AI for Museums and Cultural Heritage: Build Knowledge Systems, Not Generic Chatbots
The useful question is not where to add a chatbot. It is which knowledge is difficult to retrieve, verify, connect or reuse—and whether AI can improve that system.
Start with institutional knowledge, not a model
Museums already contain high-value knowledge: collection records, research notes, catalogues, conservation documents, transcripts, oral histories, images, video, educational material and staff expertise. An AI project becomes useful when it creates better access to those sources instead of replacing them with generic model knowledge.
Ground answers in traceable sources
For research, collections and public interpretation, an answer without a source is often not enough. Retrieval-augmented generation can combine language models with approved institutional material so that responses are grounded in a defined corpus and can link back to the documents, records or media segments that support them.
Treat archives and media as searchable knowledge
AI can help beyond question answering. Audio and video can be transcribed and segmented. Documents can be classified and enriched. Images and records can be connected through metadata. The result can be a research interface, staff assistant, editorial workflow or public discovery tool depending on permissions and audience.
Design permissions, review and logging before automation
Cultural organisations hold material with different levels of sensitivity, authority and publication status. A reliable architecture therefore needs access control, source boundaries, logging and explicit review points. Agentic workflows are most useful when they automate bounded tasks while keeping responsibility visible.
AI should strengthen the information architecture
The long-term value is not the novelty of a conversational interface. It is a better-organised knowledge layer that can support search, editorial work, research, visitor services and future interfaces. If the system improves how information is structured and connected, it remains useful even as individual models change.