What is an LLM, how a large language model works and how an AI voice operator uses it
Sado AI team · · 5 min read
In short
An LLM (large language model) is an artificial intelligence program trained on huge amounts of text: it understands what a person wrote or said and writes a fitting reply. ChatGPT, Claude and Gemini run on such models. An LLM understands meaning but doesn’t know your prices or hours until you give them to it. In an AI voice operator the LLM is the “brain”: speech becomes text, the LLM picks a reply from the company’s facts, and the reply becomes voice again.
How an LLM works, in plain words
An LLM has seen billions of sentences and learned how words follow each other. Ask it something and it builds the best-fitting answer word by word, so it understands one question asked many ways:
- “How much?”, “Narxi qancha?”, “Сколько стоит?” are one question;
- “Not tomorrow, the day after?” — it gets that the day changed;
- it pulls out what matters: a name, a date, an address, how many people;
- it understands mixed Uzbek-Russian speech too.
An LLM doesn’t hear voices; it works with text only. On the phone it needs two more parts: speech recognition (ASR) turns voice into text, and speech synthesis (TTS) turns text into voice.
Example: a travel agency in Tashkent
A travel agency gets dozens of calls a day, each phrased differently (the name “Safar Vaqti” is invented):
- The caller says: “Turkey in August, two adults and a child of seven. Antalya, by the sea.”
- A phone menu or keyword bot gets nothing out of this; it needs “press 1”.
- The LLM breaks it down: Antalya, August, 2 adults and a child (7), by the sea.
- It looks for a match in the agency’s facts (tours, dates, prices). If there is none, it says a manager will call back.
What the agency gains: the whole request is written down in one line, and the manager doesn’t ask everything again.
What an LLM doesn’t know and where it goes wrong
- It doesn’t know your prices, address or offers; you give them to it as text.
- It can confidently make up what it doesn’t know (“hallucination”).
- It can’t see your free slots or stock: that data must come from your system.
- It doesn’t hear sound: the caller’s words must first become accurate text.
So a “bare” LLM is never put on the phone. It gets a strict rule: answer only from the company’s facts; if you don’t know, don’t invent.
Where the LLM sits inside an AI voice operator
| Part | What it does | On the call |
|---|---|---|
| Speech recognition (ASR) | Turns the caller’s voice into text | “Antalya in August…” becomes text |
| LLM | Understands meaning, picks the next sentence | Picks out place, month, people |
| Company facts | The LLM answers only from these | Tours, prices, opening hours |
| Voice (TTS) | Turns the reply into a natural voice | Reply in Malika’s or Shohruh’s voice |
| Portal | Keeps the outcome | Recording, transcript, request |
An AI operator for incoming calls works along this chain: it answers in about a second, speaks in the company’s name and asks the company’s questions. If the answer isn’t in the facts, it says a manager will call back and creates a task. More: what an AI voice operator is.
The same call with an AI operator
Sado: Assalomu alaykum, Safar Vaqti, this is Malika, go ahead.
Caller: Hello. We want to go to Antalya in August. Two adults and a child, seven years old. By the sea.
Sado: Okay: August, Antalya, two adults and a child. Roughly which dates would you like to fly?
Caller: After the fifteenth, for a week.
Sado: Got it. Our manager will check the price with the hotels and call you today. Your name, please?
Caller: Dilnoza.
Sado: Thank you, Dilnoza. Request made; the manager will call this number.
The operator didn’t invent a price: the agency’s facts had none for those dates, so it opened a request. Staff see it under “Requests” in the portal.
What a business owner should know about LLMs
- However “smart” the model, the answers are only as good as your facts: prices, address, hours, common questions.
- When facts change, update them, or the old price will be quoted.
- Now and then, read call transcripts.
- Think about your customers’ language: Sado speaks Uzbek and Russian (AI operator in Uzbek).
You pay only for talked minutes: 800 so‘m a minute. Try it live or call 78 778 01 23. More answers: questions about Sado.
Questions
Are an LLM and ChatGPT the same thing?
No. ChatGPT is OpenAI’s chat app, which runs on an LLM. LLM is the general name: Claude, Gemini and others are LLMs too.
Does an LLM understand Uzbek?
Modern large models understand Uzbek, but that’s not enough for the phone: you also need accurate speech recognition and a natural Uzbek voice. Sado speaks Uzbek and Russian, in the caller’s language.
Can the AI operator quote a customer the wrong price?
It answers only from the company’s facts. If a fact is missing, it says a manager will call back and creates a task. When prices change, update them in time.