An AI voice agent is a virtual representative that talks on the phone. It calls your contacts or answers the people who call your company, listens to the reply, understands what was meant and responds on the spot, with a natural voice. It is not a recording, it is not a "press 1" menu and it is not a voicebot reading canned lines: every answer is shaped by what the person just said.

In practice, Vozis Voice takes on the work that fills hours of your team's day: calling leads back, qualifying prospects, confirming appointments, sending payment reminders, winning back inactive customers and answering calls after hours. All with the same tone of voice, the same script and the same care, from the first call to the thousandth.

What sets Vozis apart is how we deliver. You don't get software to learn and configure. Our team builds the agent, writes the script, connects your systems, tests everything and tracks the results. You see it all in the platform and decide what comes next.

Every call follows the same path, from the moment a contact enters the operation to what happens after they hang up. Knowing this path shows you where each setting makes a difference:

1. The contact comes in. A lead fills out a form, joins a campaign or calls your company's number.
2. The agent calls or answers. With the voice, language and display name you chose.
3. The conversation happens. The agent follows the script, draws on your company's knowledge and handles questions and objections naturally.
4. The data is captured. While it talks, the AI records the details that matter, like objections and the best time for a callback.
5. The call is classified. Meeting booked, interested, qualified, disqualified or whatever makes sense for your business.
6. The next step fires on its own. A WhatsApp message, a text, a CRM update or a new attempt through the cadence flow.

Every agent starts with a simple choice: will it make calls or take calls?

• Outbound (you call them): the agent calls your contacts. Think lead follow-up within seconds, outbound calling for prospecting, appointment confirmations and collections.
• Inbound (they call you): the agent answers the people who call your company. It replaces the traditional IVR and handles inbound calls 24 hours a day, with no queue and no menu.

The same platform also builds chat agents for WhatsApp, which share the same tone of voice and the same knowledge. So someone who started the conversation on a call can pick it up by message without repeating a thing.

The identity settings define how the agent introduces itself. There are four fields:

• Role: the agent's job in the operation, for example "Northgate Realty Agent" or "Appointment confirmation". It is the internal name, which keeps things organized when a company runs several agents.
• Display name: what the agent calls itself on the call. "Hi, this is Julia from Northgate."
• Type: outbound or inbound.
• Language: the language of the conversation. Vozis Voice speaks more than 30 languages, always with a native voice and a natural accent, including English, Brazilian Portuguese, Arabic and Hindi.

Calling businesses with an IVR

When the agent calls clinics, offices or other businesses, sometimes a person picks up and sometimes it is an automated menu ("press 1 for..."). The May be answered by an IVR option prepares the agent to recognize the menu, choose the right option and only then start talking with a person.

Behind the voice is a large language model, the same family of technology behind the best-known AI assistants. It understands what the customer says and decides how to respond. The Vozis team picks the provider and model for each operation, balancing answer quality and speed. On a phone call, speed is everything: the model in the example below responds in about 700 milliseconds, roughly the time a person takes to think before speaking.

Two fine-tuning settings round it out:

• Max tokens: the maximum length of each reply. On a call, short replies sound more natural, so this limit is usually kept low.
• Temperature: how much the agent varies the way it speaks. Lower, it is more predictable and sticks closer to the script. Higher, it sounds more spontaneous.

The voice is the first thing the customer notices. That is why it gets the same care as a professional recording. The team chooses between female and male voices, listens to each one with the preview button and sets the background sound: a light office ambience, for example, makes the call feel like a real conversation and avoids the artificial silence between sentences.

Then come the voice settings, each with a clear purpose:

• Voice model (v1 or v2): the generation of the voice technology. v2 is the newest and most natural.
• Speaking speed: from slower to faster. Different audiences call for different paces.
• Voice stability: how much the intonation varies. Less stable sounds more expressive; more stable sounds more consistent.
• Clarity and similarity: how faithful the voice stays to the original tone, with no distortion.

The English voices are named Chloe, Sophia, Grace, Hannah and Lily, and Ryan, Gabriel, Lucas, Jake and Thomas. For other languages, the voice is chosen and tested with the client before going live.

The voice can be yours alone. Beyond our ready-made voices, Vozis creates an exclusive voice for your company. It can be the voice your brand already uses in its campaigns, a spokesperson's or the business owner's own voice, cloned from a short recording and always with the permission of the person lending it. Your customers hear the same voice in your ads, on your service line and in every follow-up.

