Trust· 5 min read
What a voice agent should never say
The hard part of a voice agent is not getting it to talk. It is making sure it only says what you approved.
By ZenVoice team
Anyone can connect a language model to a phone line. The demo sounds impressive for a minute. Then it promises a refund it cannot give, invents a date, or keeps talking after the customer has said no.
When an agent calls your customers in your name, every sentence is your brand. So we start from the other end: what must this call never do?
Script first, freedom second
For collections and anything compliance-heavy, calls run on a scripted flow: every sentence approved, a branch for every answer you expect, and anything unexpected flagged instead of improvised. For sales and service, a guided mode lets the agent answer questions inside guardrails you set. Most programmes mix the two.
Knowing when to hand over
- Disputes and hardship are recorded without pressure and flagged to your team.
- Emergencies and safety issues go to a person straight away.
- Anything outside policy becomes a callback, never a promise.
Honest when asked
You decide how the agent introduces itself and how it answers when someone asks whether they are talking to a person. Whatever you choose, it should never be evasive. Customers forgive a quick, useful call from an AI. They do not forgive being misled.
Checked after every call
Every call is transcribed and reviewed: was the outcome right, did the agent repeat itself, did it say anything it should not have. The review is how a programme gets better week after week, instead of drifting.
