Bringing a Museum Collection to Life with Conversational AI

Inside the London Transport Museum’s AI Digital Poster project — a practical case study in using conversational AI to make museum collections speak, in 60 languages, for under £10,000.

What if a museum poster could talk back? What if visitors could ask it anything — in their own language — and get an accurate, curator-approved answer in seconds?

That’s no longer a thought experiment. At the London Transport Museum in Covent Garden, an Art Deco poster from 1927 now holds conversations with visitors as an AI-powered Digital Museum Guide. Launching this June in the Global Poster Gallery, the project is one of the clearest, most replicable examples yet of how conversational AI can widen access to heritage collections without giving away a museum’s most valuable asset: its knowledge.

At the Museum AI Summit 2026, the team behind the project explained how it was built, what it cost, and the lessons any museum can take from it.

The big idea: a poster that converses

The London Transport Museum has always looked for ways to “ignite curiosity and shape the future,” as Engagement lead Wesley Salton puts it. The spark for this project came from an unexpected place — the museum’s “padded cell,” one of its most unusual collection spaces. Looking at the mannequins at the back of the room, the team asked a simple question: what if these could talk?

That curiosity grew into something bigger. Rather than animating a single object, the team turned to the museum’s Art Deco gallery — a space already rich in design, colour, and storytelling — and asked: what if one of these iconic posters could speak, and visitors could have a conversation with the collection itself?

The chosen poster is Hearing the Riches of London by Frederick Charles Herrick, designed in 1927 as part of a series of five. It now functions as a multilingual, interactive Digital Museum Guide, answering questions about the poster, the Art Deco movement, and the wider history of London Transport.

Five reasons this approach works for museums

Dave Thomas, who led the project from the museum side, frames the value in five points that any museum considering AI should weigh up.

1. It’s quick and affordable. The experience can be delivered with equipment for under £10,000, in weeks rather than months, and costs roughly £6–£8 an hour to operate.

2. It performs. The guide interacts in natural language, responds quickly, speaks 60 languages, and is built to give accurate answers — with guardrails for visitors who try to catch it out.

3. It creates the “wow.” No human guide could learn the entire history of the Underground and deliver it in 60 languages. The AI augments visitor services rather than replacing them.

4. It drives action. After the “wow” moment, the experience can prompt a call to action — make a donation, visit the shop, explore the website, or complete a visitor evaluation form.

5. It controls the data. Crucially, the system uses large language models for language, not for knowledge. The museum’s expertise stays its own and isn’t handed to third parties — protecting its biggest long-term asset.

The technology, explained simply

The build is a partnership between the London Transport Museum, Microsoft, and the Reply group (Avvio Reply and Open Reply), with governance support from Transport for London.

At the core is GPT-realtime, a large language model that processes what a visitor says and generates a response. A second AI algorithm animates a still image of the poster into video, matching lip movements to the spoken answer to create a realistic talking portrait.

The knowledge itself is handled through Retrieval-Augmented Generation (RAG). Curator-authored documents are uploaded to a vector search database (Azure AI Search). When the guide is asked something, it retrieves the most relevant documents and uses them as context to answer — meaning responses are grounded in the museum’s own verified content, not the open internet.

As Daren Ward of Open Reply explains, the beauty of RAG is that it’s locked down and secure: the material is the museum’s content and IP, kept restricted to their use. The team also delivered a web application so museum staff can update the image, voice, greeting and closing messages, and drag and drop new knowledge (PDF or DOCX) into the database themselves — even while the guide is running live.

On safety, the system relies on Azure’s built-in guardrails to prevent hate speech and offensive language, with an additional custom layer for double protection.

Curators in the driving seat

What separates this from a generic chatbot is the curatorial rigour behind it. Georgia Morley, who shaped the knowledge base, worked alongside the curatorial department to gather authoritative information — collection database records on Herrick and the poster, the history of Underground stations, Art Deco exhibition interpretation text, a purpose-written exhibition label, and curator-authored stories from the museum’s website, including one on the representation of women in Art Deco posters.

As an educational charity, the museum needs the stories it tells to be both historically accurate and enriching. By building the knowledge base exclusively from curator-led, verified sources, the team ensured the guide speaks in the authentic voice of the museum.

Testing follows a staged approach, beginning with an alpha phase involving only the curating team, who validate every response for accuracy, consistency, and tone before any public release.

Why multilingual access matters

For project manager Ru Rowe, one of the most pressing problems was language. The museum welcomes visitors from around the world, yet most on-site content is in English. The AI guide’s ability to converse in 60 languages directly addresses that gap, letting far more visitors engage fully with the collection and the rich history of London transport.

A model for the wider sector

Lauren Sager Weinstein, Chief Data Officer at Transport for London, sees the project as a way of making AI concrete. It takes something abstract and lets people walk up, ask questions, and even try to break it — which she did herself, probing for unintended consequences. Strong, curated, high-quality data, she notes, is what makes the model’s outputs reliable.

Chi Ukachuku, Technology Advisory Manager at TfL, highlights the governance lesson: innovation in a large organisation doesn’t mean total freedom. By shaping the problem early, working iteratively, and keeping governance running alongside the work rather than bolting it on at the end, the team moved fast without losing control.

See it for yourself

The AI Digital Poster goes live in June at the Global Poster Gallery, London Transport Museum, Covent Garden. Visitors can step up, ask the poster anything, and watch a 1927 design respond in their own language.

For museums wondering “why am I not doing this?”, the project offers a reassuring answer: with the right partnerships, curatorial discipline, and a focus on controlling your own data, bringing your collection to life is more achievable — and affordable — than you might think.

–

This case study was shared at the Museum AI Summit 2026. For further examples of museums using artificial intelligence from this conference, visit the Museum AI section of the MuseumNext blog.