If the AI does the work, who gets the income?

The question is not just who loses their job

The promise of an economy in which machines do almost everything seems to offer two opposing outcomes: shared abundance or extreme concentration of income. Both appear in discussions about artificial intelligence. None automatically follows from a model's ability to write, program, or converse. Between a technological demonstration and a lasting change in the standard of living there are companies, contracts, markets, infrastructure and political decisions. This dossier investigates this passage, without assuming that the end of the work has been demonstrated.

Our central question is distributive: when an organization delivers more value per hour of human work, who can transform this gain into income, time or security? The answer may involve workers, consumers, business owners, technology providers and governments. These groups also overlap. A person can receive a salary, pay for cheaper services, and own a small stake in a business. Therefore, boiling down the debate to humans versus machines hides the mechanisms we need to examine.

We organize the analysis into three levels. First, the technical exposure of tasks to AI. Then, the results observed in work environments and in income transfer experiments. Finally, institutional choices and illustrative scenarios. An exposure index helps locate potential changes; does not measure layoffs. An experiment informs about a population and an intervention; it does not reproduce an entire economy. A simulation makes assumptions visible; does not transform these premises into prediction.

Rovemark's editorial thesis is that productivity and distribution need to be studied together. This does not mean that we already know which policy solves the problem, nor that private gains are illegitimate. It means that an implementation can only be considered successful after also asking about quality, complete costs, working conditions and the destination of the value created. The same technology can underpin very different organizations and contracts.

Exposure to AI does not equal unemployment

The International Labor Organization and NASK estimated in 2025 that about a quarter of global employment is in occupations with some degree of exposure to generative AI. The index combines task evaluation, worker and expert contributions, and model-assisted classification. Participation reaches 34% in high-income countries and 11% in low-income countries. These percentages describe occupational composition and potential technical capacity; they are not individual probabilities of losing your job.

At the highest level of exposure, the index reports 3.3% of global employment. In this specific section, the proportions are 4.7% among women and 2.4% among men. The denominator matters: we compare the participation within the employment of each group, and not the division of all exposed occupations by sex. The graph allows you to alternate the selections, keeping some degree of exposure and the maximum category separate. We should not add up the panels or treat them as a temporal trajectory.

An occupation brings together activities with different demands. Producing a first draft can be automatable while negotiating an exception, verifying a decision, or assuming responsibility still requires human participation. Even when multiple steps become technically possible, adoption depends on cost, reliability, and integration. A company may choose to serve more customers with the same team, reorganize functions or reduce hiring. Identifying the technical possibility is the beginning of the investigation; Identifying which response occurred requires other data.

It would also be a mistake to conclude that lower exposure represents complete economic protection. A worker can be in an activity that is not very digitalized and still suffer indirect effects on demand, prices or supply chains. Conversely, a highly exposed function can become more productive and valued. Our interpretation is to use the index as guidance to study concrete changes, including groups that may need support. The source does not offer a date for the disappearance of professions.

Real productivity: evidence that does not fit into a single promise

In a study published in the Quarterly Journal of Economics in 2025, Brynjolfsson, Li and Raymond analyzed the staggered introduction of an assistant into an operation with 5,172 customer service agents. The average gain reported was approximately 15% in problems solved per hour. The benefits varied depending on experience and skill. This is evidence in one company and one occupation, with staggered implementation and econometric analysis; not from a representative sample of the entire economy nor from a trial with an individual draw.

The result demonstrates why it is worth studying AI within the work process. However, a gain in hourly resolutions does not automatically mean a 15% increase in profit, salary or national production. Implementation, review, software and organizational change costs need to be included in the account. The quality of the resolution also needs to be preserved. Transferring a percentage between these measures produces apparent precision that the study design does not provide.

A counterpoint comes from METR's experiment with 16 experienced developers and 246 tasks on known projects, using early 2025 tools. The availability of AI increased execution time by 19%, while participants expected a reduction in 24% before tasks and perceived a reduction in 20% afterward. The contrast measures expectations and perception regarding the result observed in that experiment, not a contradiction with the service study. Populations, tools, tasks, and metrics are different.

The METR update from February 2026 is essential to interpret this historic finding. The researchers began to consider that the selection of participants and tasks, changes in remuneration and difficulties in registering work with simultaneous agents hampered the interpretation of the new results. The 95% confidence interval reported for the original estimate ranges from +2% to +39% in time. We do not use the result of 2025 as a diagnosis of current tools. The applicable lesson is to measure the effect in the context of use, rather than relying solely on the sensation of speed.

Saving time does not determine who makes money

Imagine a hypothetical process that requires ten human hours per delivery and now requires eight, maintaining the same quality. The direct savings are two hours per delivery, but we still don't know the financial result. Maybe there is demand to sell more. Maybe the price will fall due to competition. Perhaps the company will maintain the price and increase its margin. Perhaps the team will absorb other tasks, or part of the gain will be negotiated as shorter hours. These are possible answers, not percentages estimated in this publication.

