What problem does it solve?
The KU Leuven Association team that tested an AI assistant in its library search heard a split among respondents. One respondent wrote that they were not sure the assistant added value, being used to filters and advanced search, while others used it to get a quick first impression of a topic, to find ideas for search terms and to avoid information overload.
The job of the assistant is to help a user start. It takes a question in natural language, returns a short overview and selected sources from the discovery system's search results, and points to related questions and the full results list. The libraries in the evidence stress that it is a starting point and a complement to the usual search, never a replacement for it.
The evidence so far comes from libraries that tested the vendor's feature in beta or early rollout, and it is cautious. At KU Leuven users valued it for exploration and librarians were the most critical group, and the common problems are answers that sound more certain than they should, sources missing from the assistant's scope and users who expect a general chatbot.
How does it work?
- The user asks a question. The question is typed in plain language into the library search page rather than as a list of keywords.
- The assistant works from search results, not from the whole library. It runs the question against the content that the discovery system searches for it. Westminster notes that certain collections are excluded from the assistant's scope, and the KU Leuven slides list a "lack of search in local sources" next to "good results even with only CDI", so coverage differs from the library's full holdings.
- A language model writes an overview. The answer has a brief overview and a list of selected sources, with links to open the items in the library's holdings.
- The user is pointed onward. The page offers related research questions and a link to more results, so the user can move into a normal search. The KU Leuven slides show that most users did not look beyond the first overview.
- Library staff own the scope and the message. They decide what is switched on, label it as a beta or supplementary tool, train staff to explain it, and put help where users need it.
- Audience
- Customer facing
- Autonomy
- Assist
- Adoption
- Emerging
- Channels
- Web chat
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | 6798 | 1 | 1 organization |
Value drivers: Customer experience, Inclusion and access.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
KU Leuven Association (LIBIS, KU Leuven Libraries)
Belgium · Education · 2025
LIBIS, the library services unit that serves the KU Leuven Association's libraries, ran a beta trial of the Primo Research Assistant from February to May 2025, open to all users of the KU Leuven university libraries and the university colleges through an invitation to test, with a user survey. The trial recorded 6,798 test sessions and 104 survey responses. The team's IGeLU 2025 conference slides report positive overall appreciation, with students the most positive and librarians the most critical.
The slides state that the assistant excels in exploration and inspiration, and users described it as a stepping stone and a way to bookend a literature search. Most did not look beyond the first overview, users were unsure how the tool works or expected it to behave like ChatGPT, and the university colleges missed search in local sources.
- Interactions handled: 6798, beta trial, February to May 2025
"Total RA Test sessions: 6 798"
Claimed by: organization
University of Manitoba Libraries
Canada · Education · 2025
University of Manitoba Libraries assessed the Primo Research Assistant through a multidisciplinary working group that posed varied research questions to test usability, accuracy and fit with user needs, and then rolled it out. The group found it useful as a starting point for research, with limited local relevance and overly confident responses as the main weaknesses. Rollout focused on positioning the tool as a supplementary resource, training library staff and telling users plainly what it can and cannot do. The libraries presented the case at the 2025 ELUNA conference, with lessons on managing expectations, academic integrity and keeping traditional library resources alongside the assistant.
No outcome disclosed.
University of Westminster
United Kingdom · Education · 2025
The University of Westminster enabled the Primo Research Assistant, an AI feature of the Ex Libris Primo discovery system, on 4 February 2025. Its library guide says the tool was introduced in beta to explore its potential and understand how students and staff may use it, that it is not an alternative to Library Search, and that certain collections are excluded from its scope.
The guide lists a small number of other UK universities that have made the assistant available to students, including Cardiff, Stirling and the Open University, and links the Open University's published user testing and feedback, the CARLI summary for US libraries (which, it says, found that most do not yet find the tool ready for full implementation) and a lesson plan from the vendor. Westminster publishes no usage or satisfaction figures.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- The library's search index with clean metadata for the collections in scope
- A list of the collections and content types that are excluded, written in plain words for users
- A baseline of reference questions and search behaviour from before launch
Systems to integrate
- The library discovery system and its search interface
- Authentication or proxy access for licensed e resources
- The library's help pages and chat or reference service for handover
- Analytics on queries, clicks and feedback
Complexity: Low
Most libraries switch on a feature of their discovery system and then decide the scope, the label and the user guidance, with staff training and a plan for gathering feedback. A custom build also needs a search interface to the library index and the right to pass licensed content summaries through a language model.
- 1
Launch as a labelled beta
Westminster introduced the assistant in beta to explore its potential and understand how students and staff may use it, and states that it is not an alternative to Library Search. Say the same on the page.
