---
id: "lib-2026-07-00-aggregate-gains-from-ai-and-their-distribution"
type: "library"
title: "Aggregate Gains from AI and Their Distribution: Global Evidence from Usage Data (WP/26/147)"
lang: "EN"
org: "Unknown"
source_type: "Paper"
date: "2026-07-01"
week: "2026-W27"
url: "https://nosov.ai/library/lib-2026-07-00-aggregate-gains-from-ai-and-their-distribution/"
original_url: "https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026147.pdf"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "IMF working paper by Ha Nguyen (IMF) and Rachel Yuting Fan (ex-World Bank) building two new measures from five public waves of Anthropic Economic Index data, with stated assumptions, sensitivity ranges and replication code."
summary_basis: "full-text"
original_minutes: "60"
tags: ["productivity", "macro", "ai usage", "distribution"]
frames: ["solow-paradox"]
---

# Aggregate Gains from AI and Their Distribution: Global Evidence from Usage Data (WP/26/147)

Tier A (Strong evidence): IMF working paper by Ha Nguyen (IMF) and Rachel Yuting Fan (ex-World Bank) building two new measures from five public waves of Anthropic Economic Index data, with stated assumptions, sensitivity ranges and replication code.

Summary based on the full text.

## In short

IMF Working Paper 26/147 (July 2026) asks how large the labour cost saved by AI is and who gets it. Using five releases of the Anthropic Economic Index (January 2025 to February 2026), it values time saved at each country's wages and builds an AI concentration index (ACI) of how far gains tilt to high-paid occupations. Gains are large and rising; they remain concentrated in a professional enclave in developing countries but are broadening.

## Key numbers

- $2.7 trillion: annual labour cost equivalent of time saved by AI, about 3.4% of the 86 sample countries' GDP
- $1.4 to $5.5 trillion: range of the labour cost equivalent under alternative scaling assumptions
- 25 to 47: countries with a falling AI concentration index over six months (23% to 42%)
- 5.5%: fall over 13 months in the usage-weighted average wage of occupations served by AI conversations

## What it reports

- Labour cost equivalent (LCE) of about $2.7 trillion a year, 3.4% of combined GDP for 86 countries ($79.2 trillion), with a range of $1.4 to $5.5 trillion.
- ACI near 1.0 in developing economies such as Uganda and Cambodia; high-income countries average 0.4 to 0.6.
- Concentration is falling: countries with a declining ACI rose from 25 (23%) to 47 (42%) in six months.
- AI is spreading from software into education, sales and office work; the usage-weighted average wage fell 5.5% over 13 months, yet total LCE rises because those occupations employ far more people.
- Income and regulatory readiness predict the starting level; official English-language status predicts faster broadening.
- A footnote estimate of enterprise API value (about $1.14 trillion a year) is excluded from the headline.

## How they got it

Each AEI wave samples one million Claude conversations mapped to O*NET tasks; country geocoding from wave 3. Conversations are crosswalked to ISCO occupation groups and ILO wages and employment, scaled up by Claude's estimated market share. Regressions relate ACI and LCE to income, regulation and language.

## Limits

- Data from one provider's consumer product; scaling to the whole market depends on market-share assumptions.
- LCE values time saved; it is not GDP impact and says nothing on whether workers or firms capture it.
- Classification relies on Anthropic's own task classifier; enterprise API use is excluded.

## Why it matters for companies

For multinationals, the finding that AI use in lower-income countries is confined to a small professional group means rollouts there may widen internal gaps unless local-language and role-specific support is added. The English-language effect is a practical point for firms operating in non-English markets.

Original: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026147.pdf