Skip to main content

​
Queue-level AI ops digest

Per-call AI summaries answer “what happened on this call.” The queue-level digest answers “what is this queue doing right now.” Open Voice → Call Queues → AI digest in the Devotel Orbit console to see, per queue, the sentiment mix, the most common topics, the most common action items, and the latest per-call summaries — rolled up from the summaries post-call synthesis already cached for every handled call. The same aggregate is available over the API at GET /voice/queues/digest, with a window of up to seven days.

​
What the digest does not do

The digest never calls the LLM. It aggregates the per-call summaries that post-call synthesis already produced at call wrap-up and cached, so reading the digest — from the dashboard or the API, at any cadence — does not re-bill the model. If you want it fresh, the answer is to summarise more calls, not to poll the digest harder. A call joins a queue’s rollup when two things hold:
  • it was routed through the queue, so the call log carries a queue tag, and
  • post-call synthesis completed for it (or an on-demand call summary was generated).
Calls without a queue tag appear in per-call surfaces but are not attributed to any queue here. Untagged queues keep their identifier in place of a name.

​
What the dashboard shows

Open Voice → Call Queues and switch to the AI digest tab. Each queue is one row over the selected window:
  • Sentiment meter — the positive / neutral / negative split across summarised calls, as a segmented bar with the share of each bucket below it. This is the queue-health headline: is the billing queue running hot, is support trending negative this afternoon.
  • Top topics — the most common conversation subjects across the queue’s calls, ranked by frequency. “Billing” climbing to the top of the support queue is a staffing or product signal you get without reading a transcript.
  • Top action items — the most common follow-ups the per-call summaries recorded, ranked the same way. If the same item repeats thirty times, it is a process gap, not thirty individual loose ends.
  • Latest summaries — the three most recent per-call summaries for the queue, so you can jump from a queue trend to specific calls.
The window selector covers the last 24 hours, 3 days, or 7 days. Use 24 hours for shift-level triage and 7 days for coaching and staffing decisions, where one bad batch of calls should not move the numbers. Queues with no summarised calls in the window show as empty — “no data” and “neutral” are different states, and the digest does not merge them.

​
Reading the digest

  1. Pick the window first. All breakdowns below belong to that window.
  2. Scan the sentiment meter per queue. A queue with a negative share that outruns its peers is where triage starts; a queue running uniformly negative is a product or routing problem, not an agent problem.
  3. Check top topics and action items. A topic mix you did not expect (“returns” on the sales queue) reroutes you to the right fix before you open a single transcript.
  4. Open the latest summaries. The three newest per-call summaries per queue are your bridge from trend down to evidence.

​
API access

The console’s digest tab reads this same endpoint:
Response shape (trimmed, one queue shown):
  • hours: window in hours, default 24, maximum 168 (7 days). Non-numeric or out-of-range values fall back to the default.
  • queues is sorted by summarised-call volume descending.
  • top_topics and top_action_items return at most the top 5 entries each, ties broken alphabetically.
  • sample carries at most the 3 most recent per-call summaries.
  • The endpoint is tenant-scoped, read-only, and answers over the cached summaries — safe to poll on the same cadence you poll other read endpoints.

​
See also