The accent, too. The agent can sound like your region: American, British or Australian English, or, in Portuguese, the accents of Rio, São Paulo, Minas Gerais, the South or the Northeast. The voice feels familiar to whoever picks up, and the conversation starts with trust.

A natural conversation depends on details almost nobody notices, until they are missing: when to speak, when to wait, what to do when the person interrupts or goes quiet. Conversation dynamics controls exactly that.

Call length and end of speech

The maximum call duration ranges from 10 to 60 minutes and prevents endless calls. Smart end-of-speech detection lets the AI tell on its own when the person has finished speaking, instead of waiting a fixed amount of time. When it is turned off, the wait time before speaking takes over, from 0.3 to 5 seconds.

Interruptions

People interrupt, and the agent needs to handle that gracefully. The words to interrupt setting defines how many words the customer has to say for the agent to stop and listen, so a simple "uh-huh" doesn't cut it off mid-sentence. The wait after interruption sets how long the agent pauses before picking up again.

Silence

If the customer goes quiet, the agent waits for the silence before check-in time (5 to 30 seconds) and then says something like "Hello?". The max repetitions setting limits how many times it tries before ending the call.

Opening and closing

The start mode defines who speaks first: on outbound calls, usually the agent; on inbound calls, it may be the other way around. The end-call trigger is the phrase that signals the end of the conversation, such as "Talk soon".

Knowledge is the text that teaches the agent to be part of your company. This is where it learns who it is, who it works for, what the call should achieve and how to speak. Good knowledge usually runs 5,000 to 10,000 characters and is organized in blocks:

• Identity and positioning: name, company, role and speaking style, including everyday expressions that make the conversation feel more human.
• Goal: the ideal outcome of the call and the priorities, like booking a visit and, if that doesn't work, sending material over WhatsApp.
• Company information: industry, region, hours, address and channels.
• Lead qualification: what the agent needs to find out, like the purpose of the purchase, price range and urgency.
• Differentiators and talking points: what to highlight and how.
• Objection handling: what to say when the customer says "I don't have time", "it's too expensive" or "just send it to me on WhatsApp".

This text is fully customized. Every Vozis client gets knowledge written by our team, based on meetings, company materials and the first real calls, and it is refined with every results cycle.

If knowledge is what the agent knows, the script is the order in which the conversation unfolds. It starts with the opening greeting, the first line of the call, which can already use the contact's name through a variable like {{first_name}}. Then come the stages: opening, qualification, presentation, offer and close.

Each stage includes instructions for the agent (like "build rapport first, then explain why you're calling"), sample lines and [WAIT FOR REPLY] markers, which remind the agent to listen before moving on. The script also maps out different paths: if the customer says yes, it moves to the next stage; if they say no, the agent uses the response prepared for that case.

The script doesn't make the conversation rigid. The agent follows the order, but it can step away to answer a question using its knowledge and then return to where it left off. That is how a call keeps a clear goal without feeling like a questionnaire.

Your agent can also reach into your other systems while the call is happening. These live connections are called tools. With them, it checks an external system and answers with real information: open slots on the calendar, an order status, an invoice balance, whether a property is still available. It can also log something in your CRM or ERP without waiting for the call to end.

Each tool is linked to the agent and referenced in the script with @, at the exact point in the conversation where it should be used. The Vozis team builds these integrations to fit your business, with CRMs like HubSpot, Salesforce, Pipedrive and Zoho, calendars like Google Calendar, and your company's own systems.

The knowledge base stores the materials the company wants its agents to use: catalogs, rules, hours, menus, photos and PDFs. It is organized into groups, like "Property Catalog", "General" or "Menu", and each group can be linked to one or more agents. That way, the same material serves both phone calls and WhatsApp without being added twice.

There are two kinds of item:

• Document: free text that goes into the agent's knowledge. Use it for business hours, mandatory rules ("never promise an extra discount"), policies and FAQs.
• Media: PDFs and photos. Each media item has a description that tells the agent when to use it, for example "send when the customer wants to see photos of the house". If the customer asks for it during the call, the agent sends the file over WhatsApp.

For each item, the team sets which channels it applies to (voice agents, WhatsApp agents or both) and whether it is active.

An operation running hundreds of calls a day can't depend on someone listening to every recording. Data collection takes care of that: you tell the AI what information you want from each conversation, and it fills it in on its own at the end of every call.

Each piece of information is a property, with a name and a plain description of what should be recorded. For example:

• objections: the objections the customer raised during the call.
• follow_up_date: the date and time the customer asked to be contacted, if any.

