AI chatbot cost in Bangladesh: what you actually pay for

A Gulshan boutique owner asks three companies for an AI chatbot price. One quotes a monthly fee with "unlimited AI". One quotes a setup fee plus a smaller monthly fee. The third says "it depends on your usage". She has no way to compare them, because each quote is hiding a different part of the bill.

AI chatbot pricing looks confusing because it is really four bills wearing one name. Once you split them apart, you can compare any two offers and estimate your own monthly cost in ten minutes.

This guide breaks down each cost, shows how to estimate AI usage, explains what Meta charges on WhatsApp, and lists ways to keep the bill down.

What are the costs of running an AI chatbot?

There are four costs: the software, the AI model's usage, the messaging channel's fees and your team's time. Different vendors bundle them differently, but every chatbot has all four.

When a vendor says "unlimited AI", ask what happens at 10,000 chats a month. Someone pays for those tokens. Either the price is high enough to cover heavy users, or there is a fair-use limit in the contract.

How do AI token costs work?

AI providers bill by tokens, which are small chunks of text. Every reply has two parts: the tokens the model reads (input) and the tokens it writes (output). Output tokens usually cost several times more than input tokens.

The surprise for most people is how much the model reads per reply. To answer "XL ache?", a chatbot sends the model:

• your instructions (tone, rules, what never to promise)
• a few snippets from your knowledge base that match the question
• the last several messages in the chat
• the new message itself

So a two-word question can mean a few thousand input tokens, while the reply itself is short. Language matters too: Bangla script usually breaks into more tokens per word than English, so Bangla-heavy chats cost a little more.

How do you estimate your monthly AI bill?

Use this formula with your own numbers:

monthly replies   = chats per month x bot replies per chat
input cost        = monthly replies x input tokens per reply  / 1,000,000 x input price
output cost       = monthly replies x output tokens per reply / 1,000,000 x output price
AI cost           = input cost + output cost

A worked example with round, made-up prices (not a quote from any provider):

Swap in your provider's current prices from OpenAI, Anthropic or Google Gemini. A larger model can cost ten times more or worse for the same traffic, which is why model choice matters more than almost anything else.

How much does WhatsApp charge for chatbot replies?

Meta charges per WhatsApp message through the official WhatsApp Cloud API, and the rules change on 1 October 2026. Here is where things stand on the date of this post:

Update, 24 October 2026: these changes took effect on 1 October 2026 as announced. Bangladesh now has its own rates on Meta's rate card instead of a regional rate, and if your WhatsApp Business Account has no payment method on file, Meta stops delivering service messages once they become charged.

Rates depend on the template category and the recipient's country. Meta updates them only on fixed dates, and the current rate card is on Meta's pricing page. Our post on per-message pricing from July 2025 explains how the per-message model started.

For a chatbot that mostly answers customers who wrote first, the service-message change is the one to watch. A shop with 3,000 chats a month and five bot replies each sends far more than 1,000 service messages, so part of those replies will carry a Meta fee from October.

Bundled AI credits or your own API key: which costs less?

Bundled credits are simpler; your own key is usually cheaper at volume and more transparent.

If you go with your own key, set a monthly budget limit in the provider's dashboard on day one.

What hidden costs should you watch for?

Most surprises come from the parts of the bill nobody mentioned.

• Setup fees for "training the bot", which usually means uploading your price list.
• Per-agent seats that make adding a second staff member expensive.
• Markup on Meta fees, where the vendor bills WhatsApp messages at more than Meta's rate.
• Overage charges once you pass a message or contact limit.
• Unofficial WhatsApp tools that are cheap until the number gets banned and you lose your customer list.
• Your own time fixing wrong answers because the knowledge base was rushed.

How do you keep AI chatbot costs down?

Make each reply cheaper and send fewer replies through the AI.

1. Use a small model. For shop questions, the small, fast tier of each provider is usually enough.
2. Answer fixed questions with keyword replies. "Delivery charge" or "location" can be a fixed reply that costs no AI tokens.
3. Keep instructions tight. A 3,000-word instruction block is read on every single reply.
4. Send fewer knowledge snippets. Three good snippets often answer as well as ten.
5. Write clear FAQs. A good FAQ answer means a shorter, more accurate reply.

Keyword rules and AI replies work best together; our comparison of AI vs rule-based chatbots explains where each fits.

What does an AI chatbot cost with SaleMag?

SaleMag's AI chatbot runs on your own OpenAI, Anthropic (Claude) or Google Gemini key, so the provider bills AI usage to you directly, without a SaleMag markup on tokens. SaleMag checks the key, lists the models it can use and recommends the newest small model of that provider by default. It logs tokens per run, and the test playground shows the tokens behind each test reply, so you can measure your own "tokens per reply" before going live.

Keyword, welcome and away rules run before the AI, so fixed answers do not use tokens. WhatsApp messages are billed by Meta to your WhatsApp Business Account. SaleMag's own plan prices are on the pricing page.

Next steps

Pull last month's chat count, pick a model, and run the formula above with your provider's current prices. Then ask any vendor you are considering to split their quote into software, AI usage and Meta fees. To measure your real tokens per reply, start free and test your top 20 questions in the playground.