Most people reading this already own a rule-based chatbot. It was built in an afternoon in a flow builder, it works, and it has been quietly growing for two years. The tree is now two hundred nodes, somebody maintains a list of synonyms for the word "refund", and half the conversations still end at "Sorry, I did not get that".
That is the ceiling, and it is structural. A flow answers the questions you predicted. The complaint is always about the ones you did not.
The same three messages, both ways
Your customers do not type menu options. They type this.
The first message is the one worth sitting with. The flow wins it outright, because somebody wired that branch to the order system, and no amount of retrieval invents data it was never given. Retrieval is not a superset of rules.
What each one is actually doing
A rule-based bot matches what the visitor typed against triggers you defined, then walks the branch you drew. Every answer is a string somebody wrote, which is why it is completely predictable and completely finite.
A retrieval assistant searches your documents by meaning, hands the best passages to a model, and gets back an answer that cites them. Nobody wrote that specific sentence, which is why it covers questions nobody anticipated, and why the citation matters.
Where rules are still the right tool
Three cases, and they are real: a transactional lookup or action, where the answer lives in a system rather than a document; a regulated script that must be said in exactly the approved words; and a short qualification form, which is a flow, not a conversation.
Most teams that "move to AI" keep all three, and should. The mistake is keeping the other hundred and ninety nodes with them.
If you decide to move
The migration is mostly a deletion, which is why it goes faster than people expect.
The list of intents is the valuable part of the old bot. Somebody spent two years discovering which questions customers actually ask, and that list is exactly the order to write documents in. Export it before anybody deletes the account.
The most common failed migration takes the decision tree and writes it out as a giant instruction: if they ask about refunds say this, if they ask about shipping say that. You have rebuilt the flow with more expensive infrastructure and lost the predictability that made it work. The answers belong in documents that retrieval finds, not in a prompt.
Running both while you test
You do not have to cut over on a Monday. A reasonable two weeks looks like this.
- Point the assistant at your existing help content and put it on one page, not the whole site.
- Leave the flow in place wherever it does a transactional job.
- Read the list of questions the assistant could not answer, twice a week, and write what is missing.
- Compare the two on the same week's volume: how many conversations ended without a person, and how many ended at a fallback.
- Retire the tree branch by branch, starting with the ones that are only there to route.
The measurement matters more than the timeline. A flow builder gives you a completion rate that looks excellent because every path was drawn to complete; the number that counts is how many people got what they came for.
welcomeai.devHow WelcomeAI answers from your own materialQuestions people ask
Export the intent list from your flow builder this week, before anything else. It is the only part of a rule-based bot that is genuinely hard to rebuild, and it is the plan for everything that replaces it.
welcomeai.devWhat is RAG? Retrieval-augmented generation explained