Other common examples: interest in the product, available budget, preferred contact channel, reason for canceling, satisfaction score. These answers show up as columns in the campaign, ready to filter and export, and they can flow straight into your CRM.

At the end of every call, the agent classifies the outcome. Classifications are created for each operation, with a name, a color and a description that tells the agent when to use each one. A few examples:

• Meeting booked: the customer agreed and the meeting went on the calendar.
• Interested: showed interest and moves on to the next stage.
• Qualified: met the qualification criteria.
• Disqualified: does not meet the criteria, and the call was ended.
• Possibly interested: showed interest, but the call dropped before the end.
• Callback requested: couldn't talk and asked to be called back at another time.

Classification is what connects the conversation to the rest of the operation. It drives the campaign results, tells the cadence flow what to do next and lets automated actions know who should get a message.

Actions define what the agent does when the call ends. The most common are sending a WhatsApp message and sending a text (SMS). Each action has filters: it can fire for a status (for example, unanswered calls) or for a classification (for example, "Interested").

The message can use variables from the conversation, like {{first_name}}, as well as the data captured on the call. For example: "Hi, {{first_name}}! As mentioned on the call, I'm sending you this message. Shall we continue the conversation here?" That way, the customer gets the follow-up on WhatsApp seconds after hanging up, while the topic is still fresh.

No agent goes straight to your customers. Next to the settings sits a test panel: the team picks the call route, enters a destination phone number and the variable values (like first name) and clicks Call. The phone rings and the conversation really happens, just as it will with the customer. The same panel also tests the agent over chat.

Every change starts as a draft. Only once the test is approved is it published as a new version. The version history shows what changed and lets you roll back to an earlier version, so the agent can be fine-tuned with no risk to the operation.

The call route matters too. The Vozis Phone route uses carrier-assigned numbers to reduce the risk of the call being flagged as unwanted, and your company can also call from its own numbers.

A campaign is a set of contacts an agent will call. It is how Vozis Voice makes hundreds or thousands of AI phone calls: the list goes into the campaign (by spreadsheet, from the CRM or through the contacts API) and the agent calls within the business hours you set, for example Monday to Saturday, 8 am to 7 pm.

In the campaign list, each row shows the agent in charge, the hours, the status (active or paused) and the progress: how many contacts have been worked and how many attempts were made. From the action buttons, the team can monitor, pause or resume the campaign, download the report, check the contacts API, edit settings and redial anyone who didn't answer.

Every contact in detail

When you open a campaign, each contact shows the call status (answered, unanswered, do-not-call list), the classification, an AI-written summary and the AI-collected data columns. You can listen to the recording, read the transcript and call a specific contact again. Anyone who asked not to be called goes on the do-not-call list and is never contacted again.

Not every contact picks up on the first call. The cadence flow organizes what happens next, mixing channels into a single path: calls, WhatsApp, RCS, waits and conditions. It is the operation's "if they don't answer, do this", laid out on a visual board.

A flow starts at Contact enters the flow, and each step is added with a click:

• Call: the voice agent calls, with a maximum number of attempts, a wait between retries and a set calling window.
• WhatsApp: sends a Meta-approved message template and tracks the reply.
• RCS: an official branded message with the company's name, logo and verified badge.
• Condition: a gate before the next step, like "no answer" or "no WhatsApp reply within 1 hour".
• Wait: a scheduled pause, for example waiting 24 hours before the next message.
• Webhook: notifies your CRM, ERP or other system when the contact is classified.

A common example: the agent calls; if the contact doesn't answer, they get a WhatsApp welcome message; if they don't reply within an hour, they get another call at the end of the day. Waits respect the customer's messaging hours and never fire early.

An AI voice agent delivers the most where there is a high volume of repetitive calls and every minute of delay costs money:

• Lead follow-up: the lead fills out the form and gets a call within seconds, while interest is still hot.
• Outbound prospecting: the agent calls your list, presents the offer, qualifies the prospect and books the meeting on the sales rep's calendar.
• Answering without an IVR: callers are helped right away, through a conversation instead of a menu of options.
• Scheduling and confirmations: appointments, visits and meetings confirmed by voice, with fewer no-shows.
• Collections and payment plans: payment reminders and negotiation handled with courtesy, without awkwardness.
• Customer reactivation: inactive customers get a call with a reason to come back.
• Satisfaction surveys: every customer's feedback, captured right after service.

The industries that use it most are healthcare (clinics, hospitals and health plans), real estate agencies and developers, and financial services and collections companies. But any sales or customer service operation with frequent calls can benefit.