We believe that a post-Artificial General Intelligence (AGI) economy could plausibly have a large share of new employment in important jobs like nurses, doctors, childcare workers, physical therapists, teachers and home health aides. 

Think about the possibilities. We could live in a world where kindergarten class sizes are 6; where it’s affordable for most people to regularly see a physical or mental therapist; where physicians and nurses have time for thorough diagnostics and intervention; where new parents could spend the first years of their children's lives caring for them directly if they so choose; or where we get to regularly experience live music. And, best of all, we could live in a world where the people producing these services are highly respected, well paid, and working reasonable schedules.

To be clear: even modest steps toward that world would be labor-intensive. Reducing average class size by even 20 percent at today’s enrollment levels would require about 920,000 additional teachers. Serving the bottom quartile of K–12 students with 90 hours of high-touch tutoring per year would take about 400,000 tutors if tutoring happens in groups of three.  This illustrates how quickly even incremental improvements in the intensity of education and care could translate into employment on the scale of major American industries.

At the same time, there is substantial causal evidence that this kind of human attention can generate large returns when it is well designed. A 2024 meta-analysis of 89 randomized tutoring studies found an average achievement gain of 0.29 standard deviations, roughly equivalent to moving a student from the 50th to the 61st percentile. The study saw especially large effects when tutoring was delivered by teachers or paraprofessionals, occurred at least three days a week, and took place during the school day. 

The Tennessee Project STAR experiment, which randomly assigned 11,571 students to classes averaging about 15 or 22 students, found that smaller classes improved academic achievement and increased college attendance at age 20 by 1.8 percentage points from a baseline of 26.4 percent. Labor-intensive services can produce measurable and durable gains when workers provide sustained, high-quality attention.

There is a polarized debate about what jobs will exist in a post-AGI world. One theory is that machines perform most cognitive labor initially, and eventually physical labor, leading to evaporation of human work. Governments would then theoretically respond with universal basic income or some other mechanism for distributing the proceeds, while people devote themselves to leisure or volunteer activities. To many, that sounds deeply unappealing.

But this view treats AGI as the end of the economic story when it may really be the beginning of a new one: a world in which people continue to work, but a growing share of them work in different types of roles. If AGI creates broad material abundance and reallocates workers from some existing occupations, one of the best uses of that capacity would be to expand human-centered care. This includes more work devoted to education, healthcare, child and elder care, disability support, mental health and end-of-life care. This might also include many forms of care that today are uncompensated and occur within the family.

The argument proceeds from what is already happening to what AGI could make possible. We first show that care is already claiming a growing share of employment and spending, then explain why AGI could accelerate that shift, what it might look like in practice, and the choices required to ensure that technological abundance produces more and higher quality human care.

The care economy is already growing and likely accelerating

This vision is rooted in demographic forces that are already reshaping the economy. In July 2026, health care added 22,000 jobs, even when the number of jobs contracted overall. The pattern extends beyond a single month: over the prior 12 months, health care averaged gains of 36,000 jobs per month and social assistance about 15,000—a combined pace of roughly 51,000 jobs per month. 

Forecasters expect this growth to continue. The Bureau of Labor Statistics expects health care and social assistance employment to grow by nearly two million jobs between 2024 and 2034. Home health and personal care aides alone are projected to add about 740,000 jobs, which represents nearly three times the projected increase in software developers.

The spending data point in the same general direction. The United States spent $5.3 trillion on health care in 2024, equal to 18 percent of gross domestic product. The federal government projects that amount will reach $9.0 trillion by 2034—an average annual growth rate of 5.4 percent—and account for 20.6 percent of GDP. Across other affluent countries, health and education likewise occupy similarly large and growing portions of public budgets and household spending, driven in large part by a demographic shift to an older population on average.

Home health and personal care aide employment is projected to reach about 5.1 million by 2034, roughly 17 percent more than in 2024. Expanding access, reducing workloads, or improving standards of care would require millions more.

Private-sector health care and social assistance employed 23.9 million people in June 2026, about 15 percent of nonfarm payroll employment. The care economy is one of the largest pillars of the labor market.

Of course, the recent growth of the care economy does not, by itself, tell us what will happen under AGI; it also has not translated into either better pay or working conditions for these workers or higher quality care for individuals. The Bureau of Labor Statistics projects about two million additional health care and social assistance jobs over a decade, much less than the role that care could play in a major AGI transition. In an AGI-enabled world, we might expect this not to evolve as a modest continuation of current trends, but instead to change more markedly, expanding education, health, elder care, childcare, counseling, and social support.

