We have a lot of registered users. Some are active, others tried the service once and quietly lost interest. Hence the task: how to win back users who never complained about anything and simply stopped logging in.
The standard answer is a discount campaign or a “how do you like our service” survey. We’ve been there: low open rates, few replies, and worst of all, it bores us. And if the sender is bored, the user on the other end is even more so.
So we took a different route: we described the task to a neural network and asked it to come up with the mechanics itself. Our voice AI agents are built in a regular chat: you explain whom you’re calling and why, and you get a ready agent. People usually do this to avoid writing long prompts by hand. We wanted to see something else: could the neural network handle a task with no script at all.
Spoiler: the mechanics were invented by the neural network, not us. We only approved them and clicked “start”. Below is the whole journey: the dialogue, the full prompt, real numbers and two unedited transcripts of live calls.
Step 1. A conversation where we didn’t write the call script
Creating an agent looks like an ordinary chat: you write what you want from the call and whom you’ll be calling, and attach files if you like. The whole conversation took a few messages.
The word “boring” was our only contribution as the client. And the neural network asked about prizes on its own. We hadn’t mentioned them and hadn’t planned any: the giveaway mechanics raised a question the script couldn’t be built without.
The result was an agent named Saken: a male voice and an “upbeat, positive, competitive” personality.
Step 2. The instructions the neural network wrote for itself
If you were looking for a real call campaign script example for a voice agent, feel free to use it as a template.
## Who you are Saken is the male voice of AutoCall.kz. You call users to run a short quiz about telecom and communications. Personality: upbeat, positive, friendly, competitive. ## Call goal Get to know the user, find out their name and source (how they heard about AutoCall.kz), and run a 5-question quiz about telecom. ## Conversation flow 1. First line EXACTLY: “Hello? Hi!” — and NOTHING ELSE. Then you stay silent and wait for a response. 2. If they respond, introduction and purpose EXACTLY: “This is Saken from AutoCall.kz. I’m calling to run a short quiz with prizes for our users and lift your mood. Do you have a minute and a half?” 3. If they are silent after the first line, ask once: “Hello, can you hear me?”. If it’s silent again, say a polite goodbye and end the call (end_call). 4. If they agree: - “Great! What should I call you?” (save subscriber_name) - “Nice to meet you! Tell me, how did you hear about us: search engines, social media ads, or did someone recommend us?” (save lead_source) - Ask 5 questions one by one (after each answer, immediately record it as right/wrong and add 1 point to quiz_score): 1. “First question, an easy one: What do the letters "AI" stand for in modern technology?” (Correct: Artificial Intelligence) 2. “Second question: What single emergency number works from a mobile phone worldwide, even without a SIM card?” (Correct: 112) 3. “Third question: How many days are there in a regular, non-leap year?” (Correct: 365) 4. “Fourth question: Which phone is still the subject of legends that it’s indestructible, and that if you drop it from a balcony onto asphalt, you should worry about the asphalt, not the phone?” (Correct: Nokia 3310) 5. “And the final, hardest question! In what year was the very first mobile phone call in history made: 1973, 1968 or 1982?” (Correct: 1973) 5. Wrap-up: - Sum up: “You scored [quiz_score] out of 5!” - Say goodbye and mention that a manager will contact the winners. ## What to collect - subscriber_name: The person’s name - lead_source: How they heard about AutoCall.kz - quiz_score: Number of correct answers (0 to 5) ## Boundaries - Don’t push if the person is busy or declines to take part. - Don’t name a specific prize on the call.
We would never have come up with the fourth question, about Nokia and the asphalt, ourselves. That’s where the survey stopped being a survey: people laughed and answered more willingly after it.
Note item 1: “First line EXACTLY: ‘Hello? Hi!’ — and NOTHING ELSE”. This isn’t about style, it’s deliberate. A person who picks up isn’t listening yet for the first half second. A short greeting and a pause let them tune in, and only then comes the content.
Step 3. What happened over a hundred calls
We didn’t launch on the whole list: we picked 100 service accounts at random.
Let’s go line by line.
58 out of 100 answered. On cold lists our usual answer rate is 25–35%. Here it was almost twice as high, but the explanation is simple: the list was warm, details in the note below.
39 out of 58 agreed to play. That’s two thirds of everyone who answered. For a phone survey that’s a lot: usually most people drop off at “do you have a minute and a half”. “Quiz”, “lift your mood” and prizes work noticeably better than “rate our service”.
4 out of 39 answered all five questions correctly. So the quiz was real, not a giveaway. Almost everyone stumbled on the last question, about 1973.
18 people contacted our support on their own after the call. That number is what surprised us. Almost half of the players went looking for us in the Telegram chat to ask: “so what did I win and how do I get it?”. No SMS campaign and no call with a recorded message had ever gotten that kind of response before.
An honest caveat about the sample. We didn’t call a random hundred users. We called people who actively used our service from 2018 to 2021.
How it actually sounded
Two transcripts of live calls. We didn’t cherry-pick the best ones or clean up the awkward moments, because those are the most interesting part.
Here the agent misheard the name and called the person “Sangas”. The contact corrected it, the agent apologized and repeated the question with the right name: it didn’t freeze and didn’t start over. This is exactly the situation where a regular button-based scenario breaks, because it has no “I’ve been corrected” branch.
That’s four turns in a row that aren’t in any script: the contact asked to repeat the question, the agent cut it down to the essentials, the contact remembered the bit about asphalt, and the agent picked it up and steered the conversation back. There are no prepared branches here at all. This is exactly what voice agents are paid for: a button-based scenario would simply loop here.
What went wrong
The wording about prizes was off. In its final line, the agent said: “a manager will contact you and tell you which prize you won”. That reads as “everyone won”, hence 18 people in support asking “what do I get?!”. But that was our mistake, not the agent’s: we approved the prompt without noticing the ambiguity. It should have said “a manager will contact the winners”. The prompt version above already has this fixed.
Names are the hardest to recognize. “Salmas” instead of “Zhalgas” is a typical case: Kazakh names in a narrow telephone band (8 kHz at 64 kbps) are the hardest for speech recognition, because models are trained on better-quality recordings. The agent handled it, but if the name matters to you, it’s better to pull it from your contact list than to ask for it by voice.
We underestimated the response. A hundred calls brought 18 support requests. Had we launched on the whole list of several thousand people right away, we would have had several hundred people waiting for a prize and commitments we weren’t ready for. The test sample saved us from our own success.
How it ended
We credited all 39 players with a bonus to their dashboard balance. The powerbank and fitness watch went to the winners, those who answered every question correctly.
How to do this yourself
The mechanics aren’t tied to a quiz. The same approach works for satisfaction surveys, event invitations, subscription renewal reminders, anywhere you need a live conversation instead of a notification.
- Describe the task to the AI agent in plain words. Here is how to create a voice agent without a single line of code. Don’t be afraid to say “boring”, “let’s try something else” or “meh, not great”: that’s what it’s built for.
- Read the generated prompt in full, especially what the agent will promise. Our 18 support requests grew out of a single unchecked phrase.
- Call yourself first. The agent card has a “Test” button: the call comes straight to your browser, and you talk to the agent before the live campaign.
- Launch on a few contacts, not the whole list. Look at the numbers and the reactions, fine-tune the agent, and only then scale up.
And most importantly, call people who know you and have agreed to receive calls from you.
How our voice AI agents work, what a minute of conversation costs and how a neural network on a call differs from a push-button robot: all on a separate page.