When I build an agent that talks to customers, I start with what it should say when it doesn't know something. Then I write the rest.
It sounds backwards, but it forces the important decisions early:
- What topics is it allowed to answer? (hours, prices, booking)
- What does it do with everything else? ("I'll have someone call you back today. What's the best number?")
- How does it avoid making things up? Prices and availability come from a lookup, never from the model's memory.
Most embarrassing voice agent moments come from an agent that tried to be helpful outside its lane: a guessed price, an invented opening time, a promise the business can't keep.
A short, honest "I don't know, but here's what happens next" builds more trust than a clever answer.
What's your agent's fallback line when it can't answer?