## wpiea2026147 — 2.1 Data

## Source details

**Canonical URL:** [wpiea2026147 — 2.1 Data](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026147.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026147.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026147.pdf.json)

---

### AEI source, coverage, and sampling
- Five releases of the Anthropic Economic Index (AEI) used: January 2025 to February 2026 (waves R1–R5).  
- Each release samples one million Claude AI conversations.  
- Anthropic’s classifier maps each conversation to an O*NET occupational task (observed AI usage).  
- R1 and R2: global only (no country-level breakdown).  
- R3 onward geocoded to countries (R3 in August 2025, R4 in November 2025, R5 in February 2026).  
- R5: 176 countries receive a geographic code; 117 meet Anthropic’s minimum-conversation threshold and enter the analysis.  
- Data cover consumer-stream Claude.ai web conversations only; exclude enterprise API usage.  
- Generalizability and bias notes:
  - Single-provider data may not fully generalize; conversation volume is scaled by Claude’s market share when estimating total-market LCE.  
  - ACI is not adjusted for Claude’s user-base occupational composition (historically skewed toward software engineers), which may affect ACI level and declining-ACI trend.  
  - Including enterprise API usage would likely strengthen concentration findings because enterprise traffic is heavily skewed toward high-wage occupations (Computer & Math ≈ 51% of API conversations in R5); consumer-data ACI likely a lower bound on true concentration.

### Sample structure (R5 and waves)
- R5 (February 2026) textual summary:
  - 1,000,000 sampled conversations → 972,636 classified by the O*NET classifier → 929,714 map to an occupational task (global).  
  - Country-level: 176 countries receive geographic code (805,387 conversations), of which 117 receive O*NET task enrichment (801,151 conversations).  
- AEI releases used (selected):
  - R1: ∼Jan 2025 — Global only — ∼1 million (task percentages only; N*country n/a).  
  - R2: ∼Mar 2025 — Global only — ∼1 million (task percentages only; N*country n/a).  
  - R3: Aug 4 to 11, 2025 — Global + 114 countries — 964,494 global total; N*country = 602,175.  
  - R4: Nov 13 to 20, 2025 — Global + 117 countries — 999,875 global total; N*country = 609,954.  
  - R5: Feb 5 to 12, 2026 — Global + 117 countries (176 receive geographic code) — 1,000,000 global total; N*country = 566,418.  
- Classified totals and country-level coverage:
  - Global classified N*: R3 941,239; R4 972,146; R5 972,636.  
  - ISO countries with O*NET data—R3 114, R4 117, R5 117. N*country (classified) R3 602,175; R4 609,954; R5 566,418.  
  - ACI analysis baseline: 86 sample countries — countries in ACI analysis: R3 83, R4 85, R5 86; Countries in all 3 waves: 81.

### Occupation mapping and mapping caveats
- Each conversation maps to a single O*NET occupational task.  
- AEI’s 22 SOC-2 occupation groups crosswalked to nine ISCO-08 one-digit groups (consistent ILO employment and wage data).  
- Note on SOC→ISCO mapping: SOC places technician-type roles inside broader professional SOC groups, which map to ISCO-2 (Professionals); ISCO-3 (Technicians, ∼17% of employment) contributes no AI conversation share in the data, biasing the ACI slightly upward (conservative bias making AI gains look more unequal).

### Time estimates (hours saved per conversation) and illustrative values
- R4 and R5 report task-level time estimates: time to complete a task without AI and time with AI. Hours saved per conversation h_t = human-only − human-with-AI.  
- R3 lacks these estimates; R4 values applied to R3. Hours saved are different but stable across waves; Pearson correlation ρ = 0.80 across ≈3,200 common tasks between R4 and R5.  
- Section 3.4 bounds influence of these model-based estimates on aggregate LCE, which remains within $2.5 to $2.9 trillion.  
- Illustrative R5 h_t examples (Table 1):
  - Modify existing software (Computer & Math): Human only 3.54 hrs; With AI 17.8 min; Saved 3.24 hrs.  
  - Assist students with coursework (Education): Human only 2.82 hrs; With AI 17.6 min; Saved 2.53 hrs.  
  - Write press releases (PR / Media): Human only 1.54 hrs; With AI 14.5 min; Saved 1.30 hrs.  
  - Advise on drug interactions (Healthcare): Human only 1.19 hrs; With AI 10.4 min; Saved 1.02 hrs.  
  - Answer customer questions (Customer service): Human only 0.35 hrs; With AI 3.4 min; Saved 0.29 hrs.

