We sell AI agents, so the honest place to start is here: some businesses are perfectly well served by a rule-based chatbot. If yours is one of them, an AI agent is money you do not need to spend. This guide is about telling the two cases apart.
What each one is
A rule-based chatbot is a decision tree with a chat interface. You write the branches: if the customer taps "Opening hours", it shows the hours; if they tap "Pricing", it shows the link. The bot never says anything you did not write, because it cannot.
An AI agent understands natural language. The customer writes the way they would write to a person — with context, with two questions in one message, with typos — and the agent interprets it and holds the thread of the whole conversation, not just the last message.
That is the entire difference. Everything else follows from it.
What the decision tree gets right
Let us be fair to the tree, because its virtues are real:
- It is predictable. It never makes anything up, because it never generates anything. Every reply was approved by you.
- It is cheap. Affordable tools exist and setup is a matter of hours, not weeks.
- It is enough for closed questions. Opening hours, address, "do you deliver?", order status with a reference number: for fixed-answer questions, it works.
Where it breaks
The tree breaks at the exact moment the customer goes off-script. And customers go off-script constantly.
"Hi, we've got an office around 800 square feet, wondering if this would work for us — also do you do Saturdays?" That message — completely ordinary — fits no button. The tree replies "Sorry, I didn't understand that, please choose an option", and the customer has just learned there is nobody home.
There is a second problem that shows up later: maintenance grows with every branch. Each new question you want to cover is a new branch. Each new branch interacts with the existing ones. A year in, you have a diagram nobody on the team dares touch, and out-of-date answers stay in because merely finding them has become a project.
What the AI agent gets right — and its risk
The AI agent solves precisely that: it understands the 800-square-feet message, answers both questions, and asks for what is missing. You do not maintain branches; you maintain your business information, and the agent converses over it.
But the other half needs saying: a language model without controls can hallucinate — state something untrue with complete confidence. A price that does not exist, a service you do not offer, a lead time nobody promised. For a business, that is not a minor defect. It is an employee who lies convincingly.
Which is why, in our view, an AI agent is only fit to face customers if it has deterministic hard floors: checks that do not depend on the model and that stop it inventing prices, stock or products, no matter what the prompt asks. And if it knows how to say "I don't know" and hand the conversation to a human with full context. We go deeper in our piece on the EU AI Act, which also covers why the agent should introduce itself as an AI from the first message.
When a rule tree is all you need
Honestly:
- If 90% of your enquiries are the same five fixed-answer questions.
- If customers come to you to look things up, not to buy — there is no sale to work in the chat.
- If your catalogue and terms almost never change.
- If out-of-hours volume is negligible.
In that scenario, build a simple tree, keep it short, and spend the money elsewhere.
When you need an agent
- If enquiries arrive in free-form language and every customer asks in their own way.
- If the conversation is part of the sale: there is qualifying to do, a case to understand, something to propose.
- If people message you when you are not there — the evenings and weekends we wrote about in the weekend problem.
- If the tree you already have has grown into a monster nobody wants to maintain.
That is where an agent with serious guardrails changes things. It is what we built with Zatio: a flat monthly fee, replies in seconds on WhatsApp and web chat, and a widget that installs with a single line of script. The cost details live on the pricing page — we would rather you see them there than have us summarise them here.
The right choice depends on your situation, not on the fashion. But it deserves to be made looking at both columns in full: what the tool costs, and what every conversation lost to "Sorry, I didn't understand that" costs.