Building conversational AI with Rasa means a developer writes and maintains your assistant in code. That gives you full control, and it ties the assistant's future to whoever understands that code. This article is for a company of 10 to 250 people that already runs a Rasa assistant, or has been quoted for one, and wants to know what keeping it involves. It covers the questions to ask, the risks that rarely appear in the quote, and how to move the same conversations onto a configured agent on the levelpartner.ai platform if you decide to.
How a developer-built assistant differs from a configured one
Rasa conversational AI is a developer framework. Your assistant's behaviour lives in project files: training examples, dialogue rules, custom actions and connections to your other systems. Changing what the assistant says or does means changing those files and redeploying.
A configured agent works the other way round. Behaviour lives in written instructions and a knowledge base that a non-technical person can edit, and actions are switched on rather than coded. Neither approach is wrong. They suit different organisations.
We do not describe Rasa's features, editions or pricing here, because they change and your developer will know the current position better. The questions below apply to any developer-built assistant.
The questions to ask about an existing Rasa assistant
- Who wrote it, and are they still available to you?
- Where is it hosted, who pays for that and who applies security updates?
- How long does it take to change one answer, from request to live?
- Which of your systems does it connect to, and what breaks if one of them changes?
- Can your customer service lead see and review its conversations without help?
If one person holds all the answers, you have a key-person risk, not just a chatbot.
Before you start a move
- An export of the assistant's training examples, responses and custom action list.
- A sample of real conversations from the last few months.
- Your account with the channels the assistant currently covers.
- A cut-over date when traffic moves from the old assistant to the new agent.
How to set it up
- Go to AI Agents > Conversation AI and click + Create Bot.
- Select Prompt Based Bot.
- Choose Start from Scratch, since your intents are already mapped.
- Continue into the bot editor.
- Configure Bot Settings.
- Configure Bot Training with your approved response wording.
- Configure Bot Goals, replacing each custom action with a platform action.
The preparation happens on paper first. Group the Rasa intents by business outcome: answer a question, book, update a record or hand over. Each group becomes a paragraph of plain instructions in the prompt. A custom action that called an in-house system usually maps to a workflow the agent triggers, with a webhook action inside it that you document for whoever maintains that system. Replay your sample conversations against the new agent and note every reply that differs in substance. On the cut-over date, move one channel across, leave the old assistant on the rest for a week, then retire it.
Suppose a contractor built the agency's assistant two years ago. It handles viewing requests and maintenance reports, but nobody has updated it since the contractor left, so its answers still mention a branch that has closed. A move to a configured agent might take a week of content work. From then on, the lettings manager could change answers herself in minutes rather than hunting for a developer.
The trade-off you accept
You give up fine control over dialogue logic and the ability to host everything yourself. You gain an assistant your operations team can run, with conversations stored on the contact record next to calendars and pipelines. For most companies that are trying to grow without hiring, the second matters more.
If you have a full-time engineering team and conversational AI is part of your own product, keep the developer route. The conversational AI open source article covers that cost comparison in detail.
Companies migrating from a coded assistant often try to recreate every intent one to one. Many intents existed only because older bots needed them. Group by outcome, and you will usually end up with far fewer instructions.
Running time after the move
Plan for an hour a week in the first month. Reading hand-offs and adding missing answers is the main task. There is no deployment step, so a correction made at ten o'clock applies to the next conversation.
Troubleshooting
Strict multi-step forms behave loosely. A coded assistant enforced a fixed order. Use a form or survey for fixed data capture, and let the agent send the link.
Some replies are worded differently from before. The agent paraphrases. Put the required wording into the knowledge base and tell the agent to use it verbatim for those topics.
An integration stops working after cut-over. The old custom action was the only link. Rebuild it as a workflow webhook and test with real data.
Frequently asked questions
Should we rebuild our Rasa assistant or replace it?
Rebuild if you have a developer who will own it for years and a requirement only code can meet. Replace it if the assistant mainly answers questions, books appointments and records leads, which a configured agent covers.
Can we keep the old conversation history?
Export it before you retire the assistant. Historic transcripts can be stored as files or notes on contacts where they matter.
Will customers notice the change?
They may notice faster, more natural replies. Keep the same greeting and channel so the switch feels continuous. For how agents carry chat across channels, see AI agents and chatbot conversational AI. For examples of strong replies, read conversational chatbot examples.