Real-World Case Studies: AI in Libraries
From the course AI for Librarians
Leading library systems around the world are already deploying AI with measurable results — here is what they did and what they learned.
Case Study 1: New York Public Library — AI Reference Chatbot
NYPL launched an AI-powered chat assistant across its branch websites to handle the volume of after-hours reference queries that staff could not answer in real time.
What they built: A retrieval-augmented chatbot trained on NYPL's own LibGuides, catalog data, and FAQ corpus. The bot handles questions about hours, holds, renewals, and basic subject reference.
Outcomes after 12 months:
- 34% reduction in routine phone and email queries
- After-hours chat satisfaction score of 4.1/5
- Librarian time freed for complex research consultations rose by 22%
- Spanish-language mode added in month 6 with no retraining cost
Key lesson: Narrow the chatbot scope tightly at launch. NYPL initially allowed open-ended research questions and accuracy suffered; restricting scope to holdings and policies first built patron trust before expanding.
Case Study 2: British Library — AI-Assisted Cataloging of Historical Maps
The British Library holds over 4 million maps. Manually cataloging uncatalogued 18th and 19th-century maps would take decades at current staffing.
What they built: A computer vision pipeline using fine-tuned vision models to extract place names, dates, cartographer signatures, and scale bars from high-resolution scans. Catalogers review and approve; they do not create records from scratch.
Outcomes:
- Processing speed increased from ~30 records/day to ~400 records/day per cataloger
- 97% accuracy on place name extraction for printed maps (lower for manuscript)
- Backlog of 280,000 uncatalogued maps reduced by 60% in 18 months
Key lesson: Human review remains essential for manuscripts and edge cases. AI handles volume; experts handle ambiguity.
Case Study 3: Toronto Public Library — AI for Collection Weeding
TPL used an AI model trained on circulation patterns, publication dates, condition reports, and ILL request data to flag materials for weeding review.
Outcomes:
- Identified 180,000 items across 100 branches as weeding candidates in 6 weeks
- Manual weeding of the same scope would have taken 3 years
- Patron complaints about removed materials dropped vs. previous manual weeding cycles because AI flags were more defensible with data
Key lesson: AI weeding recommendations must be filtered through community context. A branch serving a large print-reading senior population needs different thresholds than a tech-heavy urban branch.
What These Libraries Have in Common
- They piloted in one department before scaling
- They maintained human review for all patron-visible outputs
- They measured outcomes in staff time, patron satisfaction, and collection quality — not just cost
- They published what they learned openly, contributing to the profession
Key Takeaways
- •NYPL's scoped reference chatbot freed 22% more librarian time for complex consultations
- •British Library increased cataloging throughput 13x by using AI for extraction and humans for review
- •Toronto Public Library reduced a 3-year weeding project to 6 weeks with AI-flagged candidates
- •Successful AI deployments pilot in one area, maintain human review, and measure real outcomes
Progress is kept in this browser only.