How We Calculate Risk
A task-level capability coverage index measuring how capable AI is across all occupational tasks, weighted by real employment data. Every score is reproducible. No black boxes.
Capability Coverage Index
How capable is AI across all occupational tasks today?
What It Measures
The Capability Coverage Index is a single number (0–100%) that answers: “across every task performed in the economy, weighted by how many people do that work, how much can AI do today?”
Unlike sentiment-based or news-driven indices, this number is grounded entirely in task-level AI capability scores for each of the ~20,000 tasks across ~1,100 occupations in the O*NET database, weighted by BLS employment data to reflect the actual structure of the workforce.
The Formula
The index is computed in two steps:
Step 1: Per-Occupation Coverage
For each occupation, multiply every task's time share by its AI capability score. This gives the fraction of that job's work that AI can currently handle.
Step 2: Employment-Weighted Global Average
Weight each occupation's coverage by its total employment. An occupation employing 1 million people counts 10x more than one employing 100,000.
Data Pipeline
Three data inputs combine into a single index score
O*NET Tasks
~20,000 tasks
Every task performed across 1,100+ occupations in the US economy
AI Capability
0–100 per task
How autonomously AI can perform each task today, from observed usage data
Time Fractions
% of workday
How much of each job's time is spent on each task, from O*NET surveys
Per-Occupation Coverage
What % of each job can AI handle?
Coverage Index
Single score from 0–100%
Updated daily with new capability assessments
Key Properties
Task-Grounded
Every number traces back to a specific task's AI capability score. No sentiment analysis, no news classification, no opinion weighting.
Employment-Weighted
Occupations are weighted by how many people actually work in them. A job employing millions moves the needle more than a niche role.
Reproducible
Computed via a single SQL query over the task database. No AI judgment calls at index time — the capability scores are pre-computed.
Hype-Resistant
No news sentiment, social media buzz, or CEO predictions. The index moves only when AI capability scores or employment data change.
Phase Thresholds
Each score level maps to observable labor market conditions
Per-Occupation Risk Scoring
Individual displacement scores for 39,000+ jobs using O*NET task data
AI Capability: Two Data Sources
Every task's AI capability score comes from one of two sources, used in strict priority order:
Real-world AI usage data from millions of Claude conversations. Measures how autonomously AI actually performs each task (1–5 scale), weighted by both consumer (Claude.ai) and production (API) usage. Covers ~3,200 O*NET tasks.
Source: Anthropic/EconomicIndex — release_2026_01_15
For the ~17,000 tasks not covered by Anthropic data, Gemini evaluates theoretical AI capability based on current technology, commercial products, and benchmarks. Less grounded than observed data, but better than leaving tasks unscored.
How Anthropic Scores Are Computed
Anthropic's raw data provides per-task usage counts and AI autonomy means from both Claude.ai and API traffic. These are combined via count-weighted averaging, then mapped to our 0–100 scale:
The autonomy scale is 1–5 (1 = fully human, 5 = fully autonomous). The formula maps this to 0–100.
Task Risk Formula
Every task is classified into one of five categories (determining its base risk) and scored for AI capability (from Anthropic or Gemini). These combine to produce a per-task risk score:
Tasks are weighted by estimated time fraction (% of workday) to produce the raw occupation score:
Task Categories & Base Risk
Higher base risk = more automatable task type
Occupation-Level Exposure Fallback
When Anthropic task-level coverage for an occupation is below 50%, the per-task calculation is replaced by Anthropic's occupation-level observed_exposure metric. This aggregate combines theoretical feasibility, observed usage frequency, and automation weighting — a more reliable estimate than sparse per-task data.
Protective Factors
Five factors can reduce the raw score by up to 55%, reflecting real-world barriers to displacement:
Empathy, negotiation, reading social cues
Novel ideation, artistic expression, innovation
Ambiguous judgment, ethical reasoning, strategic calls
Licensing, legal requirements, safety standards
Dexterous physical tasks, precision work
Risk Tiers
75 – 100
Critical Risk
50 – 74
High Risk
25 – 49
Medium Risk
0 – 24
Low Risk
Explore the data yourself
Search 39,000+ jobs, compare risk scores, and see exactly how each job is scored.