- 2
Test with a working group first
Manitoba's multidisciplinary group posed varied research questions to judge usability, accuracy and fit before rollout, and found it useful as a starting point but limited in local relevance and prone to overly confident responses.
- 3
Train library staff before users
Manitoba's rollout centred on training library staff and on clear communication of what the tool can and cannot do. KU Leuven's survey found librarians the most critical user group, so their questions are worth using to shape the wording of the guidance.
- 4
Put help at the point of need
KU Leuven found that online help was rarely consulted and that users were unsure how the tool works, some expecting it to behave like ChatGPT. Show scope and limits on the results page and replace vague link labels such as "view more results".
- 5
Run a time boxed pilot with a survey
KU Leuven opened an invitation to test to all users from February to May 2025 and ran a user survey, which let it compare students, staff, librarians and the university colleges.
- 6
Review by user group and decide the scope
Read results by group: KU Leuven found students most positive and librarians most critical, and the university colleges missed local sources. Decide what to keep, expand or switch off.
Guardrails
- A visible label that the answer is AI generated and a beta or supplementary tool
- Sources only from the discovery system's search results, each with a link to the item
- A plain language statement of the collections and content types outside the assistant's scope
- A clear message instead of an answer when the question needs data the assistant cannot search, such as filtering for theses
- No use of user queries to train a model without a documented basis
KPIs to instrument
- Sessions and repeat users by user group, such as undergraduates, graduate students and staff
- Share of sessions in which the user opens at least one source
- Share of users who find the answer helpful, partly helpful or not helpful
- Librarian rated accuracy of a sample of answers each term
- Queries that the assistant could not serve because of scope
Human in the loop
Librarians own the scope, the wording of the guidance and the review of the tool. They read a sample of queries and answers each term, answer questions from users who doubt a result and decide whether the tool stays on. Users remain responsible for reading the sources and for citing them correctly, and the library should say so in its guidance on academic integrity, a theme Manitoba emphasized in its lessons learned.
Common failure modes
- Answers that sound too sure
- Manitoba found overly confident responses. Show the sources next to the overview, tell users to read them and sample answers for accuracy.
- Mismatched expectations
- KU Leuven users expected a ChatGPT style conversation and asked for things the tool cannot do, such as searching the library's theses. Explain scope in the interface.
- Local and specialist sources missing
- Local content can fall outside the assistant's scope, which the KU Leuven university colleges saw as a major hindrance. Check coverage of local and multilingual sources before launch.
- Shallow use
- Most KU Leuven users did not look beyond the first overview. Track whether users open sources and pair the tool with information literacy teaching.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
A conversational assistant that interacts with people falls under the transparency duty of Article 50: users must be told they are dealing with an AI system unless that is obvious. A research assistant that finds and summarises sources is not an Annex III use, as long as it is not used for the education purposes Annex III lists, such as deciding admission, evaluating learning outcomes, assessing the level of education a person can access or monitoring students during tests.
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). People must be informed when they interact directly with an AI system, unless that is obvious from the circumstances.
- Guidance for generative AI in education and research (UNESCO, Global). UNESCO's first global guidance on generative AI in education and research; it proposes mandating the protection of data privacy and an age limit for independent conversations with GenAI platforms.
Controls to put in place
- AI disclosure on the results page, with a link to the scope and limits
- A data protection impact assessment covering query logs and any personal data in them
- Contract terms with the discovery vendor on content use, model training and data location
- Staff briefing and academic integrity guidance for students and staff
- A termly review of usage, feedback and answer accuracy with a named owner
Frequently asked questions
- What do libraries find when they test an AI research assistant?
The KU Leuven Association's beta, with 6,798 test sessions across KU Leuven and its university colleges and 104 survey responses, found positive overall appreciation, with students the most positive and librarians the most critical. Its slides say the assistant excels in exploration and inspiration, and that most users did not look beyond the first overview. Manitoba found it useful as a starting point but limited in local relevance and prone to overly confident responses.
- Does an AI research assistant replace librarians or the normal library search?
No library in the evidence positions it that way. Westminster says the tool is not an alternative to Library Search and runs it as a beta, and Manitoba positions it as a supplementary resource and trains library staff to explain what it can and cannot do.
- Is a library AI assistant high risk under the EU AI Act?
Usually not. It falls under the transparency duty of Article 50, so users must know they are talking to AI. It would move toward Annex III if it were used for purposes listed there, for example to evaluate learning outcomes, to decide admission or to monitor students during tests.
How to cite this page
Blits.ai AI Use Case Library, "AI research assistant in university library search", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/library-ai-research-assistant. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published