What an AI voice agent actually does on a call

"AI voice agent" gets used for four different things: a recorded message, an IVR menu, a text-to-speech bot reading a script, and a system that actually holds a conversation. Only the last one is worth paying for, so it is worth being precise about what happens between the ring and the hang-up.

The loop, in order

1. The line rings. A number you own — on a carrier account, a SIP trunk, or a gateway — is pointed at the agent instead of at a queue.
2. The agent speaks first. Its opening line is fixed and written by you. This is the single highest-leverage sentence in the whole system, because it sets what the caller thinks they are allowed to say next.
3. The caller replies, in words. Speech becomes text.
4. A model decides what to do. It has your prompt, the conversation so far, and whatever you attached as knowledge.
5. It may call a function. Look up an order, check availability, create a contact, hold a slot. This is the step that separates a demo from something useful.
6. It answers, in speech. With a voice you chose, per language.
7. It ends, or it hands over. Either the job is done, or a condition you named fires and a human takes it.

Steps 2, 4, 5 and 7 are yours. Steps 3 and 6 are infrastructure. Most of the difference between a good deployment and an embarrassing one lives in the parts you control.

What it can genuinely do

• Answer at 2am. No shift, no queue, no voicemail nobody checks.
• Qualify. Ask your four questions, write the answers onto the contact record, move the lead stage.
• Book. Read real availability and hold a real slot — provided it is reading a calendar and not guessing.
• Look things up. Order status, opening hours, whether you cover a postcode.
• Take the boring outbound attempts. The second and third calls nobody on your team has time to make.
• Log itself. Every call lands on the contact's timeline with an outcome, so the next person is not starting cold.

What it cannot do, and you should stop hoping

• Know things you never told it. If your pricing lives in someone's head, the agent will invent something or refuse. Both are bad; the refusal is less bad.
• Handle a caller who is furious. It can be polite. It cannot make the situation fair. Name that as a handover condition.
• Be certain. The output is probabilistic. Test it before pointing a live number at it, and keep the guardrails narrow at the start.
• Replace your team. The realistic outcome is that the same team stops losing the calls it was never going to get to.

The useful metric is not "calls handled by AI". It is "calls that ended without a human and without a complaint".

The four decisions that actually matter

Ignore the feature lists for a moment. When you set one of these up you are making four choices, and everything else is detail.

Which line does it answer?

A dedicated after-hours number is the low-risk start. Your main line is the high-risk one.

What is its job, in one paragraph?

If you cannot write it in a paragraph, the agent cannot do it either. More on writing that paragraph.

What may it do, not just say?

The functions you enable. An agent that can only talk is a very expensive FAQ.

When does it stop?

Name the conditions in advance. Handover deserves its own thought.

Where the cost goes

Three meters run at once: the carrier's per-minute charge for the call, the speech processing, and the model. Vendors bundle these differently, which is why per-minute comparisons rarely compare the same thing. Ask what a ninety-second call costs all-in, and whether that number changes when the caller speaks a second language.

In Convarza AI, calling minutes are part of the plan — 600 a month on Starter through 20,000 on Scaler — and the carrier relationship stays yours, so you can see both halves. The plans are here.

A sane first month

1. Point the agent at the calls you are currently losing — after hours, or overflow when everyone is busy. The bar is "better than voicemail", which is low.
2. Give it one job. Not three.
3. Read the first fifty transcripts yourself. You will learn more about how your customers speak than from any workshop.
4. Only then widen the scope, or move it to a line that matters.

See how the pieces fit on the AI voice agents page, or bring a number to a demo and ring it yourself.