Writing scripts for voice agents: short turns, clear choices

Updated 2026-09-19 · 4 min read · By the Convo360 team

Why voice scripts are not chat scripts

A WhatsApp message can be re-read. A spoken sentence disappears the moment it's said. This single difference should shape every line you write for a voice agent, if a caller has to ask 'sorry, what was that?', the script has already failed once.

Chat scripts can afford buttons, lists and menus the customer scans visually. Voice scripts only have sequence: one thing said, one thing heard, one thing answered. Write accordingly.

It also helps to remember who is usually calling and from where, someone walking on a busy street in Pune, a shopkeeper answering between customers, a parent with a toddler in the background. None of them are giving your script their full, undivided attention the way a reader might. Simplicity isn't a nice-to-have here; it's what makes the call usable at all.

Turn length and one question at a time

Keep each agent turn to roughly 10-15 spoken words wherever possible. If you find yourself writing 'and also' or 'additionally' inside a turn, that's usually a sign it should be two turns, not one.

• Weak: 'What's your name, and which city are you calling from, and do you have an existing account with us?'
• Better: Three separate turns, each waiting for an answer before moving to the next.

This matters more on a phone call than in chat, because the caller cannot scroll back to see what was asked.

Numbers, dates and amounts — always read back

Misheard numbers cause more real-world problems than any other voice bot failure. A ₹15,000 amount misheard as ₹50,000, or the 4th misheard as the 14th, causes a booking or payment error that's expensive to fix after the fact.

• Speak numbers in full: 'fifteen thousand rupees', not 'fifteen K'.
• Read dates with the day name attached: 'this Thursday, the 11th', not just 'the 11th'.
• Confirm before acting: 'So that's a service booking for Thursday the 11th, between 11 AM and 1 PM, shall I confirm that?'

This read-back step should be a hard rule in your script, not an optional nicety, anywhere money, dates or quantities are involved. It costs the call an extra five seconds and saves a rescheduling call, a refund, or an unhappy customer later, a trade worth making every time.

Phone numbers deserve the same care, especially in India where callers often give a mobile number quickly and expect it to be captured correctly the first time. Reading it back in groups of two or three digits, rather than one long string, is easier for the caller to verify by ear.

Designing for interruption

Real callers interrupt. Someone might answer 'Thursday' before the agent finishes listing options, or say 'no, cancel' mid-sentence. A script that assumes the caller will wait patiently through every word breaks the moment a real person calls in.

Write scripts so each turn's important content comes early, not buried at the end, if the agent gets cut off after the first few words, the caller should still have heard the key part. On Convo360, the Let the caller interrupt toggle in Call behaviour controls this at the platform level; your script should assume it's on unless the line is genuinely noisy. See the voice docs Call behaviour section for the noise and interruption settings.

Building the script into a Convo360 flow

1. Draft each agent turn as a short line in your AI agent's system prompt, or as individual gather steps if you're using a structured flow.
2. For fixed choices, like 'press 1 for sales, 2 for support', use a DTMF menu block instead of open speech, since it's faster and more reliable on noisy lines.
3. Add a confirmation turn after any number, date or amount, before the flow moves to transfer, booking or payment steps.
4. Set Acknowledgement style to a short natural word so the caller knows the agent heard them, without it feeling scripted.
5. Test the whole script with Preview / Try-it before going live, reading it aloud yourself to catch anything that sounds unnatural spoken.

Once live, check Flow Logs and call transcripts regularly. If you notice callers repeating themselves at the same point in the script, that's usually a sign the turn before it was too long, too vague, or asked two things at once, go back and split it.

Common script mistakes to avoid

A few patterns show up repeatedly in early voice scripts, and they're worth checking your own draft against:

• Writing like a brochure: phrases like 'we are pleased to offer' sound stiff spoken aloud. Say it the way you'd say it to someone standing in front of you.
• Burying the ask: if the agent needs an answer, ask directly and early in the turn, not after three sentences of context.
• No fallback for silence: if the caller doesn't respond, the script should have a gentle re-prompt rather than repeating the exact same line, which can feel like the agent is stuck.
• Overusing the caller's name: saying it once at the start is warm; saying it in every turn starts to sound robotic.

Reading the script aloud to a colleague before launch catches most of these faster than reading it silently ever will.

Step by step

1. Draft turns at 10-15 words eachWrite every agent line short enough to say naturally in one breath, avoiding stacked information.
2. Split multi-part questionsTurn any 'and also' question into two separate turns, each waiting for its own answer.
3. Add read-back confirmationsFor every number, date or amount, write a confirmation turn that repeats it back before the flow proceeds.
4. Front-load key informationPut the most important word or number early in each turn so it survives an interruption.
5. Test the script aloudRead the full script out loud yourself, then run it through Preview / Try-it before publishing.

Example conversation

Frequently asked questions

Related reading

Available across India — see WhatsApp & AI calling automation by state and city.