Most teams start an AI voice agent script the way they would write a radio advert: every line the caller will hear, in order. Then the first real call leaves the page. The caller changes the subject, gives a postcode before being asked, or asks something the script never covered. The agent either freezes or reads the next line anyway, and the caller hangs up feeling that nobody was listening.
Write instructions, not a word-for-word script. Describe who the agent is, what each call is for, how to open, how to confirm names, dates and numbers, and what it must never do. Keep those rules in a separate block from the personality, so you can change the tone without touching the rules that protect the business and the caller.
Why a word-for-word script breaks on a live call
A script is one path through a conversation, and real callers take many paths. Imagine a clinic's booking script asks, "Which day suits you?" The caller replies, "Tuesday, but not the morning, and my son has to come too." A scripted agent has one line for that answer, so it asks about the day again. The caller has just given three facts and hears that none of them was heard.
The fix is not a longer script with more branches. Every branch is another place for the script to break. It works better to tell the agent what the call is for, what it must find out and how to talk, and let it work out the next line from there.
What should an AI voice agent script include?
Good conversation design starts with the same four parts, whatever the call is for. Keep them short enough that a new team member could read them in five minutes. The instructions you write become the voice agent prompt, so keep them in plain words.
- Who it is. The name the agent uses, the business it works for, and what it can and cannot do. The name can fit your brand, but the agent must never claim to be a person.
- What the call is for. The goal, such as booking an appointment, changing a delivery or answering a question about an order, and what counts as done.
- How to talk. Short sentences, one question at a time, a warm but brief tone and no jargon the caller would not use.
- The rules. Things that must always or never happen, such as never confirming a refund amount that is not in the policy.
The name and tone will change as the brand evolves. The rules should change slowly, and only when someone who owns them signs off the change.
Write the greeting to answer two questions
Callers decide within seconds whether they have reached the right place and whether they can trust what they hear. A good opening answers both: who is speaking, and why they are calling. For a clinic, it might sound like this: "Hello, this is the booking line for the clinic. I can help you book, move or cancel an appointment. What can I do for you?"
Keep the greeting short, and test it on people who did not write it. The opening deserves its own attention, which we cover in the first ten seconds of a call.
If a caller asks whether they are speaking to a person or an AI, the agent must answer truthfully. Disclosure rules vary by country and must be followed, so check the requirements for each market before you write the greeting.
Writing for the ear
A written sentence can be long because the reader can go back over it. A spoken sentence has to land the first time. Write the way you would speak to someone standing at a counter.
- Put the answer first, then the detail.
- Ask one question per turn.
- Offer no more than three options at once.
- Say dates in words, such as "Thursday the fifteenth at ten", not "15/10".
- Avoid bullet points, brackets and phrases that only work on paper, such as "please be advised".
The table gives AI phone script examples: a common written line, and a spoken version that works better on a call.
| Written line | Spoken line | Why it is better |
|---|---|---|
| Please be advised that your appointment has been rescheduled. | I have moved your appointment to Thursday at ten. | Starts with the change and uses plain verbs |
| Could you kindly provide your full name, date of birth and postcode? | What is your full name? | One question at a time |
| We apologise for any inconvenience this may have caused. | Sorry about that. Let me fix it. | Short and specific |
| Is there anything else I can assist you with today? | Anything else I can help with? | A shorter close |
| Your reference number is 472918. | Your reference is four seven two, nine one eight. | Read in small groups |
Confirm names, dates and numbers out loud
Speech recognition makes mistakes on names, numbers and addresses, and a caller will not always notice. The script should build in a check. After a name, the agent reads it back. After a phone number, it reads the number in groups and asks whether that is right. If a name is unusual, the agent asks the caller to spell it and never guesses.
Confirm the details that cost money or time to get wrong: the appointment slot, the delivery address, the amount and the reference. Keep each check short, because confirming everything turns a booking into an interrogation.
For how these checks hold up with different accents and noisy lines, see speech recognition and accents.
Keep the rules apart from the personality
Put the rules in their own block, away from the greeting and the tone. Tone changes often, with new campaigns and new staff. Rules should change slowly, with one owner for each.
Rules include what the agent may book or change, which details it must verify first, when it must hand over to a person, and what it must never promise. A rule written only as one sentence among many can be missed on a difficult call. For the rules that matter most, the system should enforce them outside the model. Writing rules your agent can keep shows how to write them so they hold up on a live call.
Plan for the caller who goes off script
Callers will interrupt, change their mind and ask about things that are not on your list. Give the agent a clear goal and a few examples of what to do when a question falls outside what it knows:
- If the caller asks something the agent does not know, it says so and offers to take a message or pass the call to a person.
- If the caller asks for a person, the agent does not argue. It hands over with a summary.
- If the caller sounds upset, the agent slows down, acknowledges the problem once and moves on.
How the agent handles interruptions depends on the phone line and the system around it, as well as on the words. Barge-in and turn-taking explains what should happen when a caller talks over the agent.
How do you test an AI voice agent script?
Reading a script aloud is not a test. Test on real phone lines, with people who did not write the script, in the accents and background noise your callers actually have. Then follow a short cycle:
- Run 20 test calls covering the common paths and two or three awkward ones.
- Read each transcript and mark the moments the agent repeated itself, went quiet, misheard something or gave a line that sounded wrong.
- Group the failures. Most will fall into a few patterns, such as one question asked twice or a long reply to a simple question.
- Change the instruction or rule behind each pattern, not the single call.
- Run the same calls again before you change anything else.
Put a version number and an owner's name at the top of the instructions, so you always know which version took a given call. Repeat the cycle until the common paths work well, then widen the test to the awkward cases.
How VoiFlow handles this
VoiFlow agents are set up from instructions like these. The hard rules, such as permissions, limits, consent, redaction and the audit trail, are enforced outside the model. Every call is recorded, transcribed, scored and replayable, so you can find the exact line that went wrong. When the script runs out, staff can listen in, whisper to the agent or take over the live call. The conversation platform page explains how this works in practice.
Frequently asked questions
Should an AI voice agent read from a word-for-word script?
Usually not. A fixed line works for one short, predictable step, such as a notice that must be read exactly. For the rest of the call, give the agent a goal, the facts it needs and a few examples, because callers change direction.
How long should an AI voice agent script be?
Short enough that a new team member can read it in about five minutes. If it runs longer, the extra material probably belongs in the knowledge base, not in the instructions.
Does the AI agent have to say it is an AI?
When a caller asks, the agent must answer truthfully and never claim to be a person. Disclosure rules vary by country and must be followed, so check the requirements for each market you serve.
Who should write the script?
Someone who knows the calls well, such as an experienced agent, a front-desk lead or a supervisor, working with the person who sets up the agent. Start from real call recordings or notes, not from a blank page.
Want to hear how a set of instructions sounds on a live line? Try a call with one of our demo agents, or contact us about setting up your first call type.
