How to write an AI voice bot script: prompt and conversation plan

An online school calls its old database — people who attended a webinar a year ago and did not buy. If the AI voice bot script is vague, the bot speaks in generalities, gets tangled in its rules and never gets to booking a consultation. The minutes are paid for, and the lead is lost.

The instruction (prompt) is the bot's script and character: who it introduces itself as, how it speaks, which facts it gives and what goal it leads to. You don't need to write it from scratch. The assistant bot studies your website, asks about your business and builds a selling script in a couple of minutes. And after every save, the account itself looks for rules that conflict with each other.

Watch the two-minute video, and below you will find step-by-step setup.

Why your business needs a well-designed call script

• Online school: the assistant finds out the goal of the call and builds the plan around it. The bot handles objections with facts and books a call with a curator.
• Real estate: stop criteria (“who we don't work with”) filter out non-target clients. The bot says goodbye politely, and the manager gets only warm leads.
• Clinics and salons: the “Incoming” call type and a greeting in the “Starting phrases” turn the bot into a receptionist who books clients even at night.
• B2B sales: the assistant suggests at least six hook options — opening lines that hold attention. The bot reaches the person in charge and books a meeting.
• Head of sales: backups keep working versions of the script, so a bad edit is easy to roll back.

What a voice bot prompt consists of

The bot keeps the whole instruction and the whole conversation in memory, so it does not forget what the client said at the start. The instruction consists of six blocks. The “Instruction” sub-tab has four of them:

• “Sales role” — who is calling, from which company, to whom and why.
• “Speaking style” — conversational speech, pauses, a casual or formal way of addressing the client; there are ready-made styles.
• “Important rules” — for example, “After ‘Hello’, wait for the client to answer”.
• “Arguments and facts” — figures, cases, prices, guarantees.

The “Plan” sub-tab has two more:

• “Starting phrases” — the bot opens with them. The first one is a short “Hello…”, after which the bot waits for an answer.
• “Conversation plan” — steps in order, one task per item. The last item is the target action.

Role and style hold up to 1,400 characters, an item in the other blocks up to 1,000. The order of items matters; you change it by dragging.

The “{...}” button in the corner of an item opens the “Add parameter” window: the client's name from the lead card, pauses, CRM fields. The “Emotion tags” tab adds laughter, a sigh or a whisper to the speech — [laugh], [sigh], [whispers]. Tags work with all voices: the bot sounds more alive, and the other person is less likely to guess they are talking to a bot.

Mark a phrase with client data, such as a name, a booking date or an order total, with the “Do not cache this block” parameter from the “CRM fields and pause tags” tab: put it at the end of the phrase. The bot never takes such a fragment from the reply cache and says it afresh every time, so the client won’t hear someone else’s name or date. Note: while the prompt contains this parameter or any other lead parameters, “AI caching” can’t be turned on: the dashboard shows the “Caching is unavailable” window.

How to write rules and facts for an AI voice bot

A rule tells the bot how to act, and a fact states what is true about the company. “We work without prepayment” is a fact for “Arguments and facts”. “If the client asks about payment, explain that a specialist will discuss the terms” is a rule for “Important rules”.

• Rule formula — “if [situation] — act like this”. Write in plain language: one situation per item.
• The answer depends on the stage of the call. The same question needs different answers before and after the offer. If the client asks “Who is this?” or “What services?” before the bot has found out it is talking to the decision maker, let it introduce itself briefly and return to the qualifying question instead of retelling the whole offer. Write it exactly like that: “If the client asks who is calling before you have presented the offer — …; after — …”.
• Describe the sign by meaning, not with a word list. “If the answer shows they do not make decisions, apologize and politely say goodbye.” You don't need to list “I'm the accountant”, “I'm the secretary”, “I don't work here”: the AI understands the meaning, including variants the rule doesn't mention.
• No direct speech. Not “say: …”, but “explain”, “clarify”, “act”. The bot quotes ready-made phrases, and its speech turns wooden.
• Changing behavior? Edit the original rule. If you decide to agree only on a day without an exact time, find the items about an exact time in the rules and the plan and rewrite them. A new rule on top of an old one creates a contradiction.
• After handling an objection, say where to go next. Add the next plan step to the rule: “…and return to offering to send the details” or “…and offer to transfer to a manager”. Otherwise the bot answers and hands the initiative to the client: instead of the next step it asks “Would you like my colleague to advise you?”.
• Stress — with a capital letter. The “Add parameter” window has “Example of intonation”: “rEcord” or “recOrd”. A word where the bot misplaces the stress is better banned with a rule.

How the bot says dates and times to the client

A reminder call saves money: a client who gets a call the day before misses fewer appointments, and a technician doesn't drive out for nothing. The bot can save the appointment date to the CRM during the first conversation and remind the client about it with a call the day before — see “Automated appointment reminder calls” for how to set this up. But the date has to be presented correctly in the instruction. If you put the field straight into a starting phrase, the bot may read it as is — “zero eight, zero seven, two thousand twenty-five” — and the client will decide it's an automated call.

