How McKinsey’s AI Transition Estimate Changes With Its Assumptions

How wide is the reported range, and what does it actually count?

McKinsey’s September 29, 2026, base case estimates that about 11 million US workers in occupations with declining labor demand may need to move to different occupations by 2035. Its tested alternatives put the total at about 6 million when automation adoption is slower and its effect on labor demand is smaller, and more than 16 million when adoption is faster and labor substitution is greater. That is a span of roughly ten million or more across the reported cases. It is large enough to make the assumptions matter, but it is not a probability interval, a precise bound on what will happen, or a count of certain layoffs. “Workforce in motion: Skills and pathways to future jobs in the United States” calls these sensitivity tests, not predictions with assigned odds. The unit being counted is narrower than a headline may suggest. McKinsey defines a transitioning worker as someone who moves into a different occupation after labor demand declines in the worker’s existing occupation. The estimate is about possible cross-occupation movement in a modeled US labor market, not every employee whose daily tasks change, every hour technology might technically automate, or everyone in a job family where some tasks are exposed. A worker who stays in the same occupation while duties change is not automatically part of the 11 million transition figure. These are modeled US-wide estimates through 2035, not observed counts or employer staffing plans. Their scale describes a possible reallocation problem across occupations, not a calendar of individual exits. The base case is the central modeled result under McKinsey’s chosen assumptions. The lower and higher figures are alternatives built by changing assumptions about adoption speed and how much automation reduces labor demand. The report says faster adoption combined with greater substitution increases transition needs; slower adoption or greater augmentation reduces them. They do not show that the endpoints are equally likely, that the true result must fall between them, or that national results translate into a personal percentage. That framing keeps three questions separate. Could a technology perform or assist with parts of a task? Will organizations deploy it in their workflows, and how quickly? After deployment, will employers need fewer workers, or use saved capacity in other ways? The transition total sits downstream of all three. It is not a direct measure of technical capability. McKinsey says automation does not translate one-for-one into job losses and distinguishes work hours potentially absorbed by automation from reduced labor demand. Axios’s September 29 coverage, “How AI could change the workforce,” leads with the roughly 11 million base estimate and reports an annualized pace of 770,000 occupational switches. Annualizing a modeled decade-scale total does not make it a timetable for an employer or a forecast of when a particular worker must move. The national total also does not say whether a destination is feasible for someone with a given wage floor, location, credentials, available time, health, or family responsibilities. For a worker deciding what to do, the span supports attention, not an immediate career pivot. A result that changes substantially under alternative assumptions is a reason to watch actual deployment in recurring tasks and evidence of how employers absorb the resulting capacity. It cannot supply a personal risk score or tell someone to discard useful experience. The careful reading is that McKinsey’s September 2026 model puts the base case near 11 million cross-occupation transitions by 2035, while tested assumptions produce about 6 million to more than 16 million. The width reveals meaningful uncertainty; it does not say which scenario will occur, when change will reach a workplace, or what one worker should do next.

Sources: Workforce in motion: Skills and pathways to future jobs in the United States; AI may cause 11 million Americans to change occupations

Which two assumptions move the estimate?

