I've been testing tl;dv's AI meeting notes against our internal sales call reviews. The feature everyone talks about is the automatic summary, but the real value for me is in the structured data extraction. Specifically, the `question_count` metric.
Our hypothesis: The best product demos are conversations, not monologues. High prospect question count correlates strongly with engagement and pipeline velocity. tl;dv can surface this, but you need to move past the main UI.
Here's the raw data structure tl;dv's API returns for a call. The key field is `topics`:
```json
{
"call_id": "abc123",
"summary": "...",
"topics": [
{
"topic": "Pricing discussion",
"speaker": "Prospect",
"start_time": 1200,
"questions": [
"What's the cost for 50 seats?",
"Is there an annual discount?"
]
},
{
"topic": "Integration capabilities",
"speaker": "Prospect",
"start_time": 850,
"questions": []
}
]
}
```
The trick is to post-process this. The UI doesn't give you a total `question_count` per speaker, so you have to calculate it. I wrote a simple script that:
1. Fetches calls for a date range.
2. Iterates through `topics`.
3. Counts all questions where `speaker` is NOT one of our sales reps.
4. Aggregates totals per call.
The workflow:
* Run this weekly over all "Demo" tagged calls.
* Filter for calls where prospect `question_count` >= 10.
* Review those call transcripts/videos first for coaching.
* Also flag calls with 0 or 1 questions for immediate intervention.
Initial results from last quarter: Demos in the top quartile for prospect questions had a 34% higher conversion rate to next stage. The correlation is stronger than talk-to-listen ratio or call length.
The caveat: This only works if tl;dv's question detection is accurate. In my testing, it's about 85% precise. You'll get some false positives where statements are misclassified as questions, but the signal is still strong enough to be useful.
If you're just using tl;dv for summaries, you're missing the point. The extracted metadata is what lets you build actual analysis. Has anyone else tried building similar filters? I'm looking at combining `question_count` with keyword triggers from the transcript next.
Show me the query.