Why AGI could or should accelerate the shift toward care

Health care and social assistance are growing today for reasons that have little to do with frontier AI, including population aging, rising chronic disease burdens, and growing demand for paid child care. AGI could reinforce and expand the shift already underway through several distinct economic mechanisms.

If AGI raises productivity broadly across the economy, it should increase real incomes and reduce the price of many goods and services. Assuming the productivity gains are broadly distributed, households will likely redirect some of this gain toward services that were previously out of reach.

Societies effectively ration access to services like childcare and elder support because those services are expensive. Many families want aid for an aging parent, individual help for a child, or regular mental-health support that they currently cannot afford. As incomes rise—or as public resources expand—much of this latent demand could become actual employment.

For care, moreover, the relevant margin is often quality rather than basic consumption. Parents may not demand twice as many years of schooling for their children, but they may want smaller classes, more individualized instruction and better support for children with disabilities. Patients may not want twice as many hospital stays, but they may value longer appointments or closer follow-up after discharge. The expansion of care may therefore appear less as people consuming more units of the same service than as institutions providing a more attentive and labor-intensive version of it.

There may also be a Jevons-like rebound effect. When technology makes a service less expensive or more productive, the result need not be less labor devoted to it. Instead, lower costs and greater capacity can increase demand. At Stanford Children’s Hospital, faster reconstruction methods have reduced some pediatric cardiac MRI exams from roughly 90 minutes to 10. That can mean fewer labor hours per scan, but also more children scanned, shorter waits, less need for anesthesia, and more findings that require explanation and follow-up. In care, productivity can expand the frontier of what institutions are able to offer.

This argument need not depend on a permanent wedge between what humans and machines can technically do. Instead, people may continue to place special value on attention from a person, especially one who belongs to a shared institution and bears responsibility for an outcome. As Alex Imas has written, this could create a larger “relational sector,” in which human presence is itself part of what people value.

The evidence is consistent with that claim, though it does not settle what people will value in a post-AGI world. A meta-analysis covering 26 other meta-analyses and 2.6 million students found that teacher-student relationships were associated with achievement, motivation, belonging and well-being. There is some evidence that this “preference for a person” extends into health care. Randomized trials find small but significant benefits from stronger patient-clinician relationships, and most studies in a systematic review associated continuity of care with lower mortality. 

Schooling has absorbed wave after wave of technology while remaining labor-intensive in part because relationship, motivation and accountability are among its outputs, not merely inputs to be engineered away. But valuing human attention is not the same as being willing or able to pay enough to support millions of well-paid jobs, particularly if AI can satisfy some of that demand or if the gains from abundance remain concentrated.

AI can reduce administrative labor without reducing demand for human-facing labor. Automating documentation may reduce the number of hours spent moving information around. It does not follow that we want fewer teachers talking with students, fewer nurses explaining medications, or fewer aides helping their clients bathe safely. If anything, more efficient administration and higher incomes may increase demand for those interactions.

A vision for the future

Education: Imagine an elementary school in a post-AGI economy. Teachers and administrators can spend dramatically increased amounts of one-on-one time with students, providing more direct instruction and individualized lesson planning. Administrators can spend most of their time talking directly to teachers and parents. In the background, these professionals are using AI systems to generate lesson plans, translate materials into every language spoken at home, identify where each student is falling behind, and produce individualized exercises almost instantly. Much of the work that now consumes teachers’ evenings—and supports layers of administrative staff—could become dramatically easier to do. Schools might reorganize to employ fewer people performing routine administrative work, but many more adults working directly with children.

Now consider an example in healthcare. Imagine a nurse caring for patients who are, or are about to be, discharged from the hospital. An AI system could review records, reconcile medications, draft discharge instructions, monitor data from home devices, and flag patients at high risk of complications. It might recognize, for example, that an older patient with heart failure is gaining weight and has not filled a prescription.

Someone will still need to call the patient, speak with the next of kin, arrange a home visit, and make sure the patient understands which pills to take. In a more labor-intensive model of care, automation would allow that team—nurses, case managers and pharmacists—to spend more time on these tasks.

In both education and healthcare, AGI may make administration more efficient. As a society, we should use these savings to achieve a higher standard of service: smaller classes, earlier intervention, longer visits and more sustained human attention. The key question is then: how do we design institutional incentives to achieve these outcomes?

How do we get there?

