Challenge
Free-form conversations were difficult to compare at scale. The output needed to be consistent enough for reporting without hiding the original context that a reviewer might need.
Approach
We designed a Llama 3.1 prompt and processing pipeline that segmented conversations, returned structured labels, and retained the supporting text for human review.
What changed
- Conversation data became searchable and report-ready.
- Reviewers could inspect the evidence behind each classification.
- The workflow created a foundation for routing and escalation rules.