Should your voice agent speak Bangla, Banglish or English?
Pick the language your customer would use to answer an unknown number, which for most cash-on-delivery buyers in Bangladesh is Bangla — and expect the call to end up mixing English words in anyway, because that is how product names, sizes and payment terms are actually said.
Start from the answer, not the script
The question is usually asked as a branding decision — which language suits us? It is not. It is a decision about the two seconds after your customer sees an unknown number and picks up.
In that moment they are not choosing a language. They are defaulting to the one they think in. For most cash-on-delivery buyers in Bangladesh, that is Bangla. If your agent opens in English, the customer spends the first turn working out what is happening instead of listening, and the call has already lost the thing a confirmation call depends on: the customer paying attention long enough to notice that the address is wrong.
So the default is Bangla, and the interesting question is what happens next.
Nobody actually speaks one language on the phone
Listen to how a real order gets discussed in Dhaka and you will hear something that is neither textbook Bangla nor English:
আপনার order টা confirm করছি — size medium, cash on delivery, total চারশো টাকা।
Every structural word is Bangla. Every commercial noun is English. That is not sloppiness, and
it is not a lesser form of either language. It is how the vocabulary of ecommerce entered the
language: order, size, delivery, courier, cash on delivery, confirm, cancel arrived
as English words and stayed English words, while the grammar holding them together stayed
Bangla.
This has a direct consequence for a voice agent. An agent that insists on pure Bangla has to translate those nouns, and the translations are either archaic or ambiguous. Ask a customer to confirm their "নগদ মূল্য পরিশোধ" and you have replaced a phrase they use daily with one they have to decode. Worse, some translations collide with brand names — নগদ is a payment service, so using it as a generic word for cash is actively confusing.
The agent should speak the way the customer speaks. In practice that means Bangla grammar with the English terms left alone.
The three settings, and what each is actually for
Nishchit exposes three language settings on an agent, and they are not three levels of the same dial.
Bangla (bn) — the agent speaks Bangladeshi Bangla and keeps established English commercial
nouns as English. This is the right default for consumer cash-on-delivery: apparel, cosmetics,
electronics accessories, food, anything advertised on Facebook to a broad audience.
Banglish (bn_en) — the agent mixes more freely and follows the customer further into
English when they lead. Useful where your buyers are urban, younger, and likely to answer an
unknown number in English themselves, and useful for categories whose entire vocabulary is
English — B2B parts, software, imported goods with model numbers.
English (en) — the agent conducts the call in English. This is a real setting for a real
case: expatriate customers, corporate deliveries, an office reception rather than a household.
It is the wrong default for a broad consumer audience in Bangladesh, and choosing it because it
felt more professional is the most common way to make a confirmation call worse.
The setting is per agent, not per account, so a store selling in two very different categories can run two agents and route orders to the right one. See Features for how agent configuration works.
What code-switching demands of the agent
Choosing "Bangla" is the easy half. The half that decides whether the call works is what the agent does when the customer switches mid-sentence — and they will.
It has to understand both without being told. A customer who answers "haan bolen" and then reads out an English address has not changed the language of the call. If recognition is pinned to one language, half of that sentence disappears.
It has to keep numbers unambiguous. Money, quantities and phone numbers are where a mis-hearing is expensive. Bangla and English digits are both in play, sometimes in the same breath — a price said as চারশো and a house number said as "twenty-two". An agent that normalises these into one internal representation can read a total back for confirmation in the form the customer used; one that does not will read back a number the customer does not recognise as theirs.
It has to stop talking when interrupted. This matters more in a mixed-language call than a single-language one, because the customer often interrupts precisely to correct a word the agent got wrong. An agent that finishes its sentence has just talked over the correction it needed.
It has to sound like Bangladeshi Bangla, not Bangla in general. Vocabulary and intonation differ across the region, and a voice that reads as foreign undermines the call for the same reason an English opening does — the customer starts listening to the voice instead of the content.
What this does not solve
Language selection will not rescue a call that is badly designed. If the agent asks four questions where one would do, or reads the full item list before establishing whether the person on the line is the right person, the customer hangs up in whichever language you chose.
It also will not decide the harder question of what the agent is allowed to do once the customer starts talking. That is a different boundary, and it is covered in what an automated agent must never do — where the language of a call matters much less than whether the agent is honest about what it is.
How to choose, concretely
Look at your own inbound support messages before you decide anything. Whatever language your customers write to you in, unprompted, is the language they will answer the phone in. If your comments and DMs are Bangla with English product nouns, that is your answer, and it is the common case.
Then listen to a real call in the setting you picked before you turn it on for a day's orders. The call demo on this site plays a scripted confirmation turn by turn with the Bangla and the English side by side, which is the fastest way to hear whether the register matches how your customers actually talk. If it does not, that is a script and voice problem you can fix — not a reason to switch the whole call to English.