OECD mapping shows AI already fully automates about 1% of US jobs and recent capability gains push that to roughly 8%, with further plausible advances potentially automating a third — and eventually nearly all — employment by the 2030s, meaning job losses should appear before large productivity or disinflation signals.

Main Finding

Using a new, capability-based mapping between OECD AI Capability Indicators and O*NET occupational demands, the authors estimate that (a) only ~1% of US employment (six occupations) is already fully automatable by current AI; (b) a first wave of capability gains dated to 2026 raises that share to ≈8%; and (c) plausible further waves of capability progress could lift fully automatable employment to ≈32% and then ≈90% (ordered stages often labelled Wave 2 and Wave 3). Job-level displacement should become visible before any aggregate productivity/disinflationary effects, and the macro signal grows non-linearly with the automatable share.

Key Points

• Framework and focus
  • The paper stresses three places AI effects could appear: task changes (invisible because task frequencies are unmeasured), job/occupation displacement (observable), and aggregate productivity/prices (observable later). It therefore focuses on full automation of entire occupations — the first observable, high‑economic‑significance change.
• OECD AI Capability Indicators
  • Nine capability domains: language; social interaction; problem solving; creativity; metacognition/critical thinking; knowledge/learning/memory; vision; manipulation; robotic intelligence.
  • Each domain uses a five‑level scale from long‑solved (1) to full human equivalence (5).
  • Baseline (OECD Nov 2024 / June 2025 publication): current AI ≈ levels 2–3 across domains.
  • Preliminary May 2026 AI-authored re‑ratings (GPT‑5.5 and Claude/Opus) suggest substantial movement: averaged placements include language ≈3.8, problem solving ≈3.4, knowledge ≈3.5, vision ≈3.6, creativity ≈3.4, metacognition ≈3.3, social interaction ≈2.6, manipulation ≈2.5, robotic intelligence ≈2.8.
• Mapping to occupations
  • Occupation set: O*NET (~879–900 occupations); employment analysis uses ~830 occupations (≈155M US workers, 2024).
  • 15 OECD economists produced consensus ratings for a sample; four large language models replicated/extended those ratings to the full occupational space. An average of the LLM outputs predicted the human consensus better than any single human rater.
  • An occupation is counted as automatable only when AI meets every capability the occupation requires (i.e., all nine gaps closed).
• Waves and non‑linearity
  • Baseline (2024): ≈1% of employment fully automatable (six occupations), concentrated in office & administrative support.
  • Wave 1 (preliminary 2026 update): ≈8% of employment becomes fully automatable (adds parts of sales, computer & mathematical, business & financial roles).
  • Wave 2 (projected stage): ≈32% automatable.
  • Wave 3 (projected stage): ≈90% automatable.
  • Occupational groups cross these thresholds at different times: office/admin first; cognitive‑professional and physical‑service groups at Wave 2; only highly interpersonal roles (community/social services, parts of healthcare and personal care) remain substantially unautomatable at Wave 3.
• Diffusion vs capability
  • Capability ≠ adoption. The authors use a rough 10‑year diffusion benchmark with an S‑shaped adoption curve as a plausible pace (but emphasize uncertainty — diffusion depends on cost, integration, regulation, and acceptance).
  • Cost arguments (e.g., projected humanoid robot costs) and digital infrastructure suggest potentially faster diffusion than many historical GPs.
• Early empirical signals
  • The six occupations already classified as automatable showed employment declines: about −4% (2023–24) and −5% (2024–25). Occupations flagged for Wave 1 show a small decline (~−1%); Wave 2+ occupations are roughly flat or growing in the latest data.
• Macro illustrations (order‑of‑magnitude)
  • Translating displaced employment into steady annual productivity/price effects under a simple 10‑year redeployment assumption yields: baseline negligible (~0.1 percentage points), Wave 1 ≈0.9 pp, Wave 2 ≈4 pp, Wave 3 >20 pp. These are illustrative, not forecasts; they show the non‑linear potential scale.