The estimate moves along two separate dimensions: how quickly organizations adopt automation in real workflows, and how much that use reduces demand for human labor. McKinsey’s “Workforce in motion: Skills and pathways to future jobs in the United States” says faster adoption paired with greater substitution raises modeled transition needs; slower adoption or greater augmentation lowers them. These are conditional alternatives, not forecasts with assigned odds. Its cases test sensitivity, not precise upper and lower bounds. It gives no independent contribution for each assumption. The adoption axis is about implementation, not simply whether software can perform an activity in a demonstration. McKinsey evaluates detailed activity and occupation combinations, Its framework considers whether an activity is codifiable, repetitive, and easy to verify; how much workflow reconfiguration is needed; and whether regulation, social acceptance, business criticality, or reputational risk may constrain deployment. These factors inform a directional adoption estimate, not an employer’s deployment date or staffing decision. McKinsey describes codifiability, repetitiveness, ease of verification, and required reconfiguration as feasibility considerations. It separately considers regulatory constraints, social acceptability, business criticality, and reputational risk. Likewise, a tool can be adopted as assistance while the employee remains accountable for exceptions and final decisions. The adoption estimate therefore describes potential use across work hours, not a simple count of tasks transferred wholesale from people to software. Adoption also depends on workflow maturity, data quality, operating discipline, leadership priorities, and business context. A demonstration shows capability; use in an actual workflow shows adoption, but not staffing impact. A pilot may cover one task while workers retain review responsibility. The second axis asks what happens after work is automated or augmented: does the employer need fewer labor hours, or does it use the released capacity another way? McKinsey separates automated hours from reduced labor demand. Saved time may reduce staffing, but it may also reduce excessive hours, support more output as productivity lowers costs, or shift effort toward checking outputs, handling exceptions, and overseeing systems. Barriers may slow head-count reductions. Thus adoption does not automatically equal job reduction; the model estimates net effects after offsets. Substitution is also sensitive to demand: if lower costs lead an organization to serve more customers, labor savings may partly support expanded output. Where demand is fixed or weak, the same productivity gain could be treated differently. The scenario spread therefore is not a direct conversion from “AI can do this” to “a worker will lose a job.” Adoption estimates the speed and breadth of feasible work entering workflows; substitution estimates how much resulting productivity change becomes lower labor demand. The assumptions interact, but the report does not assign a precise number of transitions to either independently. For a worker, ask: is automation being used in recurring work, and what is happening to the capacity it frees? Output, service, overtime, review duties, vacancies, and staffing changes can signal different ways capacity is absorbed. One sign by itself is ambiguous; a pattern over time, understood alongside the workflow, is more informative than capability alone.

Sources: Workforce in motion: Skills and pathways to future jobs in the United States

Why does adoption not translate one-for-one into fewer jobs?

In McKinsey Global Institute’s September 2026 report, “Workforce in motion: Skills and pathways to future jobs in the United States,” the base case separates two quantities that are easy to collapse. By 2035, automation could absorb about 54 percent of current US work hours, while the estimated net reduction in labor demand is about 21 percent. McKinsey says organizational and market mechanisms offset about 60 percent of the labor impact associated with adoption. These are model estimates, not measured outcomes or a prediction for any one employer. The first figure concerns work hours that technology may automate or meaningfully augment; the second concerns the portion of that change that leaves employers needing fewer human work hours. A tool completing an activity is not, by itself, evidence that a position disappears.

The report’s bridge includes four mechanisms. Saved time can ease excessive workloads or overtime where staffing is tight, instead of reducing head count. Greater productivity can lower the cost or improve the quality of a service, leading an organization to provide more of it. McKinsey gives auditing as an example and cites a study reporting a 4.3 percent increase in auditor employment after AI adoption, alongside improved audit quality. That case illustrates one possible response to productivity; it does not establish that employment rises in other fields or that AI caused every aspect of the change. Adoption also creates work around implementation, supervision, review, exception handling, and workflow redesign. Finally, regulation, contracts, labor agreements, or social expectations can slow or limit displacement. These are modeled offsets, not guarantees that workers retain their roles.

Work is also bundled. McKinsey notes that activities in an occupation are interdependent: automating one piece may leave judgment, coordination, and accountability with a person. Imagine an analyst using a system to prepare a first draft of a recurring report. If saved time goes toward checking exceptions, advising colleagues, or producing more analysis, the workflow changes without an automatic one-for-one reduction in jobs. By contrast, a high-volume activity that can be separated cleanly from the rest of a role may be easier to convert into lower labor demand. These examples illustrate the report’s mechanisms; they are not observed cases or promised outcomes. Which pattern applies depends on demand for the work, organizational constraints, and how released capacity is used.

That conversion is why the report’s roughly 6-million-to-more-than-16-million transition span should be read as conditional on adoption and substitution assumptions, not as a measure of technical exposure alone. The 54 percent figure is not a share of workers expected to be laid off, and the 21 percent estimate does not say which occupations or employers will reduce staffing. The useful question is what happens after a task becomes faster: does the employer expand output, improve service, add review duties, reduce overtime, or reduce paid hours and vacancies? Those outcomes are not interchangeable. Evidence of capability identifies a possible change; evidence about actual workflow and sustained labor demand is needed to understand employment implications.

Sources: Workforce in motion: Skills and pathways to future jobs in the United States

What sits between reduced labor demand and a worker changing occupations?

