How an AI behaves in a call
Kol is push-to-talk. Designing for it is different from designing a chat bot.
One message at a time
- The person holds the talk button and speaks. Your AI hears them live.
- They let go. That's the end of their message: there's no "are you done?" guessing.
- Your AI answers out loud, usually starting about a second later.
- If they press the button again while it's talking, it stops and listens.
So write for the ear:
- Short replies. One to three sentences, unless asked for more. Long answers are hard to follow and the person can't skim them.
- No lists, links, markdown or emojis. Nobody can see them.
- One question at a time when you need information.
- Say numbers the way people say them ("about ten grams", "half past six").
Kol's rules already ask for this; your persona should go in the same direction.
Language
The AI answers in the language it's spoken to. A Spanish speaker calling your English-language service gets Spanish answers (your knowledge can stay in one language).
Everyone can see it's an AI, and anyone can mute it
Every AI in a Kol call announces itself. On every phone in the call:
- it's drawn in a distinct (amber) color, with an "is listening" label while it can hear;
- anyone in the call can tap that label to mute it. A muted AI hears nothing: the audio is dropped before it reaches the AI model, and it doesn't speak. Anyone can unmute it again;
- in a group call, anyone can remove it from the call.
You can't turn this off, and your persona can't change it. People always know when an AI can hear them.
It always says it's an AI
If anyone asks whether it's a person, it says plainly that it's an AI. This rule comes after your persona and wins. Don't write personas that pretend to be human; see Safety.
Length and cost
Each call is metered in minutes (kol status). Calls where your AI answers for its owner are limited to a few minutes. The AI model's cost is paid by whoever's AI key runs it: the person who got your companion, or you for a service.