AI cataloging: how AI for library cataloging works, step by step

How AI for library cataloging works: five steps

The workflow below is the one BiblioScan AI follows for physical material. Each step has a different owner: the first belongs to a person with a phone, the middle three to the machine, and the last to the cataloger.

Step 1: capture

Everything starts with an image. For a printed book, a photo taken with any phone is enough: the cover, the back cover, the title page or the copyright page, which is where the ISBN, publisher and year are usually printed, plus the table of contents when you want subject suggestions. There is no need for a dedicated scanner or a special workstation. Any member of the team can capture material with the phone they already carry, and several people can photograph in parallel, which removes the physical bottleneck an expensive overhead book scanner creates.

Material that is born digital, such as a thesis PDF, a journal article or an institutional report, starts somewhere else. That flow belongs to our other tool, MetaExtract API, which reads the document itself instead of a photo; our guide to AI metadata extraction covers it. This guide focuses on the image-based flow, because it raises the most questions: it looks like the hardest one to automate.

Step 2: recognition and extraction

The photo goes through a process that combines OCR (optical character recognition) with vision and language models. OCR turns the text in the image into plain text. The vision and language models interpret that text and relate it to what a bibliographic description is supposed to contain: where the title usually sits, how an ISBN differs from a year of publication, which string is probably the author's name.

This is what makes an AI cataloging tool more than generic OCR. It does not just read the text; it classifies each piece according to the role it plays in the record. It is the same difference as between transcribing a page and cataloging it.

Step 3: mapping to MARC21 or Dublin Core

Once the data is identified (title, author, publisher, ISBN, year and subject, among others), it has to be translated into the structure your system expects. For a catalog such as Koha, that structure is MARC21, with its numbered fields and subfields; for a repository such as DSpace, it is usually Dublin Core. Extracting "Smith, Jane" is not the same as knowing that the name belongs in field 100 as the main personal-name entry and not in a note. Mapping decides whether the record that reaches your catalog is usable or creates more correction work than it saves.

Step 4: auto-filling the record in your system

With the record structured in the right standard, the fields are inserted directly into your library system. BiblioScan AI integrates with Koha, DSpace and Alma, so the cataloger never copies and pastes between an external tool and the catalog: the record appears already loaded, with its MARC21 fields filled in, in the same screen where cataloging has always happened.

In the projects BiblioLabTech runs, this is the part that surprises teams who have cataloged by hand for years. The system they use does not change; only who does the keying changes.

Step 5: human review

This is the most misunderstood point: AI does not replace the cataloger, it changes where the cataloger steps in. Instead of transcribing every field into a blank form, the librarian reviews a proposed record and applies the judgment a model does not have: normalizing names against the authorities already used in the catalog, assigning subjects and classification, deciding how to handle editions and multiple copies, and resolving cases where the photo came out blurred or the book has no ISBN.

What the AI does and what the cataloger reviews

The fields BiblioScan AI pre-fills follow the cataloging profile agreed with each library. A typical profile covers the ISBN (020), the title statement with its statement of responsibility (245), the edition (250), publication data (264), the physical description (300), the series statement (490) and notes, plus subject suggestions drawn from the table of contents. Classification and content analysis are proposed, but the decision stays with the cataloger.

How to read the accuracy figures of AI cataloging

The BiblioScan AI product page publishes an accuracy figure of 99.9%, under a scheme of 95% automatic work and 5% human review. Treat it as what it is: the product's own marketing figure, not a guarantee for your project. There is no published methodology behind it, and the real share of records that need correction depends on the quality of the photos, the condition of the material and how demanding your library's cataloging policy is.

That is why a serious rollout includes a measured pilot. Your team records its current baseline (how many records per hour it produces and how much post-correction they need) and then measures the same thing with the tool, on your own material. It is the only honest way to put a number on the gain, because every library starts from a different place and a sector average tells you nothing useful. What does hold in every BiblioLabTech project is the change of role: the cataloger moves from typing to deciding.

Materials, formats and systems AI cataloging works with

BiblioScan AI works from photos taken with any phone or camera, with no special scanner and no particular image format. It is built for printed books and, more broadly, for any item with a page of identifiable publication data. It processes material in English, Spanish and Portuguese and returns the record according to the configured MARC21 profile; for collections in other languages, test it during the pilot before committing the full workflow to the tool.

