AI-Literacy Load by Occupation: a reproducible map of where knowing AI becomes part of the job

A reproducible map of how much AI-fluency each occupation needs on top of its domain skill, built from the open Anthropic Economic Index (AEI, 5th edition) joined to the public O*NET 29.3 occupation taxonomy. The method crosses two measured inputs per occupation: AI-usage penetration (how much of the work AI already touches) and the augmentation/automation mix (what kind of touch). Crossing them gives a 2x2 of productive literacy load (high penetration, augmentation-led: learn it for leverage) versus defensive literacy load (high penetration, automation-led: learn it to stay in the loop). The derived index is kept deliberately thin and shown beside its raw inputs, so a reader can rebuild it and disagree with the construction. Headline findings (penetration-weighted, Claude usage only): among the 146 occupations with measurable Claude.ai uptake, 114 are automation-led and 32 augmentation-led. The same tasks are far more automated through the developer API than on the consumer web surface (67.2 percent versus 42.2 percent), the clearest single signal of where work is being industrialised. Software, editing, and administrative roles land defensive; tutoring, teaching, and analysis stay augmentation-led. An appendix gives eight plain-language, failure-mode AI-literacy primitives (training cutoffs, non-determinism, compaction, model tiers, live search, and verification). The work operationalises part of the Human-AI Coexistence research agenda (Caviola, Keeling, Street and Shevlin, 2026) and is independently corroborated on the augment/automate axis by the Stanford Canaries in the Coal Mine study (2025). Two zero-dependency Python scripts, the open input pointers, and all output tables (including a dropped-task transparency file) ship with the paper. Conflict of interest: the entire dataset is published by Anthropic and measures Anthropic's own model (Claude), and the assisting AI model is also made by Anthropic. Every figure is Claude-usage, not all-AI; the paper leads with the automation numbers and labels this throughout. CC BY 4.0. ORCID 0009-0003-4213-7769.