This poster chronicles an iterative, hands-on project to automate the generation of structured MARC holdings fields (853/863) from 866 textual holdings statements — and the challenges of making that process portable across institutions. Beginning with a colleague's existing automation script, the presenter adapted the tool to their own branch collection through self-directed learning, without a formal programming background. When it became clear that the pattern-matching rules at the heart of the tool would need to be rewritten from scratch for every institution's unique holdings formatting conventions, the presenter turned to AI — exploring several distinct approaches: using AI to write and refine those matching rules directly from sample data; using AI to survey and cluster the range of formatting patterns present in the holdings data; using AI to build a reusable pattern-detection tool; and using AI to interpret holdings statements record by record, bypassing hand-crafted rules entirely. The presentation evaluates the trade-offs of each approach and reflects on what these experiments reveal about where AI assistance is genuinely useful in technical library workflows — and where human cataloging expertise remains essential.