None of this transition is guaranteed. Even if AGI makes more of these human services economically possible, ordinary market demand is unlikely to provide enough of it—or distribute it to the people who need it most. Care is not just another form of consumption. The need for it is often greatest precisely where the ability to pay is weakest. Economists often call this a market failure. Children cannot borrow against their future earnings to purchase better early education. Older people may develop substantial care needs only after their savings have been depleted and they are no longer able to work to earn money.

Care also produces benefits that extend beyond the person receiving it. Early education builds skills that benefit future employers and communities. Preventive healthcare can reduce later illness and public expenditure. Disability support can increase independence and allow both recipients and family caregivers to participate in the labor force. Because families do not capture all of these gains when deciding what they can afford, private markets are likely to provide less care than is socially valuable.

We must incentivize institutions to convert productivity gains into service quality and worker benefit rather than merely higher caseloads or lower staffing. It will also require incentives that reward follow-through, prevention, and long-term patient outcomes rather than only the volume of billable activity. Hospitals should work to give clinicians more time with each patient instead of more patients. Public purchasers could tie shared savings to access and outcomes; staffing and continuity standards could protect the human-facing part of the service; wage pass-throughs and gain-sharing could return some savings to frontline workers. The right mechanisms will differ by sector, but the principle is the same: efficiency gains should not flow only to lower headcount or higher margins.

Several policy questions deserve study now. Which reimbursement models turn faster documentation into longer visits rather than more visits? When medical imaging capacity expands, which additional uses improve health and which are waste? Do AI-supported schools add adult mentoring or simply cut staff? Which home-care models improve continuity, worker retention, and independence at sustainable total cost? Researchers should follow the productivity gains throughout the organization: who gains time, who gains access, and who bears new risks.

Second, financing systems must ensure that care is broadly available. Even in a world where AI increases abundance, greater redistribution and public provision will likely still be needed. In particular, if most gains from AI accrue to a narrow group of owners, demand for boutique education, concierge medicine, and highly customized elder care may soar while basic provision remains inadequate. The United States may have a less comprehensive welfare state than many other rich countries, but its largest and most durable social commitments are already concentrated in health, retirement, and education. Public schools, childcare subsidies, Medicare, Medicaid, disability programs, paid leave, and support for home- and community-based care are all potential channels through which an AI dividend could be converted into more widely shared services. The relevant policy questions will both be around how much income to transfer and which institutions can turn additional resources into reliable care.

Third, a labor surplus will not automatically become a care workforce. A displaced financial analyst cannot immediately become a nurse, teacher, or therapist. Many care roles require training, licensing, supervision, and interpersonal or physical skills that cannot be acquired overnight. Some people will not want to perform this work, and many existing care jobs are too poorly paid or precarious to attract workers with other options.

The goal should therefore not simply be to create more jobs in the care industry, but to create good jobs. The median hourly wage for home health and personal care aides was $17.21 in May 2025, equivalent to about $35,800 for full-time, year-round work. A successful transition would require higher wages, stable schedules, safer working conditions, and genuine career ladders. It would also require accessible training and pathways from entry-level support roles into nursing, teaching, therapy, and care management.

This is also a question of status. Care work has historically been performed disproportionately by women, immigrants, and workers of color, and its social importance has often exceeded the compensation and prestige attached to it. A post-AGI economy should not merely create more low-paid care jobs. It should recognize skilled caregiving as central economic infrastructure.

Fourth, the distributional challenge is global as well as domestic. Rich countries may spend their AI dividend on marginally useful medical procedures and highly customized services while poorer countries continue to lack vaccines, primary schools, and basic maternal care. Because much care must be delivered locally, technological abundance in one country will not automatically produce teachers, nurses, or clinics elsewhere. International financing and institution-building will still matter.

The choice abundance creates

AGI may make intelligence abundant. It will not make attention automatic.

A faster MRI can allow for more scans, shorter waits, richer diagnostics, and fewer sedated children—or simply a leaner cost center. An AI tutor can support a teacher who knows six students deeply—or justify putting sixty students in a room. An AI-powered ambient scribe can give a clinician back an evening—or add one more appointment to every hour. The technology can support any of these outcomes. Institutions choose among them.

The question is not whether machines will leave us anything to do. It is whether we use what they make possible to do more for one another. We think this is possible and worth achieving for the reasons outlined above, but it is not guaranteed. The deepest promise of this technology is not a world with no need for people. It is a world in which no one is denied the attention of another person simply because that attention costs too much.