McKinsey’s base accounting estimates that automation could reduce labor demand by the equivalent of about 36 million jobs by 2035. It then estimates that growth within workers’ existing occupations could offset the effect for roughly 25 million, leaving about 11 million who may need to move to another occupation. This is a modeled chain of labor-demand changes, not a tally of 36 million people receiving layoff notices. The distinction matters: reduced demand measures fewer full-time-equivalent workers needed for work, while a transition count concerns workers whose current occupation is modeled to decline after the report nets several forces together. The report describes these figures as model outputs under base assumptions, not observed changes. [McKinsey, “Workforce in motion: Skills and pathways to future jobs in the United States”](https://www.mckinsey.com/mgi/our-research/workforce-in-motion-skills-and-pathways-to-future-jobs-in-the-united-states).

The arithmetic makes two adjustments visible. Labor demand can fall in an occupation without every worker crossing into a new one: growth from other sources may absorb some automation effects, even as tasks change. Conversely, demand in expanding parts of the economy does not automatically cancel a decline in a particular occupation. The 11 million is the remainder after within-occupation offsets, not a direct count of automated activities. A “transitioning worker” moves to another occupation after demand declines in the current one. That is narrower than everyone whose tasks change and does not mean each modeled worker is involuntarily displaced.

Timing complicates this long-run balance. A task may be automated before an employer expands output, adds complementary work, or absorbs freed capacity. In that interval, hours, vacancies, or staffing can change. Workers may also move through ordinary turnover, retraining, or a change of employer rather than one abrupt event. The report estimates employment composition from 2025 to 2035; it gives no individual calendar or guarantee that offsets arrive before a worker must decide. Its national accounting describes possible reallocation, not the sequence at any one workplace.

A growing occupation is not automatically a workable destination. McKinsey evaluates pathways using destination demand, skill overlap, wage preservation, and time to acquire legally required credentials. It estimates almost half of workers may face an unpaved pathway, while only about one in seven has a direct pathway requiring no or very limited retraining. These are modeled assessments, not observed placements or guarantees. An opening may be impractical if a worker lacks a required license, cannot take the training time, would earn below a household pay floor, or cannot relocate or change schedule. Prior experience may transfer skills, but fit must be checked against the role’s requirements.

For a worker weighing a response, the question is whether a plausible destination has demand locally, uses existing skills, meets essential pay and schedule constraints, and offers a training route the worker can complete. That comparison may make an adjacent move more feasible than a larger change, or show that redesigning the current role is nearer-term. These tests follow the report’s pathway dimensions; they are not individualized recommendations from its national model. The 11 million figure signals reallocation pressure at scale. A personal decision depends on reachable work and the cost of getting there.

Sources: Workforce in motion: Skills and pathways to future jobs in the United States

How should the model be compared with observed employment evidence?

The OECD’s 2023 review found little detectable negative employment effect in the earlier periods it examined. McKinsey’s September 2026 model, by contrast, estimates possible occupational transitions through 2035 under stated assumptions. These findings complicate a simple displacement story, but they do not directly confirm or refute the newer scenario range: one summarizes observed historical evidence, while the other models conditional future outcomes.

The evidence reviewed in OECD Employment Outlook 2023 largely predates the recent advances in large language models and generative AI. Its chapter reports that empirical studies using cross-country differences in AI exposure or differences across local labor markets did not find statistically significant employment decreases. It also describes a U.S. Census Bureau survey of a random sample of more than 300,000 employer businesses, covering advanced-technology adoption from 2016 to 2018. Most firms reported no employment change attributed to these technologies. Among firms adopting AI, 26 percent said AI had increased employment, while fewer than 10 percent reported a decrease. These are firm reports from an earlier period, not a count of affected workers or a controlled forecast of later generative-AI effects. The OECD explicitly cautions that the evidence does not account for recent generative-AI advances; it also notes that slow adoption, implementation time, and reliance on natural attrition could delay employment effects. Historical stability therefore limits claims about what had happened by then, not what must happen next.

The International Labour Organization’s 2025 refined global index answers a different question again. Its assessment combines nearly 30,000 occupational tasks with expert validation, AI-assisted scoring, and harmonized labor-force data to estimate potential exposure to generative AI. The ILO says transformation of tasks is more likely than full replacement, while emphasizing that exposure is not actual job loss and implementation will differ across countries and sectors. This task-based index helps describe where work may change; it does not estimate how many U.S. workers will move occupations by 2035. Nor does it establish how quickly a particular employer will adopt tools or whether adoption will reduce staffing.

