Voice agent best practices

Hard-earned guidance for building voice agents that customers trust and your team is willing to defend. Part of the Voice agents section of the TrainYourAgent documentation, written to be executed against rather than skimmed.

Start with the opening line, not the system prompt

The first six seconds of every call decide whether the customer believes the agent. Spend a disproportionate amount of time on the opening — the greeting, the disclosure if your industry requires one, the question that invites the caller to state their reason for calling. If the opening lands, you have earned the next thirty seconds. If it does not, the caller is already mentally hunting for the operator key. A good opening is short…

Pick a voice that is one notch less perfect than the best on offer

The current best TTS voices are technically incredible and sound slightly too good to be true. Customers notice. We typically pick the second-best voice in any vendor's lineup — the one with a small amount of natural breathiness, a hint of regional accent, and a slightly imperfect cadence. It pulls the agent out of the uncanny valley and into the "competent receptionist" zone. Match the voice to the brand. A premium home services company should…

Disclose AI when your industry requires it

Some industries — healthcare, legal, insurance, finance, anything with a fiduciary or regulated relationship — require you to disclose that the caller is speaking with an automated system. Get a clear answer from your counsel on what is required for your jurisdiction and your industry, then bake the disclosure into the opening line. Done well, the disclosure does not hurt conversion. Done as an afterthought, it sounds defensive.

Write prompts in production English, not prompt-engineering jargon

The system prompt should read like an onboarding doc you would hand to a new receptionist. State the role, the goal, the boundaries, the escalation rules, and a small number of concrete examples. Avoid prompt-engineering ceremony like "you are a world-class expert" or "let's think step by step". The current generation of models does not need that and the extra tokens add latency.

Section context

This article is part of the Voice agents section. Hard-earned guidance for building voice agents that customers trust and your team is willing to defend.