Book a demo
Back to news

Updated:

AI in Museum? See our Case Study!

MUZA AI modules for catalogue search, alternative text, audio description, translation and object descriptions

This case study is based on the experience that Łukasz Ściebior (MUZA) and Katarzyna Reszka (Museum of Warsaw) shared during the presentation as part of the ERC2026 programme.

It describes how AI tools were introduced into MUZA: initially through the 2024 implementation work with the Museum of Warsaw, and later through further MUZA development carried out in 2026.

AI is only one part of MUZA. The system is a broader environment for managing museum collections, documentation and processes across the organisation.

The everyday challenges of collection management

Museums work with large and diverse amounts of information: object records, images, documentation, research findings, terminology, conservation data, loans, movements and publication content.

In everyday work, teams may face:

  • incomplete or inconsistent object records;
  • information distributed across systems, spreadsheets, documents and folders;
  • repeated copying and rewriting of the same information;
  • difficult searching across large collections;
  • inconsistent terminology;
  • time-consuming cataloguing and documentation;
  • limited capacity to prepare content for different audiences;
  • separate environments for preparing, reviewing and publishing information.

These problems affect more than the inventory department. They can make it harder to respond to queries, prepare exhibitions, manage loans, publish collection information, work with digital content and share knowledge between departments.

MUZA supports museums by bringing collection data and related processes into one browser-based environment. It helps organise information, standardise terminology, support collaboration and make data easier to find and reuse.

AI extends these capabilities by supporting selected tasks within the system.

The Museum of Warsaw use case

The first AI work presented at ERC2026 grew out of the implementation of MUZA at the Museum of Warsaw, which began in 2024.

The Museum of Warsaw manages a collection of approximately 765,000 objects. Its photographic collection is particularly important: it is one of the largest parts of the collection, attracts significant public interest and often requires complex historical research.

Before the implementation, the museum had worked for many years with an older environment that did not provide sufficient terminology control, data consistency or process automation. Information was entered in different ways, making records difficult to compare, connect and search.

The migration to MUZA included approximately 300,000 records from the previous system and other sources. The process made the scale of the data-quality and cataloguing challenges more visible.

Some records required further research, completion or standardisation. Photographs could contain handwritten notes, labels and stamps that provided valuable clues but were difficult to process manually at scale.

This created a clear need for tools that could support museum professionals without taking responsibility for the final record away from them.

AI-assisted preparation of object records

The first AI module was developed as an initial implementation based on the needs identified during the work with the Museum of Warsaw.

Starting with a photograph and the available contextual information, the module can prepare proposals for:

  • object descriptions;
  • keywords;
  • materials and techniques;
  • dating;
  • possible authorship;
  • dimensions;
  • selected inventory fields;
  • controlled-vocabulary terms.

It can also suggest new vocabulary entries when a relevant term does not yet exist in the configured dictionary.

The module can work with an image supplied as a starting point. The image does not need to be a high-quality digitisation; a photograph taken with a phone can provide a useful starting point when it includes relevant context.

The proposals are not final museum information. They provide a structured starting point for further work by registrars, curators and other specialists.

Reading information from photographs

The module can use optical character recognition to prepare a proposed transcription of notes, labels and stamps visible in photographs, including information found on the reverse of an image.

A staff member can compare the proposal with the source image, correct the transcription and decide whether the information is relevant to the object record.

AI helps process the visible information. It does not decide how a historical clue should be interpreted.

Verification and professional control

The discussions with the Museum of Warsaw emphasised that museum professionals need reliable control over collection data.

AI-generated content can contain errors, omissions or unsupported statements. For this reason, the MUZA workflow includes several levels of review.

Staff can:

  • assess proposals field by field;
  • compare them with additional sources;
  • review confidence indicators;
  • accept, amend or reject individual proposals.

Responsibilities can be divided between different roles:

  • the curator assesses content and evidence;
  • the registrar checks data quality, required fields and terminology;
  • a reviewer prepares and checks the draft;
  • a publisher or authorised user approves the final content for the relevant field.

This creates a human-in-the-loop workflow. AI supports preparation and checking, but museum professionals remain responsible for the information that becomes part of the record. Content generated with AI assistance is clearly labelled.

AI-assisted search with RAG

One of the AI capabilities introduced into MUZA is an AI chat based on retrieval-augmented generation, or RAG.

RAG is not a general-purpose chatbot with unrestricted access to the internet. In this use case, it works primarily as a search and retrieval layer supported by a language model.

The process has two main stages:

  1. the system searches selected connected museum data for relevant information;
  2. the language model uses the retrieved context to prepare a response in natural language.

The implemented search can work with information held in MUZA, including object descriptions and selected metadata. Users can ask questions without knowing the exact catalogue term, find relevant objects, inspect the records used to prepare the answer and save selected objects as a working list.

MUZA’s catalogue chat links its answers to the underlying records.

This can support:

  • collection queries;
  • preparation for exhibitions;
  • research;
  • finding objects by theme, author or context;
  • reducing dependence on complex filters and exact vocabulary.

