CallInsight AI listens to every sales call your team makes — then transcribes, scores, flags and coaches automatically. Hot-lead verdicts with evidence quotes, compliance checks, and knowledge that builds itself. Minutes after hangup.
Plugs into Freshworks · Deepgram transcription · WhatsApp sync · REST API
No uploads, no tagging, no listening back. The moment a call ends in your CRM, a 17-worker async pipeline takes over.
Calls and recordings pull automatically from Freshworks & Freshcaller — deals, agents and metadata included.
sync_calls · download_audioDeepgram speech-to-text with speaker diarization, then AI-corrected speaker labels and transcript repair.
deepgram · gemini 2.5Sentiment, executive summary, compliance check, coaching opportunities and Q&A extraction — every call.
execute_ai_tasksThe 9-criteria hot-lead evaluator delivers a verdict with confidence scores and transcript evidence.
call_evaluatorYour rules engine raises severity-ranked flags and routes them to the right owner in real time.
detect_flagsQuestions cluster into FAQ candidates, knowledge gaps surface, and embeddings keep search semantic.
match_and_validateavg. wall-clock from hangup → full analysis: minutes, not Mondays
Five analysis engines run on every conversation — built on Gemini 2.5, GPT-5 and Claude via a vendor-agnostic prompt registry you can extend.
Customer and agent sentiment tracked separately, with emotional triggers and the exact moments the mood shifted.
A 3–5 sentence brief with key discussion points, the call outcome, and concrete next steps — readable in 20 seconds.
Recording consent, privacy disclosure, pricing transparency — verified on every call with a risk level and a list of what's missing.
Strengths and improvement areas per agent, with timestamped examples pulled straight from the transcript.
Every question a prospect asks — and how your agent answered — extracted, embedded and validated against your knowledge base.
Natural-language questions over your entire call library, with saved analysis recipes you can version, pin and re-run.
The call evaluator scores every prospect against nine buying-intent criteria — from a concrete operation window to financial readiness to photo-consult agreement. Each criterion carries a confidence score, an ambiguity flag, and up to two verbatim evidence quotes from the transcript. Your team opens the morning dashboard and calls the right people first.
Eleven built-in flag types plus a custom rules engine — keyword, metadata, AI-result and composite AND/OR conditions. Flags carry severity, get auto-assigned, and live on the dashboard until someone resolves them.
Recurring questions cluster automatically into FAQ candidates with drafted answers. Agent answers are validated against your knowledge base with semantic matching + NLI — contradictions become violations, unknowns become knowledge gaps.
An AI sparring partner for your agents — built from your real conversations, not generic scripts.
Scenarios by difficulty with customizable personas and objection intensity — mild, moderate or strong. Locked progression keeps it a game agents want to win.
Instant feedback after every response, six live skill meters, and an objectives tracker. Or end-of-session mode for a full debrief with per-message scores.
The live-assist module pushes sub-150-character suggestions to agents mid-call — objection counters and next moves in under 100ms, powered by Claude.
Session summaries use the same skill taxonomy as your production coaching analysis — so training scores and call scores speak the same language.
Per-agent WhatsApp sessions sync every message, media file and read receipt into the same customer record as the calls — automatically.
Calls, WhatsApp threads, notes and files in one chronological view — with health scores, contact stats and deal context attached.
Telnyx numbers for SMS and voice, Brevo for transactional email — every channel lands in the same timeline.
Everything your analysts see in the dashboard, your systems can pull over a key-authenticated REST API.
Pull every call on a deal — transcripts, sentiment, summaries, Q&A and signed audio links — into your BI stack, your CRM automations, or anything that speaks HTTP.
{
"deal_id": "31002048117",
"calls": [{
"date": "2026-06-09T14:32:11Z",
"duration_seconds": 761,
"transcript": [ … 142 utterances … ],
"analysis": {
"sentiment": { "overall": "positive", "confidence": 0.92 },
"lead_verdict": { "score": "8/9", "label": "hot" },
"summary": "Prospect compared 3 clinics…",
"flags": [ "price_concern" ]
},
"audio_url": "…/api/audio/4821?token=sgn_…"
}]
}
Schema-per-tenant PostgreSQL. Your data never shares a table with anyone else's — not row-level filters, separate schemas.
Phoenix LiveView dashboards update the second a worker finishes. No refresh button, no polling.
Recordings stream through expiring, token-signed URLs from your own S3/MinIO bucket. Audio never leaks.
Validation queues for AI verdicts, FAQ candidates and answer reviews — approve, reject, or correct with full audit history.
Connect your CRM in an afternoon. By tomorrow morning, every call your team made today is scored, flagged and searchable.