---
title: "How We Read AI Exposure by Task: Roles Method | Nexum"
description: "The data behind our role pages: O*NET tasks, the GPTs are GPTs exposure labels, Anthropic Economic Index usage and BLS pay, and how to read them."
url: https://nexumautomations.com/roles/method
updated: 2026-10-31
publisher: Nexum Automations, an AI consulting firm
---

> Nexum Automations is an AI consulting firm that builds what it recommends. It finds where AI pays off in a company, builds it on the tools the team already uses and stays until it is live and working.

Method

# How we read AI exposure, task by task

Every role page reads one occupation through published data: its tasks, how exposed each task is to AI, how much AI is already used on it, and what the job pays.

Last reviewed October 30, 2026

## What a role page shows

* The tasks
The occupation's task statements from O\*NET
* Exposure rated by people
Human labels from Eloundou et al.
* Exposure rated by GPT-4
The same rubric, applied by the model
* Observed AI use
Anthropic Economic Index, where a task has data
* Jobs and pay
BLS employment and median wage

Each role page starts from the official task list of one occupation. For 8 to 15 of those tasks, a table puts three published readings side by side.

Below the table sit the job's employment and median pay from the U.S. Bureau of Labor Statistics, with the month of the data, and a note on how to read it.

The pages live in [AI and your role](https://nexumautomations.com/roles). The older guides, such as [will AI replace jobs](https://nexumautomations.com/guides/will-ai-replace-jobs), stay where they are and link here.

## Where the tasks come from

* 923
O\*NET data-level occupations
[O\*NET Resource Center, About O\*NET](https://www.onetcenter.org/overview.html)
* 19,000
task statements, at least, linked to work activities
[O\*NET Resource Center, About O\*NET](https://www.onetcenter.org/overview.html)

O NET describes itself as the nation's primary source of occupational information. Its database is free, and O NET says it is updated quarterly.

The O\*NET-SOC taxonomy has 1,016 occupational titles, of which 923 are data-level occupations. The database links more than 19,000 task statements to detailed work activities.

Each role page uses the task statements O NET lists for that occupation, in O NET's own words.

## The exposure labels

1. E1
An LLM alone makes the task at least twice as fast
2. E2
It takes extra tools built on top of the LLM
3. E0
No exposure under the rubric

* 15%
of U.S. worker tasks faster with an LLM alone, per the paper
[Eloundou, Manning, Mishkin and Rock, GPTs are GPTs (arXiv)](https://arxiv.org/abs/2303.10130)

In the paper "GPTs are GPTs", Eloundou, Manning, Mishkin and Rock (2023) rated every O\*NET task for exposure to large language models (LLMs).

Anthropic summarizes their test: whether an LLM could make a task at least twice as fast. A task scores 1 if an LLM alone can do that, 0.5 if it needs extra tools built on the LLM, and 0 otherwise.

Behind those scores sit three labels: E1 for an LLM alone, E2 for an LLM plus tools, and E0 for no exposure. People rated the tasks, and so did GPT-4\. Role pages show both ratings.

The paper estimates that with an LLM about 15% of all U.S. worker tasks could be done significantly faster at the same quality, and between 47% and 56% with software built on top.

## Observed AI use

* 97%
of observed tasks were rated feasible by Eloundou et al.
[Anthropic, Labor market impacts](https://www.anthropic.com/research/labor-market-impacts)
* 68%
of observed usage falls on tasks an LLM alone can speed up
[Anthropic, Labor market impacts](https://www.anthropic.com/research/labor-market-impacts)

Exposure says what is possible. The Anthropic Economic Index looks at what people actually do with Claude, Anthropic's AI assistant, mapped to the same O\*NET tasks.

Anthropic reports that 97% of the tasks it observed fall into categories Eloundou et al. rated feasible. Tasks rated fully feasible for an LLM alone account for 68% of observed usage.

Its observed exposure measure asks which of the tasks LLMs could speed up are actually seeing automated use in professional settings.

Where a task has no usage in the data, the page says "no observed data". A missing value is not a zero.

## Jobs and pay

* 830
occupations with yearly employment and wage estimates
[BLS, Occupational Employment and Wage Statistics](https://www.bls.gov/oes/)

Employment and pay come from the BLS. Its Occupational Employment and Wage Statistics program produces employment and wage estimates every year for about 830 occupations.

Role pages show the median annual wage and employment with the month of the estimate, as the BLS publishes them in its Occupational Outlook Handbook.

For the wider picture, see [AI job displacement statistics](https://nexumautomations.com/statistics/ai-job-displacement).

## How to read a role page

1. Read the tasks
What the job is made of, in O\*NET's words
2. Compare the ratings
Where people and GPT-4 agree or differ
3. Check observed use
Where AI is already in use
4. Look at pay
What the work costs today

High exposure means AI could speed a task up. A job is many tasks, and people still decide, check the work and deal with customers.

The paper itself says it makes no predictions about when LLMs will be developed or adopted.

Observed use comes from the people who use one AI assistant, so it shows direction rather than the whole economy.

The two exposure columns can disagree. Where people and GPT-4 rate a task differently, the gap is worth a closer look.

## What this method does not measure

It does not forecast job losses, wages or hiring in any company.

It does not measure how good AI output is on a task. It shows whether a speed-up is possible and whether use is observed.

It does not show adoption inside a given business. That depends on its processes, data and rules.

## Sources, licenses and updates

This page was last reviewed on October 30, 2026\. Each role page shows the date of the data it uses.

* O\*NET database, U.S. Department of Labor, Employment and Training Administration: a free public database
* Eloundou, Manning, Mishkin and Rock (2023), "GPTs are GPTs": the paper on arXiv; its task labels are published in the authors' repository under the MIT license
* Anthropic Economic Index, labor market impacts data: CC-BY license for the data, MIT for the code
* U.S. Bureau of Labor Statistics: OEWS estimates and the Occupational Outlook Handbook, May 2025 data

---

Source: https://nexumautomations.com/roles/method

Free AI diagnosis (30-minute call): https://nexumautomations.com/#book - Services: https://nexumautomations.com/services - Guides: https://nexumautomations.com/guides