Rule-based chatbot or generative AI: the differences that matter

The difference, in one sentence

A rule-based chatbot follows a path decided in advance by a person. A generative AI assistant interprets what the customer wrote and builds the answer on the spot, querying the store's data.

Put that way it sounds like a difference in sophistication. In a shop it is a difference in coverage: the first works on anticipated questions, the second also on those nobody anticipated — which are most of them.

How a rule-based chatbot works

Someone writes a tree in advance: if the customer clicks «returns», show the returns text; if they type a word from the list, open the matching branch. It is deterministic, and that is its virtue: the answer is always the same, and a person approved it.

The limits show in three places. It does not know the catalogue: the returns text is written, a product price is not. It does not handle rephrasing: the same question in different words falls outside the tree. And it scales badly: every new case is a new branch to write and maintain, until the tree is harder to update than the problem it solved.

How a generative AI assistant works

There is no tree. There is a language model that interprets the request, and around it a set of tools that model can use: search the catalogue, read an order, consult the store's knowledge base, open a ticket.

The model does not invent the catalogue: it queries it. That is the distinction that matters and that often gets lost in the hallucination debate — an assistant connected to data does not «remember» products, it reads them when needed.

The differences you actually see

What about hallucinations?

They are a real risk and worth naming. A language model left free to answer about a store it does not know will invent: plausible prices, plausible return policies, plausible products. The damage is not theoretical — it is a customer arriving with a promise the store never made.

They are handled in two ways, and neither is «a better model». The first is constraining the source: answers about products, prices and orders come from reading the store, not from the model's memory. The second is permission to stop: the assistant must be able to state it does not know and hand over to a person, rather than filling the gap. Flow-e is built this way — the agent states it does not know, never invents business data, and can open a ticket carrying the context already gathered.

When a rule-based chatbot is still right

When the path is short, always identical and high-risk. A returns form with three required fields, collecting an address, recording a consent: here predictability is a virtue and interpretation is a defect.

The honest choice is not «rules versus AI» but where to draw the line: rigid paths where the outcome must be identical every time, interpretation where questions are open.

How to tell them apart from outside

Three tests, two minutes, on any store:

1. Ask something in your own words, not the site's: «something warm but light for autumn, under €80». A decision tree has no such branch.
2. Ask for a comparison: «between these two, which is better as a gift?». That requires reading two product pages and justifying.
3. Ask something off-topic. What matters is not the answer: it is whether it admits it does not know, or invents.

What Flow-e does

Flow-e is an AI Layer: it interprets the request and acts on the store — searches the catalogue, compares two products, reads an order status, answers from the store's knowledge base, opens a ticket when a person is needed. Answers about products and orders come from reading real data, not from the model's memory.