Claude.ai Automation Share (Anthropic Economic Index)

## What this measures

Of the Claude.ai conversations Anthropic could classify by how people work with the AI, the share that are automation: the person directs Claude to complete a task outright ("directive") or relays errors back until it works ("feedback loop"). The rest are augmentation: iterating on a task together, learning, or checking work.

## Why it matters

Every AI economic scenario turns on whether AI helps people do tasks or does the tasks itself. This is a direct, measured read on that split from one of the largest AI products, rather than a survey.

## How to read it

Near 50% means people hand off about as many tasks as they work through with AI. A rising share means more delegation. Points before June 2026 are single sample weeks (dated at the start of the week); from June 2026 Anthropic publishes calendar months.

## What it does not say

It covers Claude.ai consumer plans (Free, Pro, Max) only: not the API, not Claude Code, not other companies' AI. It counts conversations, not hours of work or jobs. The sampling window changed from one week to one month in 2026, and classification methods can change between releases. The API series is not published here because its share jumps from about 80% to 94% between the two formats, which looks like a definition change.

## Source

Anthropic Economic Index, huggingface.co/datasets/Anthropic/EconomicIndex. Data licensed CC BY. From the June 2026 release: `collaboration_bucket_automation_pct`, global, overall. Earlier releases: Palanor computes (directive + feedback loop) ÷ (100 − none − not classified) from the published collaboration percentages, the same definition; the source file is named in each point's metadata.