Skip to content

What an AI chat agent for a website is (and what it is not)

A website AI agent answers visitors from the client's own content, within limits set by the agency. Here is what it does well and what it does not do.

The product5 min read

Sooner or later a client asks you: “Can we put an AI chatbot on our website?” The short answer is yes. The useful answer needs more care, because “chatbot” covers very different tools. Some follow a fixed decision tree. Others pass the visitor’s question to a large language model that knows nothing about the business. An AI chat agent for a website sits somewhere else: it answers from your client’s own content, inside a frame you have defined.

This article explains what the term covers, what such an agent does well and, above all, what it does not do. Getting this right is the best way to scope a project with a client and avoid promises nobody can keep.

One agent, three ingredients

A website chat agent is built from three separate parts.

Knowledge. The agent reads the business’s content: the pages of its website, a sitemap, PDF documents. That content is split into passages and indexed. When a visitor asks a question, the agent looks for the closest passages and relies on them to answer. The technique is called RAG (retrieval-augmented generation), and we cover it in a separate article.

Behaviour. The agency describes the agent’s role and mission, picks a tone, a default language and a welcome message, and lists the topics the agent must decline. It can also give a contact e-mail address for the cases where the agent cannot help.

The language model. It writes the answer from the question, the retrieved passages and the instructions. In blabla.bot the agency picks one option from a short catalogue: GPT-6 Luna or Mistral Small, each in a “Simple” or “In-depth” mode.

Each ingredient is set on its own. You can add knowledge without touching the instructions, or switch the AI option without rewriting the behaviour.

What the agent does well

A well-fed agent handles the questions that keep coming back and whose answer already exists somewhere on the site or in a document:

  • opening hours, service areas, delivery times;
  • how a service works, the steps of a process;
  • what an offer, a product page or a manual says;
  • pointing the visitor to the right page or the right person.

It answers at any hour, and in the visitor’s language when they clearly write in a language other than the default one. In blabla.bot each answer also shows its sources: the title and link of the cited page, or the file name and page number for a PDF. Visitors can check for themselves.

What the agent is not

This is the part that matters most when you talk to a client.

It is not a general search engine. The agent does not search the internet. If it finds nothing relevant enough in the business’s knowledge, it says so instead of improvising. A serious platform makes this a rule, not a setting: in blabla.bot the platform rules tell the agent not to make things up, to cite the passages it used and to decline requests outside its mission. The agency cannot switch them off.

It is not an advisor who decides. An agent does not approve an application, grant a discount or make a diagnosis. It informs from existing content. For anything that commits the business, it points to a person.

It is not an employee working inside your tools. Today the agent does not look up a customer account, change an order or write into business software. Controlled connections to tools are among the ideas explored for later, with no date.

It is not (yet) everywhere. The available channel is the text chat widget on the website. E-mail, messaging apps and voice belong to a longer-term vision, not to the current offer.

It is not infallible. A language model can be wrong, even with good content. That is why the blabla.bot widget reminds visitors that “the assistant can make mistakes”, and why careful testing before going live is still essential.

In practice: what separates a good agent from a bad one

The table below sums up the differences you can show a client.

SituationGeneric chatbotAgent grounded in the client’s content
The answer is on the websiteAnswers from memory, sometimes wronglyAnswers from the page and cites it
The content has no answerInvents a plausible answerSays it does not know and offers a contact
A sensitive topic for the businessNo particular limitPolitely declines forbidden topics
An instruction hidden in a pageMay follow itTreats content as data, not as orders
Look and feelThe vendor’sThe client’s, with no imposed brand

That last row matters to an agency. The widget carries your client’s colours and logo. The platform does not impose its own brand, and you decide whether to show a line such as “Assistant set up by” followed by your agency’s name.

Questions to ask before starting a project

Before you offer an agent to a client, a few questions tell you whether the project will hold up:

  1. Do the answers exist somewhere? If half of the visitors’ questions are only answered in the owner’s head, start by writing an FAQ page.
  2. Is the content up to date? An agent faithfully repeats outdated information. Plan a re-index after every significant change.
  3. Who takes over? Decide which contact address is offered when the agent does not know.
  4. Which topics are off limits? Negotiated prices, medical advice, legal questions: list what the agent must decline.
  5. How will you test it? Prepare a list of real visitor questions to check the answers before going live.

In short

An AI chat agent for a website is neither a gimmick nor a replacement for the team. It is a tool that answers repetitive questions from the business’s content, cites its sources and knows when to stop because it has no answer. Well scoped, it saves time for the client and their visitors. Poorly scoped, it disappoints.

To go further, the related articles cover RAG, writing instructions and how an agent recognises its limits. Questions about how blabla.bot works are also answered in the FAQ.

Tags

  • AI agent
  • chatbot
  • website
  • RAG