Methodology

How SEDI is measured

AI usage as of May 2026 · Employment through 2026Q1 · Edition 2026Q3

SEDI shows, for every US state, metro area, industry and occupation, how much of the work AI is already used on, how that compares with similar places and industries, and what has happened to private jobs since late 2022. This page explains each number on the other pages.

Exposure

Exposure is the share of the work in a job that AI is already used on, from Claude usage in the Anthropic Economic Index. Anthropic publishes the Anthropic Economic Index, which reports how often people use Claude for each task in O*NET, the Department of Labor's lists of the tasks that make up each occupation. A task counts as having observed AI usage when it reaches at least 0.01% of US Claude.ai traffic in the edition's release. This edition reads the release of 26 June 2026, which covers AI usage as of May 2026.

Exposure share. For an occupation, the exposure share is the share of its tasks with observed AI usage. It runs from 0% to 100%, and the 0.01% floor is the only cut.

Occupations with no observed usage. The Economic Index lists only tasks with enough usage to report, and a task it does not list counts as having no observed usage. In 424 occupations, holding 30.4% of jobs, none of the tasks is listed, so the exposure share is 0% and their jobs count as not exposed. Most are hands-on work, such as freight movers, janitors and truck drivers. Some are office work, such as medical secretaries, claims adjusters and insurance underwriters, where usage on their tasks was too small for this release to list.

Exposed jobs and exposed payroll. An occupation's exposed jobs are its jobs times its exposure share: work that AI is already used on, counted in full jobs. Exposed payroll is its payroll times the same share. A place or an industry adds these up over its occupations.

Exposed share of jobs. This is the headline figure on every page: exposed jobs divided by the jobs in occupations that have an exposure score. The exposed share of payroll sits beside it.

Occupations published as one code. The employment data publishes some occupations as one code that O*NET rates in parts, such as home health and personal care aides, which O*NET rates as home health aides and as personal care aides. Such a code takes the average exposure share of its parts, each part scored on its own O*NET tasks, and its page lists those tasks.

Occupations without a score. Some occupations have no usable match to the task data, mostly the "all other" codes, which O*NET does not rate. They are left out of both the exposed jobs and the jobs they are divided by, so a missing score never counts as zero. Nationally the score covers 98.7% of jobs.

Share of AI usage that carries out the task. Usage is also coded by how people work with Claude. This share is the share of an occupation's classified usage in which Claude carries out the task directly (directive use, or a feedback loop with the user), against the share in which the user works through the task with it (task iteration, learning and validation).

BLS exposure tier. The Bureau of Labor Statistics publishes its own AI exposure categories for occupations (Low, Moderate, High and Very high). Occupation pages show the tier as an independent reference.

Exposure means AI is already used on these tasks somewhere. It does not mean the jobs will go.

One source of usage data

The AI usage data comes from one vendor: Anthropic's Economic Index, which covers usage of Claude. Usage of other AI tools is outside the data.

As a check, the occupation exposure shares are compared with a second vendor's occupation measure, OpenAI Signals: for each occupation, the mean share of US work messages across its O*NET work activities. Across occupations, the rank correlation between the two measures is 0.49 in this edition.

Places and industries

Places. The pages for states and metro areas use local occupation counts. The BLS Occupational Employment and Wage Statistics (OEWS, May 2025) count employment and wages for about 800 occupations in each state and metro area. Each occupation's jobs and payroll are multiplied by its exposure share and added up. Some local occupation figures are not published by the source; they are left out and never counted as zero. Jobs and payroll on a place page are the OEWS figures, which cover all employers.

A second estimate, from each place's industry mix with national staffing, is computed as a check, for places whose industries cover at least 50% of the place's private jobs. Nationally the two must agree within 3 percentage points before an edition is published, and any metro area where they differ by more than 10 points is reviewed.

Industries. An industry's exposure comes from its national staffing pattern (the OEWS national industry files, May 2025). Its jobs are QCEW private employment in the latest quarter, and its payroll is QCEW private payroll over the latest four quarters.

Industries in a metro area. Each metro page lists each 3-digit industry with its local private jobs. Their exposed share is the industry's national share, so their exposed jobs are estimates that assume national staffing.

Comparisons

Every state, metro area and industry is ranked nationally on its exposed share of jobs. Rank 1 is the highest share. The percentile gives the share of the others it sits above, and the national median is shown beside it.

Peer groups. Metro areas are also compared with metro areas of similar size by jobs: under 100,000; 100,000 to 250,000; 250,000 to 1 million; and 1 million or more. Industries are compared with industries at the same level (3-digit or 4-digit) in the same 2-digit sector. States have no peer group. A peer group with fewer than 5 members is not shown.

What has moved

What is counted. Private employment from the QCEW, as the three-month average for each quarter. For a state it is the BLS state total, for a metro area the sum of its member counties, and for an industry the national total.

Growth since late 2022. Average private jobs over the four quarters to 2026Q1 against the average over the four quarters to 2022Q3. Each average holds one of each quarter of the year, so the seasons cancel out.

Pre-pandemic pace. The growth the unit's own 2015 to 2019 trend gives over the same two periods, extended forward. The trend is fitted on the log of employment with a term for each quarter of the year. 2020 and 2021 are left out of every fit. A place page states its growth against this pace.

The chart. Each point is the average of the four quarters ending there. The jobs line and the trend line both read 100 over the four quarters to 2022Q3, so the gap between them at the latest quarter is how far growth sits from the pace.

