How a large language model (LLM) works on a phone call, and how to keep it in check
Sado AI team · · 4 min read
In short
A large language model (LLM) is what systems like ChatGPT are built on: it understands text and writes an answer. In a voice operator it sits between speech recognition and the voice: it works out what the caller needs, answers from the company’s information and picks the next action: a booking, a callback or a transfer to staff. On the phone it must be limited: answers only from company facts, short sentences, a list of things never to promise, and a manager takes over when it doesn’t know.
What an LLM is and what it does on a call
An LLM (large language model) is a neural network trained on a huge amount of text. It knows how words relate, so it understands free speech. If a caller says “Can I come tomorrow after work, after six, I’ll have my child with me”, the model picks out the day, the time and that two people are coming.
In a voice operator the chain is: speech recognition turns voice into text, the LLM understands it and writes the answer, the voice says it. On the first and last links: speech recognition and TTS voices.
What the LLM does before each answer
- Takes the text of what the caller said and everything said before: what was asked, what was agreed.
- Reads the company’s instructions: operator name, the goal of the call (for example, booking an appointment), tone, what it must not say.
- Finds the facts it needs: price, address, hours, free slots.
- Writes the answer: one or two sentences and one question.
- Acts when needed: books the customer, saves the order, schedules a callback or transfers to staff.
Chat vs phone: what changes for an LLM
| What | In a chatbot | On the phone |
|---|---|---|
| Answer length | Long text and lists are fine | 1–2 sentences, no lists or links |
| Numbers | Written as digits | Said as words, in groups |
| Speed | A couple of seconds goes unnoticed | Every pause is heard |
| Mistakes | The customer rereads and checks | What’s said becomes a promise |
| Language | Usually one | Uzbek and Russian words mixed |
What can go wrong and how to limit it
- Making things up: with no facts, it states a guess confidently. Fix: answers only from company information, plus the rule “if unsure, a manager calls back”.
- Extra promises: a discount, an exact delivery day, a guarantee. Fix: a list of things never to promise.
- Long answers: it explains as in a chat. Fix: the instruction “keep it short, ask one question”.
- Going off topic: politics, other companies, small talk. Fix: politely steer back to the goal.
- Medical, legal or financial advice. Fix: leave those questions to a specialist.
Example: three questions in one sentence
Caller: Hello. I saw a grey corner sofa on Instagram, how much is it, and do you deliver to Yunusabad?
Sado: Hello! The grey corner sofa is 6 million 400 thousand so‘m. Delivery within Tashkent is free, Yunusabad too. When do you need it?
Caller: Could it be by Saturday? But I want to see it first.
Sado: Yes. Our showroom is in Chilanzar, open daily 9 to 8. Would you like to come tomorrow?
Caller: Yes, tomorrow around five.
Sado: Okay, I’ve booked you for five tomorrow. May I have your name?
The caller asked about price, delivery and area in one sentence. The model separated them, answered only from the shop’s information and moved the call to its goal: a showroom visit. The sofa and price are an invented example.
How it works in Sado
You don’t tune the model yourself: Sado’s team sets up the operator with you, covering the script, questions, prices and hours. The operator answers only from your information; if it doesn’t know, it says a manager will call back and creates a task. After the call, the portal shows the recording, transcript, a short summary and the next step.
More: AI operator for incoming calls and FAQ. Try it live or call 78 778 01 23.
Questions
Is a voice operator just ChatGPT?
No. ChatGPT is a general chat. A voice operator works on the phone: it hears, answers from company information, books the customer and sends the result to the CRM. The language model is only one part of it.
Do LLMs understand Uzbek?
Modern models work in Uzbek too, but test it on the phone. Sado speaks Uzbek and Russian and understands mixed speech: Uzbek-speaking AI operator.
How is company data protected?
Sado stores keys encrypted, keeps data for a limited time, and you have the right to have it deleted. Details: privacy.