AI in libraries: beyond the buzz, what actually helps

Our take on the Tosca 2026 study, and the line we hold at Loupe.

In 2026, almost every library software puts artificial intelligence front and center. The Tosca Consultants study published this year confirms it: AI has moved from talking point to everyday use. Assisted cataloging, record deduplication, enrichment, patron help. What served as a demo argument two years ago has become a buying criterion.

That is good news. It is also a trap.

The trap is AI treated as a checkbox: an assistant added so you can say you have one, without asking what it is for or where its answers come from. In a French market that has been flat for three years (around 41 million euros, concentrated among a handful of vendors, according to Archimag), the temptation is strong to follow the trend rather than real use.

The context does not help. Budgets stay under pressure, and AI arrives fast, sometimes faster than the means to assess it calmly. All the more reason to choose tools that focus on what is useful, and that own what they do not do.

At Loupe, we take it differently. The question is not "do you do AI," but "does this actually help the person in front of you, and can they trust it."

What actually helps

Take search. Most search bars assume the user already knows the exact title or the right keyword. Many people do not. A search that understands a question asked in plain language, and that surfaces the right documents even without a perfect phrasing, saves time for the user and for the front desk alike. Here, AI answers a real need, not a sales pitch.

Another example: helping a team read its own data. Understanding peak hours, how spaces are used, footfall, so decisions rest on facts rather than gut feeling. In that case, AI is only worth it if it stays legible, verifiable, and in service of a human decision.

What we refuse to do

Holding a line also means knowing how to say no.

No to an AI that invents records or answers and presents them as facts. In a library, a wrong but well-phrased piece of information does more damage than no answer at all.

No to using patron data to train third-party models. That data belongs to the library and its users. Loupe is hosted in Switzerland, and that is a choice, not a detail.

No to the black box. A useful answer is one whose source you can trace. The team has to be able to understand, and explain, what the tool puts forward.

Our conviction

AI in libraries does not need to be spectacular. It needs to be accurate, restrained, and in service of people: the users who search, and the teams who help them. The rest is decoration.

If you are weighing these questions for your institution, we are happy to talk. A short conversation is often enough to tell what actually helps from what just looks good in a demo.