### LCE methodology and formula
- LCE formula (Eq. 1):
  - LCE = sum_c sum_s sum_{t ∈ s} ( n_{c,t} / N*_{country} × C_{total,w} × h_t × W^{monthly}_{c,s} × 12 / 2,080 )
- Definitions used:
  - n_{c,t} / N*_{country}: fraction of country-level classified conversations on task t (N*_{country} = 566,418 in R5). Multiply by C_{total,w} for estimated actual conversations. Source: AEI.  
  - h_t: hours saved per conversation on task t (human-only minus human-with-AI). Source: AEI. Global: R4 and R5 observed; R3 imputed from R4.  
  - W^{monthly}_{c,s}: ILO mean monthly wage in nominal USD for ISCO-1 occupation s in country c (matched-peer fallback for 22 countries). Source: ILOSTAT.  
  - 2,080: annual working hours (40 hrs × 52 weeks).  
  - C_{total}: total AI conversations per week across all platforms; baseline approximation = 200M/0.25 ≈ 800 million. Public data; global scope; approximate.
- Interpretation and caveats:
  - LCE annualizes and values hours saved at occupation hourly wage; an indicative measure of productivity gains implied by current AI usage rather than a strict bound. Valuing every classified conversation at full estimated time saving tends to overstate realized gains; excluding enterprise API usage tends to understate total; net direction ambiguous.

### Parameters, scaling, and not-classified handling
- Not-classified rates:
  - Global-level notclassified rate: 2.7% (27,364 out of 1,000,000) not classified at global level.  
  - Country-level notclassified rate: 29.4% of O*NET-enriched country-level conversations (235,696 out of 801,151 across 117 countries). Of these, roughly 208,000 were successfully classified at the global level but not at the country level.  
  - Untagged country-level conversations cannot enter LCE or ACI; analysis scales from country-tagged conversations only (N*_{country}), redistributing untagged volume proportionally.  
- Scaling from Claude to all platforms:
  - Total Claude.ai conversations per week ≈ 200 million (derived from 18.9 million monthly active users).  
  - Claude’s share of professional AI conversations ≈ 25%.  
  - C_{total,w} ≈ 800 million/week in R5 (200M / 0.25 = 800M), scaled proportionally across earlier waves using SimilarWeb traffic.  
  - Scaling assumptions affect only LCE level; ACI and regressions (depending on relative distribution) are unaffected.

### Key assumptions underlying LCE (Box 1)
- A1. h_t is task-level; within-task heterogeneity absorbed into average.  
- A2. h_t estimated by Claude, not measured; both time variables are AI estimates.  
- A3. Each hour saved valued at the occupation’s hourly wage.  
- A4. W_{c,s} held constant.  
- A5. C_{total} varies by wave. For R5, C_{total} ≈ 800 million AI conversations per week across all platforms (200M Claude / 25% market share). For earlier waves: R3 ≈ 414M/week; R4 ≈ 489M/week.

### Worked example (Box 2): “modify existing software” in the US (R5, one week)
- US conversations on task O*NET 15-1132: 4,722 of 566,418 categorized country-level conversations (0.834%).  
- With C_{total} = 800 million:
  - Estimated US conversations this week: 0.00834 × 800,000,000 = 6,672,000.  
  - Hours saved per conversation: 3.24 hours (3.54 − 0.30).  
  - Total hours saved: 6,672,000 × 3.24 = 21,617,000 hours.  
  - Hourly wage (US Professionals, ISCO-1 group 2, ILO): $8,171 × 12 / 2,080 = $47.14/hr.  
  - Dollar value this week: 21,617,000 × $47.14 = $1.02 billion.  
  - Annualized: $1.02B × 52 = $53 billion.

### AI Concentration Index (ACI): definition and interpretation
- ACI formula (Eq. 2):
  - ACI_{c,w} = 2 × sum_s [ g_{c,s,w} × ( R_{c,s} − 1/2 ) ]  
  - g_{c,s,w} = ( sum_{t ∈ s} n_{c,t} h_t ) / ( sum_{all t} n_{c,t} h_t ) — share of hours saved by AI conversations in occupation s.  
  - R_{c,s}: occupation s’s fractional rank in the country’s own wage distribution (cumulative employment share up to occupation midpoint).  
- Properties:
  - Range: −1 to +1. Positive when gains tilt toward high-wage occupations; negative when toward low-wage occupations; zero when time savings proportional to employment.

### Aggregate AI gains: wage index and LCE (key statistics)
- Usage-weighted wage index (US occupation-level wage across global AI conversations):
  - Declines from $82,353 in January 2025 to $77,845 in February 2026 (5.5%).  
  - Decline driven by occupational share shifts: Computer & Math lost 4.9 percentage points (pp) of global AI conversation share; Education gained +4.1 pp; Sales +2.7 pp; Office/Admin +1.6 pp.  
- LCE for 86 sample countries:
  - Moves from $1.2 trillion annualized in R3 to $2.7 trillion in R5 (3.4% of GDP).  
  - Divergence between falling wage index and rising LCE reflects employment-scale effect: lower per-hour value but much larger numbers of workers in lower-wage occupations.  
- Margins driving LCE growth:
  - Extensive margin (growth in AI conversations) explains ~72% of increase R3→R4 and ~84% of increase R4→R5.  
  - Intensive margin (change in value per conversation) explains the remainder: ~28% then 16%.

### Scaling scenarios and sensitivity (LCE level)
- Median scenario (baseline): 200 million Claude conversations/week and 25% market share → baseline LCE ($2.7 trillion in R5 for 86 countries).  
- Conservative scenario: 150 million conversations and 35% share → LCE = $1.4 trillion.  
- Aggressive scenario: 250 million conversations and 15% share → LCE = $5.5 trillion.  
- Scaling assumptions affect only LCE level, not occupational distribution or ACI results.

