Three kinds of chatbot
Rule-based bots follow a fixed decision tree: buttons, menus and predefined replies. They are predictable and cheap to run but cannot handle questions outside the tree. RAG chatbots use a language model to answer questions from your own documents, so they handle free-form questions on topics you have written down. Agents go further and can take actions, such as creating a booking or looking up an order, by calling your systems.
How to choose
Match the type to the risk and the variety of questions.
- Few, predictable options and no tolerance for error, such as a booking flow: rules.
- Many varied questions answered by existing documents, such as FAQs, policies and product info: RAG.
- Tasks that need data lookups or changes in your systems: an agent, with narrow permissions and logging.
- In practice most good business bots combine them: RAG for questions, rules for sensitive steps like payment, and a handover to a person.
Handover is the feature that matters most
The best chatbots know their limits. Define when a conversation moves to a person: the customer asks for one, the bot is unsure, the topic is a complaint or a refund, or the same question is asked twice. Pass the conversation history across so the customer does not repeat themselves. A bot that never hands over frustrates customers; one that hands over well makes your team faster.
Channels and languages
Build for where your customers already are: WhatsApp, Telegram, Instagram, web chat. In the Gulf and Iraq that is often WhatsApp or Telegram, with customers writing in Arabic, English or a mixture within one message. Test with real messages, including dialect and mixed scripts, and make sure Arabic displays correctly right-to-left in every channel you support.
Keeping answers accurate
Restrict the bot to your approved sources, have it say it does not know when the documents do not cover a question, and log every conversation where a customer was unhappy or a person corrected the answer. Treat those as a to-do list: usually the fix is updating a document, not changing the model. Review them weekly for the first month.
What it costs to run
Running costs come from model usage per conversation, any messaging platform fees such as WhatsApp business conversation charges, and hosting. Caching common answers, using a smaller model for simple questions and limiting answer length keep usage predictable. We estimate monthly run cost for your expected volume as part of the scope, so the build price is not the only number you see.