Anatomy of a call-analysis agent

Gong did meeting intelligence first. Dench's call-analysis agent does it second, on top of an open data model that the rest of the CRM can read.

The loop:

1. Ingest — transcripts pulled from Zoom/Meet/Phone via Composio.
2. Tag — objection categories detected with a small classifier (~92% F1 on our internal set).
3. Aggregate — objections × deal stage × month, joined to the deal's actual outcome.
4. Recommend — pricing tweaks, content gaps, or packaging changes that move the win-rate.

The trick is step 4. Pure aggregation produces dashboards; recommendations require the agent to look at outcome data and propose interventions. We seed prior interventions and their results as memory, so the recommendations get better over time.