AI voice agent vs IVR: what actually changes

The difference is not the voice

It is tempting to describe the difference as "the IVR sounds robotic and the AI agent sounds natural". That is true and it is the least of it, because an IVR with a well recorded voice track sounds better than plenty of synthetic agents.

The structural difference is elsewhere: an IVR is a tree, and an AI agent is an interpreter.

In an IVR, someone drew every possible route before the first call ever rang. Each branch exists because a person thought of it. When a caller says something that is not in the tree — "you called me a while ago and I do not know why" — the IVR has nowhere to go. It repeats the menu, and by the third repetition the caller either hangs up or hammers the zero key.

An AI voice agent does not walk a tree: it takes what it was told, decides what to do, and acts. The routes are not drawn in advance. That is what lets it absorb the unplanned, and it is also what makes its failure mode completely different.

The failure modes differ, and that changes how you supervise

An IVR fails in a boring, visible way: people get stuck at the same point in the menu, and that point shows up in the report. You fix it by redrawing the branch.

An AI agent fails in an interesting, invisible way: it answers something untrue with confidence, or reads a denial as a confirmation. The call does not get stuck anywhere, so it never appears in an abandonment or repeat-call report. It ends well, with the wrong note written into the system.

From that comes the most useful operating rule we know: with an IVR you supervise where people get stuck; with an AI agent you supervise a sample of conversations that ended well. If you only listen to the ones that went badly, you will never find the failure that matters.

Latency: where the bar actually sits

Here is a number almost nobody uses and everybody should. In human conversation, the gap between one speaker finishing and the next starting has a median around 200 milliseconds, and it is remarkably similar across very different languages. It is not a cultural preference: it is the cadence we are calibrated to.

That sets the bar. A voice agent that answers in 300 ms reads as someone paying attention. One that answers in 900 ms reads as someone who was not listening, however good the answer is. And the caller starts doing what we all do in the face of silence: repeating themselves, or talking over the top.

The latency budget is split across recognition, decision, synthesis and network. The network is not free: the classic telephony recommendation is to keep one-way delay under 150 ms for a conversation to feel normal, and that is spent before the model has thought about anything.

The practical consequence: if the agent feels slow, measure it by segment before swapping the model. The expensive segment is rarely the one you assume.

Where the bot opens and where the person steps in

The operational question is not "bot or person", it is when the turn changes. Three criteria that work:

1. Stated intent. Anyone who asks to speak to a person, speaks to a person. No toll, no "tell me the reason first". It is the cheapest rule and the one that prevents the most complaints.
2. The agent's own confidence. If the agent is not sure it understood, it passes the call. An agent that guesses rather than interrupt is the one that produces the invisible failure described above.
3. What is at stake. Confirming an appointment and closing a payment negotiation are not the same conversation, even if the model could hold both.

What does not work is a turn counter. "After five exchanges, hand off to a human" cuts short the conversations that were going badly and the ones that were going well alike.

When an IVR is still the right answer

Worth saying, because almost nobody says it: there are cases where the tree wins.

• When the option set really is closed and short. Three options that never change do not need an interpreter.
• When volume is huge and the task is identical. A balance, an opening time. The cost per call of a tree is hard to beat.
• When the wording has to be exactly the same every time. A recorded legal notice does not tolerate paraphrase, and a model paraphrases.

In PulseDial the two run side by side on the same campaign, and the handover is configured by the three criteria above rather than by a turn counter, which is what we saw fail most often.