Guide
AI chat in your own language, right in Telegram: eight casual messages and what the agent replied
A good AI chat understands you the way you actually type: lowercase, half a question, a follow-up that leans on the last one. We sent our test agent in Telegram eight messages as Emma, who flies to London on Friday with her four-year-old son. Below is the real chat: ears on the plane, Tower of London prices with a source, a proofread email, puns, and a switch to Spanish and back.
The short answer
An AI chat in your own language is an agent you write to like a person: in your words, with typos, by voice, in fragments. It understands casual language, answers in the same language without sounding translated, keeps track of what you said three messages ago and looks up fresh facts with a source. On WAM Agents it is your own agent, with the name and avatar you choose, and it lives in an ordinary Telegram chat, with no separate app.
There is no language setting to find: the agent picks it up from your first message. In our test, eight messages in English, from blocked ears on a plane to a proofread email, cost about 20 , under half of what Free gives you in the first month.
The test: Emma, Friday, London
Emma lives in Edinburgh and flies to London on Friday with her son, who is four. Before the trip her head is full of the usual: ears on the plane, what keeps a child busy, where to go, an out-of-office email, a family chat for photos. She writes it all to the agent the way she would text a friend: lowercase, no greetings, sometimes by voice.
We ran the chat on 30 October 2026 with our test agent, and the times in the chats are real. Emma, her son and her colleagues are made up. The agent's replies are real and quoted as they came, sometimes shortened, never added to. We ran the same eight messages in Russian, Turkish and Arabic too, each in its own chat: the versions of this article in those languages show those conversations.
1. Write the way you type
The first message is a test of how human the chat feels. No "Hello, could you please help me with the following question". Lowercase, no capitals on London, a question and a plan squeezed into one line.
The answer came in 15 seconds, and the agent didn't correct the style or ask what Emma meant. First the cause in three sentences: the pressure in the middle ear stops matching the cabin, and the Eustachian tube has to catch up. Then what works, as a list, with the one tip most people miss: stay awake for landing. Then when it's time for a doctor. The last line offers to help with the trip instead of signing off.
Look at the language. "Walk off the plane with ears blocked for hours", "real pain rather than a pop": that's a friend who flies a lot, not a leaflet. This is the first thing that separates a chat that speaks your language from one that merely supports it.
2. A follow-up without repeating yourself
A real conversation is one where the second question doesn't restate the first. Emma doesn't write "how can a four-year-old child deal with ear pressure on a plane". She writes "and my son is 4, what does he do?".
Ten seconds. The agent knew who "he" was and used both details: four years old and no gum yet. So: sips through a straw, a lollipop, "open wide like a lion" instead of the Valsalva trick from the first answer, which it now tells Emma to skip for a child who would "just blast". Waking him before descent grew out of the first reply, where the agent said descent is the hard part.
On medicines the agent stayed careful: saline spray, and for anything stronger, a pharmacist and a child's dose. Treat it the same way: the agent gives you a starting point, the decision on a child's medicine is yours and your doctor's.
3. A voice note while packing
Hands full of clothes is exactly when typing stops. Emma records the next question as a nine-second voice note: what goes in the carry-on so her son isn't bored or cranky for two hours.
The agent took two hours from the voice note, remembered from the second message that he is four, and built the answer around both: fewer things, new things, wrapped to unwrap one at a time. Then the part that isn't on most lists: timing beats toys, a proper meal before boarding and he will likely sleep. A voice note costs 1 a minute to transcribe, so this one was a fraction of a magnet.
4. A search with a source
Prices and opening days change, so an answer from memory isn't good enough. Emma asks the agent to look it up and say where it came from.
The agent went to the operator's own site, Historic Royal Palaces, and came back with what a family needs: £38 for an adult, under-fives free ("so your son costs nothing"), no family ticket, and the optional 10% donation spelled out so it doesn't surprise anyone at the gate. Closed days: four a year, with winter hours for a November trip.
The useful part is "Where this is from". The agent named the page, said it checked it today, and added two caveats you would only find after opening five tabs: online booking is usually a little cheaper, and resellers quote different numbers.
5. Proofread, don't rewrite
Most people asking an AI to "check" a text get a new text back. Emma asks for corrections only, and a list of what changed.
Ten seconds, three fixes, each with a one-line reason: "its" and "it's", a comma splice turned into a semicolon, "who's wrote" into "who's written". The agent kept Emma's sentence as one sentence "since that's how you wrote it", and checked that 9 November 2026 really is a Monday. Wording, order and tone stayed hers.
If you want the opposite, a shorter version, a different tone, a post, that's a separate job. We covered ten of those in AI chatbot in Telegram: 10 real tasks we gave an agent, with its replies.
6. Puns that work in your language
A pun doesn't survive translation. If a chat thinks in one language and translates into another, wordplay is where it shows. So the sixth message asks for funny names for the family photo chat.
"The Crown Jpegs", "Mind the Snap", "Tower Bridge Over Troubled Water": every one leans on something a British reader hears instantly, the Crown Jewels, the Tube announcement, a song. The agent tied them to the London trip without being told to, and picked a favourite with a reason: it "gets funnier the more badly the photos came out".
