Capability
Call structuring
Turn the conversation into schema-valid output: what was discussed, what was disclosed, what was committed, what happens next.
The least glamorous capability and the easiest to prove. Agent time saved on after-call work is directly measurable, which makes this a good first deployment even when the conversational capabilities are still being evaluated.
Script matters here more than anywhere. A CRM field that expects Devanagari and receives Latin transliteration is silently corrupt data. Structured extraction has to be script-aware per field, not per response.
What we measure. Field-level extraction accuracy against human-labelled calls, schema validity rate, and script correctness per field.