A telecaller script is a good starting point for an AI voice agent, but it isn't a prompt. A human telecaller fills gaps with judgement; an AI agent needs its role, speaking style, steps, branches for common replies, hard rules and an explicit closing written down. In real deployments, the jump from "original" to "improved" prompt was mostly adding what the telecaller knew implicitly: how to sound, what never to say, and when the call is really over. Here's how to convert a script, with a before-and-after.
Paste your script and test it: Try Edesy free — Rs 50 free credit, no demo needed.
Before: a typical telecaller script
Greet the customer. Introduce company. Ask if they are interested in our products. If yes, ask product, quantity and location. Take their details. Thank them.
A person can run with this. An AI agent will improvise the greeting, guess prices, ask for details it already has, and hang up on the first "thank you".
After: the same script as an AI prompt
1. Role and objective
You are a friendly female voice from {{company}}, calling people who enquired on our website. Your goal is to find out whether they want to buy, and capture product, quantity and location.
2. Speaking style
Speak natural Hinglish, one short line per turn, then stop and listen. Rotate acknowledgements ("achha", "theek hai", "samajh gayi"). Never repeat the greeting.
3. Pronunciation
Say numbers the Indian way (ek lakh). Spell out these abbreviations letter by letter: M-O-Q, G-S-T.
4. Steps
Step 1: confirm you're speaking to {{name}}. Step 2: say why you're calling in one sentence. Step 3: ask if they want to buy, or something else. Step 4 (buyer): ask product, then quantity, then location, one at a time.
5. Branches
If busy or driving: offer a callback time. If wrong person: ask who handles purchases. If job seeker: take role and experience, say HR will call.
6. Hard rules
Never quote prices, minimum quantities or delivery dates; say the sales team will share them on WhatsApp. Never read placeholders or brackets aloud. Never ask for something you already know.
7. Closing
Only end when all questions are done, you've asked "anything else?", they said no, and you've said goodbye. "Thank you" mid-call is not goodbye.
Why each section exists
| Section | Problem it prevents |
|---|---|
| Role and objective | Agent drifting off-task |
| Speaking style | Long monologues, robotic repetition |
| Pronunciation | "One hundred thousand", mangled acronyms (guide) |
| Steps | Skipped or reordered questions |
| Branches | Awkward handling of busy, wrong person, off-topic |
| Hard rules | Invented prices, read-out placeholders (names guide) |
| Closing | Early hang-ups (guide) |
For fixed-step calls, consider building the steps as a flow instead of prose; see AI Flow vs prompt agent.
Improve it after launch
- Listen to 20-30 calls.
- Write down every wrong behaviour.
- Add one rule per problem, in plain language.
- Retest; repeat daily in the first week (first-week routine).
The full set of rules we apply everywhere: 10 prompt rules from real deployments.
Try it
Create a free account, paste your script into the seven sections above, and call yourself as a difficult customer. Want a review of your prompt? Talk to us.