Voice desk: calls on cases, callbacks, IVR tree and contact center reports
The voice desk puts the phone inside customer service.
The voice desk puts the phone inside customer service.
**Calls on a case.** In the case console, the "Calls on this case" box lists every call linked to the case, with direction, outcome, wait and talk time, the agent, the recording and the transcript when they exist. "Log a call" records a call that happened on any phone (a desk phone counts). Each call can be wrapped up with a disposition from the same catalog used to resolve cases (Settings > Wrap-up codes, scope "call" or "both"); codes marked with * require a note. "Summarize with AI" writes a short summary from the transcript and notes; it is only offered when there is text to summarize.
**Who is calling.** When a call comes into the softphone, the card under the number shows the matching contact or company, open cases, the last call and any pending callback. "Open a case for this call" creates a real case with the contact, the channel from your channel map and the transcript in the description.
**Your voice state.** The bar in the console shows Available, On a call, After-call work, Away or Offline. "On a call" is set by the call itself. After a call ends you get the after-call time defined in the policy; when it runs out you are available again without clicking. This state is mirrored to the inbound ring group, so the queue never rings someone who is busy.
**Callbacks.** "Request callback" creates a task with a due time. Task assignment rules (Settings > Routing) decide who calls: add a rule for tasks with type "call" to distribute callbacks round-robin or by least load; without rules the callback stays with whoever requested it. The callback queue is on the supervision page: take it, retry later, mark done or cancel.
**IVR tree.** In Settings > Telephony, the IVR menu is a tree: menu > key > destination (queue, person, submenu, message or offer a callback). Publish only when every option resolves; the preview lists what the current voice provider executes (key to queue) and what it does not yet execute.
**Call guide by queue.** Pick a queue, reuse a call script and add an opening checklist and steps.
**Reports.** The Analytics tab of the console and the supervision page show volume by hour and by queue, abandon rate, service level, average wait, talk and after-call time, dispositions, callbacks and per-agent handled calls, talk time and available time. Numbers without a base show as a dash, never as zero.
**Not included, on purpose.** Call transfer, supervisor listen-in and whisper, live transcription and waiting queues with hold music depend on carrier features that the current providers do not expose. The desk shows what exists instead of offering buttons that do nothing.
**After-call rules.** In Voice desk > Setup, "After-call rules" decides what happens when a call ends: summarize it, analyze what was said, open a case, request a callback, create a follow-up task or notify the agent. Conditions are written in the language of the operation (direction, outcome, queue, minimum talk time, only calls with a transcript, only calls without a case). An action that cannot run is not hidden: it is recorded as skipped with the reason, and you can read it on the call itself, under "What the after-call rules did". Rules run right after the call when the event bus is on, and in any case on the regular sweep, so you do not need the bus for this to work.
**Call quality.** Open a call in Voice desk and the quality box shows what the analysis found: the topics of the call, required phrases the agent did not say, forbidden terms the agent did say, the sentiment and a score from 0 to 100 with the factors that formed it. Two limits, stated on purpose: without a transcript nothing about what was said can be checked, so the call is marked "not checked" instead of scored down; and there is no talk ratio or silence detection, because the transcript we store has no speaker labels and no timestamps, and a number invented from it would look like a metric without being one. The sentiment is the only part that calls a model, and it only runs when the ticket AI policy of the account is on.
**What the analysis looks for.** In Setup, "What the analysis looks for" holds groups of terms in data. The type of the group says what finding it means: a topic classifies the call, a required phrase lowers the score when it is missing, a forbidden term lowers the score when it appears, and a competitor is a commercial signal that does not affect the score. A group can be limited to one queue.
**Quality reviews.** Anyone who manages settings can score a call from 0 to 100 and write feedback. The agent sees it in Voice desk under "Feedback I received" and marks it as read; the scorecard shows, by agent, how many reviews, the average and how many were never read. This is review after the fact, over the recording and the transcript. Listening in live and whispering to the agent during the call need real-time media from the carrier, which no connected provider offers.
**Policy by queue: recording, retention and overflow.** The line without a queue is the default for the account; each queue only states what it changes. Recording can be on for one queue and off for another. Retention removes the recording from this product after N days, and the transcript too if you ask; the file in your carrier account is only deleted when the provider offers that, which no connected provider does today, and the screen says so. Overflow decides what happens when nobody in the queue is available: ring another queue, and if there is nobody there either, save the caller's number as a callback and say so out loud. The message is only spoken when the callback was really created.
**Callback cadence.** A cadence says how long to wait after each unsuccessful attempt (for example 30 minutes, then 2 hours, then a day) and what happens when the attempts run out: close the callback, close it and open a case, or close it and notify. Nobody dials on their own; this is the clock of the callback queue. Without a cadence the behaviour is the old one: each attempt is rescheduled by the default callback time and never runs out.