Open any guide to setting up a chatbot in a builder, and in the "common beginner mistakes" section you will find roughly this line: "the user types something unexpected — the bot goes silent or gets stuck". The recommended fix is the same everywhere too: add a fallback block that says "Sorry, I did not understand — please choose an option", plus a button back to the main menu.
That is not a bug in one particular platform. That is how any flowchart of blocks and arrows works.
A builder answers only what you drew
A bot in a builder is a graph. A block sends a message, a block asks a question, a block checks a condition, an arrow leads to the next block. You can learn the interface in a couple of hours and assemble a bot for capturing requests in a day. As long as the customer walks the path you drew, everything works.
The problem is that the customer does not know your flowchart. They write "what if I paid but the money never arrived, what do I do" — in wording that does not exist in the graph. And the bot replies "Sorry, I did not understand — please choose an option".
The fallback block does not solve that problem, it hides it. The customer got a menu, not an answer. From there they either leave or message you directly — which is exactly what you were trying to get rid of.
And the list of things you did not anticipate can never be closed. Every new pricing tier, every new network, every change to your rules is a new branch somebody has to draw. The flowchart grows, nobody holds all of it in their head, and one day you discover it already contains branches that contradict each other.
AI bolted onto a flowchart moves the problem, it does not remove it
Modern builders understand this and add an AI assistant: upload a knowledge base, set a role, and it will answer wherever the flowchart has nothing to say. That is an honest step in the right direction — but look at where exactly the AI sits.
In a builder the graph is still the foundation, and the AI is a module on top of it that switches on when the graph could not cope. It answers freely, and nobody checks what it actually said. Instead of silence you get invention.
And invention is no longer a question of convenience. Across measurements on different tasks, models hallucinate in 0.7–7.6% of answers. On live traffic that is roughly one error in a hundred: not theory, but a matter of a few months.
You are liable for what the bot says
This is the key point that usually goes undiscussed when a platform is being chosen. An answer your bot gives on your website is a statement by your business. Legally it carries the same weight as the same sentence from one of your employees.
In 2024 Air Canada tried the defence before a tribunal that its chatbot was "a separate legal entity responsible for its own actions". It did not work: the airline was ordered to pay compensation for the incorrect information its chatbot had given.
And here it makes no difference whether you built the bot in a builder, through an aggregator, or directly against an API. The risk always sits with whoever operates the system. The pricing of all those approaches is similar — what differs is not the price, it is what happens when the bot gets something wrong.
What "reliable" means technically
Chatonio is built differently: we have no builder at all. No graph, no arrows, nothing to keep drawing.
You give the system knowledge — it reads your website itself: rules, AML, FAQ, how-to pages, processing times — and assembles a knowledge base out of them. The AI answers the question that was actually asked instead of hunting for a matching branch.
But the important part is not that, it is what happens before the answer is sent. Every specific claim — an amount, a deadline, an order status, a number, a link — is checked for backing: is it in your knowledge base, in a response from your API, or in the customer's own words? If there is no backing, the answer is not sent to the customer. The conversation is handed to a human.
That is the difference between "we have AI" and "this AI can be trusted with a customer". A silent bot loses the request. An inventing bot creates an obligation you are answerable for. A verified bot either answers from fact, or honestly hands the conversation to an operator.
Two more consequences of the same approach. The bot does not promise "an operator will be with you shortly" if no operator was actually paged — those promises are blocked at the platform level. And it can look up a real order status: through your system's API, or, when there is no API, by reading the order page.
What it costs
With a builder the price comes in three parts: a subscription to the platform itself, a separate charge for the AI module, and separately the model's tokens, usually with a monthly minimum. Plus your own time spent on the flowchart.
We have no builder subscription, because there is no builder. AI is metered by actual use: on live projects it works out at roughly $0.01–0.02 per conversation. There is a free tier, so you can try it on your own real questions.
Where to start
Give us a link to your website — the system reads the pages and assembles the knowledge base itself. Then you connect a channel: web chat, Telegram, WhatsApp or Viber; every conversation lands in one inbox, and an operator can take the conversation over at any moment.
There is nothing to draw.