### ACI cross‑country examples and time dynamics
- Cross‑country ACI examples (R5, Feb 2026):
  - United States: ACI = 0.44 (Professionals capture 83.6% g share; Managers 6%; Clerical 5%; Professionals ILO wage $98,047; cumulative employment 88.3% → ACI_US = +0.44 per Box 3).  
  - India: ACI = 0.86.  
  - Kenya: ACI = 0.81.  
- Income pattern (R5):
  - Sub‑Saharan Africa and South Asia with GDP per capita below $5,000: ACIs cluster between 0.85 and 1.00.  
  - Western Europe and North America with GDP per capita above $50,000: ACIs range from 0.35 to 0.62.  
- Time dynamics (R3–R5, 81 countries present in all waves):
  - 20 countries had no ACI change across two consecutive windows because all AI conversations mapped to a single occupational group.  
  - Among 61 countries with changes, diffusion toward declining ACI increased: 37 countries (46% of total) saw ACI fall in R4→R5, up from 18 countries (22%) in R3→R4.  
  - 28 countries reversed from rising ACI through November 2025 to falling by February 2026; 12 of these 28 are middle‑income economies.

### Cross‑country regression evidence (selected results)
- Regression sets: (1) predictors of LCE as % of GDP (pooled R3–R5); (2) predictors of ACI level (pooled R3–R5); (3) predictors of ∆ACI across consecutive waves. Parsimonious (GDP per capita only) and full specifications reported.
- Selected coefficients and significance (Table 5 summary):
  - log GDP per capita:
    - Column (1): 0.384*** (SE 0.068) predicting LCE/GDP.  
    - Column (3): −0.135*** (SE 0.012) predicting ACI level.  
  - AI regulatory readiness:
    - Column (2): 8.918*** (SE 2.689) predicting LCE/GDP.  
    - Column (4): −1.619*** (SE 0.383) predicting ACI level.  
  - English official:
    - Column (4): −0.005*** (SE 0.002) predicting ACI level.  
  - Other predictors:
    - Services VA (share of GDP): Column (1) 1.941* (SE 1.077) predicting LCE/GDP; Column (6) 0.032** (SE 0.016) predicting ∆ACI.  
    - Internet users (%): Column (4) 0.002*** (SE 0.001) predicting ACI level.  
    - Gini (SWIID): Column (4) 0.005*** (SE 0.002) predicting ACI level.  
- Interpretation:
  - Regulatory readiness and economic structure (service share) explain substantial cross‑country variance in LCE/GDP; including regulatory readiness reduces the simple income effect.  
  - Richer countries and those with stronger regulatory frameworks have more broadly distributed AI gains (lower ACI).  
  - Higher income inequality (Gini) and higher internet penetration (conditional on GDP per capita) associated with higher ACI.  
  - Official English status predicts faster declines in ACI (faster diffusion), interpreted as greater representation in English‑dominant LLM training data enabling broader occupational usefulness.  
- Model fit:
  - R^2 ranges: LCE/GDP specifications 0.291 and 0.440; ACI level specifications 0.673 and 0.787; ∆ACI specifications 0.065 and 0.099.

### Robustness: bounding hours saved and distributional stability
- Strict bounds constructed from AEI 95% confidence intervals:
  - h_lower_t = max(0, h_human-only_CI-lower − h_human-with-AI_CI-upper / 60).  
  - h_upper_t = max(0, h_human-only_CI-upper − h_human-with-AI_CI-lower / 60).  
- Aggregate effects:
  - Task‑level mean h_t swings from 1.80 hours (−33%) to 3.81 hours (+41%).  
  - Aggregate LCE moves within $2.5 to $2.9 trillion (baseline $2.7 trillion). Rationale: high‑volume tasks dominate LCE and have tight CIs; low‑volume tasks have wide CIs but little aggregate weight.  
- Distributional robustness:
  - No country’s ACI shifts by more than 0.5 percentage points under either bound.  
  - Cross‑country ranking of ACI values remains unchanged.  
  - Bounding shifts hours saved roughly proportionally across occupations, leaving shares of AI time savings (and thus ACI) nearly identical.

### Extension: expanded sample with peer‑borrowed wages (Section 3.5)
- Expanded sample: baseline 86 countries + 31 countries lacking direct ILO ISCO-1 wages/employment → expanded 117-country sample.  
- Peer-borrowing method: missing employment or wages borrowed from country closest in log GDP per capita among reporters; employment and wages matched separately. Appendix Table A8 lists peer assignments.  
- Aggregate LCE on expanded sample:
  - R5 LCE = $2.9 trillion, or 3.1% of these countries’ combined GDP.  
  - Baseline 86-country LCE was $2.7 trillion (3.4%).  
- Regression re‑estimation (Table 6; pooled R3–R5; 117 countries):
  - AI regulatory readiness remains strongly significant in LCE/GDP regressions and retains negative sign for ACI level (significant at 10% in expanded sample).  
  - English‑language and income slopes retain signs and significance.  
  - R2 values somewhat lower for ACI columns (example: 0.53 versus 0.79 with all controls), consistent with measurement error from peer-borrowed data.  
  - Table 6 reported values (selected):
    - Observations: 337 / 332 / 337 / 332 / 219 / 216 (columns (1)–(6)).  
    - R2: 0.256 / 0.421 / 0.394 / 0.527 / 0.053 / 0.085 (columns (1)–(6)).  
    - Example coefficients:
      - log GDP per capita: 0.367*** (0.062) in column (1); −0.107*** (0.016) in column (3).  
      - AI regulatory readiness: 8.012*** (2.201) in column (2) on LCE/GDP; −0.871* (0.447) in column (4) on ACI level.  
      - English official: 0.244 (0.192) in LCE/GDP column; −0.083** (0.040) in ACI column.  
    - Significance: * p < 0.1, ** p < 0.05, *** p < 0.01.
- Robustness and limitations:
  - Qualitative results preserved in expanded sample.  
  - Peer-borrowed data introduce measurement error, reducing explanatory power for some specifications (notably ACI regressions).  
  - Columns with full controls include 115 countries (some columns include full 117 where covariates available).

