An AI chatbot for your website can now do a lot more than write answers or generate images. You can talk to it, it understands the page you have open, it knows which other pages you went through, it takes you to the section that answers your question and it can leave the form half filled in for you. All of that comes from one idea: giving the model tools.
This post explains how a website chatbot with tools and voice works, what it already does on the sites we build, and what to decide before you put one in front of your customers.
What changes when the chatbot can use tools
A language model on its own takes text in and gives text back. Tools are functions the model can ask to run during the conversation: read the page content, search the catalogue, open a section or type into a field. The model decides when to call them and the website carries out the action. At that point it makes sense to talk about AI agents, because the assistant can act within the limits you set.
On our AI chatbots and agents page we describe it as a four-step path: converse, connect knowledge, navigate interfaces and act with permission. You can start with conversation alone and add the rest when it makes sense.
Talking instead of typing
Voice is the part people notice first. A voice AI agent listens, answers out loud and lets you interrupt mid-sentence, as you would on a call. At LetBrand we work with two engines and choose per project: ElevenLabs, when a distinct voice and a tailored voice experience matter, or GPT-Live-1, the model that listens and speaks at the same time, handles interruptions and delegates reasoning and tool use to a backend model. If neither fits, we build it custom.
Voice and text can live in the same assistant, so each visitor chooses whether to type or talk.
It understands the page in front of you
When someone asks “does this include maintenance?”, the “this” is the page they are looking at. Our agents read the content of the current page and answer about it, so the visitor does not need to copy or explain anything.
You can see this in our presentations with an agent: we turn a deck into a microsite and the agent answers from the approved material. If information is missing, it says so.
Which pages you have already seen
The same question means different things depending on what the person has already seen. Someone who reaches pricing after reading three ecommerce case studies is not asking what a visitor straight from asks. Our agents take into account the pages on the site the visitor has passed through during the visit and use that path as context.
That lets the agent accompany a purchase or a research process without asking the same questions twice. If someone has compared two plans, the agent can pick up that comparison.
It takes you to the exact spot on the site
A long answer in a chat bubble means reading twice. It is more useful when the agent opens the section that answers and points to the passage behind the answer. That is how the demo microsite works: the agent opens the relevant section and highlights the passage supporting its answer.
On a commercial site, the same approach can recommend products from your catalogue and leave the person right where they can move on, whether that is a product page, a pricing table or a contact form.
It fills in the form with you
If the visitor has already given their name, company and what they need during the conversation, asking them to type it all again into a form is a poor experience. Our agents pre-fill the fields with what was said, and the person reviews them before sending.
That review matters. An agent that submits forms by itself can get a detail wrong and nobody notices until it is too late. On the chatbots page we put it as “complete authorized tasks with human oversight”.
Other use cases that can be built
The same pieces (voice, page context, the visitor’s path and tools) can support other uses. We would build these per project, outside what our agents do as standard:
- Comparing products or plans with data from your catalogue, so the visitor does not need five open tabs.
- Booking an appointment or a demo against the real availability in your calendar.
- Guiding a new user inside a SaaS product, or answering support questions from the screen where they are stuck.
- Searching documentation and taking the person to the exact section that answers.
- Qualifying B2B leads with a few questions and handing the full conversation to your sales team.
- Post-purchase support: order status, returns or usage instructions.
- Offering voice as an alternative to typing, for people who prefer to speak or find a keyboard hard to use.
- Serving visitors in their own language even if your site is only in one or two.
Several of these build on what our agents already do. An appointment or a lead qualification, for example, ends in a form the agent can pre-fill for the person to review.
What people who have met one say
Opinions on X go both ways, and a few help us think about design. @ZacharyDeWitt called a plumber late on a Sunday, from a site that “looks like it is from 2005”, and expected nobody to pick up: “I was wrong, an AI voice agent answered, took my info, diagnosed the issue and scheduled me”. His verdict: “Voice AI is incredible”. That was a phone call. We think a website can offer the same experience when the agent does more than answer.
Others add a caveat. @jburkespraker suggested “a website chatbot to improve the booking process, with an agent making the booking on your behalf”, then added: “But maybe the process just sucks and needs an update.” We agree. An agent layered on a confusing booking flow still leaves a confusing flow, and sometimes the first job is to simplify the form.
And when it breaks, people notice. @MrJackyBoy told Openreach: “your website chatbot not working so that was a waste of time trying to contact. The irony of it.” That is why the agent should know when to hand over to a person, and why there should always be another channel.
ElevenAgents in brief
ElevenAgents is ElevenLabs’ agent platform. According to its product page, its agents speak more than 70 languages and can switch languages mid-conversation. The documentation describes a knowledge base with documents and RAG, a choice of language model (or your own) and several kinds of tools. Client tools run in the browser and, according to ElevenLabs, can interact with the page or send the user to another one. Webhook tools call external APIs, for example to schedule a meeting. It plugs into any site through a widget that supports voice, voice plus text, or chat only.
ElevenLabs, GPT-Live-1 or a custom AI agent
The choice depends on what weighs most in your case. If a brand voice comes first, ElevenLabs usually fits. If a very fluid dialogue with interruptions and delegated tasks matters more, we look at GPT-Live-1. And if you need integrations with your own systems or fine control over every step, a custom AI agent gives more room; our guide to building an agent from scratch covers the basics.
Our multi-model layer lets us choose the language model by cost or task, and provider keys stay on the server, never in the browser. You can use your own credentials or let us manage access.
What to decide first
A conversational AI agent with tools needs clear limits. Decide which actions it can take alone and which need the person’s confirmation. Explain on your site what data the agent uses, including the path through your pages, and check with whoever handles privacy how that fits your cookie policy. And define when it hands over to a human.
Then there is cost. Beyond development, model, voice and infrastructure usage is paid per use. Our pricing page explains that we estimate and itemize it separately for the chosen setup.
Where to start
A website chatbot works best when it solves one concrete question from day one. Pick the page that gets the most questions, test what an agent would do there and expand later. If you want to see one in action, the presentations with an agent page is itself a demo, and on chatbots you can tell us what yours should be able to do.