rising real income shifts expenditure toward the relational sector. If relational work is also more labor-intensive, employment can shift there too.

That is the clean version of the model. The intuition is older. As societies become richer, they do not just buy more of the same things. Spending baskets change. Agriculture shrinks as a share of employment. Manufacturing rises and then falls. Services grow. Some of this is price change; some of it is income change. Comin, Lashkari, and Mestieri’s work on structural change is important here because they find that income effects explain a large share of historical sectoral reallocation. ³

Hubmer’s work points in a related direction. Higher-income households spend relatively more on labor-intensive goods and services. In a world where growth raises income, that can push consumption toward labor-intensive sectors. ⁴

Those are real supports for Imas. They are not the same as proof.

The key assumption is that relational goods have higher income elasticity over the relevant range. That is plausible in many cases. Rich households do spend on private education, hospitality, high-touch services, coaching, care, culture, restaurants, travel, performance, personal advice, and status goods. But “labor-intensive” is not identical to “relational.” A service can be labor-intensive because it is hard to automate, because regulation requires labor, because public budgets ration capital, or because productivity growth is slow. None of those automatically means human provenance is part of what consumers value.

The model is also about expenditure and employment, not wages. Imas is careful about this in the technical note. Employment in the relational sector can rise if spending shifts there and the sector uses more labor. But wages depend on supply constraints, institutions, bargaining power, credentials, public budgets, platforms, and who captures the surplus.

That is where the argument has to slow down.

Importance Is Not Wage Power

The tempting version of the argument is simple: AI makes cognitive output cheap; relational work becomes scarce; relational workers get paid more.

That is too quick.

Relational work becomes better paid only when several things line up.

Demand has to rise. That can happen because people have more income, because AI makes other goods cheap enough to free up spending, or because the relational component becomes more salient in a synthetic world.

The human element has to be valued rather than merely required. A nurse, teacher, therapist, adviser, coach, performer, or community builder may provide something AI cannot fully substitute. But a call-center worker following a rigid script may be treated as a cost even if the interaction is technically human.

Supply has to be constrained. The constraint can be time, trust, credentialing, reputation, location, tacit judgement, language, cultural context, or a durable relationship. If anyone can provide a superficial relational wrapper, wages will not necessarily rise.

Someone has to be willing and able to pay. This is obvious in elite advising, private education, high-end care, coaching, and professional services. It is much harder in publicly funded care, elder services, disability support, early childhood education, and social work. Public budgets can cap the wage even when social value is enormous.

Workers have to capture the surplus. If a platform owns the customer relationship, controls discovery, sets prices, and treats providers as replaceable supply, demand for relational work can grow while workers remain squeezed.

Miss one of these conditions and the Imas story can still produce more relational work without producing a wage premium. That is why the payout question should be conditional. Relational work may become more economically important where AI makes output cheap. Some relational workers may be paid more. But the broad claim that relational jobs automatically become better paid is not supported.

BLS projections are useful here only as a current, non-AI-specific employment baseline. They are not wage evidence, and they are not evidence of an AI-induced relational shift. For 2024 to 2034, BLS projects healthcare and social assistance to be the fastest-growing major industry sector, adding about 1.98 million jobs. Healthcare support occupations are projected to grow 12.4 percent; community and social service occupations 6.6 percent; services for elderly people and people with disabilities 21.0 percent. ⁵

Those numbers fit a world where care-heavy work grows. They do not prove Imas’s AI mechanism. Aging, chronic illness, disability services, public funding, insurance systems, and demographic change are doing much of the work. They also do not prove wage growth. A sector can add many jobs because demand is real and still pay poorly because budgets, institutions, and bargaining power are weak.

So the honest claim is narrower and more interesting:

If AI makes many cognitive and commodity-like outputs cheaper, the human component of some services may become more important. But whether that importance becomes income depends on market structure and politics.

Different Scarcity Questions

The useful comparison is not who is optimistic. It is what each theory treats as scarce.

David Autor’s AI argument is about expertise. Autor argues that AI could, if used well, extend the reach and value of human expertise. It might let more workers perform judgement-heavy tasks that are currently reserved for elite professionals: some kinds of medical decision-making, legal work, software work, and teaching. His hopeful case is not that AI replaces expertise, but that it helps more people act with expert support. ⁶

Imas overlaps with Autor because both resist the simple displacement story. Both care about human complements. But they locate the complement differently. Autor’s scarce thing is expertise and authorized judgement. Imas’s scarce thing is relational value: human presence, trust, provenance, attention, and accountability as part of the product.

