Transforming legacy card catalogs with AI

Legacy card catalogs constitute a major barrier to discovery in research libraries, as they remain largely disconnected from modern digital infrastructures. This case study presents a large-scale implementation at Uppsala University Library, Gothenburg University Library and Lund University Library, where a hybrid AI-driven pipeline has been developed to transform digitized catalog images into structured metadata information, integrated into Libris (the Swedish union catalog). To date, 770,000 holdings have been added or updated in Libris enabling streamlined request workflows, and each record linked to the original catalog image via persistent identifiers. This paper presents the pipeline which includes scanning and analyzing the catalog, segmentation and transcription of text, metadata extraction, matching existing bibliographic records, and adding holding information. We discuss methodological choices, technical challenges, evaluation strategies, and implications for research accessibility, and focus on four key contributions: (1) handling heterogeneous and historically complex catalog data, (2) a flexible pipeline combining AI tools and classic library and systems skills, (3) large-scale metadata matching against Libris using multiple matching strategies, and (4) systematic quality evaluation using ground truth and sampling. The study demonstrates how AI can be operationalized within library infrastructures beyond experimental settings, combining programming and prompting skills with core librarian competencies in cataloging and metadata.