How AI phone calls work

The four steps

• Listening. Speech recognition turns what the person says into text as they speak.
• Knowing when you’ve finished. A turn-taking model decides whether a pause is the end of your sentence or you’re just thinking.
• Thinking. An AI model writes the reply, following the business’s approved facts and the rules for the call.
• Speaking. A text-to-speech voice reads the reply out, starting as soon as the first words are ready.

Why the gap matters

In a normal conversation, people reply about half a second after the other person stops. Leave two or three seconds and it feels robotic: the caller says “Hello?”, or starts talking again just as the AI does. Getting the gap down is the single biggest thing that makes an AI call feel natural.

What we changed to get it under a second

• We measured each step. Most of the delay was the AI model writing its first sentence, not the voice.
• We moved to a faster model with its thinking switched off for calls. The first sentence was ready in about 0.4 seconds instead of more than 1.
• The voice starts speaking as soon as the first sentence is ready, not when the whole reply is written.
• Turns are short by design: one thing per turn, then stop and listen.

What else makes a call sound human

• Waiting for the person to say hello before speaking.
• Ending each turn with a question or a clear hand-over, so there’s no awkward silence.
• Reacting to what was said, in the person’s own words, before moving on.
• A natural local voice. Ora uses a New Zealand English voice by default.

You can read a whole Ora demo call word for word, or listen to it at the top of the Ora page.