How do you write a script for an AI phone agent?

Do not write a word-for-word script. Define a short list of jobs the agent may do, the information each job needs, what done looks like and what is out of bounds. Add the business facts it needs, the things it must never say and a plan for interruptions, corrections and side questions. Build that from your real calls rather than from the calls you assume you get.

Why do some AI voice agents sound robotic?

Usually because the conversation was not designed for the ear. Long sentences, several questions at once, lists of many options, numbers read the wrong way and asking for details the caller already gave all make an agent sound mechanical, even with a natural voice model.

What should an AI phone agent confirm with the caller?

Anything that is expensive to get wrong: phone numbers, addresses, dates, times and names. The agent should read these back and wait for confirmation. Hard-to-spell details like email addresses are often better collected by texting the caller a link.

Voice AI Conversation Design: How to Script an AI Phone Agent

TL;DR: A good AI phone agent is not a script read aloud. It is a short list of jobs the agent is allowed to do, the information each job needs, the rules it must never break and a written plan for every way a caller can wander off. Design it from your real calls, not from a whiteboard. Write for the ear: short sentences, one question at a time, no menus of five options. Read back anything that is expensive to get wrong, like phone numbers, addresses and dates. And close every call with a clear next step the caller can repeat back. Most voice agents that "sound robotic" are not suffering from a bad voice model. They are suffering from a conversation nobody designed.

When a business puts an AI agent on its phone line, attention usually goes to the voice, the model and the platform. But what callers actually experience is the conversation, and that design job often goes to whoever writes the first prompt.

Voice AI latency covers how fast the agent responds, the human handoff covers how it exits, and voice agent monitoring covers how you improve it once it is live. Conversation design is everything in between: what the agent says, what it asks, in what order, and what it does when the caller does not cooperate.

A voice agent is a set of jobs, not a script

The old phone tree (press 1 for sales, press 2 for support) failed the moment a caller went off the tree.

Modern voice agents run on language models, so callers can speak naturally and you no longer write every line. But "the model will figure it out" is not a design. Without structure, an agent improvises: quoting prices you never approved, promising callbacks nobody will make, or asking for the same detail three times.

The structure that works is a short list of jobs. For each one, write down:

What the job is. Book a new appointment. Reschedule an existing one. Take a service request. Answer questions about hours and location.
What information it needs. Name, callback number, the reason for the visit, a preferred time.
What "done" looks like. An appointment in the calendar, a confirmation text sent, a ticket created.
What is out of bounds. Anything the agent must decline or hand to a person.

That list, plus the business facts the agent needs to answer questions, is the real specification. The prompt is just how it gets expressed.

Start from real calls, not from a whiteboard

The most common design mistake is building the agent around the calls the business thinks it gets. The calls it actually gets are messier.

Before writing anything, gather what you can: call recordings if you have them, notes from whoever answers the phone, the questions that come up every week. Sort them into call types and ask three things about each:

How often does it happen? The top few call types usually cover most of the volume. Design those carefully first.
Can the agent finish it alone? Booking, rescheduling and basic questions usually yes. Disputes, complaints and anything needing judgment usually no.
What does a bad outcome cost? A wrong opening hour is an annoyance. A wrong address on an emergency job is a lost customer.

This inventory decides the scope of the first version, and it becomes your test set later. If you are still deciding whether the phone line is worth automating at all, the missed-call revenue leak is the place to start.

The opening: get to the reason fast

The first few seconds set the caller's patience for the rest of the call.

Keep the greeting short. The business name and an open question. A long welcome with a marketing line is time the caller spends waiting to speak.
Be honest that it is an AI where your rules or your judgment say so. Callers who feel tricked tend to get angry when they work it out. Disclosure and recording consent also carry legal requirements in some places, covered in voice AI call recording consent.
Let the caller lead. "How can I help?" works better than reading out a list of services. The agent can work out the job from what the caller says and ask a clarifying question if it is unclear.

Collecting information without interrogating

Most jobs need details from the caller. How the agent asks for them decides whether the call feels like a conversation or a form.

One question at a time

People on the phone can hold one question in their head. "Can I get your name, phone number and the address of the property?" invites the caller to answer the first part and forget the rest. Ask for one thing, wait, acknowledge it, move on.

Do not ask for what you already have

If the caller said "I need to move my Thursday appointment" in their first sentence, the agent should not then ask "Is this a new booking or a change?" Listening to what was already said is the cheapest way to sound competent.

Read back what is expensive to get wrong

Speech recognition is very good and still not perfect, especially with names, street names, email addresses and long numbers over a poor line. Decide which fields matter enough to confirm, and have the agent read them back: "That is 4 1 2, 5 5 5, 0 1 9 8, is that right?" Confirmation adds a few seconds and saves a lot of failed callbacks.