Most and least exposed industries. The 3-digit industries are ranked by exposed share of payroll and cut into four groups that each hold about one fourth of 2019 national private employment. The top group is the most exposed fourth of industries and the bottom group the least exposed; a fourth is a group holding a fourth of private jobs. For a place, the same changes are shown for its industries in each group. The total and the two groups come from different sums, so the groups need not add up to the total. An industry page shows the change for the industry nationally, and each metro page shows the change for each of its industries.

Size and series. A unit with fewer than 1,000 private jobs in 2019 shows no change figure. Some 3-digit industries were recoded in 2022 and have no consistent series; they show the reason in place of a figure.

These figures show what has happened to jobs since late 2022. They do not show what caused it.

The tracker

The tracker page asks whether AI is showing up in jobs yet. It follows jobs and job postings in the most and least exposed work and sets each against the pace it held before the pandemic. Exposure is AI usage as of May 2026 in every quarter and month shown, because exposure has no earlier history.

Jobs growth and pace. The 3-digit private industries are ranked by exposed share of payroll and cut into four groups of equal employment, as above. For each of the most exposed and least exposed groups, growth is the average of private jobs in the latest four quarters over the average of the four quarters to 2022Q3, less 1. The pre-pandemic pace is the growth the group's own 2015 to 2019 trend gives over the same span. The lines on the page set both to 100 at late 2022. Rail transportation and the postal service are not shown, because private jobs describe only a fragment of that work.

Places. Metro areas with a change figure are cut into four groups of equal employment by exposed share of jobs, so the groups differ in size: the most exposed fourth holds a few large metro areas and the least exposed fourth many small ones. Growth and pace are built as for industries. The scatter shows every metro area, and the slope is the line through each metro area's growth against its pace on its exposed share of jobs, in points for each extra percentage point of exposed share, weighted by jobs. Both lines are set to 100 at late 2022, as for industries.

Job postings. Indeed Hiring Lab publishes a daily index of job postings for its occupational categories. Each category is matched to occupations by SOC code, and categories are ranked by their exposed share of jobs and cut into four groups of equal employment. A group's monthly value is the jobs-weighted mean of its categories' monthly means, shown only for complete months, and each group is rescaled to read 100 over July to September 2022. Postings start in February 2020, so there is no pre-2020 pace and no 95% range. The index counts every posting on Indeed, so it carries no sampling range.

Ranges. Two figures carry a 95% range, shown under "How this is measured". For industries it is the difference between the two groups' growth against pace, from how far each group's jobs strayed from its own trend in 2015 to 2019, allowing for the two moving together. For metro areas it is the slope, with a range that allows for metro areas differing in how widely they vary. The range of the difference is given only once the latest four quarters lie wholly after the four to 2022Q3.

What the tracker does not show. It describes jobs and postings. It says nothing about what caused a difference, and the groups hold industries and occupations that differ in many ways besides exposure.

Suppressed figures and lower bounds

The BLS does not publish a figure that could reveal a single employer. A figure that is withheld, uncollected or unmatched is shown as "not published by the source" and is never counted as zero.

Lower bounds. A metro page lists each 3-digit industry with at least 1,000 private jobs in the counties that publish a figure for it. When some counties' latest figures are withheld, the listed jobs and exposed jobs are sums over the counties that publish, so the true figure is at least that large. These figures are marked "≥". The same marks appear on the industry pages' lists of metro areas.

Counties left out of a change. A change compares quarters, so it needs the same counties in every quarter. A county counts toward a metro area's change when it reports jobs in every quarter the change uses: 2015 to 2019 for the pre-pandemic pace (2020 and 2021 are left out of the fit), and every quarter from 2021Q4, the first of the four quarters to 2022Q3, to 2026Q1. When the counties left out hold more than 5% of the metro area's jobs in the four quarters to 2022Q3, the metro area gets no change figure; at or under that share, the change is summed over the counties that remain and a note says which share was left out. The same rule applies to a metro area's rows for each industry and to a state's industries.

Connecticut. In 2024 Connecticut replaced its eight counties with nine planning regions, so county series break that year. Metro areas in Connecticut show no change figure until a consistent series exists. The state of Connecticut uses the BLS state total, which runs through the change, so the state page has one.

What the data does not show

  • Exposure means AI is already used on these tasks somewhere. It does not show that the jobs will go, or that employers in a given place use AI.
  • Usage is observed for one vendor's tool, Claude. Task usage is US traffic; the automation split is global, the only level it is published at. A task counts once its usage reaches the floor, however much more it gets.
  • Local occupation figures are annual (May 2025). The industry estimates and the metro industry rows assume national staffing.
  • The QCEW covers private employers that report to unemployment insurance. Self-employed workers and the public sector are outside the change figures.
  • Employment figures arrive about two quarters after the quarter they describe, and a four-quarter average is centered about one and a half quarters before its last quarter.
  • The change figures describe what happened to jobs. No figure on this site says what caused a change.
  • Job postings from the Indeed Hiring Lab appear on metro pages as context and in the tracker's postings figure. No exposure figure is derived from them.

Sources

Each source links to the page where it lives.

Rules for each edition

The rules and parameters are fixed before each edition is computed; this edition uses profile 4.3.0.

About 8421 Labs

8421 Labs operates a venture studio. Studio activity does not inform the index's rules or parameters.

Data and methodology under CC-BY-4.0. Copyright 8421 Labs.

Share this page