AI, work, and three possible futures

Start with the work inside a job

A job title is a poor unit for deciding what to automate. An accountant might identify a duplicate invoice, resolve an ambiguous contract, contact a supplier, and explain a disputed balance. Those activities have different inputs, consequences, and standards of completion. A reliable extraction system does not automatically become a reliable negotiator.

O*NET’s task-based occupation search provides a useful starting vocabulary for breaking work into activities. For an actual deployment, validate that vocabulary with the people who do the job. Record frequency, time, exception rate, and who accepts the result. Count the review work too: automating entry can create a new queue of difficult exceptions for senior staff.

The scroll illustration below follows Anthropic’s nurse example, moving from today’s task bundle through augmentation, automation and new responsibilities. The squares explain categories, not a measured distribution of nurses’ work. A task can move between categories as the tool improves, the input changes, or the organization gives it different permissions. Document that change instead of assuming a permanent label for an entire profession.

Five assumptions deserve five separate questions

The scenarios span different historical scales. The modest case describes a gradual productivity improvement comparable in scale to past digital technologies. The substantial case reaches roughly half of knowledge-work tasks technically, while actual deployment covers only part of them. The extreme case assumes rapid adoption, very high autonomy and essentially no new knowledge tasks for people. The source associates that last path with recursively improving AI; it is a conditional experiment, not a claim that such a path will occur.

The original explorer asks about capability, adoption, autonomy, productivity, and adjustment. Its design is valuable because these are different decisions. A system may complete a demonstration successfully without being deployed widely. A widely deployed assistant may still require a human to approve every consequential step.

In our interpretation, a planning team should translate those dimensions into evidence it can inspect. Capability needs a representative acceptance test. Adoption needs the share of eligible work actually completed with the tool. Autonomy needs a clear account of retained human decisions. Productivity needs end-to-end effort per accepted result. Adjustment needs evidence about what people do when their original tasks disappear.

Do not move all five assumptions together merely because a new model scores well. A document assistant could improve its answers while adoption remains constrained by inaccessible records. A service team could adopt quickly while review costs absorb the time saved. The comparison below shows the published scenarios. The five-question simulator on this page lets you change your own assumptions, recalculate every result and compare your answers with the source survey.

Output: separate the level from the growth rate

“More output in 2030” and “faster annual growth entering 2030” answer different questions. The Bureau of Economic Analysis explains GDP as a measure of production; inflation-adjusted comparisons separate output changes from price changes. The dollar amounts in our comparison use the source’s 2025 price basis.

When presenting a scenario to a board, put the comparison baseline and the date beside the number. A cumulative gain against an alternative economic path should never be presented as an annual return on an AI investment. National output is also not a forecast of your company’s sales: competition, pricing, capacity, and market access sit between the two.

For an internal business case, construct a separate operating model. Start with accepted units of work, realized demand, and the resources needed to deliver them. If a team finishes twice as many drafts but customers buy the same number of finished products, the extra drafting capacity has not automatically created twice the revenue. Preserve that distinction in both the dashboard and the budget.

Employment: exposure is not an unemployment rate

The branching diagram uses a different denominator from the unemployment charts: people employed in 2026, normalized to one hundred. In the substantial scenario, the knowledge-work group starts at 62.2 and ends at 59.7. Of the 2.5 displaced, 1.8 move into other occupations and 0.7 still need to move by 2030. These are aggregate changes, not a record of individual workers’ journeys. Normal job-market churn is excluded so that the additional transition remains visible.

Skills and vacancies matter on both sides of the move. A role can disappear before an alternative employer has a suitable opening. A former office worker may need training, certification or a move to another location. The model therefore distinguishes the demand for work from how rapidly people reach the jobs that are available.

Task exposure, positions removed, people changing occupations, and unemployment are different quantities. Under the BLS measurement framework, unemployment concerns people without work who are available and actively seeking it, with specified exceptions. People outside the labor force belong to a different category. An exposed task is not itself an unemployed person.

This distinction changes what an employer should monitor. Count workers moving into another internal role separately from departures, reduced hours, and vacancies left unfilled. Follow the transition after a training course ends. A completed course is an input; a durable role with suitable pay and conditions is an outcome.

