With Terentia in place and an AI project grant from Microsoft in hand, the Theodore Roosevelt Presidential Library was ready to bring its digital initiatives to life.

“We were already using AI on site—thinking about ways to optimize our green and sustainability certifications, how we’re ordering food and supplies,” Matt explains.

Next up was harnessing AI in their museum efforts. “We wanted to be able to surface collections in more interesting ways,” he says.

Building a trusted AI digital repository

It’s no exaggeration to say the planned AI initiatives are set to reshape the modern visitor experience. And at the center is what staff calls the GPT Reading Room.

Matt explains how it works: “It’s a GPT engine that’s privately trained on an LLM built from our materials, as well as primary source materials from 33 different collections, 18 different institutions, and secondary sources like biographies to adjust for historical bias.”

Unlike AI tools that draw from the entire internet—like ChatGPT or Perplexity— this one relies on Terentia as its trusted AI digital repository, ensuring every response is grounded in verified data. No external sources or unreliable data.

“The aim is to be able to ask [the GPT] any type of question and it responds as a virtual historian, then goes deeper into things and links into the collection assets,” Matt explains.

The beauty of the system is its scalability. For instance, a researcher can request all correspondence between Theodore Roosevelt and J.P. Morgan. While students working on a project could ask about TR’s life and get high-level answers with suggested prompts for more exploration.

The AI will extend into the physical Library as well. In an exhibit called Leadership in Action, visitors can step into the Cabinet Room and speak with an avatar of the president, powered by the same trusted GPT corpus.

As an exciting departure from written prompts, this AI experience is fully voice-activated—with the TR avatar speaking directly to the room and calling on guests by name, thanks to their RFID bracelets.

© Theodore Roosevelt Presidential Library

Transforming archival workflows with AI

Behind the scenes, artificial intelligence is already accelerating the team’s cataloguing work.

Take the Archivist App, created in a collaboration between Terentia, Microsoft, and OpenAI. Matt describes it as a “web interface” that combines Terentia’s OCR with AI to generate metadata and transcripts for records—all output reviewed by staff before entering the system.

“We didn’t fully trust AI,” Matt admits. “But what we’ve been shocked about is how good it has been.”

The Archivist App assigns confidence scores to its work, allowing users to filter and batch-approve records scoring above a certain threshold, while manually reviewing lower-confidence items. “The vast majority of transcripts and AI confidence numbers coming back are 90% or higher,” Matt says.

AI is also proving invaluable for tagging visual assets. Through AI facial recognition, the platform can help identify unknown people in photographs and videos, plus determine when and where they were taken and create links to related correspondences or events.

“It’s finding things that we as archivists maybe weren’t thinking about tagging,” Matt says.

“That ability to surface an image and tell you where it was and what other things are around that same time—that’s a huge advantage, which just doesn’t exist in current collections across different institutions.”

Solving a decades-long problem in no time

Matt points to his experience at another institution as a telling comparison. “They began digitizing and annotating their papers in 1969. That project will finish in 2030 and has cost $30 million so far,” he says.

Before AI, this was the unavoidable reality of delivering access to large-scale archives.

But with Terentia, the Presidential Library dramatically compressed that timeline—and without the hefty costs of long-term digitization and annotation.

“There’s no reason AI can’t generate those annotations on the fly,” Matt explains. “To be almost where they are in years, instead of decades, is huge.”

The quality holds up, too. Every AI response is source-linked and exportable as a PDF with full citations, making it ready for use by staff and external stakeholders like researchers, academics, and teachers.

It’s a win on multiple fronts: faster access, lower expenditure, and digital collections that begin delivering value years sooner.

© Theodore Roosevelt Presidential Library

Keeping AI ethical and responsible

Central to the Presidential Library’s approach is having both human oversight and full say over what data enters the system.

Terentia allows the team to mark sensitive materials as excluded from AI processing—such as personal letters from Roosevelt family members—and keep them safe from public access.

“There’s an ability to say this is off limits,” Matt explains. “We can have it in our collection, but don’t expose it to the web or the AI model. That access control is really huge.”

The Archivist App’s human review process also prevents inaccurate data from entering the trusted repository.

“We didn't want to feed bad data into the LLM because that’s really difficult to back out,” Matt says. When the AI doesn’t get something quite right, staff can make corrections that train the system to avoid repeating the same mistakes.

Following a human-in-the-loop approach increases staff’s confidence in the AI outputs while freeing them up from much of their previous busywork.

“We were expecting that AI might be able to help us a little bit—but AI is going to help us a lot. It’ll dramatically speed up the access to our records and collection items,” concludes Matt.