Data & Methods

• Capability measures
  • OECD AI Capability Indicators: nine domains, five‑level qualitative scales. Baseline placements from OECD (Nov 2024/June 2025).
  • Preliminary 18‑month update (May 2026) produced by prompting two frontier systems (GPT‑5.5; Claude/Opus family) to rate the scales against recent literature and benchmarks; the two system outputs were averaged. Authors note these are draft inputs, not yet expert‑validated.
• Occupational demands
  • O*NET descriptions for ~900 US occupations provide the mapping of capability demands.
  • Human rating procedure: 15 OECD economists rated a sample of occupations across the nine domains to establish a consensus reference.
  • Scaling to full occupation list: four LLMs performed the same task on the remaining occupations; the ensemble average best matched human consensus and was used to extend ratings across O*NET.
• Definition of automation
  • An occupation is called fully automatable only if AI meets every capability required (no remaining gaps across all nine domains). This is a strict, conservative definition relative to measures of partial task automation.
• Wave construction and projections
  • Wave 1: uses published baseline plus preliminary May 2026 update.
  • Waves 2 and 3: projected by adding the observed average progress (≈0.8 of a level) evenly across all nine scales per wave. The even‑advance assumption is intentionally simple and treated as ordered capability stages rather than precise-year forecasts.
• Employment and macro exercises
  • Employment data: OEWS (US) used for weights; employment‑weighted shares calculated.
  • Macro illustration: assume displaced share is automated and redeployed evenly over ~10 years; productivity gain approximated by d/(1−d) for displaced share d, with prices moving inversely. Authors emphasize this is a back‑of‑envelope exercise to show scale, not a prediction.

Implications for AI Economics

• Measurement priorities
  • Focus empirical monitoring at the occupational level (where data exist) rather than relying principally on task‑level frameworks that lack task‑frequency measurement.
  • Maintain and regularly update the OECD AI Capability Indicators and the occupation mapping to track which occupations become automatable next.
  • Pair capability updates with close, high‑frequency monitoring of employment, wages, and hiring in flagged occupations to estimate diffusion speed — the key determinant of realized impact.
  • Validate and refine occupational demand ratings with expert review and improve adoption models using empirical data on costs, regulation, and acceptance.
• Modeling and forecasting
  • Macroeconomic models calibrated to historical parameters (small productivity gains, slow diffusion) risk understating possible discontinuities; models should allow for non‑linear jumps when capability thresholds are crossed and be linked to occupation‑level automation thresholds.
  • Scenario analysis should separate capability progress from diffusion, include the possibility of rapid adoption once AI becomes objectively superior, and allow for disinflationary episodes coinciding with large productivity gains.
• Policy implications
  • Early policy action is warranted at the occupational (and sectoral) level: prepare targeted retraining and active labor market policies focused on occupations flagged as newly automatable.
  • Anticipate fiscal, social insurance, and labor‑market challenges for large‑scale displacement: in high‑displacement, high‑productivity, low‑price environments, policy mixes (income support, stimulus timed to low inflation, sectoral transition assistance) need design and simulation now.
  • International and distributional concerns: countries’ exposure depends on occupational mixes and openness to adoption; policy responses must account for uneven exposure across regions, sectors, and demographic groups.
• Research agenda for AI economics
  • Build real‑time monitoring infrastructure linking capability updates, occupation maps, and labor market outcomes.
  • Empirically estimate diffusion speeds by occupation and sector, and study determinants (costs, IT integration, regulation, user acceptance).
  • Investigate complementarities and substitution between AI and human labor at the occupation and task level (including partial automation cases), and the extent to which new tasks created are AI‑assignable rather than human‑assignable.

Reference / provenance: OECD Analysis (Elliott et al.), posted Aug 27, 2026 (CC BY 4.0). DOI: https://doi.org/10.21203/rs.3.rs-10384172/v1.