Our recommendation is to give every proposed workflow change a transition owner and a review date. Identify the skills needed in the destination role, the supervised practice available, and the capacity to absorb people there. A redeployment plan without available work is not yet an operational plan. These are organizational recommendations, not results measured by Anthropic.

Wages: an average can conceal opposite experiences

The source explains this divergence through demand and the time needed to change occupations. Faster design and permitting can support additional building projects, increasing demand for construction workers even while fewer people are needed for some office tasks. Until enough people can move into the growing occupations, their wages rise. In the substantial case, knowledge-worker wages are approximately flat; in the extreme case they fall by 11.5%, while wages in the other group rise by 33.6%.

The technical report separates cognitive occupations from other work. In its extreme case, their wage changes have opposite signs. Our comparison preserves both groups because a single average would make the distribution hard to see. The research paper’s Table 3 is the source for the detailed rates, wage changes, and income shares.

For a company, compare pay outcomes by role and experience as well as the overall payroll average. If junior positions disappear while higher-paid specialists remain, average pay can rise without any continuing employee receiving a raise. If the remaining specialists inherit more exception handling, nominally improved pay can also coexist with a more demanding job.

Report hours and employment alongside wages. A worker paid more per hour but offered fewer hours may earn less overall. Likewise, the wage of someone still employed tells us little about the income of someone between jobs. This is a reason to retain several measures, not a reason to reject all averages.

Distribution: more value needs a route to people

The model begins with labor receiving 60% of income and capital 40%. By 2030, labor’s share is 59.4%, 56.1% or 45.2% across the three cases. Automation increases the usefulness and demand for capital, so its owners can capture a greater part of the gains. In the extreme case, a 32.4% larger economy nearly offsets the drop in labor’s share: total labor income is close to its no-AI level despite the much larger economy. That is why GDP alone is an incomplete description of prosperity.

Labor’s share is a fraction of total income. To understand it, examine the numerator and denominator together. A falling share can accompany rising labor income if total output expands enough. That arithmetic alone does not tell us which households benefit, how much bargaining power workers retain, or how stable their income becomes.

For organizational planning, write down where an efficiency gain could go: lower customer prices, shorter hours, expanded service, employee compensation, investment, or owner returns. These are choices to make explicit before claiming that “everyone benefits.” The outcome may involve several routes, with different timing and different beneficiaries.

One practical governance exercise is to attach a distribution note to every large automation proposal. State who does less repetitive work, who takes on more judgment, who bears the transition cost, and who receives the financial gain. Review that note after deployment with employees’ feedback and operating evidence. This is ZharfAI’s suggested decision process, not a policy generated by the model.

Survey answers are beliefs, not outcome frequencies

The source reports that typical survey responses imply roughly 10% more GDP by 2030 and unemployment around 5%; about one in ten respondents had expectations consistent with the extreme case. The five distributions above retain the public and site-visitor samples separately. Change an answer in the local tool to move your marker across those distributions. The visitor series is a September 24 snapshot, not a live poll of this website.

Anthropic surveyed 10,980 US adults; the typical answers produced outcomes near its substantial scenario. A respondent’s expectations are inputs to a model, not observations of an economy that has already happened. Visitors answering an online explorer also form a different population from a representative public survey.

Use a similar exercise inside a company to expose disagreement. Ask operations, engineering, and employees to answer independently before discussing a shared scenario. When views differ, trace the disagreement to an assumption that can be investigated. Perhaps one team expects clean digital inputs while another knows how many records arrive as photographs or incomplete messages.

Avoid treating the most popular scenario as a commitment. The useful output is a list of assumptions, owners, and observable signals that would change the plan. Keeping that record lets a team learn from its own expectations instead of rewriting its original story after events unfold.

What this model leaves outside the frame

The paper excludes important channels including business cycles, policy responses, financial disruption, and advanced robotics. It also uses broad occupation groups. That scope is a boundary around the exercise; it cannot settle every question about AI’s social consequences.