That distinction matters when interpreting McKinsey’s range. A model can explore what follows if adoption proceeds faster and labor demand falls more, even when earlier observations showed little employment change. Conversely, the existence of technically exposed tasks does not show that a modeled transition will occur: organizations may augment work, expand output, or retain workers while redesigning tasks. The OECD’s historical evidence and the ILO’s exposure index are useful constraints on interpretation, not rival estimates of McKinsey’s transition total. They measure different periods, populations, and outcomes.

The fairest conclusion is neither that large-scale displacement is inevitable nor that earlier employment stability rules it out. The OECD constrains what can be claimed about observed effects in its evidence window; the ILO clarifies that exposure often signals task transformation rather than whole-job replacement; McKinsey tests conditional future labor-demand mechanisms. None determines what a specific workplace will do. For an individual decision, current local evidence about production use, workflow redesign, hours, vacancies, or sustained staffing changes is more relevant than treating any national measure as a personal forecast.

Sources: Artificial intelligence and jobs: No signs of slowing labour demand yet; One in four jobs at risk of being transformed by GenAI, new ILO–NASK global index shows; Generative AI and jobs: A refined global index of occupational exposure

What does a different labor-market measure add—and not add?

BLS’s 2025–35 projections add a broad labor-market baseline: the U.S. economy is projected to add 5.9 million jobs overall by 2035. The release also introduces categories comparing occupations by theoretical and observed AI exposure. These measures place McKinsey’s scenarios beside general employment direction, but they do not estimate how many workers will need to change occupations because of AI or test McKinsey’s adoption and substitution assumptions. BLS projects employment levels and changes; McKinsey models cross-occupation moves under specified assumptions. A projected net increase can coexist with declines in some occupations because totals combine gains and losses across different work. Neither result alone tells a particular worker whether their position will change or whether a growing occupation is reachable. That difference matters when interpreting a national transition total.

The distinction is especially clear in BLS’s AI exposure categories. In “Artificial Intelligence (AI) exposure categories,” BLS combines five external measures: three theoretical assessments of how AI capabilities match occupational work and two measures based on observed AI use mapped to tasks or activities. BLS explicitly says exposure does not imply job loss, productivity gains, automation probability, or wage effects; it also does not measure potential AI-related employment impacts. They therefore describe a different question from McKinsey’s modeled movement between occupations. A high relative exposure category is not an estimate of a worker being replaced, while a low category is no guarantee that an occupation will remain unaffected.

The projections also show why growth should be read through its causes and scale. In “Employment Projections: 2025–2035 Summary,” BLS attributes much projected healthcare and social-assistance growth to population aging and the increasing prevalence of chronic health conditions. It expects utilities to grow quickly in percentage terms, but says that means 58,800 additional jobs because the sector is small; nearly all the projected increase is tied to electricity demand, including AI power needs. They do not establish that a worker whose current tasks are changing can move into those jobs. The projection is national and broad; access still depends on local openings, required credentials, pay, work conditions, and the time and cost of acquiring missing skills.

BLS cautions that long-term projections are uncertain and advises readers to focus on direction and relative size rather than treating precise values as certain outcomes. A national net total can hide occupation-level contraction, and an exposure classification can identify task applicability without showing adoption or staffing effects. For a career decision, use the BLS outlook to identify occupations projected to grow or decline and examine their requirements; use exposure data to frame which tasks may intersect with AI; then look for local evidence about actual workflow changes and labor demand. Neither broad growth nor exposure labels identifies an individual transition outcome.

Sources: Employment Situation Summary; Artificial Intelligence (AI) exposure categories

What does this sensitivity range leave fixed or outside its test?

The reported range tests alternatives for adoption speed and the extent to which automation reduces labor demand. It does not represent every uncertainty in McKinsey’s 2025–35 model, and its endpoints are not a probability distribution over possible futures. The report says its purpose is to estimate how the composition of US employment might change, rather than independently forecast total employment. It anchors aggregate employment growth to a Bureau of Labor Statistics projection of about 3.1 percent over the decade, then estimates how demand may be distributed across occupations. The transition figures are conditional on that overall growth backdrop.

Other modeled forces also shape that distribution. In “Workforce in motion: Skills and pathways to future jobs in the United States,” McKinsey describes labor-demand increases tied to demographic change and rising living standards, construction and infrastructure investment, technology spending and AI infrastructure, and possible new AI-related occupations. Those drivers are part of the wider model, while the published sensitivity statement identifies adoption pace and labor-demand reduction as the assumptions varied for its alternative transition cases. The spread shows how results move under those tested changes; it does not cover every combination of demand, policy, or future work, or establish that outcomes beyond the cases are impossible.

