What your AI support agent should answer (and what it should hand to a human)

Open your Messenger inbox and read the last fifty chats. You will notice something quickly: most of them are not really conversations. They are the same small set of questions, asked in slightly different words, at all hours of the day.

"Price koto?" "Delivery kobe hobe?" "Do you open on Friday?" "Where is my order?"

Someone on your team answers each one by hand. When that person is busy, asleep or on leave, the customer waits, and a customer who waits on social media usually goes and buys from the next page in their feed.

An AI support agent fixes that, but only if you set it up with clear limits. This post walks through how to split the work between the AI and your team.

Sort your chats into three piles

Before you think about tools, go through a week of real conversations and put each one into one of three piles.

Pile 1: Answers that never change

These are questions where the right answer is written down somewhere already: opening hours, delivery areas and charges, return policy, sizes, payment methods (bKash, Nagad, card, cash on delivery), how to book. An AI agent should handle all of these, every time, instantly. There is no judgement involved, only lookup.

Pile 2: Answers that need your data

"Where is my order?" "Is the blue one in stock in XL?" "Can I move my appointment to Thursday?" The answer exists, but it lives in your order system, stock sheet or calendar, not in a FAQ. An AI agent can answer these too, but only once it is connected to those systems. Without that connection it has to guess, and guessing is how you end up with angry customers.

Pile 3: Conversations that need a person

Complaints about a damaged product. A customer who is clearly upset. Custom quotes, bulk orders, refunds above a certain amount, anything with legal or medical weight. The AI's job here is not to solve the problem. Its job is to collect the details, tell the customer honestly that a person will take over, and pass the chat on with a short summary so your team does not have to scroll back and ask the same questions again.

Most businesses find that piles 1 and 2 are the bulk of their volume. That is the work you want off your team's plate.

What a good AI agent does that a canned chatbot does not

Many owners tried button-based chatbots a few years ago and gave up. Customers ignored the buttons, typed their own questions, and the bot replied with "Sorry, I didn't understand."

A modern AI agent works differently:

• It reads the question the way a person would. "Ache?" with a product photo, "deliveri charge dhaka te koto", and "How much is shipping to Dhaka?" are the same question, and it treats them that way.
• It answers from your material, not from the internet. You give it your website, price list, policies and a sample of past chats. It is instructed to answer from that and to say "I'm not sure, let me get someone" when the answer is not there.
• It replies in the customer's language. Bangla, English or Banglish, matching whatever the customer used.
• It knows when to stop. You set the escalation rules: certain keywords, certain topics, a customer asking twice for a human, or a sentiment that reads as angry.

That last point matters more than any of the others. An agent that tries to handle everything is worse than no agent at all.

The rules you should write before going live

Treat the AI agent like a new hire on their first week. You would not let a new hire invent prices or promise refunds. Write down the same boundaries for the agent:

1. Never invent a price, discount or policy. If it is not in the knowledge base, the agent says so and hands off.
2. Topics it must not discuss. Competitors, politics, medical or legal advice, anything else your business wants to avoid.
3. When to hand off. Complaints, refunds, custom orders, payment problems, and any customer who asks for a human.
4. Tone. Formal or friendly, emoji or none, "apni" or "tumi". Give it two or three example replies you like.
5. What it collects before handing off. Name, phone, order number and a one-line description of the problem.

These rules are what separate a support agent your customers trust from one that ends up in a screenshot on someone's timeline.

Where to put it

Put the agent wherever your customers already message you. For most businesses in Bangladesh that means Facebook Messenger and Instagram DMs first, then a chat widget on the website. If a big share of your customers prefer WhatsApp, the same knowledge base can power a WhatsApp agent as well; we cover that setup in our WhatsApp automation service.

One knowledge base, several channels. When you update your delivery charge in one place, every channel answers correctly the next minute.

Measure the right things

After the first two weeks, look at three numbers:

• Resolution rate: the share of conversations the agent finished without needing a person.
• Handoff reasons: why the agent passed chats to your team. If the same question keeps getting handed off, add the answer to the knowledge base.
• Customer rating or reply tone: are people thanking it, or asking for a human in frustration?

The handoff list is the most useful report you will get. It shows you exactly what your customers want to know that you have never written down.

What it does not replace

An AI agent does not replace good service. It removes the waiting and the repetitive typing, so your team has time for the conversations that actually need a person: the unhappy customer, the big order, the regular who wants a recommendation.

If you capture leads in those chats, connect the agent to your CRM so no enquiry gets lost between the bot and your sales team. Our lead capture and CRM automation work covers exactly that handover.

A sensible first step

Do not try to automate everything on day one. Start with pile 1, your fixed answers, on one channel. Watch it for a week, read the transcripts, tighten the rules. Then connect order tracking or bookings and expand to the next channel.

If you want help building the knowledge base, writing the rules and connecting your systems, that is what our AI customer service setup does. Tell us about your inbox and we will suggest where to start.