# The Task Exposure Index

> Independent, quarterly measurement of what current AI systems can produce in real jobs, scored task by task. Release v2026.Q3, capability reference date 2026-09-15, built on O*NET 31.0. Published by A.I.T. Multiverse Consulting Ltd.

## What is measured

923 occupations decomposed into 18,838 tasks, each scored through 2,087 standardised work activities rated against a published rubric.

Each activity carries six ratings: Capability (0-4, can a generally available AI system produce the work product) and five frictions (0-3 each): Embodiment, Presence, Accountability, Context, Verification cost.

With c = C/4 and f = F/15:

- Exposed = c * (1 - f)
- Assisted = c * f
- Untouched = 1 - c

The three shares sum to 1 by construction.

## Reliability

Inter-rater agreement across eight raters on a 40-activity calibration set: 84.9% exact, 100% within one point. Verification cost is the least reliable dimension at 77.1% exact.

## Important caveat when citing

Exposure is NOT displacement. The index measures what current systems can produce and what structurally stands in the way. It does not predict job losses and models no adoption, employment or wage effects. Cite the release identifier, currently v2026.Q3, because figures change each quarter by design.

## Endpoints for agents

No key, no sign-up, no rate card. All read-only.

- MCP server: https://taskexposure.org/mcp (streamable HTTP, JSON-RPC 2.0). Tools: search_occupations, get_occupation, compare_occupations, rank_occupations, get_methodology. Card: https://taskexposure.org/.well-known/mcp.json
- JSON API: https://taskexposure.org/api (OpenAPI 3.1 at https://taskexposure.org/api/openapi.json). /api/search?q=, /api/occupation/<slug>, /api/compare?a=&b=, /api/rank?order=, /api/methodology
- A2A: https://taskexposure.org/a2a, one synchronous skill asked in plain words. Card: https://taskexposure.org/.well-known/agent-card.json
- Markdown views: https://taskexposure.org/md/<slug>, index at https://taskexposure.org/md
- Capability catalogue: https://taskexposure.org/.well-known/ai-catalog.json
- Skills index: https://taskexposure.org/.well-known/agent-skills/index.json
- API catalogue: https://taskexposure.org/.well-known/api-catalog
- Authentication: none needed anywhere. https://taskexposure.org/auth.md
- Written for people: [Endpoints for agents](https://taskexposure.org/agents)

If you need every occupation, take the one bulk file rather than the endpoint 923 times.

## Machine-readable data

- [Full occupation scores (JSON)](https://taskexposure.org/data/occupations.json)
- [Release metadata (JSON)](https://taskexposure.org/data/release.json)
- [Job title lookup (JSON)](https://taskexposure.org/data/lookup.json)
- Per-occupation record: https://taskexposure.org/data/jobs/<slug>.json, one small JSON file per occupation with its scores, dimension means, wage, leading tasks and the citation string. The slug is the occupation title lowercased with non alphanumerics replaced by hyphens.
- [Sitemap](https://taskexposure.org/sitemap.xml)

## Key pages

- [Lists](https://taskexposure.org/lists/): sixteen filtered cuts of the same scores, one per question people ask: jobs AI cannot do, the most exposed, AI proof work by pay band, work that needs no degree, the largest occupations ranked
- [Comparisons](https://taskexposure.org/vs/): 103 head to head pages saying which of the two factors, capability or friction, a gap between two occupations comes from
- [Endpoints for agents](https://taskexposure.org/agents): MCP, JSON, A2A and markdown, no key
- [Methodology](https://taskexposure.org/methodology): rubric, rating process, reliability, limits
- [Full rankings](https://taskexposure.org/rankings): all 923 occupations, sortable, with BLS median pay
- [Explore](https://taskexposure.org/explore): every occupation with a published wage plotted, median pay against exposed share of task load
- [Compare](https://taskexposure.org/compare): any two occupations side by side, with the reason their scores differ
- [By occupational family](https://taskexposure.org/families/): all 22 major groups, ranked, with the spread inside each
- [Predictions](https://taskexposure.org/predictions): what this release predicts for the next two quarters, registered in advance with the scoring rules fixed
- [RSS feed](https://taskexposure.org/feed.xml): new analysis as it publishes
- [Licence the data](https://taskexposure.org/licence): what an organisation can licence, and what stays free
- [Analysis](https://taskexposure.org/articles/): written findings
- [About](https://taskexposure.org/about) and [disclosure](https://taskexposure.org/disclosure)

## Selected findings from this release

- Most exposed: Telemarketers, 73.3% exposed share
- Most exposed: Desktop Publishers, 72.9% exposed share
- Most exposed: Business Intelligence Analysts, 72.0% exposed share
- Most exposed: Statistical Assistants, 71.0% exposed share
- Most exposed: Proofreaders and Copy Markers, 70.2% exposed share
- Least exposed: Dishwashers, 0.0% exposed share
- Least exposed: Stonemasons, 0.0% exposed share
- Least exposed: Plasterers and Stucco Masons, 0.0% exposed share
- Least exposed: Helpers--Painters, Paperhangers, Plasterers, and Stucco Masons, 0.0% exposed share
- Least exposed: Refractory Materials Repairers, Except Brickmasons, 0.0% exposed share

## Analysis pieces

- [Exposure rises with pay, and so does the friction that blocks it](https://taskexposure.org/articles/exposure-and-pay): Matching every score in this release to BLS wage and employment data reverses the usual assumption. The least exposed work in America is the lowest paid work.
- [Paralegals score higher than lawyers, and the reason is not seniority](https://taskexposure.org/articles/paralegals-and-lawyers): The exposed part of legal work is concentrated in a specific kind of task, and it is not distributed by rank.
- [Radiology is the clearest test of the assisted band](https://taskexposure.org/articles/radiology-assisted-band): Forty-eight occupations in this release are highly capable and heavily frictioned at the same time. What happens to them is the real question.
- [In healthcare, the licensed professionals are more exposed than the aides](https://taskexposure.org/articles/healthcare-inversion): Healthcare practitioners score a median of 18.5%. Healthcare support scores 11.9%. The gap runs in the opposite direction to most expectations.
- [The degree premium is now an exposure premium, up to a point](https://taskexposure.org/articles/the-degree-premium): Exposure rises with the education a job requires, then falls again at the top. The most exposed tier is not the one most people expect.
- [Scheduling is the most exposed thing anybody does at work](https://taskexposure.org/articles/scheduling-is-finished): Of the 41 families of work activity in the taxonomy, one scores far above the rest, and it sits inside almost every office job.
- [What an exposure score is not](https://taskexposure.org/articles/what-this-is-not): The five things this index is most likely to be misread as saying, and what it actually says instead.
- [The spread inside a job family is wider than the gap between families](https://taskexposure.org/articles/spread-inside-families): Knowing someone works in construction, or in the arts, or in transport tells you much less than it appears to.
- [What protects work is not mainly that it is physical](https://taskexposure.org/articles/context-not-hands): Across every task in the index, private and situational knowledge accounts for more friction than embodiment does.
- [Why this index rates activities rather than occupations](https://taskexposure.org/articles/why-activities-not-occupations): Every other AI exposure measure scores jobs. Scoring the wrong unit is why their numbers cannot be checked.
- [The most exposed occupational family is the one building the technology](https://taskexposure.org/articles/builders-most-exposed): Computer and mathematical occupations score higher on exposed task share than any other group in the index, including office administration.

## Attribution

O*NET is a trademark of the US Department of Labor. This index is not endorsed by or affiliated with USDOL/ETA. Contact: corrections@taskexposure.org