Banglish chatbot replies: why bots fail ‘dam koto’ and the fix

A customer sends your page "dam koto?" at 11 pm. The bot replies: "Sorry, I didn't understand. Please choose from the menu below." The customer does not choose from the menu. They close the chat and message the next page selling the same three-piece.

That reply happens because the bot was built to match words, and Banglish does not have fixed words. This guide explains why most chatbots fail romanized Bangla, what actually works, and how to test a bot with your own customers' messages before it talks to anyone.

What is Banglish, and why do customers write it?

Banglish is Bangla typed with English letters, often mixed with English words. Customers use it because it is fast on any phone keyboard and nobody has to switch layouts.

A single shop inbox usually holds three styles at once: Banglish ("eta koto?"), Bangla script ("দাম কত?") and English ("price please"). Many people switch between them inside one chat. A bot for a Bangladeshi shop has to treat all three as normal, not as edge cases.

Why do most chatbots fail Banglish?

Most chatbots fail Banglish because they match exact keywords, and Banglish has no standard spelling. The same question arrives in a dozen shapes, and a rule written for one shape misses the rest.

Here is what one meaning looks like in a real inbox:

A keyword bot with a rule for "price" catches none of the "koto" messages. Add "koto" and it still misses "kto". Add every spelling you can think of and the next customer types "dam ta koto hobe vai?" with the keyword buried in the middle of something else, and the rule fires on the wrong product.

Why translate-first bots are not much better

Some tools translate the message into English first, then answer. That sounds clever, but translation services treat Banglish as broken English. "Ase" can come back as "ace", "jabe" as nonsense, and "apu" as a name. The bot then answers a question nobody asked.

Why button menus push customers away

The other common workaround is a menu: "Press 1 for price, 2 for delivery". It avoids the spelling problem by refusing to read text at all. Customers who already typed their question feel ignored, and many leave before they tap anything.

How does an AI chatbot understand Banglish?

An AI chatbot built on a large language model reads the meaning of a message instead of matching exact words. It has seen huge amounts of mixed-language text, so "L size ase?", "L size e ache?" and "do you have L?" land on the same intent.

Understanding is only half the job. The other half is answering correctly, and that is where retrieval comes in. With retrieval-augmented generation, the bot first looks up the relevant facts in your own data, such as your product list and delivery policy, and then writes the reply from what it found. Without that step, a model that understands "dam koto?" perfectly can still invent a price.

So a good Banglish bot needs both:

1. A model that reads intent, so spelling does not matter.
2. Your data as the only source of facts, so prices, stock and delivery charges are yours, not guesses.

Should the bot reply in Banglish, Bangla or English?

The bot should reply in the same language and script the customer used. Mirroring is the simplest rule that feels natural to customers.

A customer who writes Banglish:

Customer: apu ei kurti ta black e ase? dam koto? Bot: Ji apu, black e ase. Dam 1,250 taka. Size M, L ar XL available. Kon size ta niben?

A customer who writes in Bangla script:

Customer: ঢাকার বাইরে ডেলিভারি চার্জ কত? Bot: ঢাকার বাইরে ডেলিভারি চার্জ ১২০ টাকা, সাধারণত ৩ থেকে ৫ দিনে পৌঁছে যায়।

The prices and fees above are examples. Your bot should quote only your own numbers.

Keep replies short. A shop chat is not an email. Two or three lines with the answer first, then one question to move the sale forward ("Kon size ta niben?"), works better than a paragraph.

What should I put in the knowledge base so Banglish answers are right?

Write your facts in the language your customers use, and add the questions in the forms they actually type. The model understands Banglish either way, but matching wording helps retrieval find the right fact faster and keeps the reply's tone close to yours.

Start with these:

• Products: name, price, sizes, colours and stock, kept up to date.
• Delivery: Inside and Outside Dhaka charges and times, written plainly.
• COD and returns: whether COD is available everywhere, exchange rules, who pays return delivery.
• Top 20 questions: each as a question-and-answer pair, with a Banglish version of the question.

For example, an FAQ entry:

Q: delivery charge koto? / ডেলিভারি চার্জ কত?
A: Dhakar bhitore 60 taka, Dhakar baire 120 taka. Cash on delivery available.

Also add short tone instructions: "Reply in the customer's language. Use 'apu' or 'bhaiya' only if the customer does. Never offer a discount unless it is in the offer list."

How do I test a Banglish chatbot before customers see it?

Test with real messages, not with messages you write yourself. You will type clean sentences; your customers will not.

1. Open your inbox and copy 30 to 50 recent customer messages with their original spelling.
2. Include the hard ones: "dtls", "kto", voice-to-text mistakes, half Bangla half English, and a product you do not sell.
3. Run each one through the bot's test chat.
4. Mark each reply right, weak or wrong.
5. For every weak or wrong answer, add the missing fact or FAQ, then run the message again.

A test list might look like this:

The last line matters. A bot that cannot tell when to step aside will trap an angry customer in a loop, in any language.

How does SaleMag handle Banglish?

SaleMag's AI chatbot is instructed to reply in the language the customer writes in, so a Banglish question gets a Banglish answer and Bangla script gets Bangla. It answers from a knowledge base you build (product list, FAQs, files, web pages) and is told never to invent prices or policies. When it takes orders, the confirmation step recognises Banglish replies like "ji", "thik ache" and "order koren", and words like "kintu" or "na" stop an automatic confirmation.

One honest limit: SaleMag's built-in list of "talk to a human" phrases is English ("talk to human", "real person" and similar). If your customers write "manush er sathe kotha bolbo", add a keyword automation for your common Banglish phrases that assigns the chat to a teammate. Once a chat is assigned to a person, the bot stops auto-replying there. The "hand over when unsure" setting also passes uncertain chats to your team.

What mistakes should I avoid?

• Forcing one language. Replying in formal Bangla to someone who wrote "price plz" feels stiff; replying in English to an elderly customer writing Bangla feels cold.
• Writing only English FAQs. The bot will understand, but its replies drift toward English tone.
• Leaving stale prices. The bot repeats whatever your data says. Update the product list before every sale.
• No handover path. Complaints and bulk orders need a person, fast.

For more on the price questions themselves, see how to auto-answer 'price koto?' messages from your catalog.

Next steps

Copy 30 real messages from your inbox today and use them as your test set. Then set up a bot that mirrors the customer's language and answers only from your data. Start free to try it on your own page, or see pricing.