1. 1
   Remove the date from the starting phrases

   Keep a greeting without the date field in “Starting phrases”.
2. 2
   Add a fact

   In “Arguments and facts”, write “Appointment time:”, insert the date field with the “{...}” button and add in brackets “say the date and time in words, without the year”.
3. 3
   Add a plan step

   In “Conversation plan”: “Introduce yourself and remind the client about the appointment. Say the date and time in words, without the year. Check that everything is still on”.
4. 4
   Add a rule

   In “Important rules”: “Write numbers in words”. This way the bot also says prices and deadlines correctly.

The bot takes the date from the fact and how to pronounce it from the plan: “You're booked for the eighth of July, at twelve thirty. Is that still on?” If the bot mixes up weekdays or dates, check “Region (AI time zone)” and the “Model thinking setting” — they are covered in “How to make an AI voice bot sound human”.

How to generate a call script from your website

1. 1
   Open the wizard

   For a new bot, the “Generate prompt” wizard opens by itself (how to create a bot is covered in “AI voice bot for business”). For an existing bot: “Training AI” → “Manual training” → “Start training” → “Generate prompt”. The new instruction replaces the current one, so the account asks you to confirm. So regenerate only before you have edited the instruction by hand; after edits, add rules and facts instead.
2. 2
   Tell it about your business

   Paste a link to your website, choose the bot's language and click “Analyze the website”. To do it without the assistant, choose “Fill in the fields myself”: there are three questions.
3. 3
   Be specific

   For generic words like “quality”, the assistant asks for figures, timelines and guarantees. The more precise the facts, the more convincing the bot.
4. 4
   Generate

   Click “Confirm and generate”. Generation takes up to two minutes, then the account writes the funnel step rules. If there is no “Interest confirmed” step, choose a target step from the list. How the bot sets steps after a call is covered in “AI voice bot CRM integration”.
5. 5
   Check the plan

   On the “Plan” sub-tab, read the phrases and the plan to the end and remove anything unnecessary. The bot opens every conversation with the starting phrases. If a phrase does not fit a common client reply — say, “Great, I've reached the right person” right after “Why do you think so?” — shorten or delete it.
6. 6
   Go through the checklist and save

   The “Prompt generated 🎉” window guides you item by item. After each item, click “Save”, or the checkmark disappears.

Launch readiness checklist

Items: “AI training (bot replies)”, “Link an inbound number”, “Conversation plan”, “Training on your own materials”, “Funnel stages training”. Each “Writing to CRM” field adds its own item, and the bot's calls add the “AI training on real calls” item.

The ↗ arrow next to an item takes you to the right place, and the ⓘ icon opens a hint. Until the checklist is complete, it appears after saving and when you start calling. This way, a half-configured bot will not burn through your budget.

How to find and fix contradictions in a prompt

“Save” in the bottom panel saves the whole bot. The account checks that all six blocks are filled in and that the instruction does not mislead the client.

Then the AI looks for contradictions. A contradiction is a pair of rules that conflict with each other: for example, “answer briefly” and “describe every plan in detail”. The bot would follow them at random. If such pairs are found, the window “The instruction is saved, but it contradicts itself” opens:

1. Click “Fix the contradiction”.
2. Choose “Accept the recommendation” or “With the assistant bot” — the edit appears in the “Before / After” window.
3. Not happy with it? Write a “Comment on the fix” and click “Redo with the comment”.
4. Click “Apply” and then “Save” again.

What to do if “The instruction cannot be saved”

Before saving, the AI checks the instruction for blatant contradictions and wording that misleads the client: the same price, timeline or company name written in different ways, one phrase both forbidden and recommended, a promise that contradicts the facts. If the check finds something like this, the bot is not saved. A window titled “The instruction cannot be saved” opens with the subtitle “AI check: the instruction contains information that misleads the client.” and an explanation below.

1. Read the explanation in the window — it says which fragment the check flagged.
2. Click “Got it” and find this fragment on the “Instruction” and “Plan” sub-tabs.
3. Keep one value instead of two, remove any promise that is not in the “Arguments and facts” block, and click “Save” again.

If the explanation is just a general phrase, look for the same things: different prices, timelines or names for the same thing, and pairs of “do” and “don't” rules about one action. The contradictions window after “Saved” is different: the bot is already saved, and the conflicting rules are simply worth fixing. All the windows that stop a save are collected in “Settings not saving”.

How to restore a previous script version: backups and logs

A backup is a saved snapshot of the instruction. “Settings” → “Backups & logs” keeps the 6 latest backups. A snapshot is taken on every save and before AI training. To restore a version, click the date → “Load into the editor” → “Save”. “Save logs” show who saved the bot and what they changed.

Read also: AI voice bot for business: how to create a bot and get leads, How to train an AI voice bot to answer like your best sales rep, AI voice bot CRM integration: call results and funnel step