### Policy‑relevant interpretation (from empirical findings)
- Aggregate productivity value is large and growing, driven mainly by AI reaching occupations that employ far more workers (extensive margin).  
- Distributional pattern: gains unequally distributed — in nearly every country they initially tilt toward high‑wage occupations, with greater concentration in developing economies; tilt declining over time.  
- Predictors:
  - Income and AI regulatory readiness predict concentration level.  
  - Official English status predicts the trajectory (whether concentration is falling), interpreted as a proxy for representation in LLM training data and faster diffusion.

*Source: wpiea2026147 - 2.1 Data (chapter/section content from the supplied PDF).*

### 2.1    Data

### 2.1    Data

### Anthropic Economic Index (AEI) — source and coverage
- Five releases of the Anthropic Economic Index (AEI) are used, spanning January 2025 to February 2026 (waves R1–R5).  
- Each release samples one million Claude AI conversations.  
- Anthropic’s classifier maps each conversation to an O*NET occupational task (observed AI usage).  
- R1 and R2 report AI usage by occupation for the world as a whole (no country-level breakdown).  
- From R3 onward (R3 in August 2025, R4 in November 2025, R5 in February 2026) conversations are geocoded to countries.  
- In R5: 176 countries receive a geographic code, of which 117 meet Anthropic’s minimum-conversation threshold for task enrichment and enter the analysis.  
- Data cover consumer-stream Claude.ai web conversations only and exclude enterprise API usage.  
- Notes on generalizability and biases:
  - Because the data capture usage from a single provider, results may not fully generalize to the broader AI market; conversation volume is scaled by Claude’s market share when estimating total-market LCE.  
  - The ACI is not adjusted for the occupational composition of Claude’s user base, which historically skewed toward software engineers; this may affect the ACI level and the declining-ACI trend.  
  - Including enterprise API usage would, if anything, strengthen concentration findings because enterprise traffic is heavily skewed toward high-wage occupations (Computer & Math accounts for approximately 51% of API conversations in R5), so the ACI computed from consumer data alone is likely a lower bound on the true concentration of AI gains.

### Sample structure (R5 example and waves)
- Figure 1 (R5, February 2026) summary (textual):
  - 1,000,000 sampled conversations → 972,636 classified by the O*NET classifier → 929,714 map to an occupational task (global).  
  - Country-level: 176 countries receive geographic code (805,387 conversations), of which 117 receive O*NET task enrichment (801,151 conversations).  
- Table 2 (AEI releases used):
  - R1: ∼Jan 2025 — Global only — ∼1 million (task percentages only; N*country n/a).  
  - R2: ∼Mar 2025 — Global only — ∼1 million (task percentages only; N*country n/a).  
  - R3: Aug 4 to 11, 2025 — Global + 114 countries — 964,494 global total; N*country = 602,175.  
  - R4: Nov 13 to 20, 2025 — Global + 117 countries — 999,875 global total; N*country = 609,954.  
  - R5: Feb 5 to 12, 2026 — Global + 117 countries (176 receive geographic code) — 1,000,000 global total; N*country = 566,418.  
- Table 3 highlights:
  - Panel A global totals: R1 ∼1M, R2 ∼1M, R3 964,494, R4 999,875, R5 1,000,000.  
  - Global classified N*: R1 and R2 report task percentages only; R3 941,239; R4 972,146; R5 972,636.  
  - Panel B country-level: ISO countries with O*NET data—R3 114, R4 117, R5 117. N*country (classified) R3 602,175; R4 609,954; R5 566,418.  
  - Panel C ACI analysis sample (baseline 86 sample countries) — countries in ACI analysis: R3 83, R4 85, R5 86; Countries in all 3 waves: 81.

### Occupation mapping
- Anthropic’s classification maps each conversation to a specific O*NET occupational task; each conversation maps to a single occupation category by construction.  
- The AEI’s 22 SOC-2 occupation groups are crosswalked to nine ISCO-08 one-digit groups (the finest ISCO-08 level with consistent ILO employment and wage data across countries).  
- Note: SOC places technician-type roles inside broader professional SOC groups, which map to ISCO-2 (Professionals); as a result, ISCO-3 (Technicians, ∼17% of employment) contributes no AI conversation share in the data, biasing the ACI slightly upward (conservative bias making AI gains look more unequal).

