If you run or are opening a business in Uzbekistan from Dubai, London or anywhere else, a large share of your customers will write to you in Uzbek — some in Latin script, some in Cyrillic, many switching to Russian mid-sentence. Putting AI on the front line looks like the obvious fix, especially when nobody on your management team reads Uzbek. But that is exactly the risk: you cannot see the mistakes yourself. This article gives you an honest map of what language models handle well in Uzbek, where they fail, how to test them on your own messages in a single day, and which decisions should stay with a person.
"The model speaks Uzbek" is only half an answer
The major models — ChatGPT, Claude, Gemini — understand Uzbek and reply in it. You can confirm that in a minute. The catch is that there is far less Uzbek text online than English or Russian, so the models have seen much less of it. The weakness rarely shows up as obvious nonsense. It shows up in details a native speaker notices immediately: unnatural word order, bookish phrasing where the customer writes casually, structures borrowed from Russian, wrong suffixes on less common words.
For a business this means quality in Uzbek is uneven. Understanding what a customer wants works almost as well as in Russian. Writing text that a native speaker will not recognise as machine-made works noticeably worse. So split tasks not by language, but by what the model does with the text — reads it or writes it — and by how much a mistake costs you.
For an international team there is one more layer: if your managers read only English, every Uzbek reply the bot sends is effectively unreviewed. That changes how much you can safely automate, and it is why the testing step later in this article matters more for you than for a local team.
Where models are already reliable in Uzbek
- Detecting intent. "Narxi qancha?" (how much is it?), "yetkazib berish bormi?" (do you deliver?), "ertaga soat nechada ochiqsizlar?" (what time do you open tomorrow?). Sorting messages by topic works well when the list of topics is short and clearly described.
- Extracting data. Name, phone number, address, date, product model — even when the customer types everything in one line with no punctuation.
- Short answers from a knowledge base. Opening hours, branch addresses, delivery terms, required documents. When the answer is literally in your materials, the model finds it and phrases it in decent Uzbek.
- Summaries for your team in English. This is the most valuable use for a foreign-managed business: the model turns an Uzbek or mixed-language conversation into a short English summary. Small stylistic flaws do not matter here.
- Converting between Latin and Cyrillic. Transliteration mostly works, but check the tricky letters (o', g', sh, ch, ng) on your own vocabulary, especially product names.
Where models fail more often than you would expect
Mixed Uzbek and Russian
In Tashkent it is normal to write "Assalomu alaykum, zakaz qilmoqchi edim" and then add, in Russian, that the order is needed by Friday — often in Cyrillic, sometimes in Russian typed with Latin letters. The model understands this but often replies in the wrong language. Spell out the rule "reply in the customer's language" and test how the model chooses when a message contains two.
Cyrillic and "Latin without apostrophes"
Uzbek officially uses the Latin script, but many people still write in Cyrillic or in Latin without the apostrophes: "ozbekcha", "o`zbekcha", or the same word in Cyrillic letters. Comprehension suffers less than generation — the model may reply in a different spelling from the customer's. Decide which script your bot uses and fix it in the instructions.
Voice messages
A significant share of Telegram users send voice notes. That is two tasks in a row: speech recognition, then understanding. Uzbek speech recognition lags well behind English and Russian, especially with background noise, fast speech, regional accents and mixed language. An error at the first step silently breaks the second: the model confidently answers something the customer never said.
Terminology and formal register
Medicine, law, banking products, customs. The model may choose a word that sounds plausible but is not what the industry uses, or confuse an official term with an everyday one. For text that goes into a contract, certificate, label or public notice, that is not acceptable.
Tone and politeness
The respectful "siz" form, honorifics and the right level of formality are handled inconsistently — sometimes too dry, sometimes too flowery. With a regular customer it is a minor issue; in a reply to a complaint it is not.
Process → what AI does → what stays with a person
| Process | What AI does | What stays with a person |
|---|---|---|
| Incoming Telegram and Instagram messages | Detects topic and language, answers standard questions from the knowledge base | Unusual requests, complaints, price negotiation |
| Lead qualification | Asks follow-up questions, records the data in your CRM, writes an English summary for the manager | Deciding whether to take the client, the first sales conversation |
| Voice messages | Transcribes and summarises for the manager | Checking the transcript when an order or amount depends on it |
| Uzbek content (posts, newsletters) | Drafts and variations | Editing by a native speaker before publishing |
| Official texts (contracts, notices, labels) | Rough translation, terminology check against your glossary | Final wording and responsibility for the text |
The logic is simple: the further a text moves from internal use towards public and legally significant use, the larger the human share. This is not distrust of the technology; it is a clear split of responsibility. We apply the same principle in our own content pipeline with human approval: the assistant prepares drafts, a person decides what goes out.
