Telegram sells across borders by default. A single ad set can put your channel in front of people in a dozen countries in a week, and the messages come back in whatever those people speak. The naive response is to write the whole playbook in English and hope. The slightly less naive one is to bolt a translation step on the end, which produces something that reads as translated, which is nearly as bad.
Scalegram's bots answer in the language the message arrived in. That is one sentence and about four separate mechanisms.
Detection comes from the message, not the profile
The obvious signal is the Telegram account's language setting, and it is the wrong one to trust. People use phones set up in one language and write in another all the time: a Romanian who bought the phone in Italy, a Moroccan writing in French, an expat whose interface has been English for a decade. The account setting tells you about the handset. The message tells you about the person.
So the language is taken from what was actually written, and the reply comes back in that language. When someone opens with two words and there is not enough to go on, the profile hint is a reasonable tiebreaker, but it never overrides text that is clearly in a different language.
A request pins the thread
The interesting case is the one that breaks most implementations. A client writes in English and then says, in English, "can you write to me in Spanish". Detect language per message and you will keep answering in English forever, because every message they send is still English.
The fix is to treat an explicit request as a setting rather than as a message. Once a client asks for a language, that choice is pinned to the thread and later messages are answered in it regardless of what the detector says about any individual line. The wrong way to implement this, and we tried it, is to add an instruction to the model telling it not to use the previous language. Negative instructions are unreliable in exactly the moment you need them. Pinning the target language positively works; telling a model what not to do does not.
The same lesson shows up across the guardrails: state what the bot must do, do not list what it must avoid and hope. Filters and pinned settings hold. Prompt-level prohibitions drift.
Mixed and ambiguous messages
Real messages are messier than test cases. People switch mid-sentence, quote an English product name inside a Spanish question, or type a language in Latin characters that is normally written in another script. A detector that treats every message as a fresh independent decision will flip languages inside a conversation, which reads worse than being consistently wrong.
Stability beats accuracy on any single message. Once a thread has a language, it takes a clear signal to change it, and a client explicitly asking is the clearest signal there is.
What does not get translated
Some things must survive untouched. Links, tracking parameters, ticker symbols, account identifiers, product names and the exact figures in a price are not prose and must not be re-rendered by a language model. A translated tracking link is a broken tracking link, and a broken tracking link means the conversion event at the end of the funnel arrives with no campaign attached, which quietly wrecks the attribution the whole system was built to provide.
Numbers deserve their own care. Decimal separators and date formats differ by market, and a figure that changes meaning between locales in a commercial conversation is a dispute waiting to happen. Keep the format explicit and unambiguous.
The material follows the language too
Answering in Spanish while sending a price sheet in English only moves the problem one step down the conversation. If you sell into a market seriously enough to advertise into it, the documents that close the deal belong in that market's language, and the saved material the bot sends should be organised so the right version goes out with the right reply. This is the one part that costs you work rather than configuration, and it is worth doing for your top two or three markets before it is worth doing for the tail.
The rules do not get more relaxed in another language
This is the part that gets missed. Every constraint on what a bot may say applies in every language it speaks. It does not flirt in French. It does not promise a return in Arabic. It does not become an investment adviser in German. Guardrails that only exist in the English prompt are guardrails that vanish the moment your best-performing ad set runs in a market you do not read, and you will not find out from a complaint, because the complaint will be in that language too. The full picture is in the guardrails on what a bot may say.
The related question of voice notes and sending your own PDFs and images at the right moment is covered in their language, your voice.
What this means for the operator
You write the playbook once. You do not maintain a copy of it per market, and you do not need a native speaker on shift for every timezone you advertise into. What you do need is a review habit: pick a market, open a few real conversations in Telegram, and read them. Scalegram will not do that for you, because it stores no message content at all and there is no transcript archive to audit. The contact record tells you the stage and when someone was last in touch. The conversation itself lives in Telegram, where you can open it and read it like anyone else, which for this particular check is the right place for it to be.
"Language is not a feature you tick. It is the first thing a stranger uses to decide whether you are a real business in their market or a page they were sent to by an ad."
— Roman Onta, Executive Director, SINGUARD
Key Takeaways
- Language is taken from the text of the message, not the Telegram account setting, because people routinely write in a language their handset is not set to.
- An explicit request pins the language for the whole thread, since a client asking in English for Spanish would otherwise keep getting English.
- Pin the target language positively rather than instructing the model to avoid the previous one, because negative instructions drift exactly when they matter.
- Links, tracking parameters, symbols and figures are never translated, and every content guardrail applies identically in every language the bot speaks.
Frequently Asked Questions
How does the bot know which language to use?
From the language of the message it received. The Telegram account's language setting is only a tiebreaker when a message is too short to judge, because people often write in a language their phone is not set to.
What if a client asks it to switch languages?
The request is stored as a setting for that thread rather than treated as a one-off message, so later messages are answered in the requested language even if the client keeps writing in the original one.
Do the content rules still apply in other languages?
Yes, identically. The bot does not flirt, promise outcomes or give financial advice in any language, and those rules are enforced outside the prompt rather than being written into an English instruction.
About the Author
Roman Onta is an Executive Director at SINGUARD. He builds the Prop Firm CRM, the Broker CRM, Scalegram and CopySignals side by side with his brother Alex Onta, and he helped on the design of eTrader, the division Alex built and leads. His ground is worldwide payment processing, AML compliance and the corporate structures brokers are built on, work the two of them carry together, shaped by executive roles in the UAE and international corporates. He lives and works in Dubai for most of the year. Meet the executive duo leading Singuard's five divisions.