– Alex Martin Richmond (Labor Economist) & Ronnie Chatterji (Chief Economist)

Let’s start with what we believe.

We believe strongly that AI is transformational. Our research and work by our colleagues at OpenAI shows that AI is already:

• Performing economically valuable work at a level as good or better than trained professionals

• Expanding what individual workers can do and produce

• Driving measurable productivity gains in the field

The technology is real, advancing quickly, and is on a trajectory to become one of the most important technologies of this century.

AI’s rapid capability growth trajectory means that measurement is critical to understanding its current and future economic effects. One dynamic that makes AI harder to measure compared to previous technologies is that every subsequent model improvement expands use cases and advances capabilities in new ways that are difficult to capture using traditional measurement tools.

AI is infinitely customizable and, as capabilities advance, each user’s experience becomes more distinct. We aim to help the public understand and react to these changes. This is why we emphasize the importance of building measurement infrastructure and sharing data with policymakers and researchers. Unlike in past economic transformations (imagine being an economist during the Industrial Revolution!), we have a wealth of data that tells us what is happening in real time. That means it is our job to think critically and carefully about interpreting that data and translating it into insights for a broad audience.

Humility and curiosity are key. Although we leverage some of the most interesting data in the world to explore these questions, we will not always arrive at definitive answers. There is a rapidly growing community of scholars studying the economics of AI and new results emerge every day. We can’t expect to be the only voice on these important topics. Instead, we should contribute to the broader research conversation and enable other economists to engage in these questions.

We believe in collaboration. Whether with government organizations, like the U.S. Department of Labor, or university researchers and institutions like the World Bank and the Partnership for AI, collaboration is key. We need perspectives from academics, multilateral organizations, governments, and the private sector to get this right. No one entity has the tools and resources to tackle these questions alone. We must work together. Incorporating these perspectives helps us keep people at the center of the AI transition and learn what kinds of information are most useful to the world.

We understand the difficulty of the task before us. Accurate economic measurement and forecasting have always been difficult. These efforts are even more challenging now that we stand at the precipice of a major technological shift. Our role is to proactively measure and inform, but not overclaim. Based on capabilities and adoption patterns, we are doing our part to measure the following phenomena:

• AI’s impact on the labor market, including changes in work, layoffs, and growth: Using our data, we examine jobs, workers, skills, new work, mobility, growth, and realistic scenarios for expanded human agency in an AI economy. Alex and Caroline are addressing these topics in their Work at the Frontier Series and my colleagues Kevin and Sonny are exploring these topics in their work on entrepreneurship.

• Changes in how organizations and firms change to adopt and adapt to AI: How organizations adopt AI, where value is created or blocked, and how tools like agents reshape workflows, productivity, and entrepreneurship. Work like B2B Signals from my colleagues G and Neel show these changes are unfolding in the enterprise data.

• Understanding how changes in the technology itself can create economic transformations: In this area, we explore frontier questions about research acceleration, science and innovation, recursive self-improvement, and the distribution of economic power. One example is this paper on agentic AI’s impact on work that Drew, David, Chris, Sonny, Alex, and Ronnie recently released in collaboration with our colleagues across OAI.

We are a small team within an AI lab, so we will not be able to generate every answer or serve as the sole source of truth. We aim to provide useful information that helps people around the world make better choices for themselves and their communities.

Taken together, we aim to make the ongoing transition to the AI economy more legible: to understand where AI is creating opportunity, where it may create disruption, and what workers, businesses, and policymakers may need to respond. This is important work and there is a lot more to do.

– The OpenAI Economic Research Team

I spent much of my academic career studying innovation and entrepreneurship, including organizations like Bell Labs and Toyota that defined innovation in their time. Now I sometimes wonder whether I am working inside the modern-day equivalent. Adam Smith used the pin factory to explain the division of labor; perhaps OpenAI is my pin factory. The balance between doing traditional research or simply talking to my colleagues about what they do is a tricky one. Should I produce cool charts on AI usage or simply breathe the air?

My team gets to do new things every day, and we do it on two clocks. On one clock, we do academic research: frame questions carefully, build new datasets, run analysis, and share the results with other economists. Peer review is slow and painstaking. It is also essential if we want to change how economists understand AI. Our paper How People Use ChatGPT draws on a privacy-preserving analysis of 1.5 million conversations to show who is using ChatGPT, what they ask it to do, and how those patterns are changing. Our newest paper, The Shift to Agentic AI: Evidence from Codex, documents the rapid shift toward agentic work inside OpenAI and compares it with adoption among individuals and workers in other organizations. Our next stop is a paper on how AI is changing work inside businesses, which will be presented at the NBER Summer Institute. And we have three more papers on the way.