Daron Acemoglu is asking another question: how large are the near-term macro effects of current AI? His answer is deliberately cautious. In “The Simple Macroeconomics of AI,” he estimates modest total factor productivity gains over the next decade from current AI task automation, especially once hard-to-learn tasks and context-specific judgement are taken seriously. He also warns that AI may widen the gap between capital and labor income unless it creates genuinely useful new tasks for workers. ⁷

That does not refute Imas. It disciplines the timing. Imas is asking how relative demand and scarcity might shift if AI makes some outputs cheap. Acemoglu is saying: do not assume current AI produces huge near-term aggregate productivity gains just because the demos are impressive. Both can be true. A modest macro effect can still hide meaningful sectoral changes. But if the near-term income effect is small, the near-term relational shift should not be oversold.

Anton Korinek and Donghyun Suh are asking a stronger transition question. In their AGI scenarios, wages depend on the race between automation and capital accumulation. If automation moves slowly enough, there can always be enough work for humans and wages can rise. If task complexity is bounded and full automation is reached, wages can collapse. Wages can also fall before full automation if automation outruns capital accumulation and labor loses scarcity value. ⁸

This is where Imas faces the hardest boundary. If AI can perform all human tasks, why would human labor retain value? Imas’s answer is that people may value the human element itself. A hand-made object, a human performance, an accountable therapist, a trusted adviser, or a real teacher may retain value even when a machine can provide the functional output.

That answer is plausible, but it may not be broad enough to protect labor as a whole. In very strong AGI worlds, relational work may become a status good, a protected institutional category, or a niche for people who can afford human provenance. It might not be a general wage floor.

Trammell and Korinek mark the aggregate boundary in their transformative-AI work: full automation can raise growth while lowering labor’s share, with wage effects depending on model assumptions. ⁹ Trammell’s later essay on whether labor is a long-run luxury sharpens the new-varieties objection. Transformative AI may not merely make a fixed set of goods cheaper. It may also create new machine-produced varieties that absorb spending. Looking at what rich people buy today may not tell us what everyone will buy after an enormous wave of AI-generated products and experiences. ¹⁰

That is the strongest theoretical counterargument to Imas. The income channel can point toward relational goods, while new AI-made goods pull spending elsewhere.

The comparison matters because these researchers are not simply disagreeing about one forecast. They are asking different questions:

• What could current AI do to productivity and inequality over the next decade?
• Could AI rebuild middle-class work by extending expertise?
• What happens to wages on the path to AGI?
• What happens to labor share if production and R&D are automated?
• What becomes scarce when commodity output gets cheap?

Do not force these into one answer. The differences are the diagnostic tool. Autor shows the expertise version of complementarity; Acemoglu disciplines the near-term income channel; Korinek, Suh, and Trammell mark the strong-AI boundary. And Trammell’s new-varieties objection points to the hardest economic counterargument: AI may not only cheapen existing goods; it may invent new machine-made things that absorb spending.

That is one way the Imas story can fail. The more direct way is that AI may compete inside the relational layer itself.

The Synthetic Relation Counterargument

The uncomfortable counterargument is that AI may not need to replace real relationship. It may only need to imitate enough of it.

AI companions already show the pressure point. Common Sense Media’s 2025 teen survey found that 72 percent of teens had used AI companions at least once and 52 percent were regular users. Among teen users, 33 percent had chosen to talk to an AI companion over a real person about something important or serious. ¹¹

That is real evidence that synthetic attention can substitute for some human interaction.

But the same survey complicates the story. Among teen users, 80 percent reported spending more time with real friends than with AI companions. A third had felt uncomfortable with something an AI companion said or did. The evidence does not say “AI replaces relationship.” It says people may use synthetic relationship even while still preferring human relationship, especially when access, embarrassment, cost, convenience, or loneliness matter.

That mixed result is probably the shape of the future.

In some settings, synthetic relation will be good enough. Many people will accept an AI tutor for drill practice, an AI coach for basic planning, an AI companion for late-night reassurance, an AI customer-support agent for routine problems, or an AI therapist-like tool when human care is expensive or unavailable. In those settings, AI can commoditize not only information but the appearance of attention.

In other settings, imitation may make human provenance more valuable. The more synthetic attention floods the environment, the more people may care who is actually present, who is accountable, who knows them, who can be trusted, and who has skin in the game. Authenticity can become more salient when imitation becomes cheap.

This is not a universal law. It is a fork.

AI may erode the low end of relational work by making cheap simulated interaction available everywhere. It may raise the premium for high-trust human work where stakes, accountability, embodied presence, credentialing, or reputation matter. And it may create many hybrid forms where one human supervises many AI-mediated relationships, which could increase the value of the human layer while limiting the number of human workers needed.

That last possibility is especially important. The future may not be “AI versus human therapist” or “AI versus teacher.” It may be one human professional plus many AI systems, wrapped in an institutional brand. The human becomes more central to trust and accountability, but less numerous per interaction.

