A chatbot is software that communicates with people through text or voice, using a conversational interface to answer questions, give information, or help someone complete a task. Not every chatbot uses AI. Traditional chatbots follow scripted rules and decision trees, while AI chatbots use natural language processing and large language models to understand what someone means and respond in a more natural way.

How does an AI chatbot work?

An AI chatbot reads or hears what someone types or says, then uses natural language processing to work out the intent behind it rather than just matching keywords. A large language model then generates a reply based on that intent, pulling in extra information from connected tools, documents, or APIs when it needs to. For voice specifically, this same process sits inside a speech to text and text to speech pipeline so the exchange happens out loud.

How are AI voices made?

Modern AI voices, or text to speech (TTS), are built in three stages: a text front end that cleans up the input and works out the individual sounds, acoustic modeling that plans timing, pitch, and stress, and a neural vocoder that generates the actual sound wave. The system is trained on recordings of a real person reading text aloud, paired with that text, but once trained it generates entirely new audio that was never actually recorded. That is different from older, robotic sounding TTS, which stitched together prerecorded clips of speech.

What is a voice AI agent?

A voice AI agent is a system that holds a spoken conversation with a person, usually over the phone, and can understand requests, answer questions, and take actions like booking an appointment or looking up an order. It runs on a real time pipeline: speech to text turns the caller's voice into text, a large language model decides what to do and drafts a reply, and text to speech turns that reply back into audio. Platforms such as Retell provide the infrastructure for building these agents.

How to make an AI voice?

Making an AI voice generally means training a text to speech model on a dataset of recorded speech paired with the matching transcript, so the model learns how that voice handles timing, pitch, and pronunciation. Developers building agents usually do not train a voice from scratch. Instead, they choose from prebuilt voices offered by a TTS or voice agent platform, write the prompts and logic the agent follows, and test it against real conversations before deploying it to phone lines or chat.

What is an AI voice agency?

An AI voice agency, or AI automation studio, is a company that designs, builds, and maintains voice and chat agents for other businesses rather than building the underlying technology itself. It typically works on top of existing platforms, handling the prompt writing, conversation design, integrations, testing, and ongoing maintenance that a working agent needs. WissaSystems is one example of this kind of studio, building and maintaining chatbots, voice agents, and related automation for businesses that want the result without building it in house.

retell

Retell AI Explained: How Voice and Chat Agents Work

2026-10-24 ·WissaSystems

Retell is a developer platform for building AI voice and chat agents: it bundles telephony, speech recognition, text to speech, and conversation logic into one toolkit, so a developer or an automation studio can design an agent, test it, and put it live without stitching together separate vendors for each piece. It is not a chatbot a business switches on out of the box. It is infrastructure that something else gets built on top of.

What Retell Actually Does

Retell AI lets a team build one agent and deploy it across phone calls, web chat, SMS, and API connections, instead of rebuilding the logic separately for each channel. It supports two ways to design a conversation: a single prompt that the underlying language model reasons over on its own, or a node based flow where each step is mapped out like a flowchart.

Retell runs the telephony, speech, and language model layers itself, along with built in analytics for reviewing calls after the fact. That means the people building on it are not separately managing a phone carrier, a speech recognition vendor, and a language model provider.

How an AI Voice Agent Actually Works

A voice agent built on a platform like Retell runs on a three stage pipeline that repeats, in real time, for the length of a call.

Speech to Text

As soon as a caller starts talking, automatic speech recognition (ASR) converts their audio into text, word by word, while they are still speaking.

The Language Model in the Middle

Once there is text, a large language model reads it, works out what the caller wants, and decides what to do next. That can mean calling a tool or an API: checking an order status, looking up a calendar slot, pulling up a customer record. The model then drafts a reply in plain language.

Turning the Reply Into Sound

The final stage turns that written reply into spoken audio using text to speech, before the caller hears anything back.

In a production system, these three stages overlap and stream instead of running one after another in a strict line. That overlap is a big part of what makes a voice agent feel like a real conversation instead of a phone tree. Timing matters too: a response under roughly 800 milliseconds tends to feel natural, while a pause stretching past about 1.5 seconds is often where callers start to hang up, according to NICE's technical breakdown of voice agent architecture.

How AI Voices Are Actually Made

The text to speech step is its own small pipeline, usually broken into three stages. First, a text front end cleans up the input and works out the individual sounds, or phonemes, that make up the words. Second, acoustic modeling plans the timing, pitch, and stress of the speech, deciding where it rises, pauses, or lands hard on a syllable. Third, a neural vocoder generates the actual sound wave from that plan, according to IBM's explainer on text to speech.