On the other clock, we publish faster. OpenAI Signals and B2B Signals provide recurring public data, while shorter reports let us address questions that cannot wait for the full academic publication cycle. Recent examples include the AI Jobs Transition Framework and our report on entrepreneurship in ChatGPT. I want to put credible evidence into the world quickly because information gives people more agency: more ability to decide how AI should fit into their work, their organizations, and their lives. It also helps policymakers and business leaders act while the window for making decisions is still open.

This Substack is where I want to bring those two sides of the work together in plain language. Some posts will explain our own research. Some will use Signals data to explore a broader economic question. Others will discuss work by researchers outside OpenAI. Three principles will guide what we publish:

• Start with evidence. Use data where we have it, and separate what we observe from what we expect.

• Use economic logic to look ahead. When the evidence is incomplete, economics can help us reason about what may come next—as long as we are explicit about our assumptions and uncertainty.

• Make the work useful. Connect the findings to choices that workers, business leaders, and policymakers are making now.

For the first post, I want to start with one of the largest changes we are seeing right now: the shift from chatbots to agents. Our recent paper on agentic AI analyzes privacy-protected Codex data from individual users, organizational users, and OpenAI employees. By “agents,” I mean AI systems that can take on delegated tasks over multiple steps, use tools, revise their work, and return a completed output rather than a single answer.

That distinction matters economically. Most work is not a sequence of isolated questions and answers. It is made up of projects, workflows, handoffs, debugging, revision, coordination, and execution. Chatbots make individual moments of work easier. Agents can potentially reorganize the workflow itself.

We have seen that transition firsthand at OpenAI. We are not a typical organization: our employees have unusually early access to AI tools, strong incentives to experiment, and colleagues who can help them learn. But that is precisely why OpenAI can be a useful frontier case—a view of what may happen when capable tools are widely available and workers begin reorganizing parts of their day around delegation.

For the first few months after Codex was released publicly, ChatGPT remained the dominant AI tool for work inside OpenAI. Through August 2025, the average OpenAI worker spent less than 10% of their output tokens on Codex. Now every department—including Legal, Finance, and Recruiting—uses Codex as its primary AI tool for work. For the average OpenAI worker, Codex accounts for more than 85% of output tokens across these tools.

The more important question is not simply whether people use Codex more. It is what they are asking it to do. By May 2026, 80.6% of sampled individual users had made at least one Codex request estimated to exceed 30 minutes of human work; 70.2% had made one estimated to exceed an hour; and 25.6% had made at least one estimated to exceed eight hours. Human-equivalent task time is difficult to measure, so those estimates deserve care. But the direction is economically meaningful: as agents become more capable, users appear willing to delegate longer and more complex projects.

Adoption is spreading beyond engineering. Inside OpenAI, Legal, Finance, and Recruiting crossed into majority Codex use around April 2026. That does not mean non-developers use Codex exactly as engineers do. It means the boundary around agentic work is expanding. People outside software engineering use agents for automation, data transformation, structured analysis, documentation, and other workflows that once required more specialized technical support.

Agents may also help people cross task boundaries. More than one-fourth of Codex work by employees in business functions involved engineering or coding. A finance employee can build an internal tool. A lawyer can automate part of a document workflow. This does not eliminate the need for expertise. In many cases it makes judgment more important: someone still has to decide what to delegate, evaluate the result, and know when to bring in a specialist.

That pattern captures the larger reason economists should care about agentic AI. Productivity is not only about doing the same task faster. It is also about changing which tasks are feasible, who can initiate them, how work is divided, and where bottlenecks sit inside an organization. The Codex paper is one example of the research program I described at the outset: combine careful measurement with economic reasoning, distinguish evidence from anecdote, and ask what new capabilities mean for real decisions.

The economics of AI is still in its early days, and many of the most important questions remain open. Which tasks will be delegated, and which will remain human-led? How will adoption differ across occupations, industries, and firms? What skills become more valuable when workers can manage agents?

These are not abstract questions. Workers, business leaders, educators, and policymakers are making choices now. That is why we are starting this Substack: to share what the evidence says, use economics to think carefully about what may come next, and be honest about what we do not yet know. That is the spirit we hope to bring here. I hope you’ll join us.

– Ronnie Chatterji

Chief Economist, OpenAI

Subscribe now