How an AI pipeline made 2,000 airline support chats queryable in hours

A customer journey intelligence platform processed roughly 2,000 airline support conversations in a single scheduled batch run, enabling leadership to query results in plain language within seconds. The system pulled data from disparate sources — chat logs, voice recordings, and agent interaction logs — and used a normalization layer to stitch records into unified, chronological journey timelines before analysis. Amazon Transcribe handled voice transcription, PII redaction ran deterministically before any content reached the model, and Claude Sonnet evaluated each journey against the client's existing quality rubrics. A two-path architecture separates batch evaluation from interactive retrieval, meaning dashboards and conversational agents query stored results rather than re-scoring conversations in real time. This design keeps compute and LLM costs predictable while allowing quality teams to update evaluation criteria by editing prompts rather than retraining models.