We Read Anthropic's Data on 756 Jobs. Here's Who AI Is Already Replacing.

1 · The dataset

The Anthropic Economic Index [1] is a quarterly public release derived from sampling Claude conversations and classifying each against the O*NET 2018 task taxonomy. The March 2026 release[2] covers conversations from the week of 2026-02-05 through 2026-02-12, totaling 16,693 metric rows across geographies and analytical facets. The labor-market subset — which we use — collapses the data into two CSVs:

• job_exposure.csv — 756 SOC codes × an observed_exposure score in [0, 1], measuring the share of Claude conversations whose constituent tasks belong to this occupation.
• task_penetration.csv — 17,998 individual O*NET task statements × a per-task penetration score, plus a collaboration-pattern breakdown used to compute an automation share per task.

Both files are CC-BY licensed. The collaboration patterns are categorized per AEI's methodology document as directive, feedback loop, task iteration, validation, or learning; the first two are pooled as automation, the remainder as augmentation.

2 · Distribution

The 756-occupation distribution is heavy at the low end. 447 occupations (59%) sit in [0.00, 0.05). Above 0.05 the distribution decays roughly exponentially, with a small cluster of Programmer-adjacent occupations forming a high-exposure tail.

3 · Most and least exposed

The top of the distribution is concentrated in occupations whose core deliverable is textual or structured information. The bottom is dominated by physical-labor occupations where Claude cannot yet operate at all.

4 · The optimism-bias problem we use AEI to fix

LLM self-grading of capability is systematically optimistic. When you ask a frontier model "how much of a CTO's work can you do?" the answer trends low — RLHF training rewards hedged, non-alarming output. Empirically we measured the gap on 200 sampled profiles: the same Sonnet 4.6 pass returned automation scores 18–24 points lower than the AEI-derived value for the matched SOC.

The fix is to anchor the axis. We don't allow Sonnet to set the Automation axis below what AEI observes. Specifically: we classify the LinkedIn role into a SOC code (cached Haiku pass), look upobserved_exposure, transform it via a linear map[0, 0.75] → [10, 95], and treat the result as a floor. The scoring agent may go higher (if the user's answers indicate heavier exposure than the macro average); it may not go lower.

5 · Implications

Anchoring to AEI moves the median reading down by 4–7 months across our test cohort of 200 profiles, with the largest shifts in occupations where AEI exposure exceeds 0.40 and the LLM had previously assigned automation in the 40–50 range. The bias correction is asymmetric: low-exposure occupations are not pushed downward, because the AEI value is a floor not a fixed value.

6 · Limitations

6.1 — AEI measures Claude usage, not all AI displacement. A task with 0% AEI penetration today could still be displaced tomorrow by a different model. We use AEI as the best public proxy, not as a complete picture.

6.2 — Single-quarter snapshot. The March 2026 release covers one week of data. Anthropic publishes quarterly; we will refresh on each release.

6.3 — US SOC taxonomy. O*NET is US-centric. The occupational mappings may not generalize cleanly outside US/CA labor markets.

References

1. [1]
   Anthropic Economic Index home

   Index landing page; release notes and methodology.
2. [2]
   Anthropic Economic Index · March 2026 report

   Primary source for the 756-occupation exposure values used in this article.
3. [3]
   Anthropic/EconomicIndex dataset (HuggingFace)

   Raw CSVs (CC-BY) including job_exposure.csv and task_penetration.csv. Release subdirectory: release_2026_03_24.
4. [4]
   O*NET Database release 29.0

   Source of the task-statement to SOC mappings we join against AEI for per-occupation task lists.