When the entire calling operation rests on operators, growing the base runs into the number of people on shift. A company either expands its staff or leaves part of its contacts unprocessed. Let’s look at the case of a client who took a third path and built an AI contact center on top of their existing telephony: they handed typical calls to a voice AI agent and left operators the ones where a human is really needed.

The Starting Point: Operators Made One Hundred Percent of the Calls

The client is an insurance company that was already using NovaTalks telephony and running outbound campaigns through the Outbound Dialer. There are twelve operators in the contact center, and they handled all calls, inbound and outbound, manually.

• The inbound line. Customers called with typical questions: working hours, policy terms, application status, booking a car inspection. In the supervisor’s estimate, roughly a third of such requests had the same answer for everyone and didn’t require a person.
• Outbound campaigns. The main one is a reminder about a policy’s expiration with a renewal offer. Every month a base of about eight thousand contacts went into the work, while the whole team managed to process around two thousand a week. Because of this, the full calling cycle stretched to almost a month, and some customers received the reminder after their policy had already expired.
• It turned into a vicious circle. To call faster, you need additional operators, but their time would again go both to reference questions and to mechanically dialing numbers. That’s exactly why the company started looking for which part of the calls could be handed to automation without losing conversation quality.

What Exactly Was Automated

Instead of automating everything indiscriminately, the call flows were divided on one principle: the typical and predictable goes to AI, situations with a human factor are left to operators.

Call type	Before integration	After integration
Reference questions on the inbound line	Operator	Voice AI agent
A direct “connect me to an operator” request	Operator	The AI agent transfers to an operator
Policy expiration reminders	Operator	Automatic calling without an operator
Initial interest qualification	Operator	Voice AI agent
Discussing terms, calculating the cost	Operator	Operator
Complaints and non-standard situations	Operator	Operator

Table 1. Distribution of call types between the AI agent and operators

How the Solution Is Built

Technically, this is an integration of the NovaTalks platform with the ElevenLabs service, which is responsible for speech synthesis and understanding the other party’s replies. The calls go through NovaTalks telephony, and the outbound campaigns are launched by the Outbound Dialer – both components are mandatory for the scheme to work.

Setting up the voice AI agent requires no development. Its behavior, role, and scope of competence are set by text instructions, so the scenario can be changed in a few minutes.

• Prompt control: a text instruction describes the AI agent’s role, its tasks, and the topics it immediately hands off to an operator.
• Knowledge base: internal documents and FAQs are connected as reference context from which the AI agent draws answers.
• Access to external systems: through an API, the AI agent gets data from the client’s CRM: name, application status, policy expiration date.
• Binding scenarios to events: each campaign has its own response logic configured – separately for policy renewals, separately for promotional offers.
• Real-time AI call transcription: the conversation is turned into text and lands in the platform interface during the call or after it ends.
• A WebRTC phone in the browser: the scenario can be tested right after setup, without separate equipment.

What the AI Agent Can Do in a Conversation

The key objection to voice automation sounds like this: the customer immediately hears a robot and hangs up. That’s exactly why, at the testing stage, one of the key criteria was the naturalness of the dialogue.

• Voice quality: natural speech in Ukrainian, as well as English, French, Dutch, and other languages.
• Interrupting the conversation: the customer can cut the AI agent off mid-sentence, and it stops, processes the new reply, as in a conversation with a person.
• Tone adaptation: the AI agent recognizes the emotional coloring of a reply and responds more reservedly or more empathetically depending on the situation.
• Complex names and numbers: car makes and models like Toyota Camry, dates, amounts, and policy numbers are pronounced correctly, without distortions.

The Inbound Line: What Changed

AI processing of inbound calls became the first point of contact. The AI agent answers typical questions, and when the customer asks for a person or the topic goes beyond the scenario, it transfers the call to a free operator.

The conversation is automatically turned into text and available in the platform interface already during the call, so the operator sees what the customer talked about with the AI agent.

In the first month of operation, the AI agent independently closed about 35% of inbound requests – the same reference part that operators used to handle. The rest of the calls reached an operator already with context: they saw what the customer had reached out about.

Outbound Campaigns: What Changed

The renewal campaign was moved into an automatic-calling format without operators. The AI agent calls, reminds about the policy’s expiration, answers the first questions, and records the reaction: interested now, call back later, or a refusal. If the customer asks to be contacted at another time, the AI agent schedules a repeat call.

The AI agent started processing the base of eight thousand contacts in two to three days instead of four weeks. The operator spoke only with customers who themselves confirmed readiness to continue – about every fifth contact. Instead of mechanically dialing numbers, the operator received ready conversations with people who already know what it’s about.

Results

• About 35% of inbound requests are closed without an operator.
• The full renewal campaign cycle shrank from four weeks to two or three days.
• About 20% of the base moves to operators – these are contacts who themselves confirmed readiness to talk further.
• Policy expiration reminders arrive on time, rather than after the term has ended.
• The contact center staff stayed the same, while the volume of the processed base grew several times over.

Who This Format Suits

The scheme is optimal for small and medium businesses, targeted outbound campaigns, and quick hypothesis testing – where hiring additional operators isn’t worthwhile.

Before launch, a few conditions are worth taking into account.

• NovaTalks telephony and the Outbound Dialer are required – the integration doesn’t work without them. There is no minimum call volume, though.
• The client creates an ElevenLabs account themselves and pays for conversation minutes directly to the vendor at its rates.
• The functionality works only in the voice channel.
• The logic of accessing external knowledge bases and web services requires additional testing for a specific scenario.

Frequently Asked Questions

Does the voice AI agent replace operators?

No. It takes on typical and predictable calls: reference questions, reminders, initial qualification. Complex conversations, calculations, and complaints remain with a human, and the AI agent transfers such calls to an operator.

Will the customer realize they’re talking to a robot?

The speech sounds natural, the AI agent holds pauses and lets itself be interrupted. It’s not worth hiding the automation: it’s more correct to note at the start of the conversation that a virtual assistant is answering and to give the option to switch to an operator.

How do you implement AI in a contact center?

No complex coding is needed. The AI agent’s behavior is set by text instructions, reference context is pulled from a knowledge base, and customer data through an API from the CRM. The finished scenario is tested with calls through the built-in WebRTC phone and adjusted based on listening to the results.

Which languages does the AI agent speak?

Ukrainian, as well as English, French, Dutch, and other languages. For multilingual support, there’s no need to build a separate scenario for each language.

Does the AI agent see the customer’s data?

Yes, if you set up access to the CRM or another system through an API. The AI agent can address the customer by name, state the policy expiration date, order status, or balance.

How much does such automation cost?

The cost consists of two parts. Conversation minutes are paid directly to ElevenLabs at their rates, and the client creates their own account. On the NovaTalks side, telephony and the Outbound Dialer are needed, and their cost depends on the configuration.

Conclusion

The case shows a simple logic: the predictable part of the calling operation is what is subject to automation. Reference questions and reminders are the same for the whole base, so a voice AI agent closes them. Conversations where arguments, calculation, and handling objections are needed remain with operators – and it is for those that time is freed up.

The company doesn’t change its platform or expand its staff in the process: the telephony and the dialer are already working, and the voice AI agent is added as one more layer of automation on top of the existing scheme.

To understand which part of the load in your contact center can be handed to an AI agent, contact the NovaTalks team – we’ll assess which requests recur most often and suggest which scenario to start with.