How the simulation works.

The question

Most existing tools - based on the original Frey & Osborne (2013) analysis - give a single number per occupation: the probability that the job itself is directly automatable. That number is necessary, but not sufficient. A plumber's job is hard to automate, but their customers may be office workers whose incomes contract sharply when AI displaces white-collar tasks. This tool models that second-order cascade: where the money flowing to your occupation comes from, and how AI adoption reshapes those flows over time.

The model, in one paragraph

For each month t over the chosen horizon, we apply an S-curve to estimate how far AI adoption has progressed. We then reduce each customer industry's spending capacity in proportion to that industry's AI exposure, and the occupation's own headcount in proportion to its Frey-Osborne automation probability. The animated revenue at the centre of the graph is the sum of every customer industry's contribution at that month.

Automation probabilities - sources and synthesis

Frey & Osborne (2013) is the canonical baseline - the most-cited per-occupation probability dataset, and the foundation for every downstream tool of this kind. It is also twelve years old, from before generative AI existed in any practical form. So the simulator does not use it raw.

The model layers on top of Frey-Osborne a documented modern AI overlay with one explicit override per occupation, philosophically anchored on:

• Felten, Raj & Seamans (2018, 2021, 2023) - AI Occupational Exposure index. Reframes the question from "will this job be automated?" to "how much of this job's task content overlaps with current language-model capability?" Their 2023 update covers generative AI specifically and finds higher exposure for higher-skill, white-collar occupations - the opposite of earlier waves of automation. We use this finding to lift Frey-Osborne's probabilities for software developers, financial analysts, marketing managers, and graphic designers.
• Eloundou et al. (2023) - "GPTs are GPTs" - task-level exposure analysis estimating that ~80% of US workers have at least some tasks affected by GPT-class models. We use this for the upper bound on the fast scenario.
• IMF (2024) - Cazzaniga et al. - staff discussion note projecting ~40% of advanced-economy jobs are high-exposure to AI. We use this as a sanity anchor for the country aggregate.
• Goldman Sachs (2023) - Hatzius & Briggs - headline estimate that 18-27% of US labour hours could be automated by current generative AI. Anchors the slow / central / fast scenarios.
• BCG / McKinsey / Anthropic Economic Index (2024-2026) - newer industry analyses argue that the dominant near-term effect is augmentation rather than replacement: 50-55% of US jobs reshaped in the next 2-3 years, not necessarily removed. The simulator's ai_demand_boost field captures this for occupations whose work AI augments productively (software, electrical infrastructure, civil engineering). The customer-industry exposure values absorb the rest as demand-side change.

Each override carries a one-sentence rationale and is visible in the simulator's title strip ("modern overlay · F-O was 4%") so the reader can see both the legacy and modern numbers. The detailed list lives in the coefficient notes.

The modern overlay is expert-judgement synthesis of the literature above. As of v0.6, the simulator also ingests Felten-Raj-Seamans AIOE scores (general, language-modelling, and image-generation variants) directly via SOC code as a parallel empirical input - so each US occupation now carries both an expert overlay value and an empirical AIOE percentile. The two broadly agree for white-collar occupations and diverge where the overlay captures non-LLM automation (truck drivers, fast-food workers), which is informative rather than a problem. Other countries pick up AIOE as soon as the SOC↔ISCO/ANZSCO/NOC crosswalks land.

Formulas

The model is intentionally simple - a few logistic curves and weighted sums. Every formula, in the order it's applied, lives on the dedicated formulas page.

Observed adoption - Anthropic Economic Index

From v0.6 the simulator also ingests the Anthropic Economic Index third report (data window 2025-08-04 to 2025-08-11). AEI publishes the observed share of Claude.ai conversations classified into each SOC major occupation group, broken down by country. We join this rollup onto each US occupation as aei_soc_major_pct: an empirical "how much measured AI usage is happening in this occupation's task space, today, in this country" signal.

The numbers are striking: in the United States, Computer-and-Mathematical tasks account for 26% of all classified Claude.ai conversations, while Construction-and-Extraction tasks account for 0.13% - a roughly 200× difference. That is observation, not projection. AEI does not replace the scenario S-curves (a single snapshot is not a time series), but it gives the near-term portion of every scenario a defensible empirical floor - and a reality check on which occupations are seeing serious AI activity right now.

AEI also reports the country-level collaboration split between automation-style use (directive, feedback-loop) and augmentation-style use (validation, iteration, learning). The US leans slightly automation (49.1% vs 50.9%); the UK leans slightly augmentation (45.5% vs 54.5%). Both numbers live in each country's meta.aei.collab_split.

