Call your voice agent a "receptionist," not an "assistant.Because the word you choose pulls the model toward different patches of training data, and those patches sound very different on a phone line.
What "assistant" pulls in?
When you call the agent an "assistant," the model reaches into the part of its training where that word lives most heavily: chatbots, help desks, customer service software, virtual office tools. That register has specific habits:
- Slow and deferential. "I'd be happy to help you with that. May I please ask for your name?"
- Over-explaining. Long acknowledgements, unnecessary preambles, scripted enthusiasm.
- Help-desk pacing. Multi-sentence responses, step-by-step framing, "let me make sure I understand".
- Apologetic by default. Soft, hedged, careful. Appropriate for written chat, slow on a phone.
This register exists because chatbots historically had to compensate for their text-based, impersonal nature. They padded everything with politeness to feel less robotic. On a phone call, that padding is what makes them sound robotic.
What "receptionist" pulls in?
"Receptionist" lives in a different patch of training data, actual phone-handling roles. That register has different habits:
- Fast and transactional. "Hi, OptimizeSmart, this is Olivia — how can I help?"
- Efficient acknowledgments. "Got it." "Okay." "Sure."
- Phone pacing. Short sentences, one question per turn, brief verbal nods between.
- Friendly but not effusive. Warm without being saccharine.
A receptionist's job is to move callers efficiently, greet, identify the need, route to the right person, and get them off the line. Every second matters. The training data for that role reflects it.
If you switch an agent from "assistant" to "receptionist" without changing anything else, you'll typically see:
- 15–25% shorter average turn length.
- More natural use of contractions (assistants over-formalise).
- Fewer scripted-enthusiasm phrases ("Absolutely!", "I'd be delighted to").
- Faster opens and closes.
- Less hedging language ("I think", "perhaps", "if you don't mind").
It's one of the cheapest changes you can make to a voice agent prompt and one of the most consistently impactful. Same instructions, same flow, same model, different word, materially different conversation.
The same principle applies to other roles.
Whatever your agent actually does, name the real-world role that does it:
- Front-desk agent → "receptionist".
- Booking agent → "scheduler" or "booking agent".
- Sales qualifier → "intake coordinator" or "qualifier".
- Support triage → "dispatcher" not "support assistant".
- Survey caller → "interviewer".
Each of these has its own training-data patch with its own pacing, vocabulary, and rhythm. You get those for free if you name the role correctly. You have to fight against them if you don't.
The labels you give the agent shape its behaviour more than the instructions you give it.
Identity statements are constitutive in a way that descriptive instructions aren't.
Calling the agent a receptionist makes it act like one; telling a self-described assistant to "act like a receptionist" produces an assistant pretending to be a receptionist, which sounds exactly like that.The same logic applies to gender ("female receptionist" vs unspecified, pulls different cadence and vocal patterns), formality ("front-desk" vs "executive assistant"), and seniority ("the receptionist" vs "our senior account manager").
Each label is assigned to a different region of the model's training data. Pick the one whose region you actually want to draw from.