The pathways analysis adds another conditional layer. McKinsey evaluates potential moves using skill overlap, destination demand, wages, and time to obtain required credentials. A modeled destination is not automatically reachable: a worker’s starting education, licensing requirements, location, pay needs, and available training time can alter the practical route. These factors describe pathway quality, not guarantees of a smooth transition.

So the range is useful for direction, not completeness. It shows that assumptions about adoption and substitution affect modeled transition need, while leaving other sources of uncertainty outside that specific sensitivity test. Treating six million and more than sixteen million as comprehensive bounds would claim more than the report tests; treating the spread as meaningless would discard a real signal about how conditional the estimate is.

Sources: Workforce in motion: Skills and pathways to future jobs in the United States

What evidence can a worker observe before making a move?

A useful local check follows the model’s sequence: capability, adoption, workflow redesign, then labor demand. Evidence becomes more relevant to a personal decision as it moves down that sequence. A tool completing a bounded task in a demonstration shows capability. It does not show routine production use, changed responsibilities, or reduced paid labor. McKinsey’s “Workforce in motion: Skills and pathways to future jobs in the United States” separates adoption from reduced labor demand. Work hours absorbed by automation do not translate one-for-one into fewer workers: organizations may use capacity for more output, new activities, or less overtime, while accountability and other bundled duties remain. McKinsey describes its adoption estimates as directional and dependent on organizational context. A product launch or small pilot is therefore weak evidence about a particular team’s staffing future; workflow maturity, data quality, operating discipline, leadership priorities, and business context affect whether technical feasibility becomes routine use. For a recurring task bundle, ask what is observable at each step. Is the tool used on ordinary work, rather than only in a showcase? Has the process changed: who checks exceptions, corrects errors, handles unusual cases, and takes responsibility for the result? When time is saved, does the team complete more work, improve service, reduce backlog or overtime, or leave capacity unused? Those outcomes describe redesign and absorption of capacity; none alone proves jobs will be cut. A conversation with a manager about workflow and review duties can clarify operations without seeking confidential plans or data. Then look for sustained local labor-demand signals: recurring vacancies that are not refilled, documented changes in paid hours or workload, or staffing reductions continuing across planning periods. Treat it as a prompt to gather context, not proof that AI caused substitution. Repeated changes tied to the same workflow are more informative, though they still do not predict an individual outcome. The Bureau of Labor Statistics’ “Artificial Intelligence (AI) exposure categories” combines theoretical exposure measures with observed AI use mapped to occupational tasks. BLS says these categories do not measure employment effects, worker replacement, or adoption probability. Broad exposure is an early signal, not a decision rule. Give more weight to evidence that your own work is being deployed and redesigned, then assess whether labor demand is persistently changing and whether an alternative fits your pay needs, location, credentials, and time. When signals remain mixed, preserve options with a low-cost, reversible step before committing to expensive retraining or a move.

Sources: Workforce in motion: Skills and pathways to future jobs in the United States; Artificial Intelligence (AI) exposure categories

What is a proportionate next move while the assumptions remain uncertain?

Map one recurring task bundle before deciding whether to retrain. note which tasks have changed, where tools are used, what review remains, and whether work has shifted toward fewer hours, more output, or different service. McKinsey’s *Workforce in Motion: Skills and Pathways to Future Jobs in the United States* models national scenarios; it does not establish what a particular employer will deploy or whether an individual will lose a job. Sustained local changes in workload, vacancies, or staffing deserve more weight.

Then compare three feasible responses. An upgrade keeps your field knowledge while improving how you perform or verify changing tasks. An adjacent move carries some knowledge into related work, subject to local openings and requirements. A larger change may make sense when direct evidence points to lasting demand reduction and a credible training route fits your circumstances. Check pay, location, prerequisites, time, health, and family responsibilities before paying for training or leaving a role. The model’s sensitivity span calls for attention, not panic or complacency; revise your choice as local evidence changes. The free task checker organizes task-level change pressure; the optional roadmap compares paths against your constraints. Neither predicts displacement or guarantees an outcome.

Sources: Workforce in motion: Skills and pathways to future jobs in the United States

Questions readers ask