Uncertainty bands

The "Uncertainty at this point" chip reflects the current month, not the chosen horizon - scrubbing further into the future always increases uncertainty no matter where you've set the end.

• Low (≤ 1 year ahead): grounded in current AI deployment data.
• Moderate (1–3 years ahead): scenario S-curves are calibrated for this band.
• High (3–6 years ahead): cumulative model error becomes significant.
• Very high (> 6 years ahead): directionally indicative; not a forecast.

Horizon and uncertainty

The simulator lets you pick a 2, 5, 10 or 20 year horizon. The underlying scenario S-curves are anchored on present-day deployment observations, which means short horizons (≤ 5 years) sit on relatively firm ground. Beyond that, every additional year compounds two kinds of uncertainty: the speed of AI capability growth, and the speed of organisational adoption. By 10 years the curves are best read as directional, and by 20 years they are speculative - useful for sketching the shape of a possible future, not for prediction. The "model uncertainty" chip in the controls panel reflects the current month, not the chosen horizon: scrubbing further into the future always increases uncertainty regardless of how far you set the end.

The formulas

adoption(t) = asymptote / (1 + exp(−steepness · (t − midpoint)))

flow_i(t) = baseline_revenue · share_i · (1 − exposure_i · adoption(t)) · (1 + boost · adoption(t))

total_revenue(t) = Σ flow_i(t)

workers(t) = employment_count · (1 − automation_prob · adoption(t))

Scenarios

Data sources for United States

Every Frey-Osborne probability and wage figure is built from the public sources below. The customer-industry shares - which determine where each occupation's revenue comes from - are currently expert-judgement estimates rather than derived from input-output tables. Each modern AI overlay carries a one-sentence rationale documenting why a Frey-Osborne 2013 number was adjusted. See the coefficient notes for the full audit trail.

• BLS Occupational Employment and Wage Statistics
  Wages and US employment counts (seed values for v0.5)
• Frey & Osborne - The Future of Employment (2013)
  Direct automation probability per occupation (baseline before modern overlay)
• Internal - modern AI overlay (post-LLM)
  Documented overrides on Frey-Osborne for occupations where post-2022 generative AI moved the picture substantially
• IMF - Gen-AI: Artificial Intelligence and the Future of Work (2024)
  Adoption-speed scenario parameters
• Goldman Sachs - The Potentially Large Effects of AI on Economic Growth (2023)
  Adoption-speed scenario parameters
• Enrico Moretti - Local Multipliers (2010); The New Geography of Jobs (2012)
  Cascade multiplier and lag for second-order spending effects
• Eloundou et al. - GPTs are GPTs (2023)
  Upper-bound exposure estimates for the fast scenario
• Felten, Raj & Seamans - AI Occupational Exposure (2018, 2023 generative-AI update)
  Programmatic per-occupation exposure scores (general AIOE, Language Modeling AIOE, Image Generation AIOE) joined by SOC code; surfaced as parallel input to automation_prob and used for cross-checking the modern overlay
• Anthropic Economic Index, third report (release 2025-09-15, data window 2025-08-04..2025-08-11)
  Observed Claude.ai usage by SOC major occupation group, joined by SOC prefix; per-occupation aei_soc_major_pct field plus country-level automation/augmentation collaboration split

What this model deliberately does not do

• It is not a forecast. It is a structured way of inspecting the cascade implications of a chosen scenario. The underlying scenarios are uncertain; reasonable economists disagree about both timing and magnitude.
• It does not yet model regional variation. All values are national means. A software developer in one major city and one in a smaller regional centre appear identical.
• It models local-service spillover, not yet occupation-to-occupation cascades. A Moretti-style local multiplier is wired in: when tradeable occupations (software developers, accountants, financial analysts) lose revenue in aggregate, that drag is propagated with a 9-month lag onto local-service occupations (chefs, personal trainers, real-estate agents). What's still not modelled is direct occupation-to-occupation income cascades - e.g. shrinking the real-estate-agent revenue pipe specifically because software-developer income fell. That would need BLS Consumer Expenditure Survey shares as a separate linkage layer.
• It does not model policy interventions. Retraining subsidies, UBI, sectoral protections - none are toggleable yet. They would shift adoption curves and demand factors but are out of scope for v0.
• The customer-industry shares for v0 are eyeballed. They are bounded by industry trade-association reports and BEA Use Tables but they are not derived from a join. v0.5 fixes this.

Provenance