Warner Chappell Cuts Cloud Costs by 50% With AI Embeddings

Warner Chappell Music's Global Match system used to rely on brute-force SQL string-matching in Snowflake to identify its publishing catalog inside massive monthly CSV files from streaming platforms. That approach was fragile against typos, missing initials, and transliteration differences, and as data volume grew, both processing time (12+ hours per batch) and cloud costs spiraled out of control.

The team abandoned heavy database-layer processing in favor of BERT-based text embeddings, encoding roughly 2 million claimable catalog records into a vector space. Using an event-driven architecture with AWS Lambda and SQS, files are sharded and processed in parallel, while scikit-learn's K-Nearest Neighbors search finds top matches in memory within milliseconds. This lets semantic similarity capture variations that strict string-matching missed entirely.

The result: cloud costs down more than 50%, processing speed up over 100x, and total match coverage up 65%. The new architecture also enabled a human-in-the-loop workflow where faster manual review feeds corrections back into the model for continuous improvement. The core engineering lesson: when scale strains a system, rethinking the underlying algorithm often beats throwing more infrastructure at the problem.