A few years ago, putting a voice agent in front of real customers meant a speech team, a telephony stack, and months of tuning. That is no longer the shape of the problem. The models handle the conversation; what is left is the part that was always yours anyway — deciding what the agent is for.
Start with one job, not a personality
The instinct is to describe a character: friendly, professional, helpful. That produces an agent that is pleasant and useless. Describe a job instead. Qualify inbound leads against four criteria. Book a callback when the caller is a fit. Hand off to a human when they ask for pricing. A narrow job is testable; a personality is not.
Give it the context it would otherwise guess
Most bad agent behaviour is not the model inventing things — it is the model filling a gap you left. Your hours, your service area, the three questions people actually ask, what you do not sell. Written down, this is usually a page. Left out, it becomes improvisation.
Connect a number and listen
The first ten real calls will teach you more than an afternoon of prompt editing. You will hear where the agent talks past people, where it misses an obvious follow-up, and where it is stiff. Fix those, then listen again.
Know where it should stop
The agent that knows when to hand off is worth more than the one that tries to close. Decide in advance which turns end with a human, and make that path fast. Callers forgive a transfer; they do not forgive a loop.
None of this takes months. It takes an afternoon and the willingness to listen to your own calls.