Modern neural TTS is trained on recordings of a real person reading text aloud, paired with the transcript of what they said. Once trained, the model is not playing back anything it heard. It generates new audio for sentences that voice never actually spoke. That is the real difference from older, robotic sounding TTS, which worked by stitching together small prerecorded clips of speech.

What Retell Costs in 2026

As of 2026, pricing pages for Retell show voice agents billed by the minute, roughly $0.07 to $0.31 per minute depending on the voice and model combination chosen, and chat agents billed per message, starting from around $0.002 per AI generated message. New accounts generally get a small amount of free credit, around $10, to test the platform before paying for anything.

Treat those figures as approximate. Usage based pricing on developer platforms changes often, so it is worth checking Retell's own pricing page directly before building a budget around it.

Chatbot or Voice Agent: What Is the Actual Difference

A chatbot, in the broadest sense, is software that communicates with people through text or voice using a conversational interface, to answer questions, share information, or help someone finish a task, according to IBM. Not every chatbot uses AI. Older, scripted chatbots follow fixed rules and decision trees: if the customer types X, show response Y. AI chatbots use natural language processing and large language models instead, so they can understand intent and reply in a more natural, flexible way.

A voice agent is the same underlying idea, spoken instead of typed, running through the speech to text, language model, text to speech pipeline described above. The channel changes. The core problem it solves, understanding what someone wants and responding usefully, does not.

Platforms vs the Studios That Build On Them

Retell, and platforms like it, hand developers the raw pieces: telephony, speech recognition, language model orchestration, text to speech, and analytics. Turning that into something a business can actually run on still means writing the prompts, mapping the conversation flows, connecting the right tools and data sources, testing edge cases, and maintaining the agent once it is live.

That ongoing work is usually the job of an AI automation studio rather than the platform itself. WissaSystems, for example, is a studio that designs, builds, and maintains AI chatbots and voice agents for other businesses, sometimes using platforms such as Retell as part of the underlying stack rather than being a platform itself. A platform is infrastructure. A studio is the team that turns that infrastructure into something a business can run on day one.

For a business deciding how to get started, the practical question is less about which platform runs on the backend and more about whether the result sounds right, handles real customers, and keeps working once it is in production. That holds true whether the underlying stack is Retell, something comparable, or a custom built pipeline.

Frequently asked questions

What is a chatbot?: A chatbot is software that communicates with people through text or voice, using a conversational interface to answer questions, give information, or help someone complete a task. Not every chatbot uses AI. Traditional chatbots follow scripted rules and decision trees, while AI chatbots use natural language processing and large language models to understand what someone means and respond in a more natural way.
How does an AI chatbot work?: An AI chatbot reads or hears what someone types or says, then uses natural language processing to work out the intent behind it rather than just matching keywords. A large language model then generates a reply based on that intent, pulling in extra information from connected tools, documents, or APIs when it needs to. For voice specifically, this same process sits inside a speech to text and text to speech pipeline so the exchange happens out loud.
How are AI voices made?: Modern AI voices, or text to speech (TTS), are built in three stages: a text front end that cleans up the input and works out the individual sounds, acoustic modeling that plans timing, pitch, and stress, and a neural vocoder that generates the actual sound wave. The system is trained on recordings of a real person reading text aloud, paired with that text, but once trained it generates entirely new audio that was never actually recorded. That is different from older, robotic sounding TTS, which stitched together prerecorded clips of speech.
What is a voice AI agent?: A voice AI agent is a system that holds a spoken conversation with a person, usually over the phone, and can understand requests, answer questions, and take actions like booking an appointment or looking up an order. It runs on a real time pipeline: speech to text turns the caller's voice into text, a large language model decides what to do and drafts a reply, and text to speech turns that reply back into audio. Platforms such as Retell provide the infrastructure for building these agents.
How to make an AI voice?: Making an AI voice generally means training a text to speech model on a dataset of recorded speech paired with the matching transcript, so the model learns how that voice handles timing, pitch, and pronunciation. Developers building agents usually do not train a voice from scratch. Instead, they choose from prebuilt voices offered by a TTS or voice agent platform, write the prompts and logic the agent follows, and test it against real conversations before deploying it to phone lines or chat.
What is an AI voice agency?: An AI voice agency, or AI automation studio, is a company that designs, builds, and maintains voice and chat agents for other businesses rather than building the underlying technology itself. It typically works on top of existing platforms, handling the prompt writing, conversation design, integrations, testing, and ongoing maintenance that a working agent needs. WissaSystems is one example of this kind of studio, building and maintaining chatbots, voice agents, and related automation for businesses that want the result without building it in house.