AI agent
The dangerous thing is not an AI that cannot answer. It is one that answers confidently and wrongly.
10:02
AI Assistant
10:02
10:04
AI Assistant
10:04
Both answers came out of your knowledge base — the delivery page and the returns policy. Nothing here was invented.
The problem
A generic chatbot invents delivery charges and return policies. One wrong answer costs you a customer and a refund, and you find out days later.
The AI answers from what you wrote down. If your knowledge base doesn't cover it, it says so and calls a human instead of guessing.
Every answer carries a confidence score. Below your threshold, the conversation goes to a person — you decide how cautious it is.
List the phrases that mean 'get me a human'. When a customer uses one, the AI hands over immediately without trying to answer.
Set your hours and the AI tells customers when your team will actually be back, instead of leaving them waiting on a reply that comes tomorrow.
Everything below is built and live today — not a roadmap.
Ask it anything, first
Before it ever meets a customer. No conversation is created; it runs against the same knowledge a real reply would use.
Ask a question the way a customer would. It runs against that agent's knowledge without creating a real conversation.
Outside Dhaka usually takes 3–5 days. Delivery is 120 BDT, and cash on delivery is available nationwide.
Answered from
Before it ever meets a customer, you can ask it anything and watch what happens. Nothing is created, no conversation is opened — it just runs against the same knowledge a real reply would use.
The answer never arrives alone. It comes with a confidence score, and with the names of the sources it was built from — so when a reply is wrong, you are not guessing at a black box. You open the source that fed it and fix the sentence that was wrong.
That is the difference between an AI you can correct and one you can only switch off.
And when it cannot
The dangerous chatbot is not the one that fails. It is the one that sounds certain and is wrong.
The dangerous chatbot is not the one that fails. It is the one that invents a refund policy, sounds certain, and you find out four days later from an angry customer.
So the threshold is yours to set, and the score is shown against it rather than hidden. Below the line the reply is never sent — the conversation goes to a person with the whole thread and the reason attached, and if you use round-robin or least-busy handover it already has an owner who already has a notification.
Every handover is counted by its cause — knowledge gap, low confidence, asked for a person, quota, outage. A handoff rate you can read is a handoff rate you can lower, and this one names the sentence you need to write.
I am not able to confirm a refund for a late delivery — let me get someone from the team to help you with this.
Against your threshold
41 of 70 needed
Your refund policy does not cover a late delivery, so nothing in the knowledge base could carry this answer.
What you hold it to
An agent that answers customers in your name should be something you can read, edit and rein in.
A name, a language and a tone you choose. Bangla, English, or whichever the customer wrote in.
Medical advice, a competitor's pricing, anything you would not want an untrained new hire answering.
Several products under one team, each with its own knowledge and its own persona, answering on its own channels.
The instructions it runs on are yours to see and edit — there is no hidden prompt doing something else.
Rupa Assistant
Answers in Bangla and English · friendly
Answer only above
70Below this it hands over instead of guessing.
Hand over immediately on
Never discusses
Answers from
Delivery & Returns · Payment & bKash · rupascloset.com/faq — and nothing else.
Add your FAQ, connect one channel, and see how much the AI handles before you pay anything.
No card required · Free plan forever · Set up in minutes