Skip to main content
The voice agent answers your phone number, holds a real conversation in Arabic or English, and does the same work the chat agent does: qualify, answer, book, escalate, write to your CRM.

​
Numbers

A workspace can hold several numbers rather than one. Each number carries:
  • A role. Inbound, outbound, or both. A number customers call does not have to be the number campaigns dial from.
  • A label, so your team can tell them apart.
  • A concurrency limit, so one number cannot be saturated by a burst.
Numbers are provisioned from our pool or ported from your existing carrier. Porting takes longer than provisioning, so most workspaces launch on a pool number and port later. Outbound calls pick their caller ID from the pool by strategy rather than at random, so a customer in a given city sees a number that makes sense to them.

​
On a call

1

Answer and identify

The agent greets in your default language and switches if the caller speaks the other one. If the number is known, the caller’s history is already loaded. See Memory and state.
2

Understand and qualify

The agent asks the questions your vertical pack and your own configuration say matter, and stops asking once it has them.
3

Answer from your knowledge

Questions about services, prices, availability and policies are answered from your knowledge base and your catalog, not from general knowledge.
4

Book

Availability is checked against your calendar before a slot is offered. The agent cannot promise a time that is not free.
5

Escalate or close

Out of scope, a complaint, or anything your handoff rules flag goes to a human. Otherwise the call ends and the conversation closes itself.

​
Latency

On a healthy call the agent replies in under a second on a normal turn and under two seconds on a turn that has to call a tool, measured from the end of the caller’s speech to the start of the substantive answer. Turns that need a tool speak while the tool runs rather than going quiet, so the caller does not experience the gap as dead air.
If you benchmark this yourself, be explicit about what you are timing. Mouth-to-ear includes the endpointing window that decides the caller has actually stopped speaking, which is a separate measurement from the one above.

​
Speech behaviour

  • Bilingual mid-call. The agent answers in the language the caller used, including switching part way through. See Languages.
  • Spoken formatting. Replies are written to be heard. Prices, dates and phone numbers are phrased the way a person says them.
  • Interruption. A caller talking over the agent stops it.
  • Email readback. When the agent captures an email address it reads it back before using it, because transcription of addresses is the single most error-prone thing on a call.

​
Recording and transcripts

Calls can be recorded per workspace. Every call produces a transcript in the inbox regardless of whether audio was retained. Transcripts are retained on the schedule in Data retention and mirrored to your CRM.
Call recording carries consent obligations that vary by jurisdiction. Configure your greeting to disclose recording where you are required to.

​
Transfer to a human

When a handoff rule fires, the agent can hand the live call to a person, and it tells the caller what is happening rather than transferring silently. If nobody picks up, the conversation stays in the inbox as an escalation and your alerting rules take over. See Routing and escalation.

​
Outbound

Outbound calling is a separate surface with its own gates: consent, quiet hours, frequency caps and a do-not-call list, all of which fail closed. See Outbound.

​
Call events

Voice emits call.started, call.completed and call.failed. Completed and failed calls can drive trigger rules, so “message the customer when their call rings out” is a rule rather than a custom integration. Call events are also available as webhooks.