AI chatbot for libraries: what it is, how it works and what it costs

What an AI chatbot for libraries is, and why it is not ChatGPT with a logo

When someone asks for "a chatbot for the library" without more context, they usually picture something like the customer-service assistant of an online store. The problem is that a generic assistant, connected to a language model and nothing else, has a serious flaw in a library: it can make things up. If a student asks whether the library has a book available, a generic model with no access to the real catalog can give an answer that sounds plausible and is false, because its job is to generate coherent text, not to check an inventory.

A well-built library chatbot solves this with RAG (retrieval-augmented generation). Before answering, the system searches the library's real sources (the catalog, the services page, the borrowing rules, the subject guides) and only then drafts a response grounded in that information. It is a documented technique that academic libraries increasingly rely on. The University of Calgary Library, for example, launched its reference chatbot, T-Rex, in 2021 on a commercial product that combines a large language model with retrieval-augmented generation, trained on the library's own web content, including its research guides and opening hours, as OCLC Research reported.

How a library chatbot works inside

Simplified, the flow has three steps:

1. The user asks a question in natural language: "Do you have Stewart's Calculus available?", "What time do you close on Saturday?", "How do I renew a loan?"
2. The system searches the library's connected sources for the information relevant to that question. If it is about a book, it queries the catalog (Koha, for example); if it is about a document in the repository, it queries DSpace; if it is about hours or rules, it queries the services page and the policies.
3. The language model drafts the answer using the information it retrieved, not its general memory.

This has a practical consequence worth stressing: the chatbot only knows what the library lets it know. If the catalog is not up to date, or the borrowing rules are not published anywhere the chatbot can read, it cannot answer those questions well either. In the projects BiblioLabTech runs, the first phase is never "switch on the chatbot"; it is reviewing which information sources exist and what state they are in, because a chatbot connected to incomplete data inherits the same gaps. The configuration also defines, from day one, the topics the assistant must never try to answer and how it recognizes that it lacks enough information and hands the conversation to a person.

Library chatbot vs generic chatbot

The questions an AI library chatbot actually resolves

It is not magic, and it does not replace library staff on complex tasks. What a well-implemented AI chatbot does well is the bulk of any library's inquiry volume: repetitive, direct, high-volume questions that a person answers over and over today. The most common:

• Hours: of the library, of specific rooms, during exams or holidays.
• Availability: whether a book is on loan, which branch holds it, whether there is a digital copy.
• Loans and renewals: how to renew, how many times, what happens with a fine.
• Recommendations: reading or bibliography suggestions based on the user's interests or program.
• Research orientation: which databases to use for a topic, how to cite in a given style, where to find a digital resource.
• Procedures: how to register as an external user, how to access subscribed databases from off campus.

What a library chatbot should not try to do is replace the reference librarian's judgment on a complex research query or a sensitive conversation with a user. The realistic goal, and the one BiblioLabTech works toward, is that the chatbot absorbs the repetitive volume and frees staff for what really requires human judgment: search-strategy advice, user education and support for researchers. Questions that require reading and synthesizing academic literature belong to a different tool, the AI research assistant.

Channels: website, WhatsApp and campus platforms

The same chatbot can live in more than one channel without duplicating the training work. The two most used in libraries are the chat widget on the website, the floating window any visitor to the online catalog can open without leaving the page, and WhatsApp, which in Latin America and many other markets is the default messaging channel for a large share of users, far ahead of email or a web chat. Adding WhatsApp is not cosmetic: a student who would never visit the library website will send a message to a number if it solves the question in ten seconds. The chatbot can also be built into academic portals, virtual learning environments and existing service desks, depending on the institution.

The chatbot replies in the language the user writes in, and institutions with an international population can fix the set of supported languages. Tone is configurable too: how users are addressed, the level of formality, the institution's own terminology and the hand-off messages. That configuration is refined during the pilot with the reference team, the people who know best how the library talks to its users.

Integration with Koha, DSpace and other systems

A chatbot that is not connected to the library system can only answer generic questions. Its real value comes from live data sources: Koha for item availability, loan status and the catalog; DSpace for the content of the institutional repository, such as theses, articles and datasets; and, depending on the case, ABCD, or OJS when the library publishes academic journals. BiblioLabTech implements the chatbot as a layer that connects to what already exists, without asking the library to change its management system to get a conversational assistant.

Personal queries are a separate design decision. When the assistant does not authenticate the user, the conversation is not linked to an identified person. If the institution wants personalized answers, such as the status of a patron's loans, authentication is required and the institution's data policies apply. Purpose, retention period and who can access conversation logs are defined in the contract.

A real case: the reference chatbot at Escuela Colombiana de Ingeniería

At the Escuela Colombiana de Ingeniería Julio Garavito in Bogotá, BiblioLabTech implemented a Reference ChatBot (ChatBot Referencista) as part of the university's complete library platform, alongside Koha with BiblioScan AI cataloging, the DSpace repository with MetaExtract API and the BiblioMeta library metasearch engine, all on a dedicated university server administered by our team. The institutional corpus was prepared, rules for escalation to staff were defined and an internal pilot ran before the chatbot opened to the public. It is the pattern this guide describes: the chatbot does not work in isolation, but as one more piece connected to the institution's real catalog and repository.

What your library needs before implementing one

Check three things before you hire a chatbot, because they decide whether it works well from day one:

1. A reasonably complete, up-to-date catalog. A chatbot does not fix a badly loaded catalog; it only exposes it faster.
2. Written policies somewhere a system can read: hours, borrowing rules, procedures. If that information lives only in staff members' heads, the chatbot cannot learn it.
3. Clarity on the channel your community really uses. There is no point prioritizing a web widget if your users live on WhatsApp, or the other way round.

Beyond that, you need a functional owner in the library who validates content, access to the catalog and repository if you want live queries, and a decision on who receives escalated questions. No infrastructure of your own is required. A typical project takes six to ten weeks to public launch, including the internal pilot with the reference team; preparing the sources is the phase that varies most.

Once live, measure it with conversation volume, the share of questions resolved without escalation, the most frequent topics and the questions left unanswered. That last list is the most useful: it is the monthly to-do list for extending the knowledge base, and it often reveals gaps in the library's public information.

What an AI chatbot for libraries costs

There is no fixed rate, and anyone offering one without knowing your case is oversimplifying. As a reference, these are the implementation ranges BiblioLabTech publishes:

On top of implementation there is a recurring monthly cost that depends on conversation volume. It has two parts worth separating in any quote: what the AI model provider charges for usage, which scales with the number of conversations, and what the implementer charges to keep the chatbot running, updated and connected to your systems. If someone offers "unlimited AI" at a flat price, ask what happens when volume spikes during exam or enrollment week, exactly when the chatbot most needs to work well. WhatsApp also carries its own recurring cost, because a WhatsApp Business account with real volume has to be activated and maintained.

Besides the reference variant, BiblioLabTech offers a Sales ChatBot for bookstores and publishers and a Student ChatBot for higher education, plus custom builds. Learn more about our AI chatbots for libraries, or tell us your volume and channels and we will quote your real case.