YEMPIK / STUDIOSoftware, projects and training.

An AI voice agent for sales: a project for an energy call center

The callback was only the start. The conversation also had to draw on defined products and prices, sales scripts, objection handling, and careful voice-quality work.

By YempikPublished Updated 7 min read

The case in brief

Yempik built a voice agent for the first sales contact at an energy-sector call center. After an online enquiry, the agent started a call, checked the person’s interest, and collected the information needed for the next step. The project included a structured source for the agent to retrieve products and prices, sales scripts, objection handling, and voice refinement. An operator remained part of the process to continue the sales conversation.

Sector
Energy · customer anonymised
Yempik’s work
Conversation, sales information, and voice
Workflow trigger
Online enquiry and callback
Human involvement
The operator continued the sales conversation

Figures reported for the project

The original project page reported the following operating figures. They refer to the configuration built for this call center; they are not performance guarantees for every voice agent.

FigureWhat it refers to
About 30 secondsFrom an online enquiry to the start of the first call.
5–8 callsConversations handled at the same time.
More than 1,000 contactsContacts handled in production.
−70%Reduction in screening work for operators, as reported by Yempik.

A clear sales task before handing over to an operator

The starting point was someone who had submitted an enquiry through an online campaign. The team needed to call them back, check whether they were still interested, and collect the information needed for the next step.

The project involved more than placing a call. During the first contact, people might ask about products or prices, or raise sales objections. The agent had to address those questions while staying consistent with the offer and the purpose of the call.

The operator continued the sales conversation. A useful handoff depended in part on the information gathered beforehand: a well-delivered call helps little if the person taking over has to start again from scratch.

Prices needed to come from a defined source

Products and prices had already been defined. We therefore set up a structure from which agents could retrieve commercial information, instead of relying on the model’s generation alone.

The principle is to separate what an agent can explain from what it can offer. The model can phrase an explanation; the product, price, and terms must match the available offer.

This constraint changes how the conversation is designed. If the agent lacks information needed to give a price, it should ask for it or pass the question on. A plausible figure is not enough: it has to be the one that applies to the situation.

The same choice matters in other sectors. An agent for a distributor, a hotel, or a professional service should retrieve terms from the designated source, without turning a plausible answer into an invented offer.

A sales script gives the conversation direction

We provided the agents with sales scripts. They needed a guide for opening the first conversation, presenting the offer, and collecting the required information in a way that remained easy to follow.

A script for a voice agent also requires decisions about the sequence: which questions to ask first, when to explore a point, and when to stop. Putting a long sales text in the prompt is not enough to define those choices.

For a similar system, it helps to distinguish three things: information that must not change, the order of the questions, and wording that can adapt to the person’s answers.

ElementWhat it needs to define
Sales informationProducts, prices, and terms retrieved from the designated source.
Call outlineThe goal, useful questions, and transition to the next step.
WordingHow to explain and ask follow-up questions without changing the offer.

Handling an objection means more than repeating the offer

Part of the work focused on handling sales objections. The agent needed to keep the conversation on track when a person did not follow the script’s ideal path.

A concern about price, a request for time to think, and a lack of interest call for different responses. These are examples of situations to consider in the design: clarifying a term is not the same as pushing after a refusal.

The response has to stay within the boundaries of the offer. An objection cannot be overcome by inventing a discount, a term, or a benefit. The agent should instead work out whether information is missing, an operator should step in, or the conversation should end.

This work applies to commercial agents in any sector: before optimising how much an agent can say, define what it may promise and when it should make room for a person.

Voice is part of how the agent behaves

We put substantial work into voice quality. What the agent says and how it says it arrive together: a correct explanation can be hard to follow if the pace is wrong or prices and terms are unclear.

In a sales conversation, it is useful to listen closely to the sensitive moments: the opening, a follow-up question, a price explanation, and a response to an objection. Voice should be assessed in dialogue, not just through a demo sentence.

Useful checks include whether numbers and names are clear, answer length, pauses, response time, and what happens when the person interrupts. These should be reviewed alongside the sales content: a pleasant voice alone does not make an agent reliable.

How the parts of the system fit together

The project joined an initial trigger, a sales conversation, and a handoff to a person. The price list and script were not finishing touches; they helped define what the agent needed to do during the call.

  1. The person submits an online enquiry and the callback starts.
  2. The agent opens the conversation and checks the person’s interest.
  3. Products and prices are retrieved from the prepared structure.
  4. The script guides the questions; objection handling helps address answers outside the ideal path.
  5. The operator receives the contact and the information gathered, then continues the conversation.

What to carry into a voice-agent project in another sector

The energy sector determines the products and terms of the offer. The design work described here addresses challenges shared by many commercial voice agents: accessing accurate information, following a conversation, responding to questions, and working with people.

A useful starting point is to gather the price list and sales rules, the scripts already in use, common objections, and the conditions for handing a conversation to an operator. These materials help define the agent’s behaviour before deciding how it should speak.

Then review the whole process. How long does the callback take? Is the information given accurate? Can the person make themselves understood? Does the operator receive what they need? These questions are more useful than judging only how human the voice sounds.

Frequently asked questions

Could the agent invent prices or discounts?

The project had defined products and prices, with a structure for retrieving them. The conversation had to stay within the available sales information and could not create new terms.

Is it enough to give a voice agent a script?

A script is one part of the work. The agent also needs access to sales information, objection handling, criteria for handing over to an operator, and voice checks in dialogue.

Does this approach work outside the energy sector?

Yes. Separating sales information from how an answer is phrased, handling objections, and defining the handoff to a person are common needs. Products, rules, and conversations still need to be designed for each context.

Did the agent complete the whole sales process on its own?

The case covers the first contact and qualification, including questions and objections. The operator remained involved in continuing the sales conversation.

Sources and method

  • Project account and Yempik operating brief

    Updated in September 2026 from the team’s account and the earlier project brief. The figures are reported by Yempik and are not independently measured. Customer anonymised.