Anthropic takes a sample of Claude conversations and has a model classify each one in two ways. First: which O*NET task is the conversation about? Second: how do the person and the model interact? The second classification has five outcomes, described above the figure. Conversations that cannot be placed are dropped.

Automation is the sum of directive and feedback loop. In both, the model does the work. The person writes the request and, in the loop, relays what happened when the result was tried. Augmentation is the sum of task iteration, validation and learning. In all three the person contributes content, judgement or questions along the way. This is Anthropic's own bucketing and the one used here.

The automation share of an occupation is the share of conversations about its tasks that are automation. The share does not say how much of the work is done by AI. It says how AI is used when it is used. An occupation where Claude is rarely used can still have a high share, from few conversations. The number of classified conversations is therefore shown in the summary and in the tooltip.

Why separate the two? The two modes of use can affect employment in opposite directions. Where use is automating, the model does the task instead of the person. Where use is augmenting, the person does more or better work with the model as help. The index page shows that employment has developed differently in the two groups of occupations.

Two differences from the front page. There, only directive use counts as automation, and the numbers come from the first sample in Handa et al. (2025). Here, Anthropic's bucketing and every published sample are used.

The task × type table has been published four times. The first sample is Claude.ai conversations from December 2024 to January 2025 (Handa et al. 2025). The next three are one week each: August 2025, November 2025 and February 2026. From August 2025 the table also exists for the API, where developers and agents call the model. The ranking of occupations moves between samples. Between August 2025 and February 2026 the rank correlation in chat is about 0.7. Against the Handa sample it is lower, about 0.3, and about 0.6 for occupations with at least a thousand conversations.

Each O*NET task belongs to one or more US SOC occupations. Conversations are split equally between them. SOC goes to ISCO-08 with the BLS crosswalk, and the ISCO code is the STYRK-08 code when it exists in the Norwegian standard. The share for a STYRK occupation is the mean over the contributing SOC occupations. The chain is the same as for the Handa measure on the index page and yields 300 to 320 occupations per sample. Occupations with no Claude use on any of their tasks drop out. The task figure shows the 15 O*NET tasks with most use among the SOC occupations the code maps to, with at least ten classified conversations. Task texts are O*NET's English statements.

O*NET is the US Department of Labor's occupational database. For every occupation, O*NET has asked a sample of incumbents how important each of 41 generalized work activities is, for example "programming", "assisting and caring for others" or "operating vehicles". The same is done for 35 skills, such as "critical thinking" and "repairing". The scale runs from 1 to 5. Together this gives every occupation a profile of 76 numbers.

We do not use the task statements to find neighbours, although O*NET has them. Tasks are written for each occupation and almost never repeat across occupations. Software developers share two task statements with one other occupation, and office clerks share none. The activity and skill profiles exist for every occupation on the same scale, and those are what we compare.

The calculation: each of the 76 numbers is standardized across occupations, so activities every occupation scores high on, like "getting information", do not dominate. The similarity between two occupations is the cosine between the standardized profiles, a number from −1 to 1. The three occupations with the highest similarity are shown. Under each are the three activities or skills that both occupations score high on and that contribute most to the similarity.

The profiles sit on US SOC codes. They go to Norwegian STYRK-08 codes through the same crosswalk as the exposure measures. A STYRK occupation mapped to several SOC occupations gets the mean of their profiles. Norwegian codes that share the same SOC set, for example nurses, specialist nurses and social educators, get an identical profile and similarity 1.

What the measure does not capture: Norwegian conditions, pay, education requirements, and how easy a switch actually is. It only says that the work consists of much the same things.

The series are the same as in figure 9 on the index page: number of employees and mean pay per full-time equivalent in the private sector, ages 21–60, per four-digit STYRK-08 code, seasonally adjusted. The series is smoothed with a trailing moving average over the last 6 months (changeable above the figure) and then rescaled to 100 in February 2025, the month Claude Code launched, the same way as on the index page. The window here starts in February 2023, two years earlier. Only occupations with at least 30 employees in every month since January 2021 have a series, 358 in all. The change in the summary is the latest month with data against February 2025 in the smoothed series. Open vacancies from NAV will be added in a later version.

1. The conversations are global. They say how Claude is used on the occupation's tasks, not how Norwegian workers use AI.
2. Claude only. OpenAI publishes no comparable table.
3. The shares are of conversations, not of working time.
4. The API gives little variation across occupations. Nearly all are above 80 percent directive.
5. O*NET describes US occupations. Norwegian occupations with the same code may differ in content.