Anthropic’s newest economics paper contains a number that can look almost reassuring. In its substantial-change scenario, cognitive employment in 2030 is 3.9 per cent below its mid-2026 level. GDP is 8.3 per cent above the no-AI path, the average wage is 2.1 per cent higher, and economy-wide unemployment is 4.6 per cent. That is disruption, but it is not a picture of knowledge work disappearing overnight.1
The difficulty is that 3.9 per cent is an aggregate. It is not a staffing forecast for the organisation that intends to put AI into every major workflow. The model does not distinguish workers by tenure or skill, or firms by adoption capacity. Anthropic identifies that heterogeneity as a useful area for future work.1
That omission matters because adoption is already highly uneven. Large and knowledge-intensive firms are adopting faster. Some organisations are still using AI as an occasional assistant; others are integrating it into multiple business processes and systems. A national or economy-wide average therefore mixes organisations that may eventually look very different from one another.2345
For a manager planning to become AI-native, the relevant question is not simply whether Anthropic’s 3.9 per cent is right. It is: where will our organisation sit inside the average?
Source: Korinek et al., The Anthropic Institute, Economic Scenarios for Transformative AI, Table 3. These are scenario outputs, not forecasts; the authors assign no probabilities to them.1
Anthropic is modelling scenarios, not predicting your company
The September paper is careful about what it is. The authors construct a task-based model and feed it different assumptions about AI capability, diffusion, productivity, automation, the creation of new human tasks and labour-market adjustment. They explicitly say the scenarios are not predictions and attach no probabilities to them.1
That is one reason the paper is useful. It turns disagreement about AI into a set of variables that can, at least in principle, be measured. In the modest scenario, AI remains a relatively ordinary technology. In the substantial scenario it becomes economically pivotal. In the extreme scenario, nearly half of today’s cognitive work is performed by AI by 2030, cognitive employment falls 21.5 per cent from mid-2026 and cognitive unemployment reaches 17.9 per cent.1
The paper is not hiding the possibility of severe disruption. The less obvious issue appears when its aggregate outputs are brought down to organisational level. Anthropic treats cognitive occupations as one broad group. A 24-year-old analyst and a 48-year-old partner are not separate populations in the model. Nor is a deeply automated firm distinguished from one using AI lightly.1
That is not a flaw in a macroeconomic scenario model. It is simply the point at which a manager needs another model.
On a narrow screen, each source is shown as a separate card.
| Source | Breadth | Depth and why it matters |
|---|---|---|
| US Census Bureau, revised October 2026 | 18% of firms used AI in a business function; 32% when employment-weighted. Very large firms in information, professional services and finance reached roughly 50–60% use. | 57% of adopting firms used AI in three or fewer business functions, and 44% of AI-using firms reported augmenting worker tasks. Of firms reporting AI-driven task effects, 66% reported augmentation only. Functional breadth and operational investment were positively associated with employment decreases, but the analysis does not establish causality.2 |
| EIB Investment Survey, 2025 | Generative AI use: 37% of EU firms and 36% of US firms. | Among AI-using firms, 81% of US firms used AI in at least two business processes, versus 55% of EU firms.3 |
| UK Business Data Survey, 2026 | 41% of businesses handling digitised data reported AI use; 82% of large firms did so. | Among AI users, 21% reported integration into existing business systems; 57% of large AI-using firms did so.4 |
| Gallup, August 2026 | 53% of US employees used AI at work a few times a year or more. | 32% used it a few times a week or more and 16% daily. Worker-level use is therefore also distributed very unevenly.5 |
Reading the table: Definitions, units, bases and weights differ across these surveys, so the percentages should not be treated as a single harmonised adoption series. Their common message is dispersion: use, integration and depth vary materially by firm size, sector, geography and worker behaviour.
An average can conceal a very different firm-level reality
A simple piece of arithmetic shows the problem. Anthropic’s substantial scenario produces a 3.9 per cent decline in cognitive employment in aggregate. Suppose, purely for illustration, that the less AI-intensive part of the economy experiences only a 1 per cent decline.
If only a quarter of cognitive employment sits in the AI-intensive cohort, that cohort would have to shrink by about 12.6 per cent to produce the same 3.9 per cent aggregate result. If half the workforce sits in that cohort, the implied decline is about 6.8 per cent. If three quarters does, it is about 4.9 per cent.
Those are not forecasts. They are simply different distributions that reconcile to the same average. The arithmetic is useful precisely because it exposes something an aggregate cannot tell us: how concentrated the change may be.
Illustrative arithmetic only. Aggregate cognitive employment is fixed at −3.9%, matching Anthropic’s substantial scenario; the non-AI-intensive remainder is assumed to fall by 1%. Different distributions can therefore produce the same macro number.
The extreme scenario provides a useful opposite check. A 21.5 per cent aggregate fall cannot plausibly be produced by a tiny elite of AI-native firms unless those firms change beyond recognition. An outcome that large would require deep displacement to become much more widespread. The distribution therefore matters in both directions: modest aggregate change can hide severe local change, while severe aggregate change requires transformation to spread.