It is useful to separate three accounting questions. The first is how much resources were saved in the process. The second is how much additional value was actually sold or how much cost disappeared. The third is how this result was distributed after paying infrastructure, suppliers and taxes. Available hours are not realized revenue. An unused draft is not a sale. And a reduction in the price of a model does not necessarily represent a proportional drop in the total cost of a reliable service.

The OECD treats AI's diffusion and adoption capacity as important conditions for its possible macroeconomic effects. Sectoral structure, skills, infrastructure and economic integration change the scope of benefits. The projections discussed by the institution are the results of models, not already observed growth that we can fully attribute to AI. In this analysis, we do not transport a measured gain in service to a projection of Brazilian GDP.

Our distributive hypothesis is that property, competition and bargaining power condition the destination of the gain. A small business can gain access to previously expensive capabilities and lose margin if all competitors do the same. A worker can produce more without controlling the pay rule. A vendor may capture part of the benefit through dependence on its platform. To test these hypotheses, it would be necessary to monitor prices, contracts, remuneration and participation of suppliers over time. Just observing adoption is not enough.

Basic income: definition before adhesion or rejection

Universal basic income usually designates an individual, regular monetary transfer, without work requirements and aimed at everyone within the population defined by the program. The World Bank guide helps distinguish universality, unconditionality and cash payments. These characteristics do not specify a sufficient amount to live on, a source of financing or what benefits would continue to exist. The expression universal salary can be confusing: salary remunerates a work relationship; an income transfer follows another institutional logic.

A concrete proposal needs to answer who receives it, from what age, for how long and with what value update. It must also clarify whether the payment adds to existing services and benefits or replaces them, how high-income people contribute through the tax system and how extraordinary needs are met. Two proposals with the same monthly value can produce completely different net distributions. The name of the policy alone does not tell you who ends up better or worse.

The favorable argument can be based on autonomy, predictability and the ability to refuse bad working relationships. The opposite argument may point to the cost, opportunity of other policies and the inadequacy of a uniform benefit for different needs. We present these reasons as evaluation criteria, without assuming a single answer. A serious discussion compares specific designs, their net beneficiaries, and the services preserved. Agreeing with the importance of social protection does not require agreeing with any so-called universal proposal.

The relationship with AI is a hypothesis for organizing society, not a mandatory technical consequence. Even if automation advances, the answer may combine transfers, public services, training, social insurance, profit sharing or changes in working hours. The choice depends on the problems identified and the institutions available. In the selected sources, there is no experiment that observes an entire economy after AI eliminates the need for human labor. This scenario remains beyond the scope of the evidence gathered here.

What income experiments allow us to conclude

In the US experiment linked to OpenResearch, 1,000 low-income participants received $1,000 per month for three years, while 2,000 control participants received $50 monthly. The August revision of 2026 of the NBER working paper 32719 reports a reduction of 4.2 percentage points in labor market participation, of around US$ 1,900 annually in individual income excluding transfers and of one to two hours per week worked, relative to the control. These are different results, with different units.

The article remains identified as a working paper not subject to peer review by academic publication. The initial version cited in the first edition of this analysis had different estimates; We have now adopted the August revision of 2026 and recorded it in the bibliography. The review also informs that the moderate effects on labor supply were not offset by other measured productive activities. This does not turn less time spent on paid work into a judgment about participants’ personal worth.

The Finnish experience offers another design. Between 2017 and 2018, 2,000 unemployed people received approximately USD 647 monthly, in illustrative conversion using the ECB's annual averages of 2017 and 2018. The official evaluation describes small effects on employment and a more favorable perception of well-being. It also warns that a change in unemployment activation rules makes it difficult to isolate the effects in the second year. The well-being survey had different response rates between groups. The original benefit was denominated in EUR: the display in USD uses 560 multiplied by the arithmetic average of the ECB annual rates of 2017 and 2018 (1.15531784313725 USD per EUR), without adjusting for inflation and without reproducing the exchange rate of each payment. These limitations prevent treating the experience as a simple demonstration that any basic income always increases employment or necessarily improves all indicators.

None of these studies measure a permanent transfer to the entire population, financed by taxpayers themselves in an economy transformed by AI. A pilot can estimate individual effects of the transfer and still leave open prices, taxes, migration, service supply and company responses on a national scale. Our reading is that experiments inform decisions and correct absolute statements. They do not end the debate. The panel below presents the American design and the results of the current review without mixing participation, income and hours on a single axis.

How much would it cost? A transparent account, with limits

The gross cost of a transfer is a simple multiplication: monthly payment per person, number of beneficiaries and twelve months. The simulator uses USD as a monetary unit and allows you to change value and population. The standard of one hundred million people is a rounded didactic hypothesis, not an estimate of the Brazilian adult or eligible population. There is no automatic query for demographic data. The choice avoids presenting an invented population as if it had been measured. The scenario in USD is a new hypothesis, not a currency conversion of the previous example in BRL.

In the initial example, USD 1,000 per month for one hundred million people corresponds to USD 1.2 trillion per year. This is gross annual disbursement. It is not additional deficit, net cost or benefit recommendation. We do not discount revenue returned by beneficiaries, replaced benefits, administrative costs or behavioral changes. We also did not add a hypothetical collection about robots. All of these installments would require their own hypotheses and databases, in addition to evaluating distributional effects.

