What makes an AI agent different from an ordinary chatbot?

Tools. An ordinary chatbot only talks: everything it says comes from the text it was filled with, so it invents slots and prices. Alongside conversation, an agent has tools it calls: reading a real calendar, saving an enquiry, recommending by industry. Ask it about a slot and it does not answer off the top of its head, it checks the calendar at that moment.

How do you stop an AI agent inventing things and over-promising?

Guardrails are built into tools, not into requests to the model. A prompt is a request the model can misread; a tool is a lock. With Webi, a booking link physically cannot be produced without a real name and email, and fake details such as test@test.com are rejected at the tool level. On top of that, the agent may only offer slots the tool actually returned.

Can an AI agent speak out loud?

It can. Webi uses ElevenLabs for voice, and his brain is the same one that runs the text conversation. What matters is that answers are written for reading aloud: no special formatting, up to three sentences, one question per step. A chatbot that reads its asterisks and tables aloud sounds broken.

Does the agent book meetings entirely on its own?

Almost. Webi reads the free slots live, narrows the choice to what the visitor wants and prepares a booking link with their details filled in. The booking itself is confirmed by the user clicking, and the agent says so clearly. He may not promise 'I have booked you in', because that would promise something he does not control.

What does it cost to put an agent like this on my site?

It depends on what the agent needs to do: conversation with a knowledge base is a smaller project, while tools such as a calendar, order status or a client database raise the scope. Broadly we are talking about a range of 500 to 5,000 EUR. For AI projects there is also an EU voucher that covers 60-70% of the cost.

All postsBlog

The AI agent that works 24/7 on my site and actually books meetings (case study: Webi)

Webi is the AI agent on this site: he answers by text and voice, reads the real free slots and saves enquiries. A case study of how he was built, and why the difference between a toy chatbot and an agent that does the job comes from tools and guardrails, not the model.

Case study
AI agent
Automation

Jurica StublićFounder · ai.webica.hr15 October 2026·7 min read

There is an AI agent running on this site. His name is Webi, I built him for my own site, and he answers questions by text and by voice, recommends automations by industry, reads my calendar and saves enquiries. He also works at 11pm on a Sunday, which is when a business owner finally has time to do research.

Chatbots are everywhere now, and most of them are useless. They recite answers from a brochure, invent appointment slots, and reply to the first concrete question with "please contact us". In this article I open Webi up: what makes him an agent rather than a toy, which guardrails keep him in line, and what of it you can copy onto your own site.

The short version: the model is not what makes the difference. The tools and the guardrails around the model are.

What Webi's job is

The first rule of a good agent: it has one job, and that job is measurable.

Webi's job is to book a free 30-minute consultation or, if the person is not ready, to take their contact details and their question. Everything else — answering questions about services, explaining the voucher, recommending automations — is in service of that job.

An agent without a goal is a toy that wastes your time and the visitor's. Once there is a goal, every part of the conversation has a direction: a question about price leads towards scoping, scoping leads towards a consultation.

A chatbot talks, an agent works

The technical difference between a chatbot and an agent fits into one word: tools.

An ordinary chatbot only talks. Everything it says comes from the text it was filled with. Ask it "do you have a free slot on Thursday" and it has no idea about any calendar, so it does what language models do when they do not know: it invents a convincing answer.

Alongside conversation, Webi has four tools he calls when needed:

Automation recommendations. The visitor describes their business, the tool returns concrete automations for that industry, with examples.
Reading the calendar. The actual free slots at that moment, with filters: "after 4pm", "on Thursday", "next week".
Preparing a booking. A scheduling link with the visitor's details already filled in.
Saving the enquiry. Name, contact and question end up in my system, not in the air.

When you ask Webi about a slot, he does not answer off the top of his head. He calls the tool, gets the real state of the calendar, and offers what he got back. It sounds simple, and this is exactly where most chatbots fall down: they have no way of knowing the truth, so they act it.

Guardrails: locks, not requests

More interesting than what Webi does is what he is not allowed to do.

He may not offer a slot the tool did not return. Not even "there is probably room on Friday".
He may not say "I have booked you in". The booking is confirmed by the user clicking the link. Webi may say exactly what is true: "I have prepared a link for you, confirm the slot in the calendar."
He may not produce a booking link without a real name and email. And here is the key detail: that check is not done by the prompt, it is done by the tool itself. Type "test@test.com" or "aaa" and the tool refuses the request.
On prices he speaks only in ranges, from 500 to 5,000 EUR depending on scope, and always steers towards a conversation for a concrete figure.

The difference between a prompt and a tool is the difference between a request and a lock. A prompt is an instruction a model can misread in an odd edge case, or that someone can talk it out of. A tool that rejects invalid data does not negotiate. For everything that might go wrong, the guardrail belongs in the tool.

