Outbound voice agent for a sales team

The problem

A real estate agency in Kazakhstan was spending hours of manager time on first-contact calls, even though most of those calls follow the same script. The calls had to move to an agent, and the sales team still needed the outcome of each one.

What I built

A voice agent that takes clients from a campaign list and calls each of them.

Telephony. Calls go through Asterisk with my own extensions. The dialling processes are written in Python, with custom routing, custom queues and dynamic call distribution.

Conversation. Every client reply goes through three steps:

1. Yandex SpeechKit turns the reply into text.
2. ChatGPT, through the OpenAI API, evaluates the reply and picks the next line of the script.
3. Yandex SpeechKit voices the answer.

The loop runs until the call ends, and then the agent records the outcome for the sales team. Recognition and synthesis stay in SpeechKit, so the model only decides what each reply means and says.

The result

The agent ran the sales team's call campaigns end to end: it dialled the list, held the conversations and recorded the outcome of every call. Managers received those outcomes instead of placing each first call themselves.

For the same agency I built automated sales call scoring. How such systems are built and what they cost is on the call and document analysis page.