AI for Historical Societies: Accessions, Donor Letters and Loans
A practical guide to AI for historical societies: keeping accession records, donor letters and loan agreements searchable, cited and curator-approved.

AI for a historical society works best as a retrieval assistant over the records the staff already keep: accession files, donor correspondence and loan agreements. The assistant answers questions only from those documents, shows the source for each answer, and proposes catalog changes that a curator must approve before anything is written to the collection database.
Key takeaways
- Start with one role, such as an agent that answers questions about accession and donor records, rather than a general archive assistant.
- Every answer should name the document, page and signature status it came from, so staff can check it quickly.
- Catalog writes stay in a review queue until a named curator approves, edits or rejects them.
- Loan agreements need their own permissions, deadlines and access limits, separate from acquisitions work.
- Success is judged by fewer unanswered research questions and fewer catalog corrections, reviewed on a fixed schedule.
Why the answers are hard to find in the first place
Most historical societies hold their knowledge in several places at once. Accession ledgers may be a spreadsheet, a legacy database and a set of paper registers that were never fully merged. Donor letters often arrive by email, get printed for the file and are scanned later. Loan paperwork lives in a separate cabinet, and the condition reports that go with it may sit in a shared drive folder named after the exhibit.
When a researcher asks who gave an object, under what conditions, and whether a loan from years ago was ever returned, someone has to search all of these places. Staff who know the collection well can do it quickly. Newer volunteers and part-time staff usually cannot, and the answer depends on who is available that day. An AI agent can reduce that dependence, but only if its source material is clean, its answers are traceable and its write access is narrow.
What an agent should and should not do with the collection
A useful boundary is to decide, before any setup, which actions the agent may take on its own and which need a person. The table below is a decision guide written as a list, not a complete policy.
- The agent may search accession, donor and loan records and summarize what it finds.
- The agent may quote donor conditions, such as restrictions on publication or reproduction, with the clause and document cited.
- The agent may list loans approaching a return or renewal date and draft a reminder for staff to review.
- The agent may draft a new catalog entry or a correction, with the proposed text and its source attached.
- The agent may not write to the catalog directly, approve a deaccession, promise access to restricted material or answer from memory when no source exists.
The last point matters most. A historical society's authority rests on its records. An answer that sounds right but cannot be traced to a document is a liability, so the agent should say when it has no source and route the question to a curator.
Keeping accession records, donor letters and loans searchable
Retrieval means the agent looks up the relevant documents at the moment of the question, then writes its answer from those passages. The standard pattern is described in retrieval-augmented generation architecture guidance from Google Cloud and AWS. For a collection, the important details are the preparation steps. Each document is split into passages, and each passage is stored with metadata: accession number, object name, donor, date of the document, document type, and a restriction status copied from the file.
The metadata is what makes the search reliable. Without an accession number on a donor letter, the agent may find a passage about a similar object and attach it to the wrong item. Staff should decide the metadata fields first, then backfill the most-requested collections before the rest. Scanned documents need text recognition, and handwritten material should be flagged as lower confidence until a person has checked a sample.
Retrieval is usually a better fit than training a model on the records. Collection files change constantly: a donor clarifies a restriction, a loan is extended, a catalog entry is corrected. A retrieval index can be updated document by document, and the old version can be removed. A model trained on the files would keep yesterday's facts until it was retrained, which is slower and harder to audit.
How citations keep answers checkable
A citation should point to something a person can open. For this work, that means the accession number, the document title, the page or paragraph, and the signature or version status. A donor letter that is a draft and a letter that was countersigned are not the same evidence, and the agent should say which one it found.
Illustrative scenario: a researcher asks whether a donated photograph can be reproduced in a book. The agent finds the donor letter, quotes the clause that grants publication rights for non-commercial use, and notes that the copy in the file is unsigned. It also finds a later email that mentions a different condition. The answer tells the curator exactly what was found, which copy was used, and that the conditions may conflict. The curator checks the signed original before replying to the researcher.
The same discipline applies to prompt injection. Donor letters, emails and scanned notes can contain text that looks like an instruction, for example a sentence telling the reader to ignore the restriction. The OWASP guidance on prompt injection prevention is clear that document content should be treated as data to be read, not as commands to be followed. The agent's instructions come from the operator, not from the files it searches.