Example: a weak reply and a good one
A customer writes to an electronics store on Telegram:
Customer: Salom, Samsung A55 bormi? Rassrochka qilsa bo'ladimi, 12 oyga (Hi, do you have the Samsung A55? Can I pay in instalments over 12 months?)
Weak reply (sent in Russian): "Hello! Yes, the Samsung A55 is in stock. Instalments are available. Anything else I can help with?"
Good reply: Assalomu alaykum! Samsung A55 hozir omborda bor. Muddatli to'lov 12 oyga mumkin — shartlarini menejerimiz aniqlab beradi. Sizga qaysi rang kerak va telefon raqamingizni qoldira olasizmi? (Uzbek: "It's in stock. A 12-month plan is possible — our manager will confirm the terms. Which colour would you like, and could you leave your phone number?")
The weak reply switched language, offered no next step and promised instalments with no conditions. The good one stayed in the customer's language, did not invent terms missing from the knowledge base, and handed the instalment question to a person after collecting what the manager needs for the call. The difference is not how "smart" the model is — it is the instructions and the knowledge base.
How to test a model on your own messages in one day
- Collect 50–100 real messages from the past month: Telegram, Instagram, your website. Include Cyrillic, Latin without apostrophes, mixed-language messages and a few voice-note transcripts. Remove personal data first.
- Write the expected outcome. For each message: the topic, the data to extract, and the correct action (answer, ask a question, hand over to a person).
- Run the messages through the model with the actual instructions and knowledge base you plan to use. A bare model gives different results.
- Have a native speaker grade the replies on three points: correct meaning, correct language and script, natural phrasing. If nobody in-house speaks Uzbek, hire a freelance reviewer for this step — it is a few hours of work.
- Split errors into harmless and costly. Harmless: style, an extra word. Costly: a wrong price, an invented condition, a changed meaning. Costly errors must not happen in automatic mode — those scenarios go to a person.
- Repeat the test after every change to the instructions and every model or version switch. An update can change how the model picks language, script and tone, and only a fixed test set will show it.
After one day you have a list instead of an opinion: scenarios you can automate now and those you cannot yet. Use the same set to choose between models — it is more honest than any general leaderboard. For a broader comparison on other criteria, see ChatGPT vs Claude vs Gemini for business.
What matters specifically in Uzbekistan
Uzbekistan's consumer protection law requires product and service information to be provided in the state language, with other languages as an addition. The advertising law requires ads to run in Uzbek, with any translation no larger or more prominent than the main text. So if a bot or automated broadcast states prices, terms or product descriptions, the Uzbek version has to be correct — that is a compliance question, not a matter of style. The official explanation on lex.uz (in Russian) covers the advertising rules; for your specific case, check with a local lawyer.
The second point is where customers write. For most local businesses it is Telegram and Instagram, which makes voice notes and mixed language a normal part of the flow rather than an exception. Plan for them from the start. For the wider market picture, read AI automation in Uzbekistan, and for where to begin, the guide to AI agents for business in Uzbekistan. More in the Uzbekistan section.
FAQ
Which model knows Uzbek best?
There is no permanent answer: models update several times a year and their Uzbek strengths shift. The reliable approach is to run your 50–100 real messages through two or three models and compare the replies with a native speaker.
Can AI answer Uzbek-speaking customers without review?
For standard questions whose answers are in your knowledge base — yes, after testing on real messages. Prices, terms, complaints and anything that sounds like a promise should go to a person or use pre-approved templates.
Our management doesn't read Uzbek. How do we stay in control?
Have the system attach an English summary to every conversation it handles, log all automated replies, and have a native speaker spot-check a sample each week. You review meaning through the summaries; the speaker reviews language.
Do we still need a native speaker if AI is in place?
Yes, but the role changes: instead of writing every reply, they maintain the glossary, check templates, grade tests and proofread public texts. That is a few hours a week, not a full-time position.
If you want to know how much of your Uzbek-language customer conversation an agent can handle, see how we build AI agents or get in touch — on the first call we look at your real messages. Cost depends on volume and data, and that first conversation is where we work it out.