For our interpretation, the appropriate response is to maintain a separate risk register rather than silently assume omitted channels are harmless. Record supply interruptions, demand weakness, customer trust, and implementation failure where they matter to the business. Do not attach invented numerical adjustments to the published scenarios and continue calling the result Anthropic’s model.

Scenario labels can create false precision. “Extreme” does not mean impossible, “modest” does not mean safe for every worker, and a middle case is not automatically the most likely. A responsible decision should be understandable even if the reader disagrees with the chosen future. Prefer investments that remain useful across several plausible conditions and identify what would cause you to stop.

What reviewers challenged

External review improved the treatment of returns to capital and the divergence between wages in the two occupation groups. Those reviewers were not asked to endorse the results. Their remaining disagreements are part of the research story: exposed occupations might expand rather than contract; the model’s two broad groups cannot capture each displaced worker’s costs; and the most extreme case may be more useful as a thought experiment than as a central planning scenario.

Other open questions concern the scale of productivity effects already visible, demand generated by data-center construction, and whether AI could accelerate innovation more than the model allows. A model that omits financial instability, aggregate-demand shocks, policy reactions and advanced robotics cannot answer those questions by changing one slider. Anthropic describes this as version 1.0 and expects the framework to change as evidence accumulates. Keep that version and its boundaries attached to any exported interpretation.

Reading a US model from Iran

The US calibration is not a forecast for Iranian employment, salaries, or output. A useful local analysis would require local evidence about occupations, wages, adoption, access, and the institutions that mediate transitions. Translating the interface into Persian does not translate the calibration.

For an Iranian organization, start with a bounded workflow and accessible tools. Evaluate Persian documents, mixed numeral systems, incomplete records, and the actual permissions available in the intended environment. Record costs in the currency and payment conditions the organization really faces. Keep access assumptions explicit rather than depending on an unverified service becoming available later.

Our articles on measuring productivity through review and rework and evidence-first automation develop the operational side of this approach. They help connect an economic question to an acceptance test without claiming that a local pilot can validate a national macroeconomic scenario.

A practical response under uncertainty

For a first review cycle, we recommend three deliverables: a task inventory, a measured pilot, and a worker-transition plan. Give each one an owner. The inventory defines what can change; the pilot tests whether the change improves accepted work; the transition plan explains what happens to the people involved.

Publish a small set of outcomes together: completed work, quality failures, total human effort, customer impact, and changes in roles or hours. Set the review interval before the pilot begins. Preserve failed cases and stop conditions; otherwise the next meeting will contain successful demonstrations with no denominator.

The wider public-policy discussion is separate. Anthropic’s June 2026 policy framework discusses monitoring and responses that vary with disruption, including support for transitions and income. Those are proposals, not enacted measures or proof that a particular intervention will work everywhere. This feature provides research interpretation, not personal investment advice.

Source notes — reviewed September 24, 2026

The comparison uses the source’s September 2026 version 1.0. Exact economic rates come from Table 3, printed page 31, with outcomes at the start of 2030; annual growth covers the preceding twelve months. Dollar output uses the companion essay. The animated preset paths use monthly outputs recovered from Anthropic’s published model, checked against Table 3 at its reported precision. The local calculator independently implements the published equations and calibration; its preset and custom-input results are checked against the source engine. All five questions, custom results, worker flows and survey comparisons run within this article. This is not an independent validation of the model. Our accounting example and organizational recommendations are original analysis.

• Anthropic’s economic scenario essay and original interactive explorer: scenario presentation, dollar output, survey overview, cover artwork, and colored task marks. Original visual credits include Kelsey Nanan, Nikki Makagiansar, and Monika Tuchowska. The original cover is reproduced without alteration; task marks retain the source shapes and colors.
• Korinek, Jones, Sacher, Cotter, and McCrory: Economic Scenarios for Transformative AI, September 2026: model, assumptions, Table 3, and limitations. External reviewer feedback is not endorsement.
• Anthropic: A policy framework for AI’s impact on work, June 2026: separate policy proposals.
• O*NET task-based occupation search: task vocabulary and occupation matching.
• BEA: Gross Domestic Product: output and inflation-adjusted measurement.
• BLS: How the Government Measures Unemployment: employment-status definitions.