### Time estimates (hours saved per conversation) and illustrative values
- In R4 and R5 the AEI reports task-level time estimates: time to complete a task without AI and time with AI. Hours saved per conversation is human-only minus human-with-AI (h_t).  
- R3 lacks these estimates; R4 values are applied to R3. Hours saved are different but stable across waves; hours saved highly correlated between R4 and R5 (Pearson correlation ρ = 0.80 across ≈3,200 common tasks).  
- Section 3.4 bounds influence of these model-based estimates on aggregate LCE, which remains within $2.5 to $2.9 trillion.  
- Table 1: Illustrative hours saved per conversation (h_t), R5 (February 2026)
  - Modify existing software (Computer & Math): Human only 3.54 hrs; With AI 17.8 min; Saved 3.24 hrs.  
  - Assist students with coursework (Education): Human only 2.82 hrs; With AI 17.6 min; Saved 2.53 hrs.  
  - Write press releases (PR / Media): Human only 1.54 hrs; With AI 14.5 min; Saved 1.30 hrs.  
  - Advise on drug interactions (Healthcare): Human only 1.19 hrs; With AI 10.4 min; Saved 1.02 hrs.  
  - Answer customer questions (Customer service): Human only 0.35 hrs; With AI 3.4 min; Saved 0.29 hrs.

### Country and wage coverage
- The 117 sample countries span the full income spectrum and cover 79% of world GDP. China is a notable exclusion (Anthropic does not operate there).  
- The ACI ranks occupations by each country’s own wages (not US wages). Country-level employment shares and occupation wages are from ILOSTAT (2024 or latest available year).  
- Of the 117 countries, 86 have direct ILO employment and wage data at ISCO-1 level, covering 67% of world GDP; these 86 form the baseline regression sample.  
- Two versions of aggregate LCE provided:
  - Baseline: uses only the 86 sample countries.  
  - Expanded: covers 115 of the 117 task-enriched countries by substituting employment or wage data from a matched peer country wherever direct ILO data are incomplete. Section 3.5 re-estimates regressions on this expanded sample and reaches the same conclusions.  
- Additional country covariates and sources:
  - GDP figures: IMF World Economic Outlook (April 2026). Denominator of labor-cost-equivalent-to-GDP ratio is nominal GDP in current US dollars for 2025. Income control in regressions is GDP per capita at PPP for 2025; both are in 2025 values.  
  - World Bank WDI: unemployment rate; services value added (% GDP); share of individuals using the internet (%).  
  - Income inequality: disposable-income Gini from Standardized World Income Inequality Database (Solt, 2020).  
  - IMF: AI regulatory readiness, the regulatory sub-index of the AI Preparedness Index (Cazzaniga et al., 2024). Only the regulatory sub-index enters regressions due to collinearity of other index dimensions with income.  
  - Binary indicator for whether English is an official or de facto official language (used rather than English proficiency).

### Methodology — Labor Cost Equivalent (LCE)
- Formula (Eq. 1):
  - LCE = sum_c sum_s sum_{t ∈ s} ( n_{c,t} / N*_{country} × C_{total,w} × h_t × W^{monthly}_{c,s} × 12 / 2,080 )
  - Definitions (from Table 4 and text):
    - n_{c,t} / N*_{country}: fraction of country-level classified conversations on task t (N*_{country} = 566,418 in R5). Multiply by C_{total,w} for estimated actual conversations. Source: AEI. Observed every wave; varies.  
    - h_t: hours saved per conversation on task t (human-only minus human-with-AI; e.g., 3.54 − 0.30 = 3.24 hours for “modify existing software”). Source: AEI. Global: R4 and R5 observed; R3 imputed from R4.  
    - W^{monthly}_{c,s}: ILO mean monthly wage in nominal USD for ISCO-1 occupation s in country c (matched-peer fallback for 22 countries). Source: ILOSTAT. Country varies.  
    - 2,080: annual working hours (40 hrs × 52 weeks) used to convert annualized wages to hourly. Constant.  
    - C_{total}: total AI conversations per week across all platforms; baseline approximation = 200M/0.25 ≈ 800 million. Public data; global scope; approximate.  
- Interpretation and caveats:
  - LCE annualizes and values hours saved at occupation hourly wage. The LCE is an indicative measure of productivity gains implied by current AI usage rather than a strict bound. Valuing every classified conversation at its full estimated time saving tends to overstate realized gains; excluding enterprise API usage tends to understate the total; net direction of bias is ambiguous.

