What an AI phone agent does on a call, what it can finish itself, and how it differs from an IVR or a bolt-on voicebot.
A Karmaflow.ai voice agent is an AI that holds a real phone conversation. It answers the number you point at it, or it dials out when a workflow tells it to. During the call it looks the caller up in your CRM, reads your published knowledge, books and updates records, sends the follow-up text or email, and hands the call to a person when a person is what the caller needs.
Every call lands in your workspace afterwards as a transcript, an AI summary, an outcome, the list of tools the agent actually used and the knowledge it actually cited — attached to the contact, so the next conversation on any channel starts from what this one established.
flowchart LR
subgraph In["Inbound"]
A[Caller dials<br/>your number] --> B
end
subgraph Out["Outbound"]
A2[Orchestration,<br/>API or a person] --> B
end
B[Voice agent] --> C[Live CRM lookup<br/>by caller ID]
B --> D[Tenant Brain<br/>published knowledge]
B --> E[Tools: book, update,<br/>text, email, transfer]
B --> F[Transcript · summary ·<br/>outcome · receipts]
F --> G[Contact timeline<br/>+ every other agent]
Answering the phone. Point a provisioned number at an inbound agent and it picks up. It greets the caller in the words you wrote, identifies them from caller ID against your contacts, and works the conversation against your instructions. If the caller reached you through a PBX such as 3CX, the agent can receive the IVR selections the caller already made and skip asking again.
Calling out. Outbound agents are dialled by an orchestration step, by the outbound API, by a helpdesk journey milestone, or by a person from the agent's own Use Voice Agent page. Outbound calls detect voicemail and act on your instructions when they hit one, and they can dial an extension after the callee answers.
Talking to a person. A warm transfer puts the caller on hold, briefs the human on what the call is about, and then bridges the two. A cold transfer forwards immediately. You give the agent a list of forwarding numbers with a condition on each, and it chooses; turn on the duty roster and whoever is actually on call today is tried first.
Doing the work, not just describing it. Tools run for real during the call: creating and updating contacts, tasks and deals, creating and replying to helpdesk tickets, searching your document stores and your Tenant Brain, sending the SMS or email the caller asked for, pressing keypad tones to get through an IVR, starting an on-screen timer, recording an opt-out.
Speaking like your business. 52 voices, including ten that accept a written speaking style and a set tuned for emotional range, plus 30 Gemini voices that each speak every language where the Gemini beta is enabled. Multilingual agents carry a different voice per language and can switch mid-call — either because the caller asked, or because language auto-detection heard them change. A pronunciation pack teaches the agent how to say your product names, and you can add your own entries (paused on calls while we test them with the latest voices; see Voices, Greetings and Pronunciation).
Being tunable. Turn-taking, interruption sensitivity and the silence-before-hang-up timers are all yours to set, with three presets (Conversational, Fast-paced, Patient) for the common shapes. Every stored value is shown back to you as the value the runtime will actually use.
| What you want | How a voice agent gets there |
|---|---|
| Stop losing after-hours and overflow calls | An inbound agent on the main line or the overflow line, answering every time |
| Shorter handle time on routine calls | The agent completes bookings, lookups and updates itself; only the exceptions reach a person |
| Follow-up that actually happens | An orchestration places the call, and the agent texts or emails the summary before hanging up |
| Fewer bad transfers | Conditions per forwarding number, plus the duty roster, plus a spoken briefing |
| Compliance you can show someone | Per-agent transcript retention, recording rules, opt-out capture, and an audit trail of what the agent was told |
| Spend that cannot run away | A daily outbound call cap at the workspace level, and per-workspace billing limits |
| Phone tree / IVR | Standalone voicebot | Karmaflow.ai voice agent | |
|---|---|---|---|
| Understands an unscripted sentence | No | Usually | Yes |
| Knows who is calling | Caller ID digits only | Rarely | Resolves to the CRM contact and their history |
| Can finish the task | No — it routes | Only what its own integration covers | Runs your CRM, helpdesk, knowledge, calendar and connected tools |
| Shares memory with your other channels | No | No | Yes — one contact record across voice, chat, SMS and email |
| Briefs the human it transfers to | No | No | Yes, on a warm transfer |
| Cites where an answer came from | n/a | No | Yes — the Knowledge used panel on the call |
| Governed | n/a | Vendor's terms | Your prompt components, your retention policy, your approval gates |
The distinction that matters most: a voice agent here is not a separate product with its own copy of your data. It is one more channel on the same workspace your CRM, helpdesk and knowledge already live in — which is why the call it takes today makes the email your mail agent sends tomorrow better.
Start here:
Then the pieces you will tune:
Related: Duty Schedule · Orchestrations · Tenant Brain · CRM