This is the same logic as in my article about the website generator: AI in the middle, deterministic control at the edges. An agent is just the conversational version of the same pattern.

One conversation, start to finish

So the tools do not stay abstract, here is the real shape of a conversation:

Visitor: "How much does a chatbot for a hair salon cost?"

Webi does not fire off a price list. He asks what the chatbot should do: answer questions, take bookings, both? The visitor says bookings. Webi calls the recommendations tool, gets examples for salons and answers in two sentences, with an indicative range. Then he asks: "Would you like to go through it on a short call? I can look at free slots."

Visitor: "Sure, but only after 4." Webi calls the calendar with that filter and offers two real slots. The visitor picks one, Webi asks for a name and email, the tool checks they are real, and only then does the booking link appear. Webi tells the truth: "I have prepared the link, confirm the slot by clicking it."

The whole conversation: six messages, not one invented fact. Every fact came from a tool, and every step led towards the job.

Voice: the same brain, stricter rules

Webi also speaks. For voice I use ElevenLabs, and the brain is the same one that runs the text conversation: one system, one body of knowledge, two channels.

Voice is a merciless test of answer quality. The rules Webi writes answers by are written for reading aloud: no formatting of any kind, three sentences at most, one question per step. A chatbot that reads its asterisks, tables and seven-point lists aloud sounds broken.

Something useful happened along the way: since answers have had to pass the read-aloud test, they are better in text too. Short and concrete works in both channels.

The security side, briefly

Three dull details that get expensive when they are missing:

API keys never go to the browser. All secrets stay on the server.
For a voice conversation, the server issues a signed URL valid for about a minute. There is no permanent key for anyone to extract and use.
The number of conversations per hour is capped. An agent without limits is somebody else's free API and a bill that arrives at your address.

What happens with questions outside the job

People ask agents all sorts of things, from politics to "are you a real person". Webi has two answers to that: briefly redirect to his job, or, when the question genuinely needs a human, offer contact and save the enquiry so I see it.

That is the last and most important guardrail: an agent that knows the limits of its own knowledge. Escalating to a human is not the agent's defeat, it is its job. The same principle applies to a WhatsApp chatbot for accounting firms: the bot handles the routine, anything non-trivial goes to the bookkeeper.

What an agent really delivers, and what it does not

Let me put expectations in the right place.

An agent does not replace a salesperson and does not create demand. Nobody buys automation because a chatbot charmed them. What an agent does is simpler and more measurable: it catches people who are already looking, at the moment they are looking.

Enquiries to my site do not arrive between 9 and 5. They arrive in the evening and at weekends, when business owners finally have the peace to do research. Without an agent, that visitor reads the page and leaves, intending to "get in touch tomorrow", and tomorrow they get in touch somewhere else. With an agent, the same visitor gets an answer to their question and a slot in the calendar within three minutes.

The other thing an agent delivers: every conversation is recorded. I see what people actually ask, in what words, and where they give up. That is free market research no survey will give you, and some of the articles on this blog came out of it.

What an agent like this would look like for you

The recipe is always the same, only the tools change:

A service business or salon: a slots tool, a service price list tool
A webshop: order status, item availability, return terms
An accounting firm: answers from a knowledge base, escalation to a bookkeeper
B2B services: qualifying the enquiry and booking a call, as with mine

The questions that set the scope: what the agent may do alone, what it may do only through a tool, and what always goes to a human. Once those three answers are clear, the rest is execution.

Most owners cannot answer those three questions off the top of their head, because nobody has counted how many enquiries arrive or what they actually say. That is part of the process analysis: two meetings walking through how enquiries flow through your business today, and a written report listing what is worth automating and in which order.

As for cost, scope is the main factor. An agent that only answers from a knowledge base is a smaller project, at the lower end of the 500 to 5,000 EUR range. Every tool — calendar, order status, a connection to your system — raises the scope, but also the value: an agent without tools talks, an agent with tools works. Setting one up realistically takes a couple of weeks, most of which goes on the knowledge base and defining the guardrails, not on the technology. And for AI projects, the EU voucher with 60-70% co-funding applies here too.

Summary

Webi is the AI agent on this site: text and voice, four tools (recommendations, calendar, booking, saving enquiries) and guardrails built into the tools rather than into requests to the model. He does not invent slots because he reads them from the calendar, does not promise bookings because he does not control them, and knows when a question should be handed to a human. That is the whole secret: one clear job, tools that tell the truth, and locks on everything that might go wrong.

The best way to check whether that works: open a conversation with Webi on this page and try to book a slot. And if you want an agent like this on your own site, start with a free AI analysis.

About the author

Jurica Stublić

Founder of webica.hr and ai.webica.hr

4+ years of development and AI automation. I work alone: you talk to me from the first conversation through to launch. I bring experience from 50+ web projects at webica.hr to AI automation for companies of 5 to 50 people.

See the process analysis More about me