All articles
Artificial intelligenceNovember 12, 202610 min read

AI voice agent for sales and support calls: how it works, what it costs to run, and when it pays off

How an AI voice agent differs from an IVR and a chatbot, which calls it can be trusted with, what the running cost is made of, and how to measure a pilot.

An AI voice agent is software that holds a phone conversation by itself: it hears the caller, understands what was said, decides what to do and replies in a synthesized voice. It pays off on frequent, repeatable calls with a clear goal, such as confirming an order, reminding about an appointment or qualifying a lead. Its running cost is the sum of telephony minutes, speech recognition, language-model usage and voice synthesis, and hard negotiations are something it should pass to a person rather than try to win.

What is an AI voice agent, and how is it different from an IVR or a chatbot?

A classic IVR is a menu: "press 1 for sales, press 2 for support". It routes calls, but it does not understand free speech and cannot answer a question that is not in its tree.

A chatbot understands free-form questions, but it lives in a messenger or a website widget, where the customer can reread the answer and reply whenever convenient. We covered that format in the article on support chatbots.

A voice agent does the chatbot's job over the phone and in real time. That changes the engineering: there is no screen to show options on, the person may interrupt at any moment, the line may be noisy, and a pause of a couple of seconds already feels like a dropped call.

How does a voice agent work under the hood?

There are two common designs.

A pipeline of three steps. Speech recognition turns the caller's words into text, a language model reads that text together with the script and the conversation history and writes a reply, and speech synthesis voices it. Each part can be chosen and replaced separately, and because everything passes through text, it is easy to log, test and constrain what the agent says. The price is delay: every step adds its own, and intonation is lost once speech becomes text.

A speech-to-speech model. One model takes audio in and produces audio out. It usually reacts faster and sounds more alive, because it hears tone and not only words. In return you get less control over exact wording, and the choice of voices and languages is whatever the model's provider offers.

Neither design wins everywhere. Where exact wording matters (prices, terms, legal phrases), the pipeline is easier to keep in line. Where the conversation is loose and speed matters most, speech-to-speech may fit better. This is decided on test calls, not on paper.

What makes a call with an AI agent feel natural?

Mostly not the voice. Modern synthesized voices are rarely the problem; calls are spoiled by timing.

  • Response latency. People expect an answer almost at once. A long pause after every phrase makes the caller repeat themselves or hang up, so the whole chain is tuned for speed, sometimes at the cost of a smaller model.
  • Interruptions. When the person starts talking over the agent, it has to stop, listen and answer the new phrase instead of finishing its paragraph.
  • Silence and noise. The agent must tell the end of a phrase from a pause for thought, and a caller from a TV in the background. Cutting in too early is as bad as waiting too long.
  • Short turns. One or two sentences at a time, one question at a time. A text that reads well on screen sounds like a lecture on the phone.
  • Honest misunderstanding. Asking to repeat once is fine. Guessing and moving on with a wrong answer is not.

Which calls does a voice agent handle well?

Calls with a clear goal, a limited set of outcomes and a volume that makes manual dialing expensive:

  • Order confirmation: check the items, address and delivery time, record the changes.
  • Appointment reminders: confirm, reschedule or cancel a visit.
  • First-pass outbound qualification: find out whether the person is interested at all, then pass the interested ones to a salesperson.
  • Callbacks: return a missed call or call at the time the customer asked for.
  • After-hours inbound: answer at night and on weekends, take the request, resolve simple ones and schedule the rest for working hours.

In each of these cases the agent does not replace the sales team; it takes over the dialing that people have no time for.

Where does a voice agent not work?

  • Complex negotiation: individual terms, bargaining, several decision makers.
  • Emotionally charged cases: complaints, disputes, bad news, anything where the person needs to be heard rather than processed.
  • Calls with no script behind them: if your own staff cannot describe what a good call looks like, the agent will not find out either.

Here the sensible role for the agent is to collect the facts and bring in a person quickly.

How does the handover to a human work?

The agent needs explicit conditions for stepping aside: the caller asks for a person, the question is outside the script, the agent failed to understand twice in a row, or the conversation is getting heated.

There are two ways to hand over. A live transfer connects the caller to an operator during the same call. A scheduled callback creates a task for a manager with a short summary. In both cases the person who picks it up should see what has already been said, so the customer does not have to start from the beginning.

What ends up in the CRM after a call?