7. Into Spanish and back
The agent takes your language from your first message and keeps it for the session. Sometimes you need one piece in another language: a note for a colleague abroad, a reply to a client. Emma asks for a Spanish version and says straight away that the chat carries on in English.
The message for Lucía is in Spanish, the explanation around it is already back in English, in the same reply. And the explanation is the interesting part: the agent chose vosotros because that's how a Madrid office talks among colleagues, gave the ustedes version for a mixed or more formal group, and replaced a literal "it will be quiet from me" that doesn't work in Spanish. Andrew stayed Andrew.
Whole documents, menu photos and voice notes in a language you don't speak are a topic of their own: AI translator in Telegram: a menu photo, a voice note and a PDF, translated in one chat.
8. Explain it to a four-year-old
The last question checks two things at once: the agent is back in English, and it still remembers that "my son" is four.
The answer is in English and about this particular child. A line Emma can say word for word, built on something he has already felt: his hand out of the car window. A proof he can do himself with a sheet of paper. Where to sit him on the plane. And the last paragraph isn't about physics at all, it's about the scary follow-up question and how not to make it scarier. That's how a person answers, not a reference page.
What makes a chat feel native
Almost every AI chat today "supports" dozens of languages. The difference shows in small things, and this test had plenty of them:
- casual input: lowercase, a follow-up with no subject, half the context left to the agent, all understood first time;
- natural output: "walk off the plane with ears blocked", "buys the most time per gram", no stiff phrasing;
- the rules of the language: "its" and "it's", a comma splice, a past participle, each explained;
- cultural context: the Crown Jewels and "Mind the gap" in a family chat name, vosotros for Madrid;
- sources you can open: the Tower's own site, not a retold travel blog.
And we checked that it isn't an English trick. The same eight messages went to the same agent in Russian, Turkish and Arabic, each in its own chat. It proofread Russian commas, Turkish "de" and Arabic hamzas, made puns that only work in each language, found prices for Sochi, Istanbul and Dubai on local official and news sites, and in Arabic answered in the Gulf dialect the message was written in. Switch the language of this page to see those chats.
Three more things make this an agent rather than a chat window. It remembers the conversation: "he" in the second message and "my son" six messages later needed no reminder. It works with what you send: voice notes, photos, screenshots and PDFs. And it is your own agent, with the name and avatar you choose, in its own container, with memory between chats. The agent starts in Telegram, and other messengers can be connected. By default it runs on DeepSeek, with requests going through DeepSeek's infrastructure in China, which you confirm during onboarding. On Pro you can connect your own model, Claude or Codex.
If what you want to train is your spoken English rather than your writing, there is a voice routine for that: Speaking practice with AI: a 10-minute voice routine in Telegram.
What stays with you
Tips for your first conversation
- Write the way you always do. Typos, lowercase and "u" don't slow the agent down, and they save you time.
- Follow on, don't start over. "And if he's four?" works better than a fresh question with all the context typed again.
- Ask for the source. "Where is this from?" turns any answer into something you can check.
- Say what to leave alone. "Just fix it, don't rewrite it" gets you corrected commas, not a new style.
- Switch languages with words. "This one in Spanish, then back to English" and there is no setting to look for.
What it costs
Everything in this article works on Free: 200 in your first month and 100 every month after, no card. Pro is $13.90 a month with 400 , and $1.39 for the first month if you subscribe inside your first 7 days.
Magnets are spent only when the agent does something, never for keeping the chat open. Going by the balance in the status line, our whole conversation of eight messages cost about 20 . On the Pricing page a web search is 1.5 , a minute of voice transcription is 1 , and 20,000 tokens of text on the default model is 1 , so everyday chatting in your language fits into Free with plenty of room.
Messages worth copying
explain [what you want to know] simply, no jargon
find out how much [what] costs right now and which days it's closed. and tell me where the info is from
check this for mistakes and punctuation. just fix it, don't rewrite it, and tell me what you changed: [text]
come up with five funny names for [what], puns welcome
now this in [language] for [who]. then we carry on in English
Create your agent and write to it the way you text your friends.
Try it with your own agent
Free, no card. The agent starts working in the chat straight away.
Create your AI agentQuestions
Do I have to write properly for the AI to understand me?
No. Our first message was all lowercase, with no greeting and half the context missing, and the agent answered the actual question without commenting on the style. Write the way you text friends: short, with typos, following on from your last message.
Do I need to set my language anywhere?
No. The agent picks up your language from your first message and keeps it for the session. If you need one piece in another language, ask for it and say you want to carry on in yours. In our test it wrote the email in Spanish and the very next answer was back in English.
Where do the facts come from, and can I trust them?
When an answer depends on fresh data, the agent searches the web and names its source. In our test it took Tower of London prices from the official site. Prices, opening days and anything medical are still worth checking at the source before you go.
Can it proofread without rewriting my text?
Yes, if you say so. We asked it to fix mistakes and leave the wording alone, and it corrected the errors, kept the sentences as they were and listed each change with a one-line reason.
Is it free?
Everything in this article works on Free: 200 in the first month and 100 a month after, with no card. Our whole conversation of eight messages came to about 20 . Pro is $13.90 a month with 400 .
Which languages does it speak?
The one you write in. The agent picks up your language from your first message and keeps it for the session. We ran this same test in English, Russian, Turkish and Arabic, each in its own chat, and the agent answered natively in each.
