AI Telemarketing Agent — Automated Outbound Calling

Outbound telemarketing is mostly filtering. Of every hundred calls, a handful reach someone who wants to hear more, and a human salesperson spends the rest of the time on rejections and voicemail.

The agent takes the filtering step. It does one small job well — find out whether the person is interested, answer simple questions about the company if they ask — and then gets out of the way. Anything that needs a real conversation goes to the sales team.

1. 01Upload the contact database. Each number becomes a lead waiting to be called.
2. 02Train the agent with the company's information: who it is, what it offers, and the answers to common questions.
3. 03The agent calls the contact, introduces itself and asks whether they are interested.
4. 04If the contact asks about the company, the agent answers from its training — and only from it.
5. 05Interested: the sales team is notified straight away with the contact's number, the call recording and the transcript, so the salesperson knows exactly what was said before they dial.
6. 06Not interested: the outcome is logged and the contact is not called again.
7. 07Busy: the agent asks when a better time would be — or, if no time is given, tries again after roughly 3 to 5 hours — and calls back at the time the contact chose.

The value for the sales team is not only the lead; it is the context. Each contact has a complete history so nobody has to ask 'what did we tell this person?'

• Call outcome: interested, not interested, busy or no answer
• Audio recording of every call
• Full transcript of what the agent and the contact said
• Callback requests with the time the contact asked for
• Attempt count and timestamps for each call
• Which salesperson was notified, and when

• A narrow agent. It qualifies interest and answers company questions; it does not negotiate, quote prices or close sales.
• Grounded answers. The agent answers from the information the company trained it with and hands over when it does not know, instead of guessing.
• Respect the contact's time. A busy contact picks the callback time, rather than being redialled blindly.
• Humans get warm leads only, with the evidence attached.

I am building this as the full-stack developer at ZAMAN IT — designing and developing the system: the contact import and call queue, the training content flow, the call logic and callback scheduling, the sales alerting, and the record store for recordings and transcripts.

In progress. Work started on 8 October 2026. The voice and conversation layer currently runs on Google's Gemini Live; the model choice will be reviewed against running cost before production. This page will be updated as the system moves from build to live use; no results or performance figures are claimed yet.