Why appraisal and description stay with people
Ask will AI replace archivists and the answer lives in the task mix, not in a software demo. Two duties carry most of the weight. Appraisal decides which records are kept permanently and which are destroyed, and that choice cannot be undone later. Arrangement and description then gives a collection its order, its provenance trail and its finding aid, so a researcher in forty years can trust what they are reading.
Both tasks are judgment made on behalf of people who are not in the room yet. A model can propose a subject heading. It cannot weigh a donor’s intent, a legal restriction and a community’s interest in the same accession, then sign its name to the decision. Access rulings work the same way. Releasing a restricted personnel file or a sealed medical record is a one-way door, and institutions keep a named professional accountable for it.
Part of the work is also physical. Brittle paper, oversize maps, glass plate negatives and obsolete magnetic tape all need careful hands, and the robotics tier that would be required here is the dexterous humanoid class rather than anything on a shelf today. You can read how the three questions are scored on our methodology page.
What AI runs, what it assists, and what stays with the archivist
The slice AI can run with light oversight is 3% of task time. It sits at the repeatable end: machine transcription passes over handwritten and typed pages, draft item-level metadata, duplicate and format checks in digital preservation workflows. Output still goes back to a person, because a confident wrong reading of a name or date is worse than a gap.
Assisted work accounts for 43% of task time. Here a model shortens a long first draft: suggesting scope and content notes, proposing subject terms, flagging records that may contain personal data so a human can review them before anything goes online.
The part that stays with people is 54% of task time. That is appraisal, negotiating deposits and donor agreements, setting access conditions, and reference work where a researcher arrives with a vague question and leaves with the right box. Coverage, our answer to whether AI can do the work at all, reads 30 out of 100; the coverage method explains what goes into it.
What has actually been tested
Very little, in this job. Our evidence grade for quality parity is D, which means no study has yet put archivists and a model on the same task and scored both. So we publish no parity number for archival work, and you should treat any outside claim of one with care.
What would settle it is not complicated. A blind comparison in which archivists and a model appraise the same accession, with independent archivists judging the retention decisions. A scored test of machine-written finding aids against human ones for standards compliance and usefulness to real readers. Transcription accuracy measured on messy nineteenth-century hands rather than clean print. Until something like that exists, the honest position is that the assist is proven and the substitution is not. The quality parity method sets out how a grade moves.
When the picture could shift
Most likely between 2035 and 2049 (8 in 10 of our scenarios). That window is a forecast, not a date on a calendar, and the replacement-year method shows how it is built.
Two things could pull it earlier. Cheap, reliable handwritten text recognition would hollow out the transcription and indexing backlog that props up many entry-level and project posts. Budget pressure in small institutions could do the rest: when a vacancy opens, a records management platform with AI features is an easier purchase than a new line on the payroll.
Two things hold it back. Physical custody of fragile, odd-shaped material still needs dexterous hands in the stacks. And governance cuts the other way: the more a collection contains personal, legal or culturally sensitive records, the more an institution wants a qualified person responsible for what is released. Most archives also have years of undigitized material, so automation cannot reach what has never been scanned.
How to stay needed as an archivist
Lean into the parts that carry responsibility. Take appraisal decisions and write down your reasoning, so the record of why something was kept is as good as the record itself. Own access and privacy rulings, including the hard refusals. Build donor and depositor relationships, because collections arrive through trust, not through a pipeline.
Two skills pay off alongside that. First, data and metadata fluency: controlled vocabularies, crosswalks, and enough comfort with machine output to audit it properly. Second, AI governance in a records context, from provenance and custody through to the terms of any digitization partnership. The government sector page is useful here, since public archives sit under rules that shape adoption.
What to do: pick one workflow you already run, measure how much review time machine output really costs you, and bring that number to your next budget conversation.
Nearby work is worth a look if you are choosing a path. Compare the data behind curators, museum technicians and conservators and librarians and media collections specialists, or view the whole librarians, curators and archivists family. You can also put two of them side by side on the compare page, or see where this kind of work lands in the list of safest jobs from AI. For context on the labor market, the Bureau of Labor Statistics counts roughly 7,970 archivist jobs in the United States, with median pay of $64,550 and projected growth of 3.4% from 2025 to 2035 (BLS, 2025).