Skip to content
buzinov.tech, home page

AI for business

Voice AI agent for outbound calls

An agent that phones your customer list, follows a script you approve and writes the outcome of every call to a sheet or to your CRM.

from $700from 10 days

Packages

Prices are fixed for the described scope. Larger scope gets its own estimate.

  • Basic

    $70010 days

    • 1 outbound scenario in Russian, such as a reminder or booking confirmation
    • Outcome of every call written to a spreadsheet
    • 10 business days if your phone system fits
  • Standard

    $1,60020 days

    • Everything in Basic
    • Up to 3 scenarios in Russian
    • Outcomes written to Bitrix24
    • Transfer to a human agent
  • Premium

    $3,20040 days

    • Everything in Standard
    • Up to 5 scenarios in Russian and Kazakh
    • Inbound calls
    • Reports and call recordings

A voice agent that phones your customer list, follows a script you approve and records the outcome of every call. From $700, from 10 business days if your phone system fits.

Work and scripts run in Russian and Kazakh. English speech is possible but discussed separately. Prices are in USD for clients outside Kazakhstan.

Who this is for

  • Staff spend hours on repetitive calls: visit reminders, booking confirmations, "is this order still active".
  • You want the outcome of each call in a sheet or in Bitrix24: confirmed, declined, call back.
  • You need to work through a large customer list in a short time.
  • Premium only: you also want inbound calls answered by a script.

What you get

Basic, $700, 10 business days if your phone system fits. One outbound scenario in Russian, such as a reminder or a booking confirmation. The outcome of each call goes to a spreadsheet.

Standard, $1,600, 20 business days. Adds up to three scenarios in Russian, outcomes written to Bitrix24 and a transfer to a person.

If the customer asks for a person, or the answer does not fit the script, the call is transferred to a staff member. This does not work on every phone system, and I check yours on the second step.

Premium, $3,200, 40 business days. Adds up to five scenarios in Russian and Kazakh, inbound calls, reports and call recordings. Kazakh suits short questions with the key point repeated back. Longer Kazakh conversations go to a person.

How it works

Each caller turn goes through speech recognition, then a language model picks the next line from your script, then speech synthesis speaks it. The model decides only what to say.

  1. I write the scenario with you: what the agent says, which answers to expect, what happens on an unexpected one.
  2. I check whether calls and the transfer can be connected on your phone system, then load your call list.
  3. A test run on your own staff, then on a small part of the list. You and I listen to the recordings, and I adjust the scenario.
  4. I run the whole list and hand over documentation.

What I need from you

  • Access to your phone system and its name.
  • The conversation in plain words: what to say and which answers to expect.
  • The call list with phone numbers and what to tell each person: visit date, name, time.
  • For Standard: Bitrix24 access with the right to change deals.
  • A person who listens to the first recordings and says whether they are acceptable for your customers.
  • Accounts with a model provider and a speech-recognition provider. I help you open them, and a bank card is enough to pay.

Running costs

Monthly costs fall into four items. I do not name sums I have not measured. Before work starts I show a calculation.

  • Server. The agent runs on a small cloud server. You pay the rent, and I name the sum in the calculation.
  • Model and speech recognition. You pay the providers directly.
  • Telephony. You pay your carrier for the minutes at your own rate. I show a calculation per 1,000 calls before work starts.
  • Support. There is no subscription. After 30 days of free fixes, work is 30 USD per hour, and only when you ask.

Where the data goes

Call audio goes to the speech-recognition provider. The text of the replies goes to the model provider. Audio cannot be masked before recognition: the speech provider hears it.

Names and phone numbers in the text are masked before it reaches the model. Sales Puls, a service for sales teams that I built and keep developing with a partner, does that, and I do the same here.

You obtain your customers' consent to the call, its recording and data processing.

What the program cannot do

  • Answer instantly. On the Sales Puls simulator the reply arrives 1.8 to 2.7 seconds after the caller stops. Scripts use short lines.
  • Recognise speech without errors. About 9 words in 100 are wrong in Russian and about 17 in 100 in Kazakh, and more on a noisy line.
  • Negotiate. It suits reminders, confirmations and simple questions.
  • Decide whether you may call. Rules on automated calls and recording are yours to check. I am not a lawyer.

These figures were measured on the Sales Puls simulator. For your task I show a calculation on 20 of your recordings before work starts.

Terms

  • The schedule starts from the day I receive access and your answers. It covers the time to launch, including waiting for your answers and review, not only my hours.
  • Payment: half before work starts, half after launch.
  • Defects found in the delivered work within 30 days of launch are fixed free of charge. After that, work is 30 USD per hour.

What is not included

  • Server, model, speech and telephony fees. You pay those directly.
  • Buying phone numbers and connecting a carrier.
  • Scenarios beyond the number in the package.
  • Languages other than Russian and Kazakh.
  • Work beyond the package: 30 USD per hour.

Example

For a real-estate agency in Kazakhstan I built an outbound agent that calls clients from a list, keeps the conversation on script and records the outcome for the sales team.

Calls run through Asterisk, speech is handled by Yandex SpeechKit and a language model picks the next line. More in the write-up Outbound voice agent for a sales team. That project has no result figures.

To check calls that already happened, see AI call scoring for sales teams.

To order, fill in the contact form. I reply within a day.

How the work goes

  1. Describe the process

    A few paragraphs: which systems, what goes wrong, what the result should be.

  2. Get the approach and a price

    Within 24 hours: how I would build it, the first risk to settle and a fixed estimate.

  3. See a first working version

    On your real data, early, so we correct course before the budget is spent.

  4. Take it over

    Tests, a README and a handover call. Support after launch if you need it.

Questions

How much does a voice AI agent cost?

Basic is $700 and 10 business days if your phone system fits: one outbound scenario in Russian with results in a spreadsheet. Standard is $1,600 and 20 business days: up to three scenarios, results in Bitrix24 and a transfer to a person. Premium is $3,200 and 40 business days: up to five scenarios in Russian and Kazakh, inbound calls, reports and recordings.

How fast does it answer?

Measured on the Sales Puls simulator, a reply arrives 1.8 to 2.7 seconds after the speaker stops. Your phone line adds its own delay. The pause is noticeable, so scripts use short lines. On your task I show a calculation on 20 of your recordings before work starts.

How accurate is speech recognition?

Measured on the Sales Puls simulator: about 9 words in 100 are wrong in Russian and about 17 in 100 in Kazakh. Noise and a poor line add errors. So a scenario reads key facts, such as a date or an address, back to the caller.

Which phone systems does it work with?

My outbound agent for a real-estate agency runs on Asterisk. Transfer to a person does not work on every phone system. On the second step I check yours, and if it does not fit, I tell you before work starts.

Does it have to say it is a program?

That is your decision, and the rules differ by country. The agent can announce itself at the start of a call. I am not a lawyer, so check local rules on automated calls and call recording.

What do I pay after launch?

A server, the model and speech recognition, and the minutes your carrier bills. Before work starts I show a calculation per 1,000 calls. Defects found within 30 days of launch are fixed free of charge, then it is 30 USD per hour.

Tell me what you need to connect or automate

Describe it in a few paragraphs. You get the approach, the first risk and a fixed estimate within 24 hours.

Message me on Telegram