AI chatbot development for a business website starts with your own information. The assistant answers visitors' questions from it, and turns the ones who are ready into a lead your team can work. It runs on your site, in the visitor's own language, and it is not allowed to invent a price.
The assistant on this page, answering about this business.
A visitor asks what a project costs, how long it takes, or whether you have built something like theirs. They get an answer now, drawn from your actual business information. No form, no promise that somebody will be in touch.
Across the conversation it builds up what the visitor is actually asking for: the kind of work, the scope, whether they already have a site. Then it matches that against what you offer.
When the visitor is ready to talk it takes a name and a contact detail, validates it, and writes a lead with the whole conversation attached to it.
The demo is the product. The assistant in the corner of this page is the same system we build for clients: same turn pipeline, same safety rules, same lead path. Ask it what our packages cost and it will tell you, because it reads the same catalogue the pricing page renders from.
What differs for your business is the knowledge it is given, not how it works.
Typed in the widget, in whatever language the visitor uses.
Your live package catalogue, your published examples and a curated question-and-answer set are put together as the knowledge the assistant may use for this turn. It is assembled per turn from your live data, so a price you change today is the price it quotes today.
A single structured call, so the answer arrives in one pass. There is no agent loop and no chain of calls it can wander off inside.
This is the part most chatbots do not have. If the draft contains a currency amount the turn was not authorised to state, or names something you do not sell, the draft is discarded and the turn is routed down a written path. The wrong answer gets replaced, never edited.
Along with a suggested next step when there is an obvious one, such as looking at the examples.
The assistant reads a knowledge base assembled from your business data. There is no document upload, no vector search and no model trained on your files. We say that plainly because the word "AI" is doing a lot of work in this market.
The assistant handles English, Arabic, Russian and Uzbek, and it replies in the language the visitor wrote in. The interface, the direction of the text and the wording all switch together.
When someone writes in a language it does not support, it says so. That is a state it handles out loud, and the visitor knows where they stand.
This is a different question from whether your website is available in Arabic. Arabic and RTL layout for the site itself is separate, priced work, added as an option on top of the package you choose.
We build both, and we deliberately do not run the same technology on both surfaces.
On your website, a model is the right fit. The visitor is browsing, the question is open-ended, and the cost of a slightly clumsy answer is a follow-up question. So the assistant composes its answer.
For our WhatsApp lead-capture path, we deliberately use a deterministic flow. At that stage reliability and predictable qualification matter more to us than open-ended generation, so the bot runs a fixed sequence of questions and composes nothing.
Not a keyword. The conversation has to produce a clear picture of what the visitor needs, and that is the condition for going further.
A name and a phone number or e-mail. The contact is validated as a real, usable detail before it counts. A malformed one is re-asked a bounded number of times, and then the assistant stops asking.
One record per conversation, updated as the conversation continues. It carries what the visitor said in their own words, the scope the conversation established, and which package the business logic matched.
The lead and the notification are written in the same transaction. If the notification channel is down the job waits and retries; the lead is already safe.
Stored with the lead, readable in your admin panel and available as a CSV export whenever you want it in a spreadsheet.
This is worth more of your attention than the answers it gets right.
It does not guess a price. Naming a figure is a backend decision, and a draft containing an amount the turn was not authorised to state never reaches the visitor.
It does not invent a service. If someone asks for something outside what you offer, it says so.
When someone asks to speak to a person, it asks for a contact detail first. A request for a human with no way to reach them is a dead end for both sides. Once it has one, exactly one notification goes out and exactly one lead exists.
If the model call itself fails, the turn falls back to a narrower answering path. The visitor never sees an error message, and a conversation that has already ended does not call the model at all.
The limits are the part worth reading, and we would sooner lose the enquiry here than halfway through a project.
The website assistant ships as part of the package below, alongside the site itself and a CRM or Google Sheets integration, because an assistant that captures leads into nothing is not worth building. On a Product Build it is scoped inside the project.
Website + AI automation + integrations for growing businesses.
Not included: Arabic/RTL, Meta/WhatsApp & AI usage fees, paid CRM/SaaS subscriptions
Compare all four packagesAI model usage is billed by the provider, per use. It is listed as excluded on the package card, and it is the one running cost an assistant has. The full picture is on the pricing page, and the site it sits on is described on our website development page.
From your own business information: your live package and price data, your published examples, and a question-and-answer set we write with you during the project. The assistant assembles it fresh on every turn, so when you change a price in your admin panel that is the price it quotes on the next question. There is no second copy of your catalogue to keep in sync.
No, and the guard is structural. If a draft answer contains a currency amount the turn was not authorised to state, the draft is discarded and the conversation is routed down a written path. The wrong answer gets replaced, never edited. The model does not choose which package to recommend either; that decision is made in code.
Not today. What we build has no document upload, no vector search and no model fine-tuned on your files, and we say so plainly. We write the question-and-answer set with you instead, which for most businesses covers the questions customers actually ask better than a pile of PDFs would.
Yes, along with English, Russian and Uzbek. It replies in the language the visitor typed in, and the interface, the text direction and the wording switch together. When a visitor writes in a language it does not support, it says so out loud. That is a separate question from having your website itself in Arabic, which is RTL layout work and is priced as an option on top of your package.
Building the assistant is part of the package; running it is not. The AI provider bills per use, so the cost scales with how much people actually talk to it, and you pay that provider directly. We disclose the expected recurring cost before implementing it. The assistant also runs against a daily budget you set, and it stops for the day once that budget is spent.
On your website it captures the request and the contact detail and notifies your team, and the live conversation then continues on WhatsApp or by phone. A live handover inside the chat itself, where a staff member replies as the business while the customer stays in the same thread, is something we build on WhatsApp.
Tell us what your customers ask you over and over. That list is most of the project. We reply on WhatsApp or by email.