The A-ordningen is employers' monthly reporting of all employment relationships to the authorities; we use it via microdata.no, provided by SSB. The sample is all private-sector wage earners, ages 21–60. Figures 13–16 use the corresponding sample for the public sector. The reference period for each month is the week containing the 16th. The “Canaries sample” is the 397 STYRK-08 occupation codes that have an exposure score; it covers the large majority of private-sector employment.

Each occupation (four-digit STYRK-08) is linked to the exposure measure “beta” from Eloundou et al. (2024, Science) via the occupation classifications ISCO-08 and SOC. Beta measures the share of an occupation's tasks that can be performed substantially faster with large language models and associated software. The quintiles are equally weighted per occupation (not employment-weighted), the same construction as in the DEL dashboard. The usage groups in figures 10–12 are instead based on observed use: the share of the occupation's Claude requests classified as automating or augmenting in the Anthropic Economic Index (Handa et al. 2025), grouped as in Brynjolfsson, Chandar and Chen (2025).

From the 2026-09 release the exposure measure can be switched at the top of the page. The default is Eloundou et al. (2024), the task-based measure the index has used from the start. The alternative is the evidence-grounded measure from Mouchel et al. (2026), built on documented AI use. Both give equal-frequency quintiles over the same 397 occupations. The rank correlation between the measures is 0.94, and two-thirds of occupations fall in the same quintile. The selector applies to the headline figure and figures 1–2, for all three outcomes. The uncertainty band is bootstrapped separately for each measure. Figure 4, the occupation cases, AI usage, the occupation selector and the public sector use the Eloundou quintiles.

Employment is the number of wage earners in the group. New hires is the number of jobs with a registered start date in the window between the previous and current monthly reference date; the series is strongly seasonal, so a moving average or seasonally adjusted view is recommended. Pay is the group's average monthly cash pay, scaled up to full-time equivalent (FTE) using each occupation-age cell's average contracted hours — so a part-time position counts as its share of a full-time year and pay changes are not conflated with changes in working time; the pay series are nominal. All series are indexed to 100 in the reference month (February 2025, or November 2022 when ChatGPT is selected).

Norwegian labor market series have strong seasonal variation (among other things the January dip in the A-ordningen and the summer peak in new hires). The seasonal adjustment is a transparent X-11 core, computed separately for each series:

1. The series is taken in logarithms, so the seasonal pattern is treated as multiplicative.
2. The trend is estimated with a centered 2×12 moving average (13 months, half weight on the endpoints), which is seasonally free by construction.
3. The seasonal factor for each calendar month is the average of the deviations from the trend for that month, normalized to sum to zero — the adjustment does not shift the level.
4. The factors are estimated on January 2021–December 2024 and then frozen: new months do not change history (releases are stable vintages), and any AI effect after 2024 cannot leak into the seasonal factors and be adjusted away.

Built-in check: twelve-month changes compare the same calendar month and are seasonally free by construction; in the current release they are practically identical computed on raw and seasonally adjusted series (correlation 1.000).

The population grows differently across age groups, so pure headcount series mix labor-market changes with demographics: a group can have more employed simply because there are more people of that age. The per capita variants therefore divide the number employed by the resident population in the same age group (SSB table 07459, quarterly figures interpolated to month; for series covering all ages, the combined population aged 21–60 is used). The index then measures jobs per person — the employment rate — rather than the number of jobs. Example: young and old in combined headcount figures can look very different even when their employment rate develops the same way.

The smoothing is a trailing moving average over the last 3 or 6 months (not centered), computed in the browser. At the start of the series the available part of the window is used. The average therefore lags turning points in the unsmoothed series somewhat. The series are rescaled after smoothing, so the displayed series is exactly 100 in the reference month. The index is thus measured against the smoothed level up to the reference, not against a single month. Smoothing and seasonal adjustment can be combined freely; the default view is seasonally adjusted with a 6-month average.

DEL uses a balanced panel of ~25,000 U.S. firms with ADP payroll data; we use the entire Norwegian population of private-sector wage earners without sample selection. Our age groups are decade groups (theirs are finer). The composition is very different: in the ADP panel 66 percent of employment sits in the two most exposed quintiles, in Norway about 40 percent. Aggregate cross-country comparisons must take this into account. DEL publishes only raw series; the seasonal and population adjustments are our additions, and our raw series reproduce their method exactly.

The A-ordningen can revise recent months. Each monthly data release is therefore archived as its own vintage and is not changed afterwards; the site always shows the latest release, and all earlier ones can be downloaded.

Figure 9 draws on two long-format data packages with every four-digit STYRK-08 occupation in the private sector that has at least 30 employees (ages 21–60) in every month from January 2021, 358 occupations in all. Age groups are pooled, because age × occupation gives small and noisy cells. The series are employment and FTE-adjusted pay, raw and seasonally adjusted, indexed to 100 in the reference month. The per capita variant does not exist by occupation; it would rescale every series within an occupation by the same factor. New hires are not published by occupation, because small occupations have months with no hires. Occupations carry the same exposure quintile as in figure 1. English names are Statistics Norway's official English STYRK-08 names; the search matches English and Norwegian names and codes. The packages are listed under Download data.

From the 2026-09 release the main cuts are also published for the public sector: general government and publicly owned enterprises (institutional sector codes 1110, 1120, 1510, 1520, 6100 and 6500), ages 21–60, for employment, new hires and pay. Occupations get the same national exposure quintile as in the private sector, and the seasonal adjustment is done the same way, series by series. The occupational mix within the quintiles is nevertheless very different between the sectors. Levels should therefore not be compared across sectors, and the AI Labor Market Index is computed for the private sector only. The public-sector packages are listed under Download data.