---
title: "A voice agent for an energy call centre: the value is in the price list and the human handover."
url: "https://simonebova.com/en/runs/agente-vocale-energia"
lang: "en"
author: Simone Bova
alternates:
  it: "https://simonebova.com/md/it/runs/agente-vocale-energia"
  en: "https://simonebova.com/md/en/runs/agente-vocale-energia"
  pl: "https://simonebova.com/md/pl/runs/agente-vocale-energia"
---

# A voice agent for an energy call centre: the value is in the price list and the human handover.

A voice agent that calls an energy call centre's leads back in about 30 seconds. What matters beyond the voice: real prices, objections, the human handover.

Name: Voice agent · energy

Terrain: SME · call centre

When: 2026

Role: Technical partner, with Raffaele Zarrelli

| Fact | Value |
| --- | --- |
| Callback | About 30 seconds |
| Contacts | More than a thousand |
| Prices | From the price list |
| Closes | A human agent |

## The situation

An energy call centre got leads from online campaigns, and a lead goes cold fast. Someone had to call back right away, check they were still interested, answer questions and objections, and hand them to a human agent along with everything they'd already said. The human agents needed to spend their time selling, rather than making the first call to everyone.

## What I did

With Raffaele we built the voice agent for that first call. It dials about 30 seconds after the online request, qualifies the lead and hands it to a human agent, who closes the sale. We decided early what the agent would never do on its own: choose whom to call, make up a price, push harder than it should.

The longest work went into what a demo doesn't show: prices taken from the price list, a planned way out of every objection, what goes into the handover, and the silences in the conversation, which weigh more on the phone than in a chat.

## What changed

In production the agent handled more than a thousand contacts, with several calls running in parallel. Every lead gets a response in about half a minute and reaches a human agent already qualified. Two things were left to do, and we wrote them down: continuous checks on answer quality, and a filter for personal data that ends up in transcripts.

## What I learned

The agent called back within about half a minute, and from the outside it looked like an autonomous agent chasing leads. It couldn't call anyone on its own initiative. Every call started from a request that had really come in from the campaigns, and the system set the recipient and the topic before the model said a word. A fast callback and an agent free to choose whom to call are two separate choices, and here the second one served no purpose. The model was left only the conversation. It's the first thing we pin down when we design an agent that makes calls: which event can start one.

Products and prices were already set, and the agent looked them up in the price list every time. If a detail needed for the right price was missing, it asked or handed the call over. A script and instructions alone weren't enough: on a call nobody can reread, and a plausible but wrong figure becomes an offer. Objections worked the same way. “Handles objections” is easy to read as “keeps pushing until the customer gives in”, and the easiest shortcut is an invented discount. There were three planned exits: clarify a condition, bring in a human agent, end the call politely. No commercial condition was born inside the conversation.

Here the agent made the first call and a human agent closed the sale, so the agent's work was judged by what the human received: interest, answers given, objections still open, no question to ask again. On a related project, a client told us the voice platform marked calls as transferred when no human agent had picked up. A transfer logged by the agent doesn't prove a person received it. Since then we define what the human receives before we choose the voice, and for us a transfer counts only when someone on the other end has answered.

Looking back over the project, Raffaele wrote down two things the numbers didn't show. Monitoring mostly tracked how many leads reached a human agent, and a good transfer rate can sit alongside wrong answers. And the agent didn't ask for personal data, yet a caller can read their bill out loud: name and address end up in the transcript without anyone asking for them. Since then those are the first two questions we ask of a voice agent before production: how we measure answer quality apart from transfers, and where the data that arrives uninvited ends up.

[The case on yempik.com](https://www.yempik.com/caso-call-center-energia)

[Runs](https://simonebova.com/md/en/runs)

[simone@yempik.ai](mailto:simone@yempik.ai) · [Book a call](https://calendar.app.google/LKDrQPKGV2EAFCMN9)