AI Across the Economy: a uniform-methodology impact matrix for all 463 ISIC Rev.5 classes

A reproducible, data-model-first assessment of the economic impact of artificial intelligence — current frontier systems and their credible ten-year trajectory — on every one of the 463 four-digit classes of the International Standard Industrial Classification (ISIC) Revision 5 (22 sections, 87 divisions, 258 groups, 463 classes). The unit of analysis is the ISIC class; the atomic record is an activity–mechanism cell (4,448 in total). Contents. Each class record carries assessment cells across ten domain clusters (labour/productivity, quality/variety, employment, innovation, market structure, demand, risk/regulation, cross-sector interaction, capability–exposure reconciliation, obsolescence). Every cell records a direction (strong_negative … strong_positive), a 0–4 magnitude against a fixed value-added numéraire, a ten-year horizon, an effect type, an explicit confidence tier (A empirical / B analogical / C theoretical), and resolvable source citations. The dataset also includes: a signed inter-sector interaction graph (2,045 edges, each with a channel and sign); a harmed-via-interaction and transmits-harm view; a self-inflicted-versus-imported obsolescence-risk classification (72 at-risk classes); a per-class capability–exposure bridge reconciling language-task exposure (AIIE/AIOE lineage) against embodied automatability; and a lightweight general-equilibrium adding-up reconciliation. Provenance and honesty. Anchor date 1 July 2026 (frozen snapshot). Every quantitative claim is quoted from an individually cited source confirmed to resolve at production; missing data are recorded as missing. Approximately 95% of cells are Tier B/C inference: the dataset establishes direction, mechanism and relative ranking, and should be read as a set of structured, testable hypotheses rather than calibrated point magnitudes. The matrix is unweighted by value added; interaction edges are analyst-set structural hypotheses, not estimated elasticities. Files. master_matrix.json / master_matrix.xlsx (per-cell matrix); assessments.zip → 463 authored class records (<code>.json, primary source of truth); synthesis.json (typology, interaction edges, harmed-via-interaction, obsolescence); bibliography.json (582 sources, UTF-8); schema/*.json (validating schemas); data/domain_taxonomy.json (controlled vocabulary). Standard. British English; APA 7th references with resolving DOIs/URLs; confidence explicitly tiered. Fully reproducible from the authored records plus the published pipeline.