The funding debate requires at least two additional analyses. One is budgetary: which revenues and expenses change, over what period and with what administrative capacity? Another is economic: who supports the incidence of new taxes and how do families and companies respond? The legal payer of a tax is not necessarily the one who absorbs its final cost. We do not assume that the estimated value of a saved hour can be fully collected, nor that private wealth and public revenue are equivalent.

The OECD note on basic income draws attention to the financing and distributional implications of the reforms. This guides central care: a proposal that distributes to everyone and eliminates targeted support could harm people with greater needs. Our assessment requires showing the net result per group, not just the sum of payments. The graph makes visible the arithmetic scale of choices; it does not decide its fiscal viability or its effects on inflation. To answer these questions, a much more complete model would be needed.

Three ways to distribute the gains

The first scenario is concentration. Companies or suppliers with hard-to-replace assets retain much of the additional value. In this hypothetical path, production can grow while a portion of workers face loss of income or reduced negotiating capacity. The decisive variable would not only be the performance of the model, but the possibility of competitors and workers participating in the creation and appropriation of value. To observe this path, we would need to examine margins, prices, wages and contractual dependencies.

The second scenario is competitive diffusion. More accessible tools help small businesses, reduce prices and enable new services. Consumers can gain purchasing power, and some organizations can grow by serving previously unviable markets. This does not guarantee that every worker will receive a pay increase, nor does it eliminate difficult transitions. An assessment should monitor the entry of companies, quality, prices and professional mobility, checking whether the gains spread beyond the first organizations that adopted the technology.

The third scenario involves negotiation over time and protection. Part of the earnings can finance shorter working hours, training during working hours, participation in results or collective income and service mechanisms. There are choices about who participates, how to measure the outcome and how to distribute risks. A reduction in working hours announced without controlling the intensity of work may not improve the team's life. A transfer without adequate services may not address essential needs. These mechanisms need to be evaluated by design and execution.

The three paths are not mutually exclusive nor have they been assigned probabilities. They can coexist by sector, company and period. They serve to organize verifiable questions. If we want to know which path is prevailing, we must define in advance indicators that may contradict our preference. A hypothesis that explains any result after observation does not guide research. The editorial commitment is to monitor signals of concentration and diffusion with the same care, and to revise the interpretation when the evidence changes.

What a company can measure now

An organization does not need to solve the future of capitalism to improve its AI assessment. You can start by delimiting a process, a result and a minimum quality condition. In service, this may mean confirmed resolution with reopening rate; in documents, acceptance with human review; in operations, valid delivery without increase in incidents. Choosing only the volume of responses generated rewards activity that may not produce value. The metric should represent the completed task and the relevant consequence for the customer.

The protocol proposed by Rovemark starts with a baseline and records changes in demand and complexity. When feasible, compare similar groups or alternate usage conditions in a planned manner, without exposing people to uncontrolled critical decisions. The record needs to include review time, rework, cost of templates and tools, errors, interruptions, and training. A positive outcome that depends on exceptional oversight may not persist as the operation grows. The question is whether the gain survives everyday use.

Monitoring should include distribution within the team. Beginners and experts may respond differently; Averages can hide worsening in rare tasks or customers with specific needs. Remuneration, workload and autonomy are also results to be observed, with protection of personal data and access proportional to the need. The analysis should not become an opaque mechanism for individual assessment. Publishing aggregated metrics and usage rules helps make worker engagement more informed.

This is a proposed protocol, not a list of experiments already performed by Rovemark. To produce our own research, we still need participating organizations, advance definition of outcomes, consent and an adequate database. The intended contribution is to link operational evidence to distributional consequences. Instead of announcing that a certain tool replaces a percentage of the team, we want to show which tasks improved, which failed, what the full cost was, and who received the benefit. These are results that can be discussed and, when possible, reproduced.

What would change our conclusion?

Our position would change in the face of consistent evidence, in multiple contexts, that productivity gains spread with increases in income and quality of life without the need for additional mechanisms. It would also change if robust studies showed persistent job and income losses on a scale greater than the observed transition capacity. In both cases, it would be necessary to separate the effects of AI from changes in demand, the economic cycle and other technologies. Aggregated numbers without mechanism identification do not resolve this attribution.

On basic income, longer and broader program outcomes, with observed financing and market responses, would help reduce uncertainties that individual pilots do not address. The assessment should compare real alternatives, including improved existing systems, and show who wins and loses. We also need evidence about care, health, education and freedom of choice, without assuming that any activity outside of employment is unproductive or that every desirable effect has already been demonstrated.

The question in the title remains open-ended on purpose. The sources show that technical exposure, productivity and labor supply are measurable but different phenomena. They do not authorize a date for the end of employment or a universal value that any country can finance. Our editorial conclusion is to prioritize measuring the complete gain and its distribution. The quality of this agenda will depend on making the premises debatable and allowing new results to correct the narrative, including when this goes against the commercial interests of Rovemark itself.