Every call should finish with a record, otherwise the work is invisible. Typically it is the outcome (confirmed, rescheduled, declined, no answer, transferred), the details gathered, the next step with a date, and a recording or transcript.

The outcome goes into the customer card and moves the deal along the pipeline, so managers keep working in the system they already use. How to connect telephony, CRM and messengers into one flow is described in our article on CRM integration.

Can one agent speak several languages?

Yes, but each language is a separate piece of work. Recognition quality and available voices differ from language to language, and a script translated word for word rarely sounds right. The practical approach is to set the language by the customer's country or the number dialed, adapt the script with a native speaker, and test each language on real calls. Switching language in the middle of a call is possible, but less reliable than choosing it up front.

What does the running cost consist of?

Besides the one-off development, there are four ongoing components:

  1. Telephony: minutes of connection, which depend on the country, the carrier and the type of number.
  2. Speech recognition: billed by the amount of audio processed.
  3. The language model: billed by the volume of text it reads and writes. A long script that is resent on every turn costs more than a compact one.
  4. Voice synthesis: billed by the amount of speech generated.

Three things drive the total: the average length of a call, the share of calls that are actually answered, and the models chosen. That is why the useful figure is not the price of a minute but the cost of one result: one confirmed order or one qualified lead. Provider prices change, so we estimate them for the specific task before a pilot rather than quote a universal number.

What do you need to prepare?

  • A script or knowledge base: what the agent says, what it may promise and what it must not. Recordings of your best calls help a lot.
  • Telephony access: a number and a connection to your carrier or PBX.
  • A CRM, or at least a spreadsheet that contacts come from and outcomes go to.
  • People for test calls: someone on your side who knows the product and can say "a customer would never accept that answer".

How is a pilot run and measured?

A pilot takes one scenario and a limited list of contacts. First come internal test calls, then a small batch of real ones with every conversation reviewed, then the volume grows step by step.

Three numbers are agreed before the start:

  • Conversion to the target action: the share of calls that end with a confirmed order, a booked visit or a qualified lead.
  • Share of calls completed without a human: how often the agent closes the conversation by itself.
  • Cost per result: all running costs divided by the number of target actions.

They are compared with what you have now: the same calls made by people, or not made at all. For reviewing conversations at scale, the approach from our article on AI call analytics applies.

What about consent and calling rules?

Rules differ by country, so check the local law before launch, ideally with a lawyer. In general terms: where it is required, tell the caller at the start that they are talking to an automated assistant; call only within the hours allowed locally; follow the rules on recording calls and on processing personal data; and respect a request not to be called again. And apart from the law: a caller who was told up front has no reason to feel deceived later.

What mistakes are typical?

  • Launching on the whole contact base at once instead of a small batch. Errors found on the first calls then cost you the entire list.
  • A script written for reading, with long sentences and several questions in one turn.
  • No way out to a human. A caller stuck in a loop remembers it.
  • Measuring minutes, not results. A cheap call that confirms nothing is not cheap.
  • Nobody listens to the calls after launch. Scripts, prices and the way customers phrase things change, and the agent has to be kept up to date.

Where to start

Pick one type of call that your team makes often and by the same script, and count how many of them happen per month. With that, tell us about the task: we build such agents as part of our AI integration work and will suggest what a pilot could look like. If your customers prefer text to calls, see also the article on WhatsApp Business automation.

FAQ

Can callers tell they are talking to an AI?

Often yes, especially in a longer conversation, and it is better not to hide it. The aim is not to imitate a human but to solve the task quickly and politely.

Will a voice agent replace call center operators?

It takes over routine calls: confirmations, reminders, first-pass outbound. Complex and emotional conversations, and everything the agent hands over, stay with people.

How long does it take to launch?

It depends on how ready the script, telephony access and CRM are. A single well-defined scenario launches much faster than a general-purpose agent, which is why we start with one.

What does the agent do when it does not understand the caller?

It asks once to repeat, and if that fails again, it transfers the call to an operator or creates a callback task. Such calls are flagged for review.

Can the agent work with our existing number and PBX?

In most cases yes: the agent connects to your carrier or PBX as one more line. The details depend on your provider and are checked at the start of the project.

Can the agent handle many calls at the same time?

Yes. The number of parallel calls is limited by telephony channels and model provider limits, not by the agent itself, so volume grows without hiring.

Shall we discuss your project?

Tell us about it — we'll come back with an estimate and a proposal.

WhatsAppTelegram