That is another reason the wage claim is fragile. It also turns the question from psychology to institutions: who carries liability, who owns the customer relationship, and who decides when synthetic relation is enough?

Scarcity Moves Through Institutions

The relational-sector argument is not only about jobs. It is about institutions.

If AI makes back-office analysis, routine drafting, translation, coding, and planning cheaper, firms may compete less on producing generic outputs and more on trust, distribution, taste, service design, integration, and accountability. A consulting firm does not become valuable merely because it can produce slides or analysis. It becomes valuable if clients trust its judgement enough to act. A law firm is not only a document factory. A hospital is not only a diagnosis engine. A school is not only a content-delivery system.

The same point applies to development and geoeconomics. If more economic value moves into embodied, local, trusted, or institutionally mediated services, then countries do not benefit from AI only by gaining model access. They benefit by having care systems, education systems, public trust, professional institutions, managerial capacity, service quality, and legal accountability that can turn cheap model output into reliable human capability.

This connects back to the first article’s conversion frame. AI capability does not become economic capability by itself. It has to be converted through workflows, organizations, institutions, incentives, and people.

That conversion layer is often relational. Someone has to decide what to trust. Someone has to explain uncertainty. Someone has to notice when the model is wrong. Someone has to own the relationship with the customer, patient, student, client, team, or public. Someone has to be accountable when the output matters.

This is also where agentic data science and consulting become more interesting. If AI makes analysis cheaper, the scarce work moves up a layer. It is less about producing a chart or model and more about knowing which question matters, which evidence is trustworthy, which uncertainty changes the decision, and how an organization will actually use the result. The output becomes cheaper. The accountable judgement around the output becomes more important.

That does not make all consultants safe. It probably makes generic analysis more commoditized. But it makes the relational and institutional parts of advisory work more visible: trust, taste, synthesis, implementation, and responsibility. The next move is to state the claim narrowly enough that those differences do not disappear.

The Claim, Narrowly

The clean version of the argument is conditional:

If AI lowers the cost of many automatable outputs, and if rising real income shifts demand toward goods and services where human presence is part of the value, then relational work can become more economically important. If that work is labor-intensive, employment can shift there too. But wages rise only where demand is funded, supply is constrained, and workers or professionals capture the surplus.

That is less dramatic than “AI-proof jobs.” It is more useful.

It lets us say why some human work may become more valuable without pretending that social value guarantees pay. It lets us see why therapy, teaching, care, hospitality, advising, management, sales, performance, and local services are not all the same. It lets us compare a public caregiver, a celebrity therapist, a trusted consultant, a platform-mediated tutor, and an AI companion without forcing them into one category.

It also keeps the strong-AI question open. In a world of modest current AI, the relational sector may be one place where value shifts as some outputs get cheaper. In a world of transformative AI, relational scarcity may become more political: who gets human care, who gets synthetic care, who can pay for provenance, and which institutions protect human accountability.

That may be the more important version of the story. The labor-market question is not only whether AI can do the task. It is whether society treats human involvement as a cost, a luxury, a right, a credential, a liability, or a source of meaning.

That brings the argument back to scarcity, not safety.

What Remains Scarce

The adoption data tells us that AI is spreading faster as a tool than as a fully reorganized economic system. People are experimenting. Firms are slower. Productivity evidence is still uneven. Social pressure can move adoption before proof arrives. ¹²

Imas’s question begins where that map leaves off. If adoption deepens and more outputs become abundant, what becomes scarce?

Not intelligence in the abstract. Not “cognitive work” as a category. Not even human labor in general.

The scarce thing may be accountable judgement inside a relationship.

That phrase is deliberately narrow. It excludes a lot. It does not protect every job. It does not guarantee higher wages. It does not solve the distribution problem. It does not make synthetic relation harmless. It does not answer the AGI wage-collapse scenarios by itself.

But it does name something the simpler automation debate misses. There are parts of the economy where people are not just buying an answer. They are buying trust, responsibility, care, attention, presence, provenance, and the right to hold another person or institution accountable.

So the sorting exercise has the axis wrong. The question is not simply whether a model can perform the task. It is whether human relationship is part of the value, and whether the institutions around that relationship let workers capture anything from it.

Further reading:

• Alex Imas, “What will be scarce?”
• Alex Imas, “Sectoral Change, AI, and the Relational Sector”
• David Autor, “Applying AI to Rebuild Middle Class Jobs”
• Daron Acemoglu, “The Simple Macroeconomics of AI”
• Anton Korinek and Donghyun Suh, “Scenarios for the Transition to AGI”
• Philip Trammell and Anton Korinek, “Economic Growth under Transformative AI”
• Philip Trammell, “Is Labor a Luxury in the Long Run?”
• Common Sense Media, “Talk, Trust, and Trade-Offs”