A museum director we work with described her institution's archive: 240,000 objects, 14% on permanent display, the rest in climate-controlled storage. Most haven't been on public view in a generation. Most haven't been photographed since the 1990s. Most cataloging records are sparse.
The museum doesn't lack collection. It lacks discoverability of its own collection.
AI changes that economics — not by replacing the curator, but by making the curator's existing knowledge scale across the 86% of objects nobody ever sees.
Where AI earns its place
Catalog enrichment. The agent reads existing records, cross-references provenance documents, and proposes additional metadata fields. Curator reviews. A backlog of 200,000 records becomes tractable.
Image-based search and similarity. "Find all objects in the storage room with similar motifs to this one." Multimodal models do this well. Curators discover unexpected connections.
Translation of historical documents. Donor letters, accession ledgers, conservation reports, sometimes in 19th-century cursive or dead languages. The agent drafts; specialists verify.
Visitor-facing wayfinding and Q&A. "Where can I see Egyptian sculpture? What time does the gift shop close?" Boring questions that eat docent time.
Conservation triage. The agent reads condition reports, flags objects whose state has changed, suggests inspection priorities.
What AI doesn't do
- Authenticate. Provenance and authentication are deeply human, often legal, sometimes adversarial.
- Curate exhibitions. Curating is interpretation. The thesis is the curator's; the agent helps gather supporting material.
- Decide deaccessioning. Deaccessioning is contentious, board-level, sometimes legally constrained.
- Replace docents. Visitors come to museums in part for the human encounter. The agent supplements, doesn't replace.
A catalog-enrichment pipeline
[object record]
→ [pull existing metadata + photos + provenance docs]
→ [LLM with the museum's cataloging conventions + style guide]
→ [propose: subject tags, period, technique, comparable objects]
→ [confidence-score each proposal]
→ [curator review queue, sorted by confidence and importance]
→ [curator accepts/edits/rejects]
→ [updates to museum's CMS with audit trail]
The audit trail is mandatory. Every change should track: original record, agent proposal, curator decision, timestamp. Museums live and die by provenance documentation.
The trust constraint
Museums have higher epistemic responsibility than most institutions. Misattribution, misdating, or misinterpretation have research and reputational costs that can take decades to undo.
What we recommend:
- All AI-proposed metadata is flagged as such in the record.
- A curator-of-record approves every public-facing claim.
- The agent's training and prompt templates are documented for the IMLS or peer institutions' review.
- The agent never publishes directly to a public-facing surface without human approval.
Funding and grant pitches
Catalog AI projects make excellent grant pitches because the deliverables are measurable: number of records enriched, percentage of collection digitized, hours of curator time freed for primary research.
A typical 18-month grant pilot we've seen succeed: $250k to digitize and enrich 30,000 records, with a public-facing search interface as the deliverable. Real ROI on a granular metric.
Close
Museums are sitting on more knowledge than they've ever published. AI makes the publishing economics make sense for the first time. The curator stays sovereign over interpretation; the agent does the cataloging that would otherwise wait another generation.
Related reading
- Agents in libraries — adjacent public institution.
- RAG is a public library — the underlying retrieval frame.
- Agents in government — public-trust constraints.
We help cultural institutions put AI to work without compromising scholarship. Get in touch.