### Parameters and scaling
- Denominator treatment and notclassified conversations:
  - Global-level notclassified rate: 2.7% (27,364 out of 1,000,000) not classified at global level.  
  - Country-level notclassified rate: 29.4% of O*NET-enriched country-level conversations (235,696 out of 801,151 across 117 countries). Of these, roughly 208,000 were successfully classified at the global level but not at the country level (small-cell suppression due to privacy thresholds).  
  - Because these lack country-level task assignment, they cannot enter LCE or ACI calculations. The analysis scales from country-tagged conversations only (N*_{country}), redistributing untagged volume proportionally rather than treating it as zero.  
- Scaling from Claude to all AI platforms:
  - Total Claude.ai conversations per week ≈ 200 million (derived from 18.9 million monthly active users).  
  - Claude’s share of professional AI conversations ≈ 25% (based on enterprise market-share data).  
  - This yields C_{total,w} ≈ 800 million/week in R5 (200M / 0.25 = 800M), scaled proportionally across earlier waves using SimilarWeb website traffic.  
  - Scaling assumptions affect only the level of LCE; ACI and regressions (which depend on relative distribution) are unaffected.  
- Appendix references:
  - Appendix Section B documents full derivation and sources for scaling parameters. Appendix Table A2 reports sensitivity of LCE to alternative scaling assumptions.

### Key assumptions underlying the LCE (Box 1)
- A1. h_t is at the task level; within-task heterogeneity is absorbed into the average.  
- A2. h_t is estimated by Claude, not measured. Both time variables (human-only and human-with-AI) are AI estimates, introducing potential bias.  
- A3. Each hour saved is valued at the occupation’s hourly wage (assumes time saved is worth what a worker would have been paid for that time).  
- A4. W_{c,s} is held constant.  
- A5. C_{total} varies by wave. For R5 (Feb 2026), C_{total} ≈ 800 million AI conversations per week across all platforms (200M Claude / 25% market share). For earlier waves, scaled proportionally using claude.ai monthly website traffic as proxy: R3 (Aug 2025) ≈ 414M/week; R4 (Nov 2025) ≈ 489M/week.

### Worked example (Box 2): “modify existing software” in the US (R5, one week)
- US conversations on this task (O*NET 15-1132, Software Developers): 4,722 of 566,418 categorized country-level conversations (0.834%).  
- With C_{total} = 800 million (all AI platforms):
  - Estimated US conversations on this task this week: 0.00834 × 800,000,000 = 6,672,000.  
  - Human-only time: 3.54 hours. Human-with-AI time: 0.30 hours.  
  - Hours saved per conversation: 3.54 − 0.30 = 3.24 hours.  
  - Total hours saved: 6,672,000 × 3.24 = 21,617,000 hours.  
  - Hourly wage (US Professionals, ISCO-1 group 2, ILO): $8,171 × 12 / 2,080 = $47.14/hr.  
  - Dollar value this week: 21,617,000 × $47.14 = $1.02 billion.  
  - Annualized: $1.02B × 52 = $53 billion.

### AI Concentration Index (ACI) — definition and interpretation
- ACI formula (Eq. 2):
  - ACI_{c,w} = 2 × sum_s [ g_{c,s,w} × ( R_{c,s} − 1/2 ) ]
  - where g_{c,s,w} = ( sum_{t ∈ s} n_{c,t} h_t ) / ( sum_{all t} n_{c,t} h_t ) is the ratio of hours saved by AI conversations in occupation s to hours saved across all occupations in country c.  
  - R_{c,s} is occupation s’s fractional rank in the country’s own wage distribution: with occupations ordered from lowest to highest wage, R_{c,s} is the cumulative employment share up to the midpoint of occupation s (employment shares e_{c,s} enter through ranks).  
- Index properties:
  - Ranges from −1 to +1. Positive when gains tilt toward high-wage occupations; negative when toward low-wage occupations; zero when time savings are proportional to employment (g_{c,s,w} = e_{c,s} for every occupation).  
  - Can turn negative when AI concentrates in low-wage occupations.

### Regressions — empirical approach (brief)
- Three sets of OLS regressions estimated:
  1. Predictors of how much AI value a country captures relative to its GDP (LCE as a percentage of GDP in R5).  
  2. Predictors of the level of AI concentration (ACI pooled across the three country-level waves).  
  3. Predictors of whether concentration is rising or falling (change in ACI across consecutive waves).  
- Each regression shown in parsimonious specification (GDP per capita only) and full specification (all covariates); progressive specifications in Appendix Tables A3–A5.

### Summary of key empirical findings (as stated in Section 3 results preview)
- Aggregate productivity value: large and growing, driven not by per-conversation value increases but by AI reaching occupations that employ far more workers.  
- Distributional pattern: gains are unequally distributed — in nearly every country they initially tilt toward high-wage occupations, with greater concentration in developing economies, although the tilt is declining over time.  
- Predictors from regressions:
  - Income and AI regulatory readiness predict the level of concentration.  
  - The trajectory (whether concentration is falling) is predicted by whether a country’s official or de facto official language is English (representing whether a country’s institutional knowledge base is well represented in training data of English-dominant LLMs).

