​

Sales Conversation Intelligence

​

What each call is scored on

• Talk ratio. The share of attributed floor time the rep held, computed from the diarised transcript’s speaker segments. Attributed time excludes segments whose speaker role could not be resolved, so a music-on-hold stretch or an unlabelled diarisation never distorts the split. When no segment could be attributed to either side, the call reports its ratio as unavailable rather than inventing a number.
• Competitor mentions. A case-insensitive, whole-phrase scan of the transcript against your workspace’s tracker list — the competitor and rival-product names you configure. Each hit reports the term and its mention count.
• Next steps. The concrete commitments and follow-ups the call produced (“send the revised quote”, “book the security review”), taken from the post-call analysis that already runs on the call. A call that ended with no agreed next step is itself a deal signal. Up to six next steps are kept per call.
• Deal-risk score. A 0–100 score with a low / medium / high band, built from the signals above. Every point the score adds carries a written reason (“negative customer sentiment detected”, “3 competitor mentions on the call”, “no next steps identified”), so a rep or manager can always see why a call is flagged — it is a coaching aid, not a black box. Factors: negative or disengaged sentiment, a rep monologue (the rep holding roughly three-quarters or more of the attributed talk time), competitor mentions, and a missing next step.

​

Turning it on

curl -X PUT "https://api.orbit.devotel.io/api/v1/insights/sales-conversation-intelligence/config" \
  -H "X-API-Key: dv_live_sk_your_key_here" \
  -H "Content-Type: application/json" \
  -d '{"enabled": true, "competitor_trackers": ["AcmeVoice", "RingCentral", "Five9"]}'

• Up to 50 tracker terms, each at most 80 characters.
• Terms match case-insensitively on whole words or phrases — a Gong tracker does not fire on “gonged”.
• If you never set a list, the workspace starts with a sensible default list of CCaaS/CPaaS and sales-engagement category names. Setting competitor_trackers replaces that list wholesale — send the full list you want active.

​

Reading the rollout

curl "https://api.orbit.devotel.io/api/v1/insights/sales-conversation-intelligence?days=30" \
  -H "X-API-Key: dv_live_sk_your_key_here"

{
  "data": {
    "from": "2026-09-11T00:00:00.000Z",
    "to": "2026-10-11T00:00:00.000Z",
    "calls_analyzed": 84,
    "avg_rep_talk_ratio": 0.42,
    "risk_band_distribution": { "low": 61, "medium": 17, "high": 6 },
    "top_competitor_mentions": [
      { "term": "AcmeVoice", "count": 9 }
    ],
    "at_risk_deals": [
      {
        "contact_id": "cnt_01JC2Z9X4T",
        "calls_count": 3,
        "avg_risk_score": 72,
        "latest_risk_band": "high",
        "latest_next_steps": ["Book a follow-up demo for the procurement team"],
        "latest_analyzed_at": "2026-10-10T14:02:00.000Z"
      }
    ]
  },
  "meta": { "request_id": "req_01HZQX4E7JQ4M2E2H2DM9XE5FJ", "timestamp": "2026-10-11T12:00:00.000Z" }
}

​

API surface

​

Which tool answers which question

​

Limits

• Voice calls only, and only completed calls that carry a transcript and speaker-diarised segments. Messaging conversations do not appear; the topic and sentiment surfaces cover those.
• Language coverage matches whatever languages your transcription is configured for — competitor matching and talk-ratio run over the produced transcript, so coverage follows your speech-to-text configuration rather than a fixed list.
• No minimum call length gate. A very short call scores on the evidence it has; expect a short call with no extracted commitments to read as riskier because of it.
• No extra retention window beyond the call record itself. The rollup reads calls within your chosen days window, and the per-call result lives on the call record for as long as the call data does — it inherits the call log’s retention, it does not add one.
• Volume bounds: each summary request scans at most the 5,000 most recently analyzed calls in the window; the at-risk list surfaces at most 25 contacts, and the competitor list at most 10 terms. Tracker config is capped at 50 terms of at most 80 characters each, and a call keeps at most 6 extracted next steps.
• Talk-ratio coverage: calls whose speakers could not be attributed (for example a diarisation with generic speaker labels) report no ratio and are excluded from the average — the surface never fabricates a split.

​

See also

• API reference — sales conversation intelligence endpoints — full request/response contract.
• Conversation Intelligence — the support-QA aggregate across channels.
• Topic Intelligence — cross-channel contact-reason mining.
• Insights dashboards — the catalog of every other Insights surface.