How to design a voice AI agent's script: tone, limits and opening lines

Same clinic, same synthetic voice, same calendar integration. On a Tuesday morning, two appointment confirmation calls go out. In the 9:04 call, the agent opens with a thirty-eight-second paragraph: who it is, the centre's opening hours, the cancellation policy, the recording notice. The customer interrupts at second twelve; the agent finishes its paragraph anyway; the customer repeats the question; at three minutes and forty seconds the call dies in an "I'll just call back myself". In the 9:11 call, the agent opens in seven seconds, asks one question, confirms the answer and hangs up. Seventy-four seconds, appointment confirmed.

The only difference between the two calls is the script. And the word is misleading: a voice agent's script is not the text it recites, but the set of decisions about how it converses — the tone, the length of each turn, what it does when interrupted, which lines are fixed word for word and what is left to the model's judgment under rules. The agent's voice is the company's voice.

This article walks through those decisions one by one, from tone to testing. We assume you have already decided where the agent ends and your team begins — that limit comes before the script, and we make that argument there; what comes now is how that limit sounds in practice.

A script is not a monologue: it is a specification

The intuition inherited from classic IVRs is to write the script as a tree: if they say A, answer B. We have already explained why that approach breaks: customers step off the expected path almost every time, because they mix topic, urgency and context in the same sentence. With a language model behind it, the script changes nature: it stops being a text to be recited and becomes a specification to be met.

That specification has three layers. Identity and tone: who the agent is, how it speaks, which vocabulary it uses and which is forbidden. Conversation rules: how long a turn lasts, how it handles interruptions and silences, how it rephrases when it doesn't understand, when it hands the case over. Fixed lines: a handful of sentences written word for word that the agent always delivers the same way. The most useful design question is what goes in the third layer and what stays in the first two.

The criterion for the left column is simple: whatever carries legal or brand risk gets fixed literally. The opening, because it is the channel's first impression; privacy and opt-out, because they are commitments, not conversation — we covered the opt-out when designing appointment reminder calls, and here all that matters is that it is a fixed string, not an improvisation —; the handoff and the closing, because they are the moments the customer remembers. Everything else is rules, not paragraphs. The script doesn't tell the agent what to say at every moment; it tells it who it is and what it can never say.

Tone is a business decision, not a technical setting

The first decision in the first layer looks minor and is not: does the agent address customers formally or informally? "Good morning, I am calling from the clinic with regard to your appointment tomorrow" and "Hi, I'm calling from the clinic about your appointment tomorrow" carry the same information and describe two different companies — in Spanish, the tú/usted choice makes the split even sharper. A dental clinic is naturally informal; a firm handling inheritances probably is not. There is no universal answer — there is consistency with how your company already speaks on the phone —, but there is a universal mistake: not deciding, and letting the model choose on every call.

The decision is written down, not assumed. And it doesn't stop at the form of address: tone includes the register — warm, neutral, formal — and two short vocabulary lists. The words of your business the agent should use, and the forbidden words: internal jargon, tool codes, back-office anglicisms no customer has ever said. If your team says "incident" and your customers say "problem", the agent says "problem".

Sentence length is tone too. Long, subordinate-laden sentences sound like a recorded announcement; short ones sound like a person taking the call. Tone is not how the agent sounds: it is how your company sounds when you are not in the room.

The opening lines: ten seconds that decide everything else

What the opening must contain we already wrote when discussing how to introduce the agent and what the law requires: what it is, on whose behalf, what it can solve and how to reach a person, all four pieces in the first sentence. What was left to say is the constraint that makes it hard to deliver: the time budget.

The full opening must fit in about eight seconds — twenty-five to thirty words —, with a single idea and a single question. The 9:04 agent from the introduction wasn't failing by hiding information, but by giving all of it at the door: opening hours, policies and recording notices are things you say when they become relevant, not before the customer has had a chance to speak. A well-budgeted outbound opening sounds like this: "I'm the virtual assistant for Clínica Ribera. I'm calling about your appointment tomorrow at 10:30. Is now a good time to confirm it?".

The budget changes with context. An inbound call during business hours allows a minimal opening; an outbound call needs to justify in the first sentence why the phone is ringing; the after-hours call has its own expectation rules, which we detailed in after-hours calls. All three variants share one property: they are the fixed lines par excellence. They are written literally, read aloud and timed.

One question per turn: pace, interruptions and silences

The most common mistake when writing a script is writing it. On paper, forty words are a reasonable paragraph; on the phone they are fifteen seconds of uninterrupted audio, and the average customer starts talking over it long before that. The operating rule: turns of one or two sentences, around twenty words, and a single question per turn. Ask two, and the customer answers one, the other is lost, and the agent ends up asking again.

Interruptions are not a customer failure: they are the natural way of talking on the phone. When the customer talks over it, the agent goes quiet and listens; an agent that finishes its paragraph while being spoken to turns the call into two simultaneous monologues. Hence a design consequence that isn't intuitive: the important part goes at the beginning of every sentence, because the end may never be heard.

Silence is scripted too. On the first silence — about four seconds —, a soft re-engagement that brings the context back: "Are you still there? I was asking about your appointment tomorrow". On the second, a closed question that lowers the effort of answering: "Shall I confirm it, then?". On the third, close gracefully and leave the door open: "It seems I can't hear you well. We'll call you another time, no problem". A script that only anticipates answers is half a script.

Rephrasing without repeating: the script for misunderstandings

There will be turns the agent doesn't understand. The golden rule is never to repeat the same sentence word for word: literal repetition is what gives machines away and what exasperates people.

