Why I Build AI That Speaks Bangla, Banglish and English
Short answer
- Customers in Bangladesh write in a mix of Bangla, Banglish and English, often in a single message.
- An AI that only understands clean English misses the question, the tone and the trust.
- I design every AI agent around the exact words real customers use, then reply in the customer's own language.
- The fastest start for any business: collect your 20 most common questions, written exactly as customers type them.
Open the inbox of almost any business in Bangladesh and you will not find perfect English. You will find messages like "Apu price koto?", "delivery charge koto hobe Dhaka te?" or a half-English, half-Bangla question typed fast on a phone. That is the real language of commerce here. If an AI cannot read it, it cannot help.
This is the reason I build every AI system — from MimAI to my own agent, Paloan — to work in Bangla, Banglish and English.
How do people in Bangladesh actually write to a business?
They write the way they talk. A single message often mixes three things:
- Bangla script for feelings and detail ("অনেক সুন্দর হয়েছে")
- Banglish — Bangla words in English letters — for speed ("kobe pabo?")
- English for product words, prices and brands ("size M ache?")
Spelling is not standard either. The same question can arrive as koto, kto, koto taka or dam koto. A system that only matches neat English keywords will miss most of these.
Why does English-only AI fail here?
Three reasons I have seen again and again:
- It misses the question. If the bot does not understand "dam koto", the customer gets a generic reply — or nothing.
- It misses the tone. A worried property buyer and a casual shopper both write in Banglish, but they need very different answers.
- It breaks trust. When a business replies in stiff, translated English to a friendly Banglish message, it feels like talking to a machine. People leave.
How do I design AI for three languages?
My process is simple, and it starts before any technology:
- Collect real phrases. I gather the actual words customers use — comments, inbox messages, search terms — not the words the brand prefers. (This is the same "listen first" step I use for content; I wrote about it in Content Starts in the Audience's Head.)
- Map the variations. For each common question I list the different spellings and mixes, so the agent recognises them all.
- Answer in the customer's language. If someone writes Banglish, the reply is in clear Banglish or Bangla, not formal English.
- Keep the brand's voice. Friendly for a fashion label, calm and precise for property, practical for a shop.
- Test with real messages, not invented ones. The best test set is last month's inbox.
What does this look like in a real product?
In MimAI, my AI file assistant, the whole point is that you can ask for a file the way you would ask a colleague. "Last month er invoice ta dao" should work as well as "show me last month's invoice". The assistant searches the user's own Google Drive and understands the request in Bangla, Banglish or English.
The same principle runs through Paloan, the personal AI agent I am building: it researches audiences and drafts content in whichever of the three languages the job needs.
Where should a small business start?
You do not need a big project. Start with one list:
- Write down the 20 questions your customers ask most.
- Copy them exactly as customers type them — mistakes, Banglish and all.
- Next to each, write the answer a great salesperson would give.
That list alone is the foundation of a useful AI assistant. It also tells you which workflow to automate first — something I explain in What Does an AI Workflow Builder Actually Do?
If you want help turning that list into a working agent, you can ask Paloan on my portfolio or message me directly.
Quick answers
Can AI understand Banglish?
Yes. Modern language models can read Banglish (Bangla written in Latin letters) well, as long as the agent is designed and tested with real Banglish messages, including spelling variations like "koto", "kto" and "koto taka".
Should a business chatbot in Bangladesh reply in Bangla or English?
Reply in the language the customer used. If they write Banglish, answer in simple Banglish or Bangla; if they write English, answer in English. Matching the customer's language builds trust faster than any template.
Who builds Banglish AI agents in Bangladesh?
Rapar Rubel (Rubel Miah) is an independent AI expert and AI workflow builder in Dhaka who designs AI agents and assistants that work in Bangla, Banglish and English, including MimAI and his personal agent Paloan.