*Source: wpiea2026147 - 2.1 Data (chapter/section content from the supplied PDF).*

### Box 4 provides a simple worked example illustrating the ACI formula.

### Box 4 provides a simple worked example illustrating the ACI formula.

### Aggregate AI gains: wage index and LCE
- Wage index (usage-weighted average US occupation-level wage across global AI conversations):
  - Declines from $82,353 in January 2025 to $77,845 in February 2026 (5.5%).
  - Decline driven by occupational share shifts: Computer & Math lost 4.9 percentage points (pp) of global AI conversation share while Education gained +4.1 pp, Sales +2.7 pp, and Office/Admin +1.6 pp.
- Labor‑cost equivalent (LCE) for the 86 sample countries:
  - Moves from $1.2 trillion annualized in R3 to $2.7 trillion in R5 (3.4% of GDP).
  - Divergence between falling wage index and rising LCE reflects the employment‑scale effect: lower per‑hour value but much larger numbers of workers in lower‑wage occupations.
- Margins driving LCE growth (from Equation 1 decomposition):
  - Extensive margin (growth in number of AI conversations) accounts for about 72% of the increase from R3 to R4 and about 84% from R4 to R5.
  - Intensive margin (change in value per conversation) accounts for the remainder: about 28% then 16%.
- Income‑tilt of aggregate gains (R5):
  - AI gains as share of GDP: high‑income countries 4.2% of GDP; middle‑income roughly 0.6%; low‑income 0.1%.
  - High‑income countries account for 96% of total LCE for the 86 sample countries.

### Scaling scenarios and sensitivity
- Median scenario assumptions: 200 million Claude conversations per week and a 25% market share → baseline LCE reported ($2.7 trillion in R5 for 86 countries).
- Conservative scenario: 150 million conversations and a 35% share → LCE = $1.4 trillion.
- Aggressive scenario: 250 million conversations and a 15% share → LCE = $5.5 trillion.
- Scaling assumptions affect only the level of LCE, not its occupational distribution or ACI‑based results.

### AI concentration index (ACI): within‑country distribution
- Definition and interpretation:
  - ACI derived from AI concentration curves (Lorenz‑style): orders nine ISCO‑1 occupation groups from lowest‑paid to highest‑paid using ILO wages; ACI proportional to net signed area between curve and 45° line. Larger ACI → greater concentration of AI gains among higher‑paid occupations.
- Cross‑country examples (R5, Feb 2026):
  - United States: ACI = 0.44. US R5 shares (example in Box 3): Professionals 84%, Managers 6%, Clerical 5%.
    - US ISCO‑1 group excerpt (R5): Professionals ILO wage $98,047; capture 83.6% g share; Cum emp 88.3% → ACI_US = +0.44 (per Box 3 calculation).
  - India: ACI = 0.86.
  - Kenya: ACI = 0.81.
- Income pattern in ACI (R5):
  - Sub‑Saharan Africa and South Asia with GDP per capita below $5,000: ACIs cluster between 0.85 and 1.00.
  - Western Europe and North America with GDP per capita above $50,000: ACIs range from 0.35 to 0.62.
- Time dynamics (changes across waves):
  - Of 81 countries present in R3–R5, 20 had no ACI change across two consecutive windows because all AI conversations mapped to a single occupational group.
  - Among 61 countries with changes, diffusion toward declining ACI increased: 37 countries (46% of total) saw ACI fall in the second period (R4→R5), up from 18 countries (22%) in the first (R3→R4).
  - 28 countries reversed from rising ACI through November 2025 to falling by February 2026; 12 of these 28 are middle‑income economies.

### Cross‑country regression evidence (Table 5 summary)
- Dependent variables and samples:
  - Columns (1)–(2): LCE as % of GDP, pooled R3, R4, R5, wave fixed effects, Observations = 252, 250.
  - Columns (3)–(4): ACI level, pooled R3, R4, R5, wave fixed effects, Observations = 252, 250.
  - Columns (5)–(6): ∆ACI across R3→R4 and R4→R5 windows (first differences), Observations = 165, 164.
- Key coefficients and significance (selected):
  - log GDP per capita:
    - Column (1): 0.384*** (SE 0.068) predicting LCE/GDP.
    - Column (3): −0.135*** (SE 0.012) predicting ACI level.
  - AI regulatory readiness:
    - Column (2): 8.918*** (SE 2.689) predicting LCE/GDP.
    - Column (4): −1.619*** (SE 0.383) predicting ACI level.
  - English official:
    - Column (4): −0.005*** (SE 0.002) predicting ACI level.
  - Other notable predictors:
    - Services VA (share of GDP): Column (1) 1.941* (SE 1.077) predicting LCE/GDP; Column (6) 0.032** (SE 0.016) predicting ∆ACI.
    - Internet users (%): Column (4) 0.002*** (SE 0.001) predicting ACI level.
    - Gini (SWIID): Column (4) 0.005*** (SE 0.002) predicting ACI level.
- Interpretation:
  - Regulatory readiness and economic structure (service share) explain a substantial portion of cross‑country variance in LCE/GDP; inclusion of regulatory readiness reduces the simple income effect.
  - Richer countries and those with stronger regulatory frameworks have more broadly distributed AI gains (lower ACI).
  - Higher income inequality (Gini) and higher internet penetration (conditional on GDP per capita) are associated with higher ACI.
  - English official status predicts faster declines in ACI (faster diffusion), interpreted as greater representation in LLM training data enabling broader occupational usefulness.
- Model fit:
  - R^2 ranges: LCE/GDP specifications 0.291 and 0.440; ACI level specifications 0.673 and 0.787; ∆ACI specifications 0.065 and 0.099.

