Most voice AI platforms summarise calls a minute or two after they end. By then your rep has already moved on, the CRM is still empty, and the context is gone. We thought we could do better.
Floo's new post-call pipeline writes the executive summary, extracts action items, scores sentiment, and pushes the CRM update in under 3 seconds. Here's how it works.
The architecture
We pipe the transcript stream straight into OpenAI's gpt-4o-mini the moment the call connects. By hangup, the model has already pre-warmed on 90% of the conversation. The final pass takes 2.4 seconds on average.
- Streaming transcript hits the LLM in 200ms chunks
- Sentiment scored per-minute against a fine-tuned head
- Action items extracted via structured tool-call output
- CRM webhook fired in parallel, no waiting on the summary
Why gpt-4o-mini?
We benchmarked Claude Haiku, Gemini 2.5 Flash, and gpt-4o-mini on 200 real customer calls. Haiku produced slightly nicer prose but ran 1.8s slower. Gemini was the fastest but lost detail. gpt-4o-mini hit the sweet spot.
Tip, if your summaries need legal-grade accuracy, swap in Claude Haiku 4.5 from the model picker. The price bumps from $0.022 to $0.025 per minute, worth it for compliance-heavy use cases.
What ships in the CRM
Every call now lands in HubSpot, Salesforce, and Pipedrive with a structured payload: transcript URL, two-sentence summary, sentiment trend graph, list of action items with assignees, and any compliance flags. Open the deal in your CRM and the whole conversation is right there.
Our reps used to spend 4 minutes typing call notes. Now they spend zero. The CRM is just always right.
Beta customer, healthcare network
What's next
We're working on multi-language summaries (Hindi, Spanish, Mandarin first), custom rubric scoring for QA teams, and a 'coach mode' that flags missed opportunities in real time. All shipping over the next quarter.