Move hard-to-spell details out of the call

Email addresses spelled out over the phone are a common failure point. When a job needs one, consider having the agent text a link where the caller types it instead. The same goes for anything long, like a reference number or a payment step.

Writing for the ear

Text that reads well can sound awful aloud. Voice has its own rules.

Short sentences. One idea each. Long sentences make callers lose the thread and interrupt.
No more than two or three options at once. "We have openings Tuesday morning or Wednesday afternoon" works. Six time slots read out in a row does not.
Say numbers the way people say them. Times, prices and phone numbers need explicit rules for how they are spoken, or the agent may read them in a way no person would.
Avoid internal jargon. Your team says "PM visit". Callers say "the yearly check-up".
Acknowledge before you act. "Let me check that for you" covers the moment the agent is looking something up, so silence does not feel like a dropped call.

Planning for detours

Real callers do not follow the path. Good conversation design assumes they will not and decides in advance what happens.

Interruptions. Callers talk over the agent. It should stop, listen and respond to what they said, not finish its sentence.
Corrections. "No, sorry, I meant Friday." The agent needs to change the detail it collected, not add a second one.
Side questions mid-job. A caller booking an appointment asks whether you have parking. The agent answers and returns to the booking without starting over.
Silence and noise. A caller who goes quiet gets one gentle prompt, then a sensible close. A line full of background noise gets a request to repeat, not a guess.
Calls the agent should not take. Emergencies, complaints, legal or medical questions and anything on the out-of-bounds list go to a person according to the escalation rules.

Guardrails: what the agent must never say

Every business has things the phone line must not do. Write them down explicitly, because a language model will not guess them.

Typical examples: never quote a price that is not on the approved list. Never promise a specific technician, arrival time or outcome. Never give medical, legal or financial advice. Never confirm or discuss another customer's details. Never invent an answer when it does not know: say so and offer a callback.

Guardrails are also where vertical differences show up. A veterinary clinic, a law firm and a roofing company have very different lists. Our guides for restaurants, med spas and auto repair shops cover what changes by industry.

The close: leave the caller with a next step

How a call ends shapes the memory of it. Every successful job should close with three things:

A summary of what was agreed: "You are booked for Tuesday at 10 with the callback number ending 0198."
What happens next: a confirmation text, a callback from the team, a reminder the day before.
A chance to add anything: "Is there anything else I can help with?" asked once, not as a loop.

Then follow through: a confirmation text that never arrives undoes a good call.

Multilingual and multi-location agents

If you serve callers in more than one language, each language is effectively a second version of the conversation, not a translation of the first. Greetings, politeness norms and how people give addresses all differ. In our work on Fortell, where the voice and SMS systems help Community Action Agencies handle intake in more than 100 languages, the design work per language is as important as the speech models. Multilingual voice AI goes deeper.

Multiple locations add another layer: the agent needs to know which branch the caller wants, and each branch may have different hours, services and staff to transfer to. For CallGuard AI, which won a US chain and onboarded its first branches, per-location routing is part of the design from the start rather than a setting added later. The practical side of numbers and routing is covered in AI receptionist phone setup.

Testing the design before callers do

Before launch, run the conversation against the call inventory you built at the start. Ask the people who answer your phones today to call in as the difficult customers they remember, and work through the detour list on purpose. Read the transcripts for repeated questions, missed details and anything said that is not on the approved list. Keep those test calls: they become the regression set you rerun every time the agent changes, which is where monitoring takes over.

Questions to ask whoever designs your voice agent

Which call types is the first version built to finish, and how were they chosen?
What information does each job collect, and which fields are read back?
Where is the list of things the agent must never say, and can we edit it?
What happens on interruptions, corrections and side questions?
How does the call close, and what confirmation does the caller receive?
What test calls were run before launch, and can we see the transcripts?

If the answer to most of these is "the model handles it", the conversation has not been designed yet.

Where we fit

We build the voice agents and automations behind CallGuard AI, which answers conversations, books appointments and captures revenue around the clock, and CallSetter AI, which answers every call in under 60 seconds, qualifies leads and books appointments. Conversation design is part of every build: the job list, the information each job needs, the guardrails and the detours, all written down and agreed with the people who know the calls best.

If you are evaluating an AI receptionist, an AI appointment setter or a custom AI agent, we start from your real calls and ship a first version on the same ship-in-days approach we use across the studio.

Want a voice agent that sounds like it understands your callers? Book a demo and we will map your top call types, what each one needs and how your agent should handle the calls that go off script. See our work: CallGuard AI, CallSetter AI, Fortell AI and more, shipped in days, not months.