### Robustness: bounding hours saved (task‑level uncertainty)
- Construction of strict bounds using AEI 95% confidence intervals:
  - Lower bound per task: h_lower_t = max(0, h_human-only_CI-lower − h_human-with-AI_CI-upper / 60).
  - Upper bound per task: h_upper_t = max(0, h_human-only_CI-upper − h_human-with-AI_CI-lower / 60).
- Aggregate effects of extreme bounds:
  - Task‑level mean h_t swings from 1.80 hours (−33%) to 3.81 hours (+41%).
  - Aggregate LCE moves within a narrow band of $2.5 to $2.9 trillion (baseline $2.7 trillion).
  - Rationale: high‑volume tasks (5,000+ conversations) have confidence intervals averaging 6% of their point estimates and dominate the LCE; low‑volume tasks (10–50 conversations) have intervals averaging 89% but contribute little to aggregate LCE.
- Distributional robustness:
  - No country’s ACI shifts by more than half a percentage point under either bound.
  - Cross‑country ranking of ACI values remains unchanged.
  - Bounding shifts hours saved roughly proportionally across occupations, leaving shares of AI time savings (and thus ACI) nearly identical.

*Source: wpiea2026147 - Box 4 provides a simple worked example illustrating the ACI formula. (https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026147.pdf)*

### 3.5    Extension:  expanded sample with peer-borrowed wages

### 3.5    Extension:  expanded sample with peer-borrowed wages

### Sample construction and peer-borrowing method
- Expanded the baseline 86-country sample by adding 31 countries that lack direct ILO wages or employment at the ISCO-1 level, producing an expanded 117-country analysis sample.
- For each missing dimension (employment or wages), the missing component is borrowed from the country closest in log GDP per capita among those reporting that specific component; employment and wages are matched separately.
- Appendix Table A8 lists peer-country assignments; peer sources supply substituted data: (e) employment, (w) wage, (ew) both.

### Aggregate LCE on the expanded sample
- On the expanded 117-country sample the aggregate R5 LCE is $2.9 trillion, or 3.1% of these countries’ combined GDP.
- By comparison, the 86-country baseline LCE was $2.7 trillion (3.4%).
- Columns (1) and (3) of Table 6 cover 117 countries; columns (2) and (4) cover the 115 with a complete set of covariates.
- Standard errors are clustered by country in all columns.

### Regression re-estimation (Table 6: pooled R3 to R5; 117 countries)
- Specifications mirror Table 5; pooled R3 to R5 with wave fixed effects.
- Key coefficient outcomes (preserved qualitative conclusions):
  - AI regulatory readiness (AIPI) remains strongly significant in the LCE/GDP regression and keeps its negative sign for the ACI level, significant at the 10% level in the expanded sample.
  - English-language and income slopes retain their signs and significance.
- R2 values are somewhat lower for the ACI columns (example: 0.53 versus 0.79 with all controls in the baseline), consistent with peer-borrowed data introducing some measurement error.
- Table 6 (selected reported values):
  - Observations: 337 / 332 / 337 / 332 / 219 / 216 (columns (1)–(6) respectively).
  - R2: 0.256 / 0.421 / 0.394 / 0.527 / 0.053 / 0.085 (columns (1)–(6)).
  - Example coefficient estimates (standard errors in parentheses):
    - log GDP per capita: 0.367*** (0.062) in column (1); −0.107*** (0.016) in column (3).
    - AI regulatory readiness: 8.012*** (2.201) in column (2) on LCE/GDP; −0.871* (0.447) in column (4) on ACI level.
    - English official: 0.244 (0.192) in LCE/GDP column; −0.083** (0.040) in ACI column.
  - Significance notation: * p < 0.1, ** p < 0.05, *** p < 0.01.

### Robustness and limitations of the extension
- Qualitative results are preserved despite the expanded sample.
- Peer-borrowed data introduce measurement error that reduces explanatory power for some specifications (notably ACI regressions).
- Columns (1) and (3) (LCE/GDP and ACI level without full covariates) include the full 117 countries; columns with full controls include 115 countries.

### Connection to broader findings and interpretation (from subsequent sections)
- Expanded-sample results reinforce two policy-relevant relationships:
  - Representation of a country’s institutional knowledge in AI training data (proxied by official English-language status) is associated with faster broadening of AI gains (faster ACI declines), even after controlling for income.
  - AI regulatory readiness matters for both the aggregate value AI creates (dominant predictor of LCE relative to GDP) and for broader distribution of gains (associated with lower ACI).
- Caveats reiterated in the paper (relevant to interpretation of expanded-sample results):
  - LCE is a partial-equilibrium measure valuing time savings at current wages and does not net out costs (job displacement, wage compression, reallocation) or infrastructure/energy/subscription costs.
  - The hours-saved variable h_t is estimated by Claude rather than measured experimentally; ACI is robust to h_t assumptions because it depends on relative distribution of time savings across occupations.
  - Data limitations: Claude.ai web conversations only (exclude API-based enterprise usage and other platforms); each wave covers one week; ACI likely understates concentration due to enterprise usage skewing to high-skill occupations.

*Aggregate Gains from AI and Their Distribution: Global Evidence from Usage Data — Working Paper No. WP/2026/147*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026147.pdf_