What is a chatbot framework, and how is it different from a platform?
A framework gives you ready-made parts: handling a conversation, remembering context, calling your systems and connecting to channels such as a website or a messaging app. You still write the logic. A platform goes further, with a visual builder and hosting, and gives you less control in return.
Which chatbot frameworks are worth looking at?
Rasa is an open-source framework for building conversational AI, with paid editions for enterprise use. Botpress is a visual platform for building chatbots and AI agents, with a free tier and paid plans. The Microsoft Bot Framework SDK is deprecated and has been replaced by the Microsoft 365 Agents SDK, so check any older guide that recommends it. For chatbots that use a language model to plan steps and call tools, graph-based frameworks are an option, and we compare three of them in LangGraph vs CrewAI vs AutoGen. Some chatbots skip a framework and call a model API directly, which gives the most control and the most to maintain.
How do you choose a chatbot framework?
Ask five questions. How much control do you need, and how fast do you need to launch? Who will maintain it once it is live? Must it run in your own environment, or can it be hosted? Which channels does it need to reach? Does it only answer questions, or must it act on your systems? A framework suits a build that needs control. A platform suits a fast launch with fewer special needs.
When is a framework not enough?
When the chatbot must answer from your own documents, call internal systems, or act on its own, the framework is only one part. You also need retrieval over your documents, integrations with your systems, and approval steps, so a person signs off on any action with real consequence. Custom assistants are covered on our generative AI page.
How do you build a chatbot on a framework?
Define the job and write down what the chatbot must never do. Choose the approach: framework, platform or a direct model API. Connect your knowledge and your systems. Test with real questions from real users before launch. Deploy with monitoring, and a clear route to a person for anything the chatbot cannot answer.
What should you measure after launch?
Whether the chatbot answered the question, how often it handed over to a person, and whether its answers matched your sources. Those three show you where the content or the logic has gaps.