​

Topic Intelligence

​

What a “topic” is

• Upstream intent classification. When the conversation-intelligence pipeline has already classified the thread (the intent label), that label wins.
• Lexicon fallback. When no upstream label exists, the service falls back to a built-in contact-reason lexicon over the conversation’s summary or last message, so the surface works from day one — billing, refund, cancellation, technical issue, account access, delivery, sales inquiry, complaint, appointment, product support — with other as the catch-all bucket.

​

How topics are computed

• Current vs prior window. The comparison window is split in half: the current span and an equal-length prior span immediately before it (a 30d read compares the last 30 days against the 30 days before that). Trend is always relative to the prior span.
• Per-channel split. Each topic carries its volume broken down by channel, so you can see that “delivery” is voice-heavy while “account access” is web-chat-heavy.
• Trend direction. A topic absent from the prior window is marked new; otherwise, a swing of at least 15% up or down marks it rising or falling, and anything inside that band stays flat.

​

Metrics the surface reports

• Volume and trend. Current-window conversation count, the prior-window count, the percentage change (or new when the prior window was empty), and the trend label — rising, falling, flat, or new.
• CSAT. Mean CSAT over current-window conversations that carried a score (1–5). null when none did, so an unrated topic reads as blank rather than zero.
• Deflection rate. The fraction of resolved conversations that closed with no human assignee — self-service resolution. A conversation still open counts toward volume but not toward deflection.
• Average handle time (AHT). Mean resolution time in seconds across current-window conversations with a recorded value. Conversations the idle-conversation sweep force-closed without being worked are excluded rather than counted, so a conversation that sat untouched for a week does not report that week as handle time.

​

Reading flow

1. Pick the window. The selector offers 24h, 7d, 30d (default), and 90d. Use 24h for incident triage and 30d or 90d for staffing and routing decisions where one bad afternoon should not move the number.
2. Optionally filter one channel. The channel filter narrows the corpus to a single surface — voice, SMS, WhatsApp, email, or web chat — when a portfolio-wide spike needs localizing.
3. Read the ranked list. Topics appear highest-volume first, so the top rows answer “what is consuming our inbound load.” Scan the trend column, not just the volume: a small topic marked rising is the early-warning signal a large flat topic hides.
4. Drill into a topic. Each topic links up to five recent example conversations (id, channel, timestamp) so you can jump from a spiking reason to the raw threads behind it. Use the examples to check tag quality before acting on the trend.

​

Decisions it feeds

• Route the escalation team. A topic marked rising with high volume and low deflection is the candidate for a staffing adjustment or a dedicated queue before it saturates.
• Fix AI containment. High volume + low deflection means your flow or AI agent is escalating what it should have resolved — that is the containment fix, not a routing fix.
• Rebalance human handle time. High handle time with high deflection is expensive: the flow resolved it without a human, but slowly — a cheaper automation target.
• Audit the tagging. If the top topic is other or a topic you do not recognize, drill into the linked examples. Mis-tagged or catch-all-dominated buckets mean the contact-reason taxonomy needs attention, not that customers are confused.

​

API contract

curl "https://api.orbit.devotel.io/api/v1/analytics/topic-intelligence?window=30d" \
  -H "X-API-Key: $ORBIT_API_KEY"

{
  "window": "30d",
  "channel": null,
  "current_since": "2026-08-01T00:00:00.000Z",
  "previous_since": "2026-07-02T00:00:00.000Z",
  "total_conversations": 4120,
  "topics": [
    {
      "topic": "delivery",
      "volume": 842,
      "previous_volume": 610,
      "change_pct": 38.0,
      "trend": "rising",
      "channels": { "voice": 270, "sms": 180, "whatsapp": 240, "email": 120, "web_chat": 32 },
      "avg_csat": 2.8,
      "deflection_rate": 0.62,
      "avg_aht_seconds": 214.5,
      "examples": [
        { "id": "conv_8f04c2", "channel": "whatsapp", "occurred_at": "2026-08-30T14:05:00.000Z" }
      ]
    }
  ]
}

​

See also

• Contact reasons — the text-inbox reason classifier that Topic Intelligence generalizes cross-channel.
• Sentiment portfolio dashboard — the how customers feel counterpart to what they contact you about.
• Insights dashboards — the catalog of every other Insights surface.