AI Adoption in Finance: a human-verified, stage-classified corpus of institutional investors and its human-AI disagreement record

A public, sourced classification of where institutional investors sit on AI adoption — pensions, sovereign-wealth funds, endowments, asset managers and hedge funds — placed on a four-stage bar (exploring / piloting / scaling / embedded) from public evidence alone. An AI research agent drafts every row and proposes a stage; a human verifies every row against its cited sources before it publishes. Because each review decision is recorded, the rate at which the human and the agent disagree is published as data rather than asserted: see data/agreement.json. Those figures are anchored, non-independent agreement and should not be read as an inter-rater reliability estimate — the reviewer saw the proposed stage before deciding. Institutions assessed against the methodology whose public record did not support a stage are published by name with reasons in data/not_classified.json, so the corpus can be read as a rate rather than a highlight reel. Evidence in Chinese, Japanese and Korean is stored verbatim and rendered in English at display time. Licensing: the data files in data/ are CC BY 4.0; the code in this repository is MIT. This record is registered under CC BY 4.0 because the corpus is the citable artifact. Coverage is not a sample of any defined population — institutions enter through research passes, not a sampling frame — so every figure describes this corpus and none is an industry rate. Independence. This is an independent personal project, produced entirely in the author's personal capacity. It is not affiliated with, sponsored by, funded by, or endorsed by any employer or institution, and every view and classification in it is the author's alone. No non-public information from the author's professional work informs any classification, and institutions where the author has a professional affiliation are excluded from coverage entirely. Nothing here is investment advice.