Economy-wide employment scenarios are not organisation-level workforce plans.
A manager needs to model the organisation they intend to become
Anthropic’s framework gives us useful ingredients for doing that. It separates how much work AI can affect from how widely it is actually used. It distinguishes automation from augmentation. It allows some automated tasks to be replaced by new human tasks. And it shows that the labour-market result changes sharply when those parameters move.1
At firm level, another variable becomes unavoidable: what management chooses to do with the capacity AI releases. A productivity gain can become lower headcount. It can also become more output, better service, improved quality, lower backlog, new products or spare resilience. Task automation does not mechanically translate into job elimination.
On a narrow screen, each planner component is shown as a separate card.
| Planner component | Reader-facing question or result | Provenance |
|---|---|---|
| Current workforce | What is the current cognitive headcount? | User assumption |
| AI integration | How deeply does the organisation expect to integrate AI by 2030? | User assumption |
| Capacity choice | Will released capacity be used mainly for growth or for lower headcount? | User assumption |
| Adoption position | Does the organisation expect to be a slower, typical or leading adopter? | User assumption |
| Model mechanics | Capability × diffusion × [automation × (1 − reinstatement) + augmentation share × g/(1 + g)], followed by management’s capacity-conversion assumption. | Research-informed Andrew’s Intel heuristic. Presets are interpolations, not Anthropic outputs. |
| Indicative workforce | A scenario result and sensitivity range for discussion, not a predicted headcount or confidence interval. | Modelled output |
This is scenario planning, not fortune telling. Source-derived assumptions, user assumptions and modelled outputs must remain visibly separate.
A worked example shows the point. Start with 1,000 cognitive workers and choose a deep-integration scenario, a leading-adopter position and a balanced use of released capacity. The simple planner produces an indicative workforce of about 922 by 2030. Choose the more aggressive AI-native scenario with the same headcount/output balance and the indicative figure falls to about 781.
Neither number is a prediction from Anthropic. Both are outputs of an Andrew’s Intel heuristic that translates published scenario parameters into organisation-level assumptions. The value is not the apparent precision. The value is seeing how much the result changes when the operating model changes.
Three organisations can live inside the same economy
On a narrow screen, each organisation is shown as a separate card.
| Organisation | Likely workforce issue | Management priority |
|---|---|---|
| AI-light | Headcount changes slowly; the existing career pyramid largely survives. | Use AI to improve work without assuming the current development pipeline has become obsolete. |
| AI-transition | Junior and routine work starts disappearing before the target human organisation is clear. | This is the danger zone: do not let near-term automation accidentally determine long-term workforce shape. |
| AI-native | A large reduction in cognitive headcount may be rational across several levels, not just at entry level. | Design the smaller human organisation deliberately: which people remain essential, how many are needed, and how will they be produced? |
Boundary: These are planning contexts, not a taxonomy of firms. An organisation may move from one to another over time.
The workforce pyramid still matters—even if it becomes much smaller
This is where the earlier entry-level hiring evidence still belongs, but in a more contextual role.
Anthropic’s March 2026 labour-market work found no systematic rise in unemployment among highly exposed occupations, but it did find a tentative 14 per cent decline in the job-finding rate for 22–25-year-olds entering highly exposed work compared with 2022. The estimate was only just statistically significant.6 Later evidence strengthened the signal. Stanford’s August revision found employment among 22–25-year-olds in highly AI-exposed occupations about 19 per cent below the path implied by less-exposed peers, with the adjustment appearing mainly through reduced hiring.7 A US Census Bureau working paper reported that regression-adjusted employment among 22–24-year-olds in the most AI-exposed quintile of industry-state cells fell 12 per cent over the ten quarters to 2025 Q2, relative to broadly stable employment in less-exposed cells, with reduced hiring accounting for most of the decline.8
Those are emerging signals, not settled causality. Stanford describes its patterns as early and descriptive. The Census paper discusses pre-existing shifts, the pandemic, remote work, educational-attainment confounds and monetary-policy effects. Yale’s Budget Lab still finds no clear evidence of large-scale, economy-wide AI disruption across occupational churn, employment and wages, although its October tracker flags a volatile rise in unemployment among AI-exposed occupations that is not corroborated by the other measures.9
That is precisely why the firm-level distribution matters. The aggregate can remain calm while particular cohorts, occupations or organisations move first.
If a highly AI-integrated organisation ultimately needs far fewer cognitive workers at every level, reducing junior intake may be entirely sensible. But it still needs some mechanism for producing the smaller number of experienced humans it expects to retain.
Andrew’s Intel has dealt with that apprenticeship problem separately in AI productivity: the hidden capability tax. That analysis examines the risk of removing the routine work through which judgement, context and tacit knowledge are developed. The relevant point here is narrower: even a much slimmer AI-native organisation has to work backwards from the humans it still expects to need.10
The humans who remain may become cheaper—or much more expensive
There is a second uncertainty managers should not hide inside an average. If many organisations reduce entry-level intake but still need experienced judgement, the future market for mid-career and senior people can tighten. Importing skills then becomes more expensive. Retention costs rise. Pay compression can push pressure farther up the hierarchy as organisations try to preserve differentials and prevent the small pool of critical people from leaving.