Instead, a three-rung ladder. First, rephrase with different words: "Sorry, I didn't quite understand. Are you calling about an appointment or an invoice?". Second, narrow down to closed options: "Let's make it easy: is this about tomorrow's appointment, yes or no?". Third, stop trying: the third comprehension failure is exactly one of those triggers of the limit you decided before writing the script, and there the conversation gives way to the handoff with everything captured so far.

One wording nuance that changes the tone of the whole ladder: the fault always lies with the agent, never with the customer. It says "I didn't understand you", never "you didn't say it clearly" or "I can't hear you". And the conceptual takeaway is easy to remember: the third rephrasing does not exist; the handoff does.

The handoff line and the plan B when nobody is available

The handoff to a person has its own fixed line, and it earns that status because it does three things in one second: it announces the change, sets an expectation and removes the fear of repeating everything. "I'll put you through to a colleague on the team. She already has your details, you won't have to tell it all again". What information travels with the customer is not this article's business: it is the context package we describe in detail in the agent-to-person handoff design. The script only puts a voice to that moment — but if the voice promises "she already has your details", the package has to make it true.

That leaves the uncomfortable case: the customer wants a person and nobody is available. The rule we inherit from inbound flows: never promise what won't be kept. No "one moment, please" followed by endless hold music, no "someone will call you right now" if nobody will call right now. The honest alternative is concrete and comes with a time window: "I can't put you through to anyone on the team right now. I can ask them to call you tomorrow before noon, does that work for you?". The handoff line is the most important one in the script, because it is the one spoken when the script is no longer enough.

One script per language, not a translation

If you serve customers in Spanish, Catalan, English or Portuguese, you don't have one script: you have four. The part that resists translation the most is precisely the first layer, the tone. Spanish tú/usted and Catalan tu/vostè are separate decisions — the same clinic can be informal in Spanish and prefer vostè with part of its Catalan-speaking clientele —, and English flattens the distinction entirely: the register has to be rebuilt with other tools.

The fixed lines are written in each language and reviewed by a native speaker on the team; translating them on the fly is where the phrasings no customer would ever say come from. And the language switch itself is scripted: if the customer answers in Catalan to a Spanish opening, the agent neither ignores it nor turns it into a problem — "Si prefereixes que parlem en català, cap problema" — and continues the conversation in the customer's language.

What is shared across languages is the specification: the same intents, the same limits, the same rephrasing rungs, the same silence rules. The surface changes; the structure does not. The script gets localised; the limit does not.

The metrics that matter

The script is tuned with numbers, not impressions, and it has its own. The misunderstanding rate: what percentage of turns ends in a request to repeat or rephrase. It is the script's direct thermometer; if it rises after a change, the change gets reverted. Interruptions per call, especially the ones arriving in the first third of the agent's turn: many early interruptions mean turns that are too long or information in the wrong order. The agent's average turn length, measured in seconds per intervention rather than per call: it is the monologue detector, and it is compared against the budget of about twenty words. And the success rate per intent: what percentage of calls completes its intent's objective without a handoff or an abandonment, segmented intent by intent, because it says which part of the script is failing and not just that something is.

The four are read together. A high success rate with long turns describes a script that solves but tires: it will work until it doesn't. Many interruptions with few misunderstandings mean the customer is running ahead of the script: there is text to spare, not comprehension to gain. And early hang-ups in the first seconds are deliberately not on this list: they are a signal about the introduction, the territory of the transparency article, not of the script's body.

How BeeAgent fits in

In BeeAgent the script is not a feature added on top of the agent: it is part of its design. Tone and form of address are defined per agent and per language; the fixed lines — opening, privacy, opt-out, handoff, closing — are written literally and kept apart from the instructions; and the rules for rephrasing, silence and handing over to a person are configured as what they are, rules, connected to the escalation limits the script puts into words.

And because it is a no-code platform, adjusting the script is not a project: the operations team reads the morning's transcripts and changes an opening or shortens a turn in minutes, without a ticket to development; the full process is in the guide to building your first agent without writing code. One honest warning, though: if it is not yet clear what the agent is supposed to solve, the script won't fix it. First the process, then the words.

Your script in three weeks

The first week is for decisions and writing. You decide the tone and the form of address per language, write the fixed lines and the list of intents with their rules, and run the cheapest, most profitable test there is: the read-aloud. One person on the team reads the agent's lines to another playing the customer, stopwatch in hand; anything that sounds like a recorded announcement gets shortened.

The second week is for breaking the script on purpose. Fifteen or twenty internal test calls in which the team interrupts, mumbles, changes topic mid-sentence and goes silent, and one session with the transcripts on the table to adjust turns, rephrasings and silences. It is cheaper for the agent to meet your team in a bad mood than your first customer.

The third week is the real pilot, limited to one intent and one time window, with daily transcript reviews and the four metrics from the previous section as the criterion. The discipline that makes it work: a single change at a time, with a name, a date and a way back. A script is never finished: it is versioned.

Conclusion

The script is the part of the agent your customers hear. Everything else — integrations, routing, data — works in the shadows; the tone, the length of the turns, the rephrasing after a misunderstanding and the handoff line are the entire experience of the channel for the person on the other end of the line. That is why they are not technical settings left to the model: they are business decisions that get written, tested and corrected with real transcripts on the table.

If you are sketching the script of your first voice agent, in the use cases you can hear how real agents sound in customer service, or write to us and we will write the five fixed lines of your first flow with you.