The source records remain important. The system is intended to make collection information easier to explore, not to hide where the answer came from.

Sources that can be connected to RAG

The RAG layer can be supplied with selected sources that the institution wants to make searchable.

The current use case is based primarily on:

  • object descriptions;
  • selected metadata;
  • collection information stored in MUZA.

The solution can also be extended with additional indexed sources, such as:

  • digitised documentation;
  • bibliographic references;
  • digitised books;
  • research materials;
  • selected repositories or document environments.

Additional sources are selected, prepared and indexed for each implementation. The chat can search them once they have been configured.

Data security and infrastructure

The setup discussed during the ERC2026 preparation was designed to run locally, within infrastructure managed by the MUZA team or the institution. With a locally hosted model, the passages retrieved from collection records are processed within that infrastructure. If an external model is configured, the retrieved passages supplied to it are processed according to that configuration and the provider’s data-handling terms.

This architecture can help the institution retain control over:

  • which sources are indexed;
  • which data is made available to the search;
  • where the data is processed;
  • who can use the module;
  • how access to the system is managed.

RAG itself is not an automatic guarantee of security or accuracy. The final level of protection depends on the selected infrastructure, model, access controls, integration, maintenance and data-processing rules.

For this reason, security and privacy must be considered as part of the implementation architecture, not as a separate marketing feature.

Developing descriptions for different audiences

Further MUZA development in 2026 introduced tools for preparing descriptions adapted to different audiences and purposes.

The same object data can be used as a starting point for content intended for:

  • children;
  • adult visitors;
  • researchers;
  • educators;
  • exhibition teams;
  • people on the autism spectrum.

The system uses information already held in the object record to prepare a draft. Staff then review, edit and approve the result before using it in museum work or online collections.

The goal is to reduce repetitive writing while making existing collection knowledge easier to reuse.

Accessibility content

MUZA also includes tools supporting accessibility work.

They can help with:

  • identifying images with missing alternative text;
  • preparing draft alternative text;
  • creating transcripts for audio and video;
  • preparing accessible subtitles;
  • identifying speakers, sound effects and background sounds;
  • preparing draft audio descriptions for images and complete, time-coded audio description scripts for films.

For films, the MUZA system can generate a complete draft audio description script. It analyses the video in segments and describes what is visible at each point, with time codes so staff can jump straight to the relevant scene and check the description against the original footage. The script can be downloaded as a WebVTT file.

These tools address a practical problem: preparing accessibility content across a large digital collection can require substantial time and repeated manual work.

Translation with museum terminology

Another MUZA AI capability supports translation.

Staff can translate selected fields or descriptions, compare the original with the proposed version and review the result before publication. The workflow can preserve selected proper names or phrases that should remain in the original language.

Terminology can also be configured for the institution. This is important because a general translation may not reflect the meaning of a specialist museum term.

The process supports faster preparation of multilingual collection content while keeping responsibility for the final wording with museum staff.

MUZA is more than AI

AI is one group of tools within MUZA, not the definition of the system.

MUZA supports wider museum work, including:

  • collection and inventory management;
  • documentation and digitalisation;
  • archives;
  • conservation;
  • loans and deposits;
  • object movements;
  • exhibitions and projects;
  • portals and public access;
  • cooperation between departments;
  • data migration and data quality improvement.

The system is intended to help museums reduce the need to move between disconnected tools, avoid repeated data entry, improve search and filtering, work with shared information and reuse collection data in different processes.

AI tools support selected activities within this environment. They do not replace the collection-management system, museum workflows or professional expertise.

Beyond the experiment

The MUZA and Museum of Warsaw case shows that practical AI development can begin with specific operational problems:

  • incomplete records;
  • repetitive cataloguing work;
  • difficult searches;
  • lost historical context;
  • accessibility workload;
  • translation workload;
  • the need to prepare content for different audiences;
  • fragmented review and approval processes.

The first module grew out of the 2024 implementation work with the Museum of Warsaw. Further AI capabilities were introduced and developed as part of MUZA product development in 2026.

The right solution depends on an institution’s data, workflows, terminology, infrastructure, security requirements and professional responsibilities.

AI can help museums save time, improve access to information and make better use of existing knowledge. Human review, clear responsibilities and control over the data remain essential.

Katarzyna Reszka and Łukasz Ściebior presenting at ERC2026 in Berlin
Katarzyna Reszka (Museum of Warsaw) and Łukasz Ściebior (MUZA) at ERC2026 in Berlin.

Discuss AI and collection management in your museum

Are you considering AI for cataloguing, collection research, accessibility, translation or collection search?

Tell us what your museum needs and find out whether we can implement the MUZA system at your institution free of charge.

Contact the MUZA team:

Jan Sikora
International Business Growth Partner
jan.sikora@mobilems.pl
+48 797 418 825

ERC2026 in Berlin: the conference audience, Łukasz Ściebior at the lectern with Katarzyna Reszka, and the MUZA booth