But that wage cascade is conditional, not inevitable. If AI also automates enough expert work, demand for experienced cognitive labour can fall. Anthropic’s extreme scenario produces cognitive wages 11.5 per cent below the no-AI path even while economy-wide output is much higher.1
The paper itself points to related labour-economics work with exactly this ambiguity: automating the inexpert tasks inside an occupation can increase the expertise and wage of the human job that remains; automating expert tasks can lower both.1
That makes the workforce question harder, but also clearer. Management is not deciding whether to preserve today’s pyramid. It is deciding which human capabilities survive into the target operating model, how scarce those capabilities may become, and what pipeline will produce them.
The meeting question is: where are we inside the average?
On a narrow screen, each question is shown as a separate card.
| Test | Question |
|---|---|
| 1. Capability | What share of our cognitive work do we realistically expect AI to be capable of handling by 2030? |
| 2. Integration | How much of that capability do we actually expect to integrate into live workflows—not just make available to employees? |
| 3. Work design | Of the AI-performed work, how much will automate tasks outright and how much will augment people? |
| 4. Capacity choice | What proportion of released capacity are we deliberately planning to take as lower headcount, rather than more output, quality or service? |
| 5. Scarce humans | Which human roles remain scarce and strategically important in the target organisation? |
| 6. Pipeline | What entry, development and succession pipeline produces those people—and does today’s hiring policy still support it? |
| 7. Review | Which assumptions would have to change for this workforce plan to be wrong, and how often will we update them? |
The test is deliberately scenario-based. It does not ask managers to pretend they know the 2030 headcount. It asks them to expose the assumptions already implicit in their automation plans.
What the Anthropic number may really be hiding
The most important thing about Anthropic’s 3.9 per cent substantial-scenario decline is not whether the number proves reassuring or alarming. It is that the number belongs to an average economy, while managers make decisions inside individual organisations.
If adoption remains uneven—and every credible adoption survey currently suggests that it will—some firms can sit far from that average. The organisations most committed to AI integration may require materially fewer cognitive workers than the aggregate number implies. Others may retain something much closer to today’s structure. Both can be true at the same time.
That is why the quick productivity win is not enough. A manager should not ask only how many people AI lets the organisation remove this year. The better question is what human organisation the firm intends to need when the transition is mature, and what hiring, development and retention choices have to be made now to make that organisation viable.
The risk is not hiring fewer people. The risk is hiring fewer people on an implicit forecast of the future organisation that nobody has actually made.
Sources and evidence
The Anthropic paper is treated as a scenario framework, not a forecast. Adoption studies use different definitions and samples; they establish heterogeneity rather than a harmonised adoption rate. Early-career labour-market evidence remains suggestive rather than a settled causal estimate. Time-sensitive claims were rechecked against primary sources on 29 October 2026.
Source posture: The headline is an aggregation argument, not an accusation that Anthropic deliberately concealed evidence. Firm-level distributions and planner presets are Andrew’s Intel illustrations. Scenario bands are not confidence intervals, and none of the early-career studies establishes a simple causal path from AI to employment change.
- Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter and Peter McCrory, Economic Scenarios for Transformative AI, The Anthropic Institute Working Paper 2026-02, September 2026. Primary source for the scenario framework, Table 3 outcomes and model boundaries.
- Kathryn Bonney et al., The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks, US Census Bureau CES Working Paper 26-25R, revised October 2026; original April 2026. Firm adoption, functional breadth, task effects and descriptive employment associations.
- European Investment Bank, EIB Investment Survey 2025: European Union overview. EU and US generative-AI adoption and business-process breadth.
- UK Department for Science, Innovation and Technology, UK Business Data Survey 2026. AI use and integration by firm size.
- Gallup, Artificial Intelligence Indicator, August 2026 data. Frequency of workplace AI use among US employees.
- Maxim Massenkoff and Peter McCrory, “Labor market impacts of AI: A new measure and early evidence”, Anthropic, 5 March 2026. Observed-exposure method and the tentative young-worker job-finding estimate.
- Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence”, Stanford Digital Economy Lab, revised 12 August 2026. Descriptive payroll evidence through June 2026.
- Lee C. Tucker, “You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators”, US Census Bureau CES Working Paper 26-27, April 2026. Administrative-data evidence on early-career hiring and employment.
- The Budget Lab at Yale, “Tracking the Impact of AI on the Labor Market”, updated 14 October 2026. Economy-wide counter-signal and the tracker’s volatile, uncorroborated unemployment movement among exposed occupations.
- Andrew’s Intel, AI productivity: the hidden capability tax. Approved corpus analysis used only for the related apprenticeship and capability-development mechanism; it is not yet publicly released.