AI strategy for museums: what the work covers, step by step

An AI strategy for a museum or cultural institution should answer four questions: where AI is already being used in the building, which processes are worth starting with, which risks must be settled before anything reaches the public, and which one small experiment will prove or disprove the case. The result should be a short plan with owners, costs and the next three moves, not a long report.

The work follows four steps: listen to the people who do the work, map and rank the processes with an honest view of the risks, test one real case, then decide. The test should prove or disprove the case within weeks, so the director and the board decide on evidence rather than slides.

Why does a museum need an AI strategy now?

Because staff are already using AI, with or without one. The first UNESCO and ICOM global survey on AI in museums, published in September 2026 with responses from more than 400 museums in 90 countries, found that 57% of responding museums use AI and 55% have no internal AI policy, strategy or guidelines. It describes that use as mostly exploratory or staff-led rather than institution-wide, and names accuracy, copyright and data protection as the leading concerns.

The rules have also moved. Since 2 August 2026, Article 50 of the EU AI Act has applied (see the dates in Article 113). It requires that people are told when they are interacting with an AI system, and that deployers disclose deep fakes and AI-generated text published to inform the public on matters of public interest, unless that text has had human review and someone holds editorial responsibility for it. Article 4 asks organisations that use AI to take measures to support the AI literacy of their staff.

And the policy context is still forming. The European Commission is preparing an AI strategy for the cultural and creative sectors, expected in early 2027. Waiting for it is not a plan, but it is a good reason to build a strategy that can adapt.

What does an AI strategy engagement for a museum cover?

Here is the work, step by step, and what each step should leave behind.

Step	What happens	What you get
1. Listen	Conversations with the people who do the work: registrars, curators, education, visitor services, finance, communications. Not only with the people who describe it.	A picture of where time really goes, and where AI is already in use without anyone deciding it should be
2. Map	Each process is ranked by value, risk and how ready the data and tools are	A ranked map of where agents and automation pay off first
3. Risk view	Data protection, copyright, the AI Act and reputation, checked against the top candidates	A clear list of what must be settled before anything reaches visitors
4. Test	One small, real experiment on the most promising case	Evidence, not slides: what worked, what a person had to correct, what it cost
5. Decide	Leadership chooses what to continue, stop or scale	A short plan with owners, costs and the next three moves

Workshops and briefings for leadership, the board and teams can run alongside these steps, so the people who will approve the plan understand what they are approving.

Where should a museum start with AI?

Rarely where the first idea points. The first idea is often a visitor-facing chatbot or generated content for an exhibition. Those are possible, but they carry the most risk: they speak in the museum's name, they fall under the AI Act's transparency rules, and a wrong answer about a collection object is a reputational problem, not a typo.

The better first candidates are usually internal, text heavy and repetitive, with a clear owner:

• Administration and reporting. Funder reports, board papers, the weekly summary nobody has time to write.
• Document intake. Invoices, contracts, loan agreements and applications, read and sorted before a person checks them.
• Collection documentation. First drafts of object records from existing notes, always reviewed by a curator before they are saved.
• Research support. Summaries of literature, press and comparable exhibitions, with the sources attached so they can be checked.

These score well on value and readiness, and a mistake stays inside the building, where a person catches it. Once one of them works, the museum has the evidence, the habits and the governance to consider something public.

What should a museum's AI policy cover?

A short policy that people actually read beats a long one that sits in a folder. It should settle at least these points:

1. Who approves what reaches the public. Anything published, exhibited or said to a visitor in the museum's name is approved by a named person. An AI system can draft it, never release it.
2. Disclosure. When visitors are talking to an AI system, and when content was generated or changed with AI, they are told clearly. The Austrian Museums Association has published labelling guidelines that are a useful model.
3. Accuracy. AI output about collections, history or people is checked against sources by someone with the expertise to do it.
4. Rights and data. Which collection images, texts and visitor data may go into which tools, and which may not, given copyright, image rights and data protection.
5. Tools and training. Which tools staff may use for which tasks, and what training they get first.

The policy and the strategy should be written together. A policy without a strategy only says no. A strategy without a policy pushes risk onto whoever runs the first pilot.

How do you test an AI idea without a big project?

Pick one process from the top of the map and run a real experiment on it for a few weeks. Real means real documents, real volume and the person who owns the process, not a demo with sample data.

Decide in advance what would count as success: hours saved, fewer errors, a faster turnaround. Then track three things from the first week: what the AI did, what a person corrected and what it cost. The corrections matter most. They show whether the case holds, and they show exactly where the instructions, the data or the scope need work.

An experiment that disproves the case is a good result. It costs a few weeks instead of a failed project, and it frees the budget for the next candidate on the map.

What usually goes wrong?

• Tools first, questions later. A licence is bought before anyone has decided what problem it solves.
• Strategy by committee alone. The plan is written by people who describe the work, and the people who do it find it does not fit.
• No owner. The pilot belongs to "the digital team" rather than to the person whose process it changes, so nobody keeps it running.
• Public first. The first project speaks to visitors, so the first mistake is a public one.

None of these are technology problems. They are the reason a strategy has to rest on evidence from the institution's own work.

Where to start

List where AI is already used in your institution, by whom and for what. Pick the three processes that take the most staff time and score them on value, risk and readiness. Then run one experiment on the best of them, with a person approving every output.

Our story started with digital communities, virtual worlds and hybrid events for cultural institutions such as the Goethe-Institut, Bozar and EUNIC, including a digital programme on AI and cultural relations, and our AI strategy work rests on agents we build and run ourselves. If you are planning this for your institution, we are happy to talk it through.