EXPOSED
The description layer of archival work — writing finding aids, generating metadata, transcribing handwritten documents, subject indexing, and answering routine reference queries — is exactly what current AI does at usable quality, and backlogs are the main reason institutions will adopt it fast. What resists is the physical side: rehousing fragile paper, monitoring environmental conditions, handling deteriorating film and magnetic media, and doing on-site appraisal of a donor's basement. Appraisal decisions (what enters the permanent record and what is destroyed), provenance judgment, restricted-records and privacy determinations, and donor negotiation are consequential calls no one will hand to a model unsupervised — but they're a thin slice of a small occupation with no licensure protecting it.
Mixed — a routine tier and a judgment tier. DACS-compliant finding aids, EAD encoding, series-level description, and accession registers are structured text generation that models already produce at draft quality, and what pulls this to 9 rather than 4 is the irreducible work of opening 200 unlabeled boxes to establish original order and physically arranging series before any description can exist.
Some physical or field component. Rehousing brittle nitrate negatives, reformatting open-reel audio and U-matic tape on legacy decks, reading hygrothermograph data in stacks, and appraising a donor's flood-damaged garage put real hands on real objects — but most of the week is still spent at a workstation in a climate-controlled reading room, which is why this sits at 11 and not with the field techs at 16.
No licence, no signature requirement. No state licenses archivists; the Academy of Certified Archivists' CA credential is voluntary, rarely required outside federal and large university postings, and carries no personal legal exposure — your institution's counsel, not you, answers for a wrongful FOIA release or a botched deed of gift.
Some relationship component. Donor cultivation on a family papers collection can run years and the relationship genuinely closes the deal, but the vast majority of your output — finding aids, catalog records, digitized items — is consumed by researchers who never learn your name, holding this to 8.
Meaningful discretion. Appraisal is irreversible: deciding a records series has no enduring value means it is destroyed, and restricted-records calls under FERPA, HIPAA, or a 50-year donor closure have no lookup table — but retention schedules, institutional collecting policies, and a supervising director constrain most of those calls, which keeps this at 11 rather than in the high teens.
Has AI actually changed your work?