Answers in Banglish.
Refuses to invent.
Six in ten chats arrived after hours and most were typed in Banglish. We built an assistant that answers only from approved content, hands anything else to a person, and is measured every week on how often it invents an answer.
This is a composite scenario built from how we deliver this kind of work. The client is fictional and the figures are targets, not measured results. We will replace it with a named engagement when a client agrees to be quoted.
An online electronics retailer in Bangladesh handling about 40,000 orders a month, with a support team of twelve working two shifts
- Sixty percent of chats arrived outside support hours. Customers wrote in Bangla script, Romanised Banglish and English, often in one message. Agents copied answers from a shared document that was two versions behind the warranty policy.
- An earlier chatbot trial had invented a return window that did not exist, and the retailer had honoured it. The team's condition for trying again: the assistant must not guess.
How we approached it.
Approved knowledge only
Policies, product data and delivery rules went into a versioned knowledge base. The assistant may answer only from it.
Three scripts, one intent
Language and script detection for Bangla, Banglish and English, with replies in the customer's script.
Retrieval with refusal
If the passages do not hold the answer, the assistant says so and opens a handoff, with the chat history attached.
A test set from real chats
300 anonymised chats, including 80 the assistant must refuse, scored by two graders who did not know which setup produced each answer.
Weekly measurement
Invention rate, resolution rate, handoff time and language accuracy reported every week, with the failing chats attached.
Try it yourself.
A simplified illustration with made-up data. It plays once when it scrolls into view, and you can pause it or step through.
order ta kobe ashbe?
Kal shokal 10 tar moddhe pouche jabe.invented
Step 1 / 6 Grounding off. The assistant answers the first question from habit and invents a delivery time.
What we built.
- Assistant on the web shop and WhatsApp Business
- Versioned knowledge base with an approval step for changes
- Language and script detection with replies in kind
- Retrieval with a refusal rule and a human handoff queue
- Evaluation harness and weekly report
Outcomes (targets).
These are the results we would aim for. They are not measured results from a named client.
- Target58%chats resolved without a personafter six weeks
- Target9% to 0.7%invented answers on the unanswerable setbare prompt versus retrieval with refusal
- Target4 smedian first replyany hour
- Target< 1 minhandoff to a personduring support hours
Timeline and stack.
8 weeks to launch, then weekly reviews
What we learned.
- Banglish is the real language of support. Test sets must be written the way customers type.
- Refusal is a feature. The assistant that says 'let me get a person' earns trust.
- Measure the invention rate, not just the resolution rate.
How we would deliver it.
Want this
for your business?
Message an engineer on WhatsApp or send a brief. We agree scope and price before work begins and build in small working steps you can test.