Results are strongest on modern monographs with clear printing. With rare books, material in non-Latin scripts, serials or audiovisual material, the gain is smaller and the review heavier. For those collections, the honest recommendation is to keep cataloging by hand and use the tool where it pays off: the volume of everyday monographs. Books without an ISBN can still be processed when the copyright page carries other identifying data, but extraction confidence drops and those records usually need more review.

On the system side, BiblioScan AI integrates with Koha, DSpace and Alma, so it works whether your catalog is open source or commercial. Koha and DSpace are free software maintained by their communities (koha-community.org and dspace.org); BiblioScan AI is a BiblioLabTech product that works alongside them. At Universidad Continental in Peru, it runs on top of a commercial library platform the university had already implemented, with no migration.

Multiple copies, series and collections

A practical question comes up in every rollout: when a library holds several copies of the same title, the AI does not catalog each photo as a brand-new book. The usual workflow is to catalog the title once, with a complete bibliographic record, and then add each additional copy as an item attached to that record, with its own barcode and location. Photographing each copy only makes sense when copies really differ: a different edition, a condition that needs a note, or a donation with its own provenance.

Series and collections work the same way. The AI identifies the title and, when the book states it explicitly on the cover or back cover, the series it belongs to. Whether that series is cataloged as a higher-level record with linked volumes or as notes inside each individual record is a cataloging-policy decision for the library, not for the model. Once again, this is where professional judgment comes in and automation steps aside.

Where the time savings actually come from

Capture does not save time compared with taking a reference photo, which many libraries already did anyway. The real savings sit in steps 2, 3 and 4: recognition, MARC21 mapping and auto-fill. In a manual workflow, those steps cost minutes of keying per record; with AI, they happen in seconds and all at once. Step 5, review, still takes the time it always took to make those decisions, except that it is now the only part of the process that consumes skilled staff time.

That is also why AI cataloging changes the economics of a backlog. Every uncataloged book is a resource the institution has already paid for and nobody can find. Moving catalogers from transcription to review is what makes clearing years of accumulated material realistic for a small team.

Photos or PDFs: BiblioScan AI and MetaExtract API

BiblioLabTech approaches AI cataloging from two complementary directions. BiblioScan AI starts from a photo of physical material and auto-fills MARC21 fields in Koha, DSpace or Alma. MetaExtract API starts from a PDF and extracts its metadata (author, title, abstract, keywords and local fields) for repositories and document-management systems. When both are implemented for the same library, BiblioScan AI usually covers the print collection and MetaExtract API covers the institutional repository, on the same infrastructure.

A real example: AI cataloging on Koha and DSpace

At the Escuela Colombiana de Ingeniería Julio Garavito in Bogotá, BiblioLabTech runs exactly this workflow: Koha with BiblioScan AI for the physical collection and DSpace with MetaExtract API for the repository, with mandatory human review before each record is saved in both systems. The result, as the institution describes it, is that descriptive work moved from transcription to review, which is precisely the change of role this guide describes.

Universidad Católica Luis Amigó in Medellín took a phased route: it started with MetaExtract API on its DSpace repository and ABCD catalog, then added BiblioScan AI for print material once the first service was stable, and renewed both.

How to roll out AI cataloging in your library

When BiblioLabTech implements BiblioScan AI, the order is usually:

1. Connect and configure. BiblioScan AI is connected to your system (Koha, DSpace or Alma) and the cataloging profile is defined: mandatory fields, subject vocabularies and local rules.
2. Run a test batch. Real material from your collection is processed to calibrate extraction against your own criteria and to set the baseline for the pilot.
3. Agree the review procedure. Who reviews what, what is always checked and how consistency with existing records is maintained.
4. Open the workflow to the whole team. Only once the test batch has been validated.

There is no catalog migration: the tool connects to what you already have. A typical rollout takes six to eight weeks from assessment to deployment, including the measured pilot, and four to five weeks when the library already has a defined cataloging profile and its system exposes a standard integration. Training is light on the operational side, because taking photos is simple, and deeper on the bibliographic side: what to always check, how to handle subject suggestions and how to stay consistent with what is already cataloged.

On cost, there is no single rate. BiblioScan AI is sold as an annual service license that covers implementation, profile configuration, support and improvements during the period. The price depends on the number of branches or libraries using it, the expected processing volume and the integrations required with third-party systems, and a firm figure comes after an assessment of your system and a test batch of your own material. Every Koha project by BiblioLabTech includes BiblioScan AI at no extra cost; see our Koha implementation guide for the platform side. If you want a real figure rather than an estimate, send us a sample of your material and your system details and we will scope your case.