Curator approval for every catalog change
The approval step is where most of the risk sits, so it should be designed as a workflow rather than a reminder. A reasonable sequence looks like this.
- The agent drafts a change and attaches the source passages it relied on.
- The change enters a review queue showing the current value and the proposed value side by side.
- A named curator approves, edits or rejects the change, and the reason is recorded.
- Approved changes are written by a service permission that can update the specified fields but cannot delete records or change restriction flags.
- Every write is logged with the requester, the approver, the time and the source used.
The permission design follows the authorization principle of least privilege described in OWASP's authorization guidance. The agent should hold only the access it needs for the queue, and its write permission should be narrower than a curator's. Access to loan files and restricted donor material should be granted per role, reviewed regularly and removed when a volunteer or contractor leaves. Where systems support delegated access through OAuth 2.0, scopes should be limited to the specific actions each role requires.
Loan agreements, deadlines and permission review
Loan work is time-based. Agreements have return dates, insurance periods, renewal windows and condition report requirements. A missed date can mean an object stays on a lender's list longer than intended, or that a renewal is signed without an updated condition check. An agent is useful here because it can hold the deadlines in one place and surface them before they arrive.
It should not sign anything or send external correspondence without approval. A practical setup has the agent produce a weekly list of loans with dates in the next period, flag any file missing a condition report, and draft renewal letters for staff to edit. Permission review should happen on a fixed calendar, usually aligned with the society's own committee or audit cycle, so that access to loan files matches who currently needs it.
Governance, evaluation and what to measure
For a small organization, governance should be proportionate but written down. The NIST AI Risk Management Framework offers a structure for mapping risks, deciding who is responsible for each, and reviewing the system over time. In practice, that means a short record of what the agent does, which documents it can reach, who approves its outputs, and what happens when it fails.
Evaluation should use a fixed set of questions with known answers taken from the collection, such as the donor of a named object or the return date of a particular loan. Rerun that set after every change to the documents, the metadata or the instructions. Log which documents the agent retrieved for each answer, so that a wrong answer can be traced to a retrieval problem, a document problem or a reasoning problem. Observability concepts from OpenTelemetry apply well to this: the record of each request, its sources and its outcome.
The measures that matter are operational. How often does a curator reject a proposed change, and why? How many research questions get a cited answer on the first pass? How many catalog entries were corrected after approval? Those numbers belong to the society, and they should be reviewed by the people who own the collection.
How a managed engagement is scoped
Be Human begins by defining one useful role for the society, the memory and tools that role needs, the approval boundaries for every write, and an observable success criterion, such as cited answers to a defined set of research questions. The managed service can then include monitoring of retrieval and approval queues, recurring evaluations against the collection, maintenance of the document connections, permission reviews on a set schedule, incident response when a source or access problem appears, and controlled improvements that go through the same approval path as catalog changes. Service is remote and available in English and Portuguese.
The approach is deliberately narrow at the start. A society that begins with accession records and donor correspondence can learn how its curators review proposals before adding loan deadlines or public-facing answers. Decisions about what the agent may touch remain with the people responsible for the collection.
If you want to find the first task worth handing to an agent in your society, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent edit our collection database directly? No, not in a well-designed setup. The agent drafts proposed changes with the supporting source attached, and those drafts enter a review queue. A named curator approves or rejects each change, and only then does a narrowly scoped service permission write the update.
Do we need to digitize everything before an agent can help? No. The agent can start with the records that are already digital and the most-requested collections, then expand as scanning and metadata work progresses. Scanned handwritten material should be flagged as lower confidence until staff have checked a sample of its text recognition.
How do we know whether an answer is reliable? Each answer should cite the accession number, document, passage and signature or version status it relied on, and say so when no source exists. Staff can verify those citations directly, and a fixed set of questions with known answers should be rerun after every change to measure whether the answers are still correct.
Is it better to train a model on our records or to search them? For most collections, searching current documents is the better fit. Records change as donors clarify conditions and loans are extended, and a search index can be updated one